555.58.02

555.58
555.52.04
2026-01-27 19:49:47 +00:00 · 2024-07-01 14:52:22 +02:00 · 2024-06-27 19:21:59 +02:00 · 2024-06-05 09:39:46 +02:00 · 2024-05-21 15:12:40 +02:00 · 2024-04-25 16:24:58 +02:00
1813 changed files with 395893 additions and 216979 deletions
--- a/.github/ISSUE_TEMPLATE/20_build_bug.yml
+++ b/.github/ISSUE_TEMPLATE/20_build_bug.yml
@@ -32,6 +32,14 @@ body:
    description: "Which kernel are you running? (output of `uname -a`, say if you built it yourself)."
  validations:
    required: true
+- type: checkboxes
+  id: sw_host_kernel_stable
+  attributes:
+    label: "Please confirm you are running a stable release kernel (e.g. not a -rc). We do not accept bug reports for unreleased kernels."
+    options:
+    - label: "I am running on a stable kernel release."
+  validations:
+    required: true
 - type: textarea
  id: bug_description
  attributes:
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,7 +1,61 @@
 # Changelog

+## Release 555 Entries
+
+### [555.58.02] 2024-07-01
+
+### [555.58] 2024-06-27
+
+### [555.52.04] 2024-06-05
+
+### [555.42.02] 2024-05-21
+
+## Release 550 Entries
+
+### [550.78] 2024-04-25
+
+### [550.76] 2024-04-17
+
+### [550.67] 2024-03-19
+
+### [550.54.15] 2024-03-18
+
+### [550.54.14] 2024-02-23
+
+#### Added
+
+- Added vGPU Host and vGPU Guest support. For vGPU Host, please refer to the README.vgpu packaged in the vGPU Host Package for more details.
+
+### [550.40.07] 2024-01-24
+
+#### Fixed
+
+- Set INSTALL_MOD_DIR only if it's not defined, [#570](https://github.com/NVIDIA/open-gpu-kernel-modules/pull/570) by @keelung-yang
+
+## Release 545 Entries
+
+### [545.29.06] 2023-11-22
+
+#### Fixed
+
+- The brightness control of NVIDIA seems to be broken, [#573](https://github.com/NVIDIA/open-gpu-kernel-modules/issues/573)
+
+### [545.29.02] 2023-10-31
+
+### [545.23.06] 2023-10-17
+
+#### Fixed
+
+- Fix always-false conditional, [#493](https://github.com/NVIDIA/open-gpu-kernel-modules/pull/493) by @meme8383
+
+#### Added
+
+- Added beta-quality support for GeForce and Workstation GPUs. Please see the "Open Linux Kernel Modules" chapter in the NVIDIA GPU driver end user README for details.
+
 ## Release 535 Entries

+### [535.129.03] 2023-10-31
+
 ### [535.113.01] 2023-09-21

 #### Fixed
@@ -50,6 +104,14 @@

 ## Release 525 Entries

+### [525.147.05] 2023-10-31
+
+#### Fixed
+
+- Fix nvidia_p2p_get_pages(): Fix double-free in register-callback error path, [#557](https://github.com/NVIDIA/open-gpu-kernel-modules/pull/557) by @BrendanCunningham
+
+### [525.125.06] 2023-06-26
+
 ### [525.116.04] 2023-05-09

 ### [525.116.03] 2023-04-25
--- a/README.md
+++ b/README.md
@@ -1,7 +1,7 @@
 # NVIDIA Linux Open GPU Kernel Module Source

 This is the source release of the NVIDIA Linux open GPU kernel modules,
-version 535.113.01.
+version 555.58.02.


 ## How to Build
@@ -17,7 +17,7 @@ as root:

 Note that the kernel modules built here must be used with GSP
 firmware and user-space NVIDIA GPU driver components from a corresponding
-535.113.01 driver release.  This can be achieved by installing
+555.58.02 driver release.  This can be achieved by installing
 the NVIDIA GPU driver from the .run file using the `--no-kernel-modules`
 option.  E.g.,

@@ -74,7 +74,7 @@ kernel.

 The NVIDIA open kernel modules support the same range of Linux kernel
 versions that are supported with the proprietary NVIDIA kernel modules.
-This is currently Linux kernel 3.10 or newer.
+This is currently Linux kernel 4.15 or newer.


 ## How to Contribute
@@ -179,16 +179,19 @@ software applications.

 ## Compatible GPUs

-The open-gpu-kernel-modules can be used on any Turing or later GPU
-(see the table below). However, in the 535.113.01 release,
-GeForce and Workstation support is still considered alpha-quality.
+The NVIDIA open kernel modules can be used on any Turing or later GPU
+(see the table below). However, in the __DRIVER_VERION__ release, GeForce and
+Workstation support is considered to be Beta quality. The open kernel modules
+are suitable for broad usage, and NVIDIA requests feedback on any issues
+encountered specific to them.

-To enable use of the open kernel modules on GeForce and Workstation GPUs,
-set the "NVreg_OpenRmEnableUnsupportedGpus" nvidia.ko kernel module
-parameter to 1. For more details, see the NVIDIA GPU driver end user
-README here:
+For details on feature support and limitations, see the NVIDIA GPU driver
+end user README here:

-https://us.download.nvidia.com/XFree86/Linux-x86_64/535.113.01/README/kernel_open.html
+https://us.download.nvidia.com/XFree86/Linux-x86_64/555.58.02/README/kernel_open.html
+
+For vGPU support, please refer to the README.vgpu packaged in the vGPU Host
+Package for more details.

 In the below table, if three IDs are listed, the first is the PCI Device 
 ID, the second is the PCI Subsystem Vendor ID, and the third is the PCI
@@ -648,9 +651,12 @@ Subsystem Device ID.
 | NVIDIA T1000 8GB                                | 1FF0 17AA 1612 |
 | NVIDIA T400 4GB                                 | 1FF2 1028 1613 |
 | NVIDIA T400 4GB                                 | 1FF2 103C 1613 |
+| NVIDIA T400E                                    | 1FF2 103C 18FF |
 | NVIDIA T400 4GB                                 | 1FF2 103C 8A80 |
 | NVIDIA T400 4GB                                 | 1FF2 10DE 1613 |
+| NVIDIA T400E                                    | 1FF2 10DE 18FF |
 | NVIDIA T400 4GB                                 | 1FF2 17AA 1613 |
+| NVIDIA T400E                                    | 1FF2 17AA 18FF |
 | Quadro T1000                                    | 1FF9           |
 | NVIDIA A100-SXM4-40GB                           | 20B0           |
 | NVIDIA A100-PG509-200                           | 20B0 10DE 1450 |
@@ -658,6 +664,7 @@ Subsystem Device ID.
 | NVIDIA A100-SXM4-80GB                           | 20B2 10DE 147F |
 | NVIDIA A100-SXM4-80GB                           | 20B2 10DE 1622 |
 | NVIDIA A100-SXM4-80GB                           | 20B2 10DE 1623 |
+| NVIDIA PG509-210                                | 20B2 10DE 1625 |
 | NVIDIA A100-SXM-64GB                            | 20B3 10DE 14A7 |
 | NVIDIA A100-SXM-64GB                            | 20B3 10DE 14A8 |
 | NVIDIA A100 80GB PCIe                           | 20B5 10DE 1533 |
@@ -665,6 +672,7 @@ Subsystem Device ID.
 | NVIDIA PG506-232                                | 20B6 10DE 1492 |
 | NVIDIA A30                                      | 20B7 10DE 1532 |
 | NVIDIA A30                                      | 20B7 10DE 1804 |
+| NVIDIA A30                                      | 20B7 10DE 1852 |
 | NVIDIA A800-SXM4-40GB                           | 20BD 10DE 17F4 |
 | NVIDIA A100-PCIE-40GB                           | 20F1 10DE 145F |
 | NVIDIA A800-SXM4-80GB                           | 20F3 10DE 179B |
@@ -681,6 +689,7 @@ Subsystem Device ID.
 | NVIDIA A800 40GB Active                         | 20F6 103C 180A |
 | NVIDIA A800 40GB Active                         | 20F6 10DE 180A |
 | NVIDIA A800 40GB Active                         | 20F6 17AA 180A |
+| NVIDIA AX800                                    | 20FD 10DE 17F8 |
 | NVIDIA GeForce GTX 1660 Ti                      | 2182           |
 | NVIDIA GeForce GTX 1660                         | 2184           |
 | NVIDIA GeForce GTX 1650 SUPER                   | 2187           |
@@ -743,11 +752,18 @@ Subsystem Device ID.
 | NVIDIA H800 PCIe                                | 2322 10DE 17A4 |
 | NVIDIA H800                                     | 2324 10DE 17A6 |
 | NVIDIA H800                                     | 2324 10DE 17A8 |
+| NVIDIA H20                                      | 2329 10DE 198B |
+| NVIDIA H20                                      | 2329 10DE 198C |
 | NVIDIA H100 80GB HBM3                           | 2330 10DE 16C0 |
 | NVIDIA H100 80GB HBM3                           | 2330 10DE 16C1 |
 | NVIDIA H100 PCIe                                | 2331 10DE 1626 |
+| NVIDIA H200                                     | 2335 10DE 18BE |
+| NVIDIA H200                                     | 2335 10DE 18BF |
 | NVIDIA H100                                     | 2339 10DE 17FC |
 | NVIDIA H800 NVL                                 | 233A 10DE 183A |
+| NVIDIA GH200 120GB                              | 2342 10DE 16EB |
+| NVIDIA GH200 120GB                              | 2342 10DE 1805 |
+| NVIDIA GH200 480GB                              | 2342 10DE 1809 |
 | NVIDIA GeForce RTX 3060 Ti                      | 2414           |
 | NVIDIA GeForce RTX 3080 Ti Laptop GPU           | 2420           |
 | NVIDIA RTX A5500 Laptop GPU                     | 2438           |
@@ -800,6 +816,7 @@ Subsystem Device ID.
 | NVIDIA RTX A2000 12GB                           | 2571 10DE 1611 |
 | NVIDIA RTX A2000 12GB                           | 2571 17AA 1611 |
 | NVIDIA GeForce RTX 3050                         | 2582           |
+| NVIDIA GeForce RTX 3050                         | 2584           |
 | NVIDIA GeForce RTX 3050 Ti Laptop GPU           | 25A0           |
 | NVIDIA GeForce RTX 3050Ti Laptop GPU            | 25A0 103C 8928 |
 | NVIDIA GeForce RTX 3050Ti Laptop GPU            | 25A0 103C 89F9 |
@@ -815,6 +832,14 @@ Subsystem Device ID.
 | NVIDIA GeForce RTX 3050 4GB Laptop GPU          | 25AB           |
 | NVIDIA GeForce RTX 3050 6GB Laptop GPU          | 25AC           |
 | NVIDIA GeForce RTX 2050                         | 25AD           |
+| NVIDIA RTX A1000                                | 25B0 1028 1878 |
+| NVIDIA RTX A1000                                | 25B0 103C 1878 |
+| NVIDIA RTX A1000                                | 25B0 10DE 1878 |
+| NVIDIA RTX A1000                                | 25B0 17AA 1878 |
+| NVIDIA RTX A400                                 | 25B2 1028 1879 |
+| NVIDIA RTX A400                                 | 25B2 103C 1879 |
+| NVIDIA RTX A400                                 | 25B2 10DE 1879 |
+| NVIDIA RTX A400                                 | 25B2 17AA 1879 |
 | NVIDIA A16                                      | 25B6 10DE 14A9 |
 | NVIDIA A2                                       | 25B6 10DE 157E |
 | NVIDIA RTX A2000 Laptop GPU                     | 25B8           |
@@ -832,6 +857,8 @@ Subsystem Device ID.
 | NVIDIA RTX A2000 Embedded GPU                   | 25FA           |
 | NVIDIA RTX A500 Embedded GPU                    | 25FB           |
 | NVIDIA GeForce RTX 4090                         | 2684           |
+| NVIDIA GeForce RTX 4090 D                       | 2685           |
+| NVIDIA GeForce RTX 4070 Ti SUPER                | 2689           |
 | NVIDIA RTX 6000 Ada Generation                  | 26B1 1028 16A1 |
 | NVIDIA RTX 6000 Ada Generation                  | 26B1 103C 16A1 |
 | NVIDIA RTX 6000 Ada Generation                  | 26B1 10DE 16A1 |
@@ -840,17 +867,28 @@ Subsystem Device ID.
 | NVIDIA RTX 5000 Ada Generation                  | 26B2 103C 17FA |
 | NVIDIA RTX 5000 Ada Generation                  | 26B2 10DE 17FA |
 | NVIDIA RTX 5000 Ada Generation                  | 26B2 17AA 17FA |
+| NVIDIA RTX 5880 Ada Generation                  | 26B3 1028 1934 |
+| NVIDIA RTX 5880 Ada Generation                  | 26B3 103C 1934 |
+| NVIDIA RTX 5880 Ada Generation                  | 26B3 10DE 1934 |
+| NVIDIA RTX 5880 Ada Generation                  | 26B3 17AA 1934 |
 | NVIDIA L40                                      | 26B5 10DE 169D |
 | NVIDIA L40                                      | 26B5 10DE 17DA |
 | NVIDIA L40S                                     | 26B9 10DE 1851 |
 | NVIDIA L40S                                     | 26B9 10DE 18CF |
+| NVIDIA L20                                      | 26BA 10DE 1957 |
+| NVIDIA L20                                      | 26BA 10DE 1990 |
+| NVIDIA GeForce RTX 4080 SUPER                   | 2702           |
 | NVIDIA GeForce RTX 4080                         | 2704           |
+| NVIDIA GeForce RTX 4070 Ti SUPER                | 2705           |
+| NVIDIA GeForce RTX 4070                         | 2709           |
 | NVIDIA GeForce RTX 4090 Laptop GPU              | 2717           |
 | NVIDIA RTX 5000 Ada Generation Laptop GPU       | 2730           |
 | NVIDIA GeForce RTX 4090 Laptop GPU              | 2757           |
 | NVIDIA RTX 5000 Ada Generation Embedded GPU     | 2770           |
 | NVIDIA GeForce RTX 4070 Ti                      | 2782           |
+| NVIDIA GeForce RTX 4070 SUPER                   | 2783           |
 | NVIDIA GeForce RTX 4070                         | 2786           |
+| NVIDIA GeForce RTX 4060 Ti                      | 2788           |
 | NVIDIA GeForce RTX 4080 Laptop GPU              | 27A0           |
 | NVIDIA RTX 4000 SFF Ada Generation              | 27B0 1028 16FA |
 | NVIDIA RTX 4000 SFF Ada Generation              | 27B0 103C 16FA |
@@ -864,6 +902,7 @@ Subsystem Device ID.
 | NVIDIA RTX 4000 Ada Generation                  | 27B2 103C 181B |
 | NVIDIA RTX 4000 Ada Generation                  | 27B2 10DE 181B |
 | NVIDIA RTX 4000 Ada Generation                  | 27B2 17AA 181B |
+| NVIDIA L2                                       | 27B6 10DE 1933 |
 | NVIDIA L4                                       | 27B8 10DE 16CA |
 | NVIDIA L4                                       | 27B8 10DE 16EE |
 | NVIDIA RTX 4000 Ada Generation Laptop GPU       | 27BA           |
@@ -872,13 +911,24 @@ Subsystem Device ID.
 | NVIDIA RTX 3500 Ada Generation Embedded GPU     | 27FB           |
 | NVIDIA GeForce RTX 4060 Ti                      | 2803           |
 | NVIDIA GeForce RTX 4060 Ti                      | 2805           |
+| NVIDIA GeForce RTX 4060                         | 2808           |
 | NVIDIA GeForce RTX 4070 Laptop GPU              | 2820           |
 | NVIDIA RTX 3000 Ada Generation Laptop GPU       | 2838           |
 | NVIDIA GeForce RTX 4070 Laptop GPU              | 2860           |
 | NVIDIA GeForce RTX 4060                         | 2882           |
 | NVIDIA GeForce RTX 4060 Laptop GPU              | 28A0           |
 | NVIDIA GeForce RTX 4050 Laptop GPU              | 28A1           |
+| NVIDIA RTX 2000 Ada Generation                  | 28B0 1028 1870 |
+| NVIDIA RTX 2000 Ada Generation                  | 28B0 103C 1870 |
+| NVIDIA RTX 2000E Ada Generation                 | 28B0 103C 1871 |
+| NVIDIA RTX 2000 Ada Generation                  | 28B0 10DE 1870 |
+| NVIDIA RTX 2000E Ada Generation                 | 28B0 10DE 1871 |
+| NVIDIA RTX 2000 Ada Generation                  | 28B0 17AA 1870 |
+| NVIDIA RTX 2000E Ada Generation                 | 28B0 17AA 1871 |
 | NVIDIA RTX 2000 Ada Generation Laptop GPU       | 28B8           |
+| NVIDIA RTX 1000 Ada Generation Laptop GPU       | 28B9           |
+| NVIDIA RTX 500 Ada Generation Laptop GPU        | 28BA           |
+| NVIDIA RTX 500 Ada Generation Laptop GPU        | 28BB           |
 | NVIDIA GeForce RTX 4060 Laptop GPU              | 28E0           |
 | NVIDIA GeForce RTX 4050 Laptop GPU              | 28E1           |
 | NVIDIA RTX 2000 Ada Generation Embedded GPU     | 28F8           |
--- a/kernel-open/Kbuild
+++ b/kernel-open/Kbuild
@@ -70,14 +70,26 @@ $(foreach _module, $(NV_KERNEL_MODULES), \

 EXTRA_CFLAGS += -I$(src)/common/inc
 EXTRA_CFLAGS += -I$(src)
-EXTRA_CFLAGS += -Wall $(DEFINES) $(INCLUDES) -Wno-cast-qual -Wno-error -Wno-format-extra-args
+EXTRA_CFLAGS += -Wall $(DEFINES) $(INCLUDES) -Wno-cast-qual -Wno-format-extra-args
 EXTRA_CFLAGS += -D__KERNEL__ -DMODULE -DNVRM
-EXTRA_CFLAGS += -DNV_VERSION_STRING=\"535.113.01\"
+EXTRA_CFLAGS += -DNV_VERSION_STRING=\"555.58.02\"

 ifneq ($(SYSSRCHOST1X),)
 EXTRA_CFLAGS += -I$(SYSSRCHOST1X)
 endif

+# Some Android kernels prohibit driver use of filesystem functions like
+# filp_open() and kernel_read(). Disable the NV_FILESYSTEM_ACCESS_AVAILABLE
+# functionality that uses those functions when building for Android.
+
+PLATFORM_IS_ANDROID ?= 0
+
+ifeq ($(PLATFORM_IS_ANDROID),1)
+ EXTRA_CFLAGS += -DNV_FILESYSTEM_ACCESS_AVAILABLE=0
+else
+ EXTRA_CFLAGS += -DNV_FILESYSTEM_ACCESS_AVAILABLE=1
+endif
+
 EXTRA_CFLAGS += -Wno-unused-function

 ifneq ($(NV_BUILD_TYPE),debug)
@@ -92,7 +104,6 @@ endif

 ifeq ($(NV_BUILD_TYPE),debug)
 EXTRA_CFLAGS += -g
- EXTRA_CFLAGS += $(call cc-option,-gsplit-dwarf,)
 endif

 EXTRA_CFLAGS += -ffreestanding
@@ -107,7 +118,7 @@ ifeq ($(ARCH),x86_64)
 endif

 ifeq ($(ARCH),powerpc)
- EXTRA_CFLAGS += -mlittle-endian -mno-strict-align -mno-altivec
+ EXTRA_CFLAGS += -mlittle-endian -mno-strict-align
 endif

 EXTRA_CFLAGS += -DNV_UVM_ENABLE
@@ -123,6 +134,16 @@ ifneq ($(wildcard /proc/sgi_uv),)
 EXTRA_CFLAGS += -DNV_CONFIG_X86_UV
 endif

+ifdef VGX_FORCE_VFIO_PCI_CORE
+ EXTRA_CFLAGS += -DNV_VGPU_FORCE_VFIO_PCI_CORE
+endif
+
+WARNINGS_AS_ERRORS ?=
+ifeq ($(WARNINGS_AS_ERRORS),1)
+ ccflags-y += -Werror
+else
+ ccflags-y += -Wno-error
+endif

 #
 # The conftest.sh script tests various aspects of the target kernel.
@@ -149,6 +170,10 @@ NV_CONFTEST_CMD := /bin/sh $(NV_CONFTEST_SCRIPT) \
 NV_CFLAGS_FROM_CONFTEST := $(shell $(NV_CONFTEST_CMD) build_cflags)

 NV_CONFTEST_CFLAGS = $(NV_CFLAGS_FROM_CONFTEST) $(EXTRA_CFLAGS) -fno-pie
+NV_CONFTEST_CFLAGS += $(call cc-disable-warning,pointer-sign)
+NV_CONFTEST_CFLAGS += $(call cc-option,-fshort-wchar,)
+NV_CONFTEST_CFLAGS += $(call cc-option,-Werror=incompatible-pointer-types,)
+NV_CONFTEST_CFLAGS += -Wno-error

 NV_CONFTEST_COMPILE_TEST_HEADERS := $(obj)/conftest/macros.h
 NV_CONFTEST_COMPILE_TEST_HEADERS += $(obj)/conftest/functions.h
@@ -208,102 +233,7 @@ $(obj)/conftest/patches.h: $(NV_CONFTEST_SCRIPT)
 	@mkdir -p $(obj)/conftest
 	@$(NV_CONFTEST_CMD) patch_check > $@

-
-# Each of these headers is checked for presence with a test #include; a
-# corresponding #define will be generated in conftest/headers.h.
-NV_HEADER_PRESENCE_TESTS = \
- asm/system.h \
- drm/drmP.h \
- drm/drm_auth.h \
- drm/drm_gem.h \
- drm/drm_crtc.h \
- drm/drm_color_mgmt.h \
- drm/drm_atomic.h \
- drm/drm_atomic_helper.h \
- drm/drm_atomic_state_helper.h \
- drm/drm_encoder.h \
- drm/drm_atomic_uapi.h \
- drm/drm_drv.h \
- drm/drm_framebuffer.h \
- drm/drm_connector.h \
- drm/drm_probe_helper.h \
- drm/drm_blend.h \
- drm/drm_fourcc.h \
- drm/drm_prime.h \
- drm/drm_plane.h \
- drm/drm_vblank.h \
- drm/drm_file.h \
- drm/drm_ioctl.h \
- drm/drm_device.h \
- drm/drm_mode_config.h \
- drm/drm_modeset_lock.h \
- dt-bindings/interconnect/tegra_icc_id.h \
- generated/autoconf.h \
- generated/compile.h \
- generated/utsrelease.h \
- linux/efi.h \
- linux/kconfig.h \
- linux/platform/tegra/mc_utils.h \
- linux/printk.h \
- linux/ratelimit.h \
- linux/prio_tree.h \
- linux/log2.h \
- linux/of.h \
- linux/bug.h \
- linux/sched.h \
- linux/sched/mm.h \
- linux/sched/signal.h \
- linux/sched/task.h \
- linux/sched/task_stack.h \
- xen/ioemu.h \
- linux/fence.h \
- linux/dma-resv.h \
- soc/tegra/chip-id.h \
- soc/tegra/fuse.h \
- soc/tegra/tegra_bpmp.h \
- video/nv_internal.h \
- linux/platform/tegra/dce/dce-client-ipc.h \
- linux/nvhost.h \
- linux/nvhost_t194.h \
- linux/host1x-next.h \
- asm/book3s/64/hash-64k.h \
- asm/set_memory.h \
- asm/prom.h \
- asm/powernv.h \
- linux/atomic.h \
- asm/barrier.h \
- asm/opal-api.h \
- sound/hdaudio.h \
- asm/pgtable_types.h \
- asm/page.h \
- linux/stringhash.h \
- linux/dma-map-ops.h \
- rdma/peer_mem.h \
- sound/hda_codec.h \
- linux/dma-buf.h \
- linux/time.h \
- linux/platform_device.h \
- linux/mutex.h \
- linux/reset.h \
- linux/of_platform.h \
- linux/of_device.h \
- linux/of_gpio.h \
- linux/gpio.h \
- linux/gpio/consumer.h \
- linux/interconnect.h \
- linux/pm_runtime.h \
- linux/clk.h \
- linux/clk-provider.h \
- linux/ioasid.h \
- linux/stdarg.h \
- linux/iosys-map.h \
- asm/coco.h \
- linux/vfio_pci_core.h \
- linux/mdev.h \
- soc/tegra/bpmp-abi.h \
- soc/tegra/bpmp.h \
- linux/cc_platform.h \
- asm/cpufeature.h
+include $(src)/header-presence-tests.mk

 # Filename to store the define for the header in $(1); this is only consumed by
 # the rule below that concatenates all of these together.
--- a/kernel-open/Makefile
+++ b/kernel-open/Makefile
@@ -28,7 +28,7 @@ else
  else
    KERNEL_UNAME ?= $(shell uname -r)
    KERNEL_MODLIB := /lib/modules/$(KERNEL_UNAME)
-    KERNEL_SOURCES := $(shell test -d $(KERNEL_MODLIB)/source && echo $(KERNEL_MODLIB)/source || echo $(KERNEL_MODLIB)/build)
+    KERNEL_SOURCES := $(shell ((test -d $(KERNEL_MODLIB)/source && echo $(KERNEL_MODLIB)/source) || (test -d $(KERNEL_MODLIB)/build/source && echo $(KERNEL_MODLIB)/build/source)) || echo $(KERNEL_MODLIB)/build)
  endif

  KERNEL_OUTPUT := $(KERNEL_SOURCES)
@@ -42,7 +42,11 @@ else
  else
    KERNEL_UNAME ?= $(shell uname -r)
    KERNEL_MODLIB := /lib/modules/$(KERNEL_UNAME)
-    ifeq ($(KERNEL_SOURCES), $(KERNEL_MODLIB)/source)
+    # $(filter patter...,text) - Returns all whitespace-separated words in text that
+    # do match any of the pattern words, removing any words that do not match.
+    # Set the KERNEL_OUTPUT only if either $(KERNEL_MODLIB)/source or
+    # $(KERNEL_MODLIB)/build/source path matches the KERNEL_SOURCES.
+    ifneq ($(filter $(KERNEL_SOURCES),$(KERNEL_MODLIB)/source $(KERNEL_MODLIB)/build/source),)
      KERNEL_OUTPUT := $(KERNEL_MODLIB)/build
      KBUILD_PARAMS := KBUILD_OUTPUT=$(KERNEL_OUTPUT)
    endif
@@ -57,12 +61,15 @@ else
      -e 's/armv[0-7]\w\+/arm/' \
      -e 's/aarch64/arm64/' \
      -e 's/ppc64le/powerpc/' \
+      -e 's/riscv64/riscv/' \
    )
  endif

  NV_KERNEL_MODULES ?= $(wildcard nvidia nvidia-uvm nvidia-vgpu-vfio nvidia-modeset nvidia-drm nvidia-peermem)
  NV_KERNEL_MODULES := $(filter-out $(NV_EXCLUDE_KERNEL_MODULES), \
                                    $(NV_KERNEL_MODULES))
+  INSTALL_MOD_DIR ?= kernel/drivers/video
+
  NV_VERBOSE ?=
  SPECTRE_V2_RETPOLINE ?= 0

@@ -74,7 +81,7 @@ else
  KBUILD_PARAMS += NV_KERNEL_SOURCES=$(KERNEL_SOURCES)
  KBUILD_PARAMS += NV_KERNEL_OUTPUT=$(KERNEL_OUTPUT)
  KBUILD_PARAMS += NV_KERNEL_MODULES="$(NV_KERNEL_MODULES)"
-  KBUILD_PARAMS += INSTALL_MOD_DIR=kernel/drivers/video
+  KBUILD_PARAMS += INSTALL_MOD_DIR="$(INSTALL_MOD_DIR)"
  KBUILD_PARAMS += NV_SPECTRE_V2=$(SPECTRE_V2_RETPOLINE)

  .PHONY: modules module clean clean_conftest modules_install
--- a/kernel-open/common/inc/nv-chardev-numbers.h
+++ b/kernel-open/common/inc/nv-chardev-numbers.h
@@ -0,0 +1,43 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+
+#ifndef _NV_CHARDEV_NUMBERS_H_
+#define _NV_CHARDEV_NUMBERS_H_
+
+// NVIDIA's reserved major character device number (Linux).
+#define NV_MAJOR_DEVICE_NUMBER  195
+
+// Minor numbers 0 to 247 reserved for regular devices
+#define NV_MINOR_DEVICE_NUMBER_REGULAR_MAX         247
+
+// Minor numbers 248 to 253 currently unused
+
+// Minor number 254 reserved for the modeset device (provided by NVKMS)
+#define NV_MINOR_DEVICE_NUMBER_MODESET_DEVICE      254
+
+// Minor number 255 reserved for the control device
+#define NV_MINOR_DEVICE_NUMBER_CONTROL_DEVICE      255
+
+#endif  // _NV_CHARDEV_NUMBERS_H_
+
--- a/kernel-open/common/inc/nv-hypervisor.h
+++ b/kernel-open/common/inc/nv-hypervisor.h
@@ -37,13 +37,11 @@ typedef enum _HYPERVISOR_TYPE
    OS_HYPERVISOR_UNKNOWN
 } HYPERVISOR_TYPE;

-#define CMD_VGPU_VFIO_WAKE_WAIT_QUEUE         0
-#define CMD_VGPU_VFIO_INJECT_INTERRUPT        1
-#define CMD_VGPU_VFIO_REGISTER_MDEV           2
-#define CMD_VGPU_VFIO_PRESENT                 3
-#define CMD_VFIO_PCI_CORE_PRESENT             4
+#define CMD_VFIO_WAKE_REMOVE_GPU              1
+#define CMD_VGPU_VFIO_PRESENT                 2
+#define CMD_VFIO_PCI_CORE_PRESENT             3

-#define MAX_VF_COUNT_PER_GPU 64
+#define MAX_VF_COUNT_PER_GPU                  64

 typedef enum _VGPU_TYPE_INFO
 {
@@ -54,17 +52,11 @@ typedef enum _VGPU_TYPE_INFO

 typedef struct
 {
-    void  *vgpuVfioRef;
-    void  *waitQueue;
    void  *nv;
-    NvU32 *vgpuTypeIds;
-    NvU8 **vgpuNames;
-    NvU32  numVgpuTypes;
-    NvU32  domain;
-    NvU8   bus;
-    NvU8   slot;
-    NvU8   function;
-    NvBool is_virtfn;
+    NvU32 domain;
+    NvU32 bus;
+    NvU32 device;
+    NvU32 return_status;
 } vgpu_vfio_info;

 typedef struct
--- a/kernel-open/common/inc/nv-ioctl-numa.h
+++ b/kernel-open/common/inc/nv-ioctl-numa.h
@@ -25,14 +25,12 @@
 #ifndef NV_IOCTL_NUMA_H
 #define NV_IOCTL_NUMA_H

-#if defined(NV_LINUX)
-
 #include <nv-ioctl-numbers.h>

-#if defined(NV_KERNEL_INTERFACE_LAYER)
-
+#if defined(NV_KERNEL_INTERFACE_LAYER) && defined(NV_LINUX)
 #include <linux/types.h>
-
+#elif defined (NV_KERNEL_INTERFACE_LAYER) && defined(NV_BSD)
+#include <sys/stdint.h>
 #else

 #include <stdint.h>
@@ -81,5 +79,3 @@ typedef struct nv_ioctl_set_numa_status
 #define NV_IOCTL_NUMA_STATUS_OFFLINE_FAILED         6

 #endif
-
-#endif
--- a/kernel-open/common/inc/nv-ioctl-numbers.h
+++ b/kernel-open/common/inc/nv-ioctl-numbers.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2020-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2020-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -39,5 +39,6 @@
 #define NV_ESC_QUERY_DEVICE_INTR     (NV_IOCTL_BASE + 13)
 #define NV_ESC_SYS_PARAMS            (NV_IOCTL_BASE + 14)
 #define NV_ESC_EXPORT_TO_DMABUF_FD   (NV_IOCTL_BASE + 17)
+#define NV_ESC_WAIT_OPEN_COMPLETE    (NV_IOCTL_BASE + 18)

 #endif
--- a/kernel-open/common/inc/nv-ioctl.h
+++ b/kernel-open/common/inc/nv-ioctl.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2020-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2020-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -142,4 +142,10 @@ typedef struct nv_ioctl_export_to_dma_buf_fd
    NvU32       status;
 } nv_ioctl_export_to_dma_buf_fd_t;

+typedef struct nv_ioctl_wait_open_complete
+{
+    int rc;
+    NvU32 adapterStatus;
+} nv_ioctl_wait_open_complete_t;
+
 #endif
--- a/kernel-open/common/inc/nv-kthread-q-os.h
+++ b/kernel-open/common/inc/nv-kthread-q-os.h
@@ -0,0 +1,62 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2016 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef __NV_KTHREAD_QUEUE_OS_H__
+#define __NV_KTHREAD_QUEUE_OS_H__
+
+#include <linux/types.h>            // atomic_t
+#include <linux/list.h>             // list
+#include <linux/sched.h>            // task_struct
+#include <linux/numa.h>             // NUMA_NO_NODE
+#include <linux/semaphore.h>
+
+#include "conftest.h"
+
+struct nv_kthread_q
+{
+    struct list_head q_list_head;
+    spinlock_t q_lock;
+
+    // This is a counting semaphore. It gets incremented and decremented
+    // exactly once for each item that is added to the queue.
+    struct semaphore q_sem;
+    atomic_t main_loop_should_exit;
+
+    struct task_struct *q_kthread;
+};
+
+struct nv_kthread_q_item
+{
+    struct list_head q_list_node;
+    nv_q_func_t function_to_run;
+    void *function_args;
+};
+
+
+#ifndef NUMA_NO_NODE
+#define NUMA_NO_NODE (-1)
+#endif
+
+#define NV_KTHREAD_NO_NODE NUMA_NO_NODE
+
+#endif
--- a/kernel-open/common/inc/nv-kthread-q.h
+++ b/kernel-open/common/inc/nv-kthread-q.h
@@ -24,13 +24,14 @@
 #ifndef __NV_KTHREAD_QUEUE_H__
 #define __NV_KTHREAD_QUEUE_H__

-#include <linux/types.h>            // atomic_t
-#include <linux/list.h>             // list
-#include <linux/sched.h>            // task_struct
-#include <linux/numa.h>             // NUMA_NO_NODE
-#include <linux/semaphore.h>
+struct nv_kthread_q;
+struct nv_kthread_q_item;
+typedef struct nv_kthread_q nv_kthread_q_t;
+typedef struct nv_kthread_q_item nv_kthread_q_item_t;

-#include "conftest.h"
+typedef void (*nv_q_func_t)(void *args);
+
+#include "nv-kthread-q-os.h"

 ////////////////////////////////////////////////////////////////////////////////
 // nv_kthread_q:
@@ -85,38 +86,6 @@
 //
 ////////////////////////////////////////////////////////////////////////////////

-typedef struct nv_kthread_q nv_kthread_q_t;
-typedef struct nv_kthread_q_item nv_kthread_q_item_t;
-
-typedef void (*nv_q_func_t)(void *args);
-
-struct nv_kthread_q
-{
-    struct list_head q_list_head;
-    spinlock_t q_lock;
-
-    // This is a counting semaphore. It gets incremented and decremented
-    // exactly once for each item that is added to the queue.
-    struct semaphore q_sem;
-    atomic_t main_loop_should_exit;
-
-    struct task_struct *q_kthread;
-};
-
-struct nv_kthread_q_item
-{
-    struct list_head q_list_node;
-    nv_q_func_t function_to_run;
-    void *function_args;
-};
-
-
-#ifndef NUMA_NO_NODE
-#define NUMA_NO_NODE (-1)
-#endif
-
-#define NV_KTHREAD_NO_NODE NUMA_NO_NODE
-
 //
 // The queue must not be used before calling this routine.
 //
@@ -155,10 +124,7 @@ int nv_kthread_q_init_on_node(nv_kthread_q_t *q,
 // This routine is the same as nv_kthread_q_init_on_node() with the exception
 // that the queue stack will be allocated on the NUMA node of the caller.
 //
-static inline int nv_kthread_q_init(nv_kthread_q_t *q, const char *qname)
-{
-    return nv_kthread_q_init_on_node(q, qname, NV_KTHREAD_NO_NODE);
-}
+int nv_kthread_q_init(nv_kthread_q_t *q, const char *qname);

 //
 // The caller is responsible for stopping all queues, by calling this routine
--- a/kernel-open/common/inc/nv-linux.h
+++ b/kernel-open/common/inc/nv-linux.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2001-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2001-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -35,6 +35,7 @@
 #include "os-interface.h"
 #include "nv-timer.h"
 #include "nv-time.h"
+#include "nv-chardev-numbers.h"

 #define NV_KERNEL_NAME "Linux"

@@ -57,14 +58,10 @@
 #include <linux/version.h>
 #include <linux/utsname.h>

-#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 32)
-#error "This driver does not support kernels older than 2.6.32!"
-#elif LINUX_VERSION_CODE < KERNEL_VERSION(2, 7, 0)
-#  define KERNEL_2_6
-#elif LINUX_VERSION_CODE >= KERNEL_VERSION(3, 0, 0)
-#  define KERNEL_3
-#else
-#error "This driver does not support development kernels!"
+#if LINUX_VERSION_CODE == KERNEL_VERSION(4, 4, 0)
+// Version 4.4 is allowed, temporarily, although not officially supported.
+#elif LINUX_VERSION_CODE < KERNEL_VERSION(4, 15, 0)
+#error "This driver does not support kernels older than Linux 4.15!"
 #endif

 #if defined (CONFIG_SMP) && !defined (__SMP__)
@@ -248,7 +245,7 @@ NV_STATUS nvos_forward_error_to_cray(struct pci_dev *, NvU32,
 #undef NV_SET_PAGES_UC_PRESENT
 #endif

-#if !defined(NVCPU_AARCH64) && !defined(NVCPU_PPC64LE)
+#if !defined(NVCPU_AARCH64) && !defined(NVCPU_PPC64LE) && !defined(NVCPU_RISCV64)
 #if !defined(NV_SET_MEMORY_UC_PRESENT) && !defined(NV_SET_PAGES_UC_PRESENT)
 #error "This driver requires the ability to change memory types!"
 #endif
@@ -406,32 +403,6 @@ extern int nv_pat_mode;
 #define NV_GFP_DMA32 (NV_GFP_KERNEL)
 #endif

-extern NvBool nvos_is_chipset_io_coherent(void);
-
-#if defined(NVCPU_X86_64)
-#define CACHE_FLUSH()  asm volatile("wbinvd":::"memory")
-#define WRITE_COMBINE_FLUSH() asm volatile("sfence":::"memory")
-#elif defined(NVCPU_AARCH64)
-    static inline void nv_flush_cache_cpu(void *info)
-    {
-        if (!nvos_is_chipset_io_coherent())
-        {
-#if defined(NV_FLUSH_CACHE_ALL_PRESENT)
-            flush_cache_all();
-#else
-            WARN_ONCE(0, "NVRM: kernel does not support flush_cache_all()\n");
-#endif
-        }
-    }
-#define CACHE_FLUSH()            nv_flush_cache_cpu(NULL)
-#define CACHE_FLUSH_ALL()        on_each_cpu(nv_flush_cache_cpu, NULL, 1)
-#define WRITE_COMBINE_FLUSH()    mb()
-#elif defined(NVCPU_PPC64LE)
-#define CACHE_FLUSH()            asm volatile("sync;  \n" \
-                                              "isync; \n" ::: "memory")
-#define WRITE_COMBINE_FLUSH()    CACHE_FLUSH()
-#endif
-
 typedef enum
 {
    NV_MEMORY_TYPE_SYSTEM,      /* Memory mapped for ROM, SBIOS and physical RAM. */
@@ -440,7 +411,7 @@ typedef enum
    NV_MEMORY_TYPE_DEVICE_MMIO, /* All kinds of MMIO referred by NVRM e.g. BARs and MCFG of device */
 } nv_memory_type_t;

-#if defined(NVCPU_AARCH64) || defined(NVCPU_PPC64LE)
+#if defined(NVCPU_AARCH64) || defined(NVCPU_PPC64LE) || defined(NVCPU_RISCV64)
 #define NV_ALLOW_WRITE_COMBINING(mt)    1
 #elif defined(NVCPU_X86_64)
 #if defined(NV_ENABLE_PAT_SUPPORT)
@@ -753,7 +724,6 @@ static inline dma_addr_t nv_phys_to_dma(struct device *dev, NvU64 pa)
 #define NV_VMA_FILE(vma)              ((vma)->vm_file)

 #define NV_DEVICE_MINOR_NUMBER(x)     minor((x)->i_rdev)
-#define NV_CONTROL_DEVICE_MINOR       255

 #define NV_PCI_DISABLE_DEVICE(pci_dev)                           \
    {                                                            \
@@ -862,16 +832,16 @@ static inline dma_addr_t nv_phys_to_dma(struct device *dev, NvU64 pa)
 #define NV_PRINT_AT(nv_debug_level,at)                                           \
    {                                                                            \
        nv_printf(nv_debug_level,                                                \
-            "NVRM: VM: %s:%d: 0x%p, %d page(s), count = %d, flags = 0x%08x, "    \
+            "NVRM: VM: %s:%d: 0x%p, %d page(s), count = %d, "                    \
            "page_table = 0x%p\n",  __FUNCTION__, __LINE__, at,                  \
            at->num_pages, NV_ATOMIC_READ(at->usage_count),                      \
-            at->flags, at->page_table);                                          \
+            at->page_table);                                                     \
    }

 #define NV_PRINT_VMA(nv_debug_level,vma)                                                 \
    {                                                                                    \
        nv_printf(nv_debug_level,                                                        \
-            "NVRM: VM: %s:%d: 0x%lx - 0x%lx, 0x%08x bytes @ 0x%016llx, 0x%p, 0x%p\n",    \
+            "NVRM: VM: %s:%d: 0x%lx - 0x%lx, 0x%08lx bytes @ 0x%016llx, 0x%p, 0x%p\n",    \
            __FUNCTION__, __LINE__, vma->vm_start, vma->vm_end, NV_VMA_SIZE(vma),        \
            NV_VMA_OFFSET(vma), NV_VMA_PRIVATE(vma), NV_VMA_FILE(vma));                  \
    }
@@ -1104,6 +1074,8 @@ static inline void nv_kmem_ctor_dummy(void *arg)
        kmem_cache_destroy(kmem_cache);     \
    }

+#define NV_KMEM_CACHE_ALLOC_ATOMIC(kmem_cache)     \
+    kmem_cache_alloc(kmem_cache, GFP_ATOMIC)
 #define NV_KMEM_CACHE_ALLOC(kmem_cache)     \
    kmem_cache_alloc(kmem_cache, GFP_KERNEL)
 #define NV_KMEM_CACHE_FREE(ptr, kmem_cache) \
@@ -1130,6 +1102,23 @@ static inline void *nv_kmem_cache_zalloc(struct kmem_cache *k, gfp_t flags)
 #endif
 }

+static inline int nv_kmem_cache_alloc_stack_atomic(nvidia_stack_t **stack)
+{
+    nvidia_stack_t *sp = NULL;
+#if defined(NVCPU_X86_64)
+    if (rm_is_altstack_in_use())
+    {
+        sp = NV_KMEM_CACHE_ALLOC_ATOMIC(nvidia_stack_t_cache);
+        if (sp == NULL)
+            return -ENOMEM;
+        sp->size = sizeof(sp->stack);
+        sp->top = sp->stack + sp->size;
+    }
+#endif
+    *stack = sp;
+    return 0;
+}
+
 static inline int nv_kmem_cache_alloc_stack(nvidia_stack_t **stack)
 {
    nvidia_stack_t *sp = NULL;
@@ -1376,7 +1365,19 @@ typedef struct nv_dma_map_s {
         i < dm->mapping.discontig.submap_count;                              \
         i++, sm = &dm->mapping.discontig.submaps[i])

+/*
+ * On 4K ARM kernels, use max submap size a multiple of 64K to keep nv-p2p happy.
+ * Despite 4K OS pages, we still use 64K P2P pages due to dependent modules still using 64K.
+ * Instead of using (4G-4K), use max submap size as (4G-64K) since the mapped IOVA range
+ * must be aligned at 64K boundary.
+ */
+#if defined(CONFIG_ARM64_4K_PAGES)
+#define NV_DMA_U32_MAX_4K_PAGES           ((NvU32)((NV_U32_MAX >> PAGE_SHIFT) + 1))
+#define NV_DMA_SUBMAP_MAX_PAGES           ((NvU32)(NV_DMA_U32_MAX_4K_PAGES - 16))
+#else
 #define NV_DMA_SUBMAP_MAX_PAGES           ((NvU32)(NV_U32_MAX >> PAGE_SHIFT))
+#endif
+
 #define NV_DMA_SUBMAP_IDX_TO_PAGE_IDX(s)  (s * NV_DMA_SUBMAP_MAX_PAGES)

 /*
@@ -1456,6 +1457,11 @@ typedef struct coherent_link_info_s {
     * baremetal OS environment it is System Physical Address(SPA) and in the case
     * of virutalized OS environment it is Intermediate Physical Address(IPA) */
    NvU64 gpu_mem_pa;
+
+    /* Physical address of the reserved portion of the GPU memory, applicable
+     * only in Grace Hopper self hosted passthrough virtualizatioan platform. */
+    NvU64 rsvd_mem_pa;
+
    /* Bitmap of NUMA node ids, corresponding to the reserved PXMs,
     * available for adding GPU memory to the kernel as system RAM */
    DECLARE_BITMAP(free_node_bitmap, MAX_NUMNODES);
@@ -1603,6 +1609,30 @@ typedef struct nv_linux_state_s {

    struct nv_dma_device dma_dev;
    struct nv_dma_device niso_dma_dev;
+
+    /*
+     * Background kthread for handling deferred open operations
+     * (e.g. from O_NONBLOCK).
+     *
+     * Adding to open_q and reading/writing is_accepting_opens
+     * are protected by nvl->open_q_lock (not nvl->ldata_lock).
+     * This allows new deferred open operations to be enqueued without
+     * blocking behind previous ones (which hold nvl->ldata_lock).
+     *
+     * Adding to open_q is only safe if is_accepting_opens is true.
+     * This prevents open operations from racing with device removal.
+     *
+     * Stopping open_q is only safe after setting is_accepting_opens to false.
+     * This ensures that the open_q (and the larger nvl structure) will
+     * outlive any of the open operations enqueued.
+     */
+    nv_kthread_q_t open_q;
+    NvBool is_accepting_opens;
+    struct semaphore open_q_lock;
+#if defined(NV_VGPU_KVM_BUILD)
+    wait_queue_head_t wait;
+    NvS32 return_status;
+#endif
 } nv_linux_state_t;

 extern nv_linux_state_t *nv_linux_devices;
@@ -1646,22 +1676,13 @@ typedef struct nvidia_event
    nv_event_t event;
 } nvidia_event_t;

-typedef enum
-{
-    NV_FOPS_STACK_INDEX_MMAP,
-    NV_FOPS_STACK_INDEX_IOCTL,
-    NV_FOPS_STACK_INDEX_COUNT
-} nvidia_entry_point_index_t;
-
 typedef struct
 {
    nv_file_private_t nvfp;

    nvidia_stack_t *sp;
-    nvidia_stack_t *fops_sp[NV_FOPS_STACK_INDEX_COUNT];
-    struct semaphore fops_sp_lock[NV_FOPS_STACK_INDEX_COUNT];
    nv_alloc_t *free_list;
-    void *nvptr;
+    nv_linux_state_t *nvptr;
    nvidia_event_t *event_data_head, *event_data_tail;
    NvBool dataless_event_pending;
    nv_spinlock_t fp_lock;
@@ -1672,6 +1693,12 @@ typedef struct
    nv_alloc_mapping_context_t mmap_context;
    struct address_space mapping;

+    nv_kthread_q_item_t open_q_item;
+    struct completion open_complete;
+    nv_linux_state_t *deferred_open_nvl;
+    int open_rc;
+    NV_STATUS adapter_status;
+
    struct list_head entry;
 } nv_linux_file_private_t;

@@ -1680,6 +1707,21 @@ static inline nv_linux_file_private_t *nv_get_nvlfp_from_nvfp(nv_file_private_t
    return container_of(nvfp, nv_linux_file_private_t, nvfp);
 }

+static inline int nv_wait_open_complete_interruptible(nv_linux_file_private_t *nvlfp)
+{
+    return wait_for_completion_interruptible(&nvlfp->open_complete);
+}
+
+static inline void nv_wait_open_complete(nv_linux_file_private_t *nvlfp)
+{
+    wait_for_completion(&nvlfp->open_complete);
+}
+
+static inline NvBool nv_is_open_complete(nv_linux_file_private_t *nvlfp)
+{
+    return completion_done(&nvlfp->open_complete);
+}
+
 #define NV_SET_FILE_PRIVATE(filep,data) ((filep)->private_data = (data))
 #define NV_GET_LINUX_FILE_PRIVATE(filep) ((nv_linux_file_private_t *)(filep)->private_data)

@@ -1689,28 +1731,6 @@ static inline nv_linux_file_private_t *nv_get_nvlfp_from_nvfp(nv_file_private_t

 #define NV_STATE_PTR(nvl)   &(((nv_linux_state_t *)(nvl))->nv_state)

-static inline nvidia_stack_t *nv_nvlfp_get_sp(nv_linux_file_private_t *nvlfp, nvidia_entry_point_index_t which)
-{
-#if defined(NVCPU_X86_64)
-    if (rm_is_altstack_in_use())
-    {
-        down(&nvlfp->fops_sp_lock[which]);
-        return nvlfp->fops_sp[which];
-    }
-#endif
-    return NULL;
-}
-
-static inline void nv_nvlfp_put_sp(nv_linux_file_private_t *nvlfp, nvidia_entry_point_index_t which)
-{
-#if defined(NVCPU_X86_64)
-    if (rm_is_altstack_in_use())
-    {
-        up(&nvlfp->fops_sp_lock[which]);
-    }
-#endif
-}
-
 #define NV_ATOMIC_READ(data)            atomic_read(&(data))
 #define NV_ATOMIC_SET(data,val)         atomic_set(&(data), (val))
 #define NV_ATOMIC_INC(data)             atomic_inc(&(data))
@@ -1783,12 +1803,18 @@ static inline NV_STATUS nv_check_gpu_state(nv_state_t *nv)
 extern NvU32 NVreg_EnableUserNUMAManagement;
 extern NvU32 NVreg_RegisterPCIDriver;
 extern NvU32 NVreg_EnableResizableBar;
+extern NvU32 NVreg_EnableNonblockingOpen;

 extern NvU32 num_probed_nv_devices;
 extern NvU32 num_nv_devices;

 #define NV_FILE_INODE(file) (file)->f_inode

+static inline int nv_is_control_device(struct inode *inode)
+{
+    return (minor((inode)->i_rdev) == NV_MINOR_DEVICE_NUMBER_CONTROL_DEVICE);
+}
+
 #if defined(NV_DOM0_KERNEL_PRESENT) || defined(NV_VGPU_KVM_BUILD)
 #define NV_VGX_HYPER
 #if defined(NV_XEN_IOEMU_INJECT_MSI)
@@ -1982,31 +2008,6 @@ static inline NvBool nv_platform_use_auto_online(nv_linux_state_t *nvl)
    return nvl->numa_info.use_auto_online;
 }

-typedef struct {
-    NvU64 base;
-    NvU64 size;
-    NvU32 nodeId;
-    int ret;
-} remove_numa_memory_info_t;
-
-static void offline_numa_memory_callback
-(
-    void *args
-)
-{
-#ifdef NV_OFFLINE_AND_REMOVE_MEMORY_PRESENT
-    remove_numa_memory_info_t *pNumaInfo = (remove_numa_memory_info_t *)args;
-#ifdef NV_REMOVE_MEMORY_HAS_NID_ARG
-    pNumaInfo->ret = offline_and_remove_memory(pNumaInfo->nodeId,
-                                               pNumaInfo->base,
-                                               pNumaInfo->size);
-#else
-    pNumaInfo->ret = offline_and_remove_memory(pNumaInfo->base,
-                                               pNumaInfo->size);
-#endif
-#endif
-}
-
 typedef enum
 {
    NV_NUMA_STATUS_DISABLED             = 0,
@@ -2067,4 +2068,7 @@ typedef enum
 #include <linux/clk-provider.h>
 #endif

+#define NV_EXPORT_SYMBOL(symbol)        EXPORT_SYMBOL_GPL(symbol)
+#define NV_CHECK_EXPORT_SYMBOL(symbol)  NV_IS_EXPORT_SYMBOL_PRESENT_##symbol
+
 #endif  /* _NV_LINUX_H_ */
--- a/kernel-open/common/inc/nv-lock.h
+++ b/kernel-open/common/inc/nv-lock.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2017 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2017-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -37,6 +37,7 @@

 #if defined(CONFIG_PREEMPT_RT) || defined(CONFIG_PREEMPT_RT_FULL)
 typedef raw_spinlock_t            nv_spinlock_t;
+#define NV_DEFINE_SPINLOCK(lock)  DEFINE_RAW_SPINLOCK(lock)
 #define NV_SPIN_LOCK_INIT(lock)   raw_spin_lock_init(lock)
 #define NV_SPIN_LOCK_IRQ(lock)    raw_spin_lock_irq(lock)
 #define NV_SPIN_UNLOCK_IRQ(lock)  raw_spin_unlock_irq(lock)
@@ -47,6 +48,7 @@ typedef raw_spinlock_t            nv_spinlock_t;
 #define NV_SPIN_UNLOCK_WAIT(lock) raw_spin_unlock_wait(lock)
 #else
 typedef spinlock_t                nv_spinlock_t;
+#define NV_DEFINE_SPINLOCK(lock)  DEFINE_SPINLOCK(lock)
 #define NV_SPIN_LOCK_INIT(lock)   spin_lock_init(lock)
 #define NV_SPIN_LOCK_IRQ(lock)    spin_lock_irq(lock)
 #define NV_SPIN_UNLOCK_IRQ(lock)  spin_unlock_irq(lock)
--- a/kernel-open/common/inc/nv-mm.h
+++ b/kernel-open/common/inc/nv-mm.h
@@ -29,27 +29,33 @@
 typedef int vm_fault_t;
 #endif

-/* pin_user_pages
+/*
+ * pin_user_pages()
+ *
 * Presence of pin_user_pages() also implies the presence of unpin-user_page().
- * Both were added in the v5.6-rc1
+ * Both were added in the v5.6.
 *
- * pin_user_pages() was added by commit eddb1c228f7951d399240
- * ("mm/gup: introduce pin_user_pages*() and FOLL_PIN") in v5.6-rc1 (2020-01-30)
- *
- * Removed vmas parameter from pin_user_pages() by commit 40896a02751
- * ("mm/gup: remove vmas parameter from pin_user_pages()")
- * in linux-next, expected in v6.5-rc1 (2023-05-17)
+ * pin_user_pages() was added by commit eddb1c228f79
+ * ("mm/gup: introduce pin_user_pages*() and FOLL_PIN") in v5.6.
 *
+ * Removed vmas parameter from pin_user_pages() by commit 4c630f307455
+ * ("mm/gup: remove vmas parameter from pin_user_pages()") in v6.5.
 */

 #include <linux/mm.h>
 #include <linux/sched.h>
-#if defined(NV_PIN_USER_PAGES_PRESENT)
+
+/*
+ * FreeBSD's pin_user_pages's conftest breaks since pin_user_pages is an inline
+ * function. Because it simply maps to get_user_pages, we can just replace
+ * NV_PIN_USER_PAGES with NV_GET_USER_PAGES on FreeBSD
+ */
+#if defined(NV_PIN_USER_PAGES_PRESENT) && !defined(NV_BSD)
    #if defined(NV_PIN_USER_PAGES_HAS_ARGS_VMAS)
-        #define NV_PIN_USER_PAGES pin_user_pages
+        #define NV_PIN_USER_PAGES(start, nr_pages, gup_flags, pages) \
+            pin_user_pages(start, nr_pages, gup_flags, pages, NULL)
    #else
-        #define NV_PIN_USER_PAGES(start, nr_pages, gup_flags, pages, vmas) \
-            pin_user_pages(start, nr_pages, gup_flags, pages)
+        #define NV_PIN_USER_PAGES pin_user_pages
    #endif // NV_PIN_USER_PAGES_HAS_ARGS_VMAS
    #define NV_UNPIN_USER_PAGE unpin_user_page
 #else
@@ -57,80 +63,83 @@ typedef int vm_fault_t;
    #define NV_UNPIN_USER_PAGE put_page
 #endif // NV_PIN_USER_PAGES_PRESENT

-/* get_user_pages
+/*
+ * get_user_pages()
 *
- * The 8-argument version of get_user_pages was deprecated by commit
- * (2016 Feb 12: cde70140fed8429acf7a14e2e2cbd3e329036653)for the non-remote case
+ * The 8-argument version of get_user_pages() was deprecated by commit
+ * cde70140fed8 ("mm/gup: Overload get_user_pages() functions") in v4.6-rc1.
 * (calling get_user_pages with current and current->mm).
 *
- * Completely moved to the 6 argument version of get_user_pages -
- * 2016 Apr 4: c12d2da56d0e07d230968ee2305aaa86b93a6832
+ * Completely moved to the 6 argument version of get_user_pages() by
+ * commit c12d2da56d0e ("mm/gup: Remove the macro overload API migration
+ * helpers from the get_user*() APIs") in v4.6-rc4.
 *
- * write and force parameters were replaced with gup_flags by -
- * 2016 Oct 12: 768ae309a96103ed02eb1e111e838c87854d8b51
+ * write and force parameters were replaced with gup_flags by
+ * commit 768ae309a961 ("mm: replace get_user_pages() write/force parameters
+ * with gup_flags") in v4.9.
 *
 * A 7-argument version of get_user_pages was introduced into linux-4.4.y by
- * commit 8e50b8b07f462ab4b91bc1491b1c91bd75e4ad40 which cherry-picked the
- * replacement of the write and force parameters with gup_flags
+ * commit 8e50b8b07f462 ("mm: replace get_user_pages() write/force parameters
+ * with gup_flags") which cherry-picked the replacement of the write and
+ * force parameters with gup_flags.
 *
- * Removed vmas parameter from get_user_pages() by commit 7bbf9c8c99
- * ("mm/gup: remove unused vmas parameter from get_user_pages()")
- * in linux-next, expected in v6.5-rc1 (2023-05-17)
+ * Removed vmas parameter from get_user_pages() by commit 54d020692b34
+ * ("mm/gup: remove unused vmas parameter from get_user_pages()") in v6.5.
 *
 */

 #if defined(NV_GET_USER_PAGES_HAS_ARGS_FLAGS)
-    #define NV_GET_USER_PAGES(start, nr_pages, flags, pages, vmas) \
-        get_user_pages(start, nr_pages, flags, pages)
-#elif defined(NV_GET_USER_PAGES_HAS_ARGS_FLAGS_VMAS)
    #define NV_GET_USER_PAGES get_user_pages
+#elif defined(NV_GET_USER_PAGES_HAS_ARGS_FLAGS_VMAS)
+    #define NV_GET_USER_PAGES(start, nr_pages, flags, pages) \
+        get_user_pages(start, nr_pages, flags, pages, NULL)
 #elif defined(NV_GET_USER_PAGES_HAS_ARGS_TSK_FLAGS_VMAS)
-    #define NV_GET_USER_PAGES(start, nr_pages, flags, pages, vmas) \
-        get_user_pages(current, current->mm, start, nr_pages, flags, pages, vmas)
+    #define NV_GET_USER_PAGES(start, nr_pages, flags, pages) \
+        get_user_pages(current, current->mm, start, nr_pages, flags, pages, NULL)
 #else
    static inline long NV_GET_USER_PAGES(unsigned long start,
                                         unsigned long nr_pages,
                                         unsigned int flags,
-                                         struct page **pages,
-                                         struct vm_area_struct **vmas)
+                                         struct page **pages)
    {
        int write = flags & FOLL_WRITE;
        int force = flags & FOLL_FORCE;

    #if defined(NV_GET_USER_PAGES_HAS_ARGS_WRITE_FORCE_VMAS)
-        return get_user_pages(start, nr_pages, write, force, pages, vmas);
+        return get_user_pages(start, nr_pages, write, force, pages, NULL);
    #else
        // NV_GET_USER_PAGES_HAS_ARGS_TSK_WRITE_FORCE_VMAS
        return get_user_pages(current, current->mm, start, nr_pages, write,
-                              force, pages, vmas);
+                              force, pages, NULL);
    #endif // NV_GET_USER_PAGES_HAS_ARGS_WRITE_FORCE_VMAS
    }
 #endif // NV_GET_USER_PAGES_HAS_ARGS_FLAGS

-/* pin_user_pages_remote
+/*
+ * pin_user_pages_remote()
 *
- * pin_user_pages_remote() was added by commit eddb1c228f7951d399240
- * ("mm/gup: introduce pin_user_pages*() and FOLL_PIN") in v5.6 (2020-01-30)
+ * pin_user_pages_remote() was added by commit eddb1c228f79
+ * ("mm/gup: introduce pin_user_pages*() and FOLL_PIN") in v5.6.
 *
 * pin_user_pages_remote() removed 'tsk' parameter by commit
- * 64019a2e467a ("mm/gup: remove task_struct pointer for  all gup code")
- * in v5.9-rc1 (2020-08-11). *
+ * 64019a2e467a ("mm/gup: remove task_struct pointer for all gup code")
+ * in v5.9.
 *
 * Removed unused vmas parameter from pin_user_pages_remote() by commit
- * 83bcc2e132("mm/gup: remove unused vmas parameter from pin_user_pages_remote()")
- * in linux-next, expected in v6.5-rc1 (2023-05-14)
+ * 0b295316b3a9 ("mm/gup: remove unused vmas parameter from
+ * pin_user_pages_remote()") in v6.5.
 *
 */

 #if defined(NV_PIN_USER_PAGES_REMOTE_PRESENT)
    #if defined(NV_PIN_USER_PAGES_REMOTE_HAS_ARGS_TSK_VMAS)
-        #define NV_PIN_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, vmas, locked) \
-            pin_user_pages_remote(NULL, mm, start, nr_pages, flags, pages, vmas, locked)
+        #define NV_PIN_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, locked) \
+            pin_user_pages_remote(NULL, mm, start, nr_pages, flags, pages, NULL, locked)
    #elif defined(NV_PIN_USER_PAGES_REMOTE_HAS_ARGS_VMAS)
-        #define NV_PIN_USER_PAGES_REMOTE pin_user_pages_remote
+        #define NV_PIN_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, locked) \
+            pin_user_pages_remote(mm, start, nr_pages, flags, pages, NULL, locked)
    #else
-        #define NV_PIN_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, vmas, locked) \
-            pin_user_pages_remote(mm, start, nr_pages, flags, pages, locked)
+        #define NV_PIN_USER_PAGES_REMOTE pin_user_pages_remote
    #endif // NV_PIN_USER_PAGES_REMOTE_HAS_ARGS_TSK_VMAS
 #else
    #define NV_PIN_USER_PAGES_REMOTE NV_GET_USER_PAGES_REMOTE
@@ -138,7 +147,7 @@ typedef int vm_fault_t;

 /*
 * get_user_pages_remote() was added by commit 1e9877902dc7
- * ("mm/gup: Introduce get_user_pages_remote()") in v4.6 (2016-02-12).
+ * ("mm/gup: Introduce get_user_pages_remote()") in v4.6.
 *
 * Note that get_user_pages_remote() requires the caller to hold a reference on
 * the task_struct (if non-NULL and if this API has tsk argument) and the mm_struct.
@@ -148,37 +157,35 @@ typedef int vm_fault_t;
 *
 * get_user_pages_remote() write/force parameters were replaced
 * with gup_flags by commit 9beae1ea8930 ("mm: replace get_user_pages_remote()
- * write/force parameters with gup_flags") in v4.9 (2016-10-13).
+ * write/force parameters with gup_flags") in v4.9.
 *
 * get_user_pages_remote() added 'locked' parameter by commit 5b56d49fc31d
- * ("mm: add locked parameter to get_user_pages_remote()") in
- * v4.10 (2016-12-14).
+ * ("mm: add locked parameter to get_user_pages_remote()") in v4.10.
 *
 * get_user_pages_remote() removed 'tsk' parameter by
 * commit 64019a2e467a ("mm/gup: remove task_struct pointer for
- * all gup code") in v5.9-rc1 (2020-08-11).
+ * all gup code") in v5.9.
 *
- * Removed vmas parameter from get_user_pages_remote() by commit a4bde14d549 
- * ("mm/gup: remove vmas parameter from get_user_pages_remote()")
- * in linux-next, expected in v6.5-rc1 (2023-05-14)
+ * Removed vmas parameter from get_user_pages_remote() by commit ca5e863233e8
+ * ("mm/gup: remove vmas parameter from get_user_pages_remote()") in v6.5.
 *
 */

 #if defined(NV_GET_USER_PAGES_REMOTE_PRESENT)
    #if defined(NV_GET_USER_PAGES_REMOTE_HAS_ARGS_FLAGS_LOCKED)
-        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, vmas, locked) \
-            get_user_pages_remote(mm, start, nr_pages, flags, pages, locked)
-
-    #elif defined(NV_GET_USER_PAGES_REMOTE_HAS_ARGS_FLAGS_LOCKED_VMAS)
        #define NV_GET_USER_PAGES_REMOTE get_user_pages_remote

+    #elif defined(NV_GET_USER_PAGES_REMOTE_HAS_ARGS_FLAGS_LOCKED_VMAS)
+        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, locked) \
+            get_user_pages_remote(mm, start, nr_pages, flags, pages, NULL, locked)
+
    #elif defined(NV_GET_USER_PAGES_REMOTE_HAS_ARGS_TSK_FLAGS_LOCKED_VMAS)
-        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, vmas, locked) \
-            get_user_pages_remote(NULL, mm, start, nr_pages, flags, pages, vmas, locked)
+        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, locked) \
+            get_user_pages_remote(NULL, mm, start, nr_pages, flags, pages, NULL, locked)

    #elif defined(NV_GET_USER_PAGES_REMOTE_HAS_ARGS_TSK_FLAGS_VMAS)
-        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, vmas, locked) \
-            get_user_pages_remote(NULL, mm, start, nr_pages, flags, pages, vmas)
+        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, locked) \
+            get_user_pages_remote(NULL, mm, start, nr_pages, flags, pages, NULL)

    #else
        // NV_GET_USER_PAGES_REMOTE_HAS_ARGS_TSK_WRITE_FORCE_VMAS
@@ -187,14 +194,13 @@ typedef int vm_fault_t;
                                                    unsigned long nr_pages,
                                                    unsigned int flags,
                                                    struct page **pages,
-                                                    struct vm_area_struct **vmas,
                                                    int *locked)
        {
            int write = flags & FOLL_WRITE;
            int force = flags & FOLL_FORCE;

            return get_user_pages_remote(NULL, mm, start, nr_pages, write, force,
-                                         pages, vmas);
+                                         pages, NULL);
        }
    #endif // NV_GET_USER_PAGES_REMOTE_HAS_ARGS_FLAGS_LOCKED
 #else
@@ -204,18 +210,17 @@ typedef int vm_fault_t;
                                                    unsigned long nr_pages,
                                                    unsigned int flags,
                                                    struct page **pages,
-                                                    struct vm_area_struct **vmas,
                                                    int *locked)
        {
            int write = flags & FOLL_WRITE;
            int force = flags & FOLL_FORCE;

-            return get_user_pages(NULL, mm, start, nr_pages, write, force, pages, vmas);
+            return get_user_pages(NULL, mm, start, nr_pages, write, force, pages, NULL);
        }

    #else
-        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, vmas, locked) \
-            get_user_pages(NULL, mm, start, nr_pages, flags, pages, vmas)
+        #define NV_GET_USER_PAGES_REMOTE(mm, start, nr_pages, flags, pages, locked) \
+            get_user_pages(NULL, mm, start, nr_pages, flags, pages, NULL)
    #endif // NV_GET_USER_PAGES_HAS_ARGS_TSK_WRITE_FORCE_VMAS
 #endif // NV_GET_USER_PAGES_REMOTE_PRESENT

--- a/kernel-open/common/inc/nv-pgprot.h
+++ b/kernel-open/common/inc/nv-pgprot.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2015 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2015-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -60,6 +60,7 @@ static inline pgprot_t pgprot_modify_writecombine(pgprot_t old_prot)
 #endif /* !defined(NV_VMWARE) */

 #if defined(NVCPU_AARCH64)
+extern NvBool nvos_is_chipset_io_coherent(void);
 /*
 * Don't rely on the kernel's definition of pgprot_noncached(), as on 64-bit
 * ARM that's not for system memory, but device memory instead. For I/O cache
@@ -119,6 +120,13 @@ static inline pgprot_t pgprot_modify_writecombine(pgprot_t old_prot)
 #define NV_PGPROT_WRITE_COMBINED(old_prot)    old_prot
 #define NV_PGPROT_READ_ONLY(old_prot)                                         \
    __pgprot(pgprot_val((old_prot)) & ~NV_PAGE_RW)
+#elif defined(NVCPU_RISCV64)
+#define NV_PGPROT_WRITE_COMBINED_DEVICE(old_prot)                             \
+    pgprot_writecombine(old_prot)
+/* Don't attempt to mark sysmem pages as write combined on riscv */
+#define NV_PGPROT_WRITE_COMBINED(old_prot)     old_prot
+#define NV_PGPROT_READ_ONLY(old_prot)                                         \
+            __pgprot(pgprot_val((old_prot)) & ~_PAGE_WRITE)
 #else
 /* Writecombine is not supported */
 #undef NV_PGPROT_WRITE_COMBINED_DEVICE(old_prot)
--- a/kernel-open/common/inc/nv-procfs-utils.h
+++ b/kernel-open/common/inc/nv-procfs-utils.h
@@ -92,6 +92,24 @@ typedef struct file_operations nv_proc_ops_t;
 #endif

 #define NV_DEFINE_SINGLE_PROCFS_FILE_HELPER(name, lock)                     \
+    static ssize_t nv_procfs_read_lock_##name(                              \
+        struct file *file,                                                  \
+        char __user *buf,                                                   \
+        size_t size,                                                        \
+        loff_t *ppos                                                        \
+    )                                                                       \
+    {                                                                       \
+        int ret;                                                            \
+        ret = nv_down_read_interruptible(&lock);                            \
+        if (ret < 0)                                                        \
+        {                                                                   \
+            return ret;                                                     \
+        }                                                                   \
+        size = seq_read(file, buf, size, ppos);                             \
+        up_read(&lock);                                                     \
+        return size;                                                        \
+    }                                                                       \
+                                                                            \
    static int nv_procfs_open_##name(                                       \
        struct inode *inode,                                                \
        struct file *filep                                                  \
@@ -104,11 +122,6 @@ typedef struct file_operations nv_proc_ops_t;
        {                                                                   \
            return ret;                                                     \
        }                                                                   \
-        ret = nv_down_read_interruptible(&lock);                            \
-        if (ret < 0)                                                        \
-        {                                                                   \
-            single_release(inode, filep);                                   \
-        }                                                                   \
        return ret;                                                         \
    }                                                                       \
                                                                            \
@@ -117,7 +130,6 @@ typedef struct file_operations nv_proc_ops_t;
        struct file *filep                                                  \
    )                                                                       \
    {                                                                       \
-        up_read(&lock);                                                     \
        return single_release(inode, filep);                                \
    }

@@ -127,46 +139,7 @@ typedef struct file_operations nv_proc_ops_t;
    static const nv_proc_ops_t nv_procfs_##name##_fops = {                  \
        NV_PROC_OPS_SET_OWNER()                                             \
        .NV_PROC_OPS_OPEN    = nv_procfs_open_##name,                       \
-        .NV_PROC_OPS_READ    = seq_read,                                    \
-        .NV_PROC_OPS_LSEEK   = seq_lseek,                                   \
-        .NV_PROC_OPS_RELEASE = nv_procfs_release_##name,                    \
-    };
-
-
-#define NV_DEFINE_SINGLE_PROCFS_FILE_READ_WRITE(name, lock,                 \
-write_callback)                                                             \
-    NV_DEFINE_SINGLE_PROCFS_FILE_HELPER(name, lock)                         \
-                                                                            \
-    static ssize_t nv_procfs_write_##name(                                  \
-        struct file *file,                                                  \
-        const char __user *buf,                                             \
-        size_t size,                                                        \
-        loff_t *ppos                                                        \
-    )                                                                       \
-    {                                                                       \
-        ssize_t ret;                                                        \
-        struct seq_file *s;                                                 \
-                                                                            \
-        s = file->private_data;                                             \
-        if (s == NULL)                                                      \
-        {                                                                   \
-            return -EIO;                                                    \
-        }                                                                   \
-                                                                            \
-        ret = write_callback(s, buf + *ppos, size - *ppos);                 \
-        if (ret == 0)                                                       \
-        {                                                                   \
-            /* avoid infinite loop */                                       \
-            ret = -EIO;                                                     \
-        }                                                                   \
-        return ret;                                                         \
-    }                                                                       \
-                                                                            \
-    static const nv_proc_ops_t nv_procfs_##name##_fops = {                  \
-        NV_PROC_OPS_SET_OWNER()                                             \
-        .NV_PROC_OPS_OPEN    = nv_procfs_open_##name,                       \
-        .NV_PROC_OPS_READ    = seq_read,                                    \
-        .NV_PROC_OPS_WRITE   = nv_procfs_write_##name,                      \
+        .NV_PROC_OPS_READ    = nv_procfs_read_lock_##name,                  \
        .NV_PROC_OPS_LSEEK   = seq_lseek,                                   \
        .NV_PROC_OPS_RELEASE = nv_procfs_release_##name,                    \
    };
--- a/kernel-open/common/inc/nv-proto.h
+++ b/kernel-open/common/inc/nv-proto.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 1999-2021 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 1999-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -25,10 +25,8 @@
 #define _NV_PROTO_H_

 #include "nv-pci.h"
-#include "nv-register-module.h"

 extern const char *nv_device_name;
-extern nvidia_module_t nv_fops;

 void        nv_acpi_register_notifier   (nv_linux_state_t *);
 void        nv_acpi_unregister_notifier (nv_linux_state_t *);
@@ -86,8 +84,11 @@ void          nv_shutdown_adapter(nvidia_stack_t *, nv_state_t *, nv_linux_state
 void          nv_dev_free_stacks(nv_linux_state_t *);
 NvBool        nv_lock_init_locks(nvidia_stack_t *, nv_state_t *);
 void          nv_lock_destroy_locks(nvidia_stack_t *, nv_state_t *);
-void          nv_linux_add_device_locked(nv_linux_state_t *);
+int           nv_linux_add_device_locked(nv_linux_state_t *);
 void          nv_linux_remove_device_locked(nv_linux_state_t *);
 NvBool        nv_acpi_power_resource_method_present(struct pci_dev *);

+int           nv_linux_init_open_q(nv_linux_state_t *);
+void          nv_linux_stop_open_q(nv_linux_state_t *);
+
 #endif /* _NV_PROTO_H_ */
--- a/kernel-open/common/inc/nv.h
+++ b/kernel-open/common/inc/nv.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 1999-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 1999-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -42,6 +42,7 @@
 #include <nv-caps.h>
 #include <nv-firmware.h>
 #include <nv-ioctl.h>
+#include <nv-ioctl-numa.h>
 #include <nvmisc.h>

 extern nv_cap_t *nvidia_caps_root;
@@ -50,9 +51,6 @@ extern const NvBool nv_is_rm_firmware_supported_os;

 #include <nv-kernel-interface-api.h>

-/* NVIDIA's reserved major character device number (Linux). */
-#define NV_MAJOR_DEVICE_NUMBER 195
-
 #define GPU_UUID_LEN    (16)

 /*
@@ -223,7 +221,6 @@ typedef struct
 #define NV_RM_PAGE_MASK     (NV_RM_PAGE_SIZE - 1)

 #define NV_RM_TO_OS_PAGE_SHIFT      (os_page_shift - NV_RM_PAGE_SHIFT)
-#define NV_RM_PAGES_PER_OS_PAGE     (1U << NV_RM_TO_OS_PAGE_SHIFT)
 #define NV_RM_PAGES_TO_OS_PAGES(count) \
    ((((NvUPtr)(count)) >> NV_RM_TO_OS_PAGE_SHIFT) + \
     ((((count) & ((1 << NV_RM_TO_OS_PAGE_SHIFT) - 1)) != 0) ? 1 : 0))
@@ -469,17 +466,9 @@ typedef struct nv_state_t
        NvHandle hDisp;
    } rmapi;

-    /* Bool to check if ISO iommu enabled */
-    NvBool iso_iommu_present;
-
-    /* Bool to check if NISO iommu enabled */
-    NvBool niso_iommu_present;
-
    /* Bool to check if dma-buf is supported */
    NvBool dma_buf_supported;

-    NvBool printed_openrm_enable_unsupported_gpus_error;
-
    /* Check if NVPCF DSM function is implemented under NVPCF or GPU device scope */
    NvBool nvpcf_dsm_in_gpu_scope;

@@ -488,6 +477,22 @@ typedef struct nv_state_t

    /* Bool to check if the GPU has a coherent sysmem link */
    NvBool coherent;
+
+    /*
+     * NUMA node ID of the CPU to which the GPU is attached.
+     * Holds NUMA_NO_NODE on platforms that don't support NUMA configuration.
+     */
+    NvS32 cpu_numa_node_id;
+
+    struct {
+        /* Bool to check if ISO iommu enabled */
+        NvBool iso_iommu_present;
+        /* Bool to check if NISO iommu enabled */
+        NvBool niso_iommu_present;
+        /* Display SMMU Stream IDs */
+        NvU32 dispIsoStreamId;
+        NvU32 dispNisoStreamId;
+    } iommus;
 } nv_state_t;

 // These define need to be in sync with defines in system.h
@@ -505,6 +510,7 @@ struct nv_file_private_t
    NvHandle *handles;
    NvU16 maxHandles;
    NvU32 deviceInstance;
+    NvU32 gpuInstanceId;
    NvU8 metadata[64];

    nv_file_private_t *ctl_nvfp;
@@ -603,6 +609,15 @@ typedef enum
    NV_POWER_STATE_RUNNING
 } nv_power_state_t;

+typedef struct
+{
+    const char *vidmem_power_status;
+    const char *dynamic_power_status;
+    const char *gc6_support;
+    const char *gcoff_support;
+    const char *s0ix_status;
+} nv_power_info_t;
+
 #define NV_PRIMARY_VGA(nv)      ((nv)->primary_vga)

 #define NV_IS_CTL_DEVICE(nv)    ((nv)->flags & NV_FLAG_CONTROL)
@@ -615,11 +630,19 @@ typedef enum
 #define NV_IS_DEVICE_IN_SURPRISE_REMOVAL(nv)    \
        (((nv)->flags & NV_FLAG_IN_SURPRISE_REMOVAL) != 0)

+/*
+ * For console setup by EFI GOP, the base address is BAR1.
+ * For console setup by VBIOS, the base address is BAR2 + 16MB.
+ */
+#define NV_IS_CONSOLE_MAPPED(nv, addr)  \
+        (((addr) == (nv)->bars[NV_GPU_BAR_INDEX_FB].cpu_address) || \
+         ((addr) == ((nv)->bars[NV_GPU_BAR_INDEX_IMEM].cpu_address + 0x1000000)))
+
 #define NV_SOC_IS_ISO_IOMMU_PRESENT(nv)     \
-        ((nv)->iso_iommu_present)
+        ((nv)->iommus.iso_iommu_present)

 #define NV_SOC_IS_NISO_IOMMU_PRESENT(nv)     \
-        ((nv)->niso_iommu_present)
+        ((nv)->iommus.niso_iommu_present)
 /*
 * GPU add/remove events
 */
@@ -764,8 +787,8 @@ nv_state_t*  NV_API_CALL  nv_get_ctl_state       (void);

 void   NV_API_CALL  nv_set_dma_address_size      (nv_state_t *, NvU32 );

-NV_STATUS  NV_API_CALL  nv_alias_pages           (nv_state_t *, NvU32, NvU32, NvU32, NvU64, NvU64 *, void **);
-NV_STATUS  NV_API_CALL  nv_alloc_pages           (nv_state_t *, NvU32, NvBool, NvU32, NvBool, NvBool, NvS32, NvU64 *, void **);
+NV_STATUS  NV_API_CALL  nv_alias_pages           (nv_state_t *, NvU32, NvU64, NvU32, NvU32, NvU64, NvU64 *, void **);
+NV_STATUS  NV_API_CALL  nv_alloc_pages           (nv_state_t *, NvU32, NvU64, NvBool, NvU32, NvBool, NvBool, NvS32, NvU64 *, void **);
 NV_STATUS  NV_API_CALL  nv_free_pages            (nv_state_t *, NvU32, NvBool, NvU32, void *);

 NV_STATUS  NV_API_CALL  nv_register_user_pages   (nv_state_t *, NvU64, NvU64 *, void *, void **);
@@ -782,8 +805,6 @@ NV_STATUS NV_API_CALL   nv_register_phys_pages   (nv_state_t *, NvU64 *, NvU64,
 void      NV_API_CALL   nv_unregister_phys_pages (nv_state_t *, void *);

 NV_STATUS  NV_API_CALL  nv_dma_map_sgt           (nv_dma_device_t *, NvU64, NvU64 *, NvU32, void **);
-NV_STATUS  NV_API_CALL  nv_dma_map_pages         (nv_dma_device_t *, NvU64, NvU64 *, NvBool, NvU32, void **);
-NV_STATUS  NV_API_CALL  nv_dma_unmap_pages       (nv_dma_device_t *, NvU64, NvU64 *, void **);

 NV_STATUS  NV_API_CALL  nv_dma_map_alloc         (nv_dma_device_t *, NvU64, NvU64 *, NvBool, void **);
 NV_STATUS  NV_API_CALL  nv_dma_unmap_alloc       (nv_dma_device_t *, NvU64, NvU64 *, void **);
@@ -810,6 +831,7 @@ void   NV_API_CALL  nv_acpi_methods_init         (NvU32 *);
 void   NV_API_CALL  nv_acpi_methods_uninit       (void);

 NV_STATUS  NV_API_CALL  nv_acpi_method           (NvU32, NvU32, NvU32, void *, NvU16, NvU32 *, void *, NvU16 *);
+NV_STATUS  NV_API_CALL  nv_acpi_d3cold_dsm_for_upstream_port (nv_state_t *, NvU8 *, NvU32, NvU32, NvU32 *);
 NV_STATUS  NV_API_CALL  nv_acpi_dsm_method       (nv_state_t *, NvU8 *, NvU32, NvBool, NvU32, void *, NvU16, NvU32 *, void *, NvU16 *);
 NV_STATUS  NV_API_CALL  nv_acpi_ddc_method       (nv_state_t *, void *, NvU32 *, NvBool);
 NV_STATUS  NV_API_CALL  nv_acpi_dod_method       (nv_state_t *, NvU32 *, NvU32 *);
@@ -833,7 +855,7 @@ void       NV_API_CALL  nv_put_firmware(const void *);
 nv_file_private_t* NV_API_CALL nv_get_file_private(NvS32, NvBool, void **);
 void               NV_API_CALL nv_put_file_private(void *);

-NV_STATUS NV_API_CALL nv_get_device_memory_config(nv_state_t *, NvU64 *, NvU64 *, NvU32 *, NvS32 *);
+NV_STATUS NV_API_CALL nv_get_device_memory_config(nv_state_t *, NvU64 *, NvU64 *, NvU64 *, NvU32 *, NvS32 *);
 NV_STATUS NV_API_CALL nv_get_egm_info(nv_state_t *, NvU64 *, NvU64 *, NvS32 *);

 NV_STATUS NV_API_CALL nv_get_ibmnpu_genreg_info(nv_state_t *, NvU64 *, NvU64 *, void**);
@@ -874,15 +896,17 @@ NvBool    NV_API_CALL nv_match_gpu_os_info(nv_state_t *, void *);
 NvU32     NV_API_CALL nv_get_os_type(void);

 void      NV_API_CALL nv_get_updated_emu_seg(NvU32 *start, NvU32 *end);
+void      NV_API_CALL nv_get_screen_info(nv_state_t *, NvU64 *, NvU32 *, NvU32 *, NvU32 *, NvU32 *, NvU64 *);
+
 struct dma_buf;
 typedef struct nv_dma_buf nv_dma_buf_t;
 struct drm_gem_object;

 NV_STATUS NV_API_CALL nv_dma_import_sgt  (nv_dma_device_t *, struct sg_table *, struct drm_gem_object *);
 void NV_API_CALL nv_dma_release_sgt(struct sg_table *, struct drm_gem_object *);
-NV_STATUS NV_API_CALL nv_dma_import_dma_buf      (nv_dma_device_t *, struct dma_buf *, NvU32 *, void **, struct sg_table **, nv_dma_buf_t **);
-NV_STATUS NV_API_CALL nv_dma_import_from_fd      (nv_dma_device_t *, NvS32, NvU32 *, void **, struct sg_table **, nv_dma_buf_t **);
-void      NV_API_CALL nv_dma_release_dma_buf     (void *, nv_dma_buf_t *);
+NV_STATUS NV_API_CALL nv_dma_import_dma_buf      (nv_dma_device_t *, struct dma_buf *, NvU32 *, struct sg_table **, nv_dma_buf_t **);
+NV_STATUS NV_API_CALL nv_dma_import_from_fd      (nv_dma_device_t *, NvS32, NvU32 *, struct sg_table **, nv_dma_buf_t **);
+void      NV_API_CALL nv_dma_release_dma_buf     (nv_dma_buf_t *);

 void      NV_API_CALL nv_schedule_uvm_isr        (nv_state_t *);

@@ -898,6 +922,8 @@ typedef void (*nvTegraDceClientIpcCallback)(NvU32, NvU32, NvU32, void *, void *)
 NV_STATUS NV_API_CALL nv_get_num_phys_pages      (void *, NvU32 *);
 NV_STATUS NV_API_CALL nv_get_phys_pages          (void *, void *, NvU32 *);

+void      NV_API_CALL nv_get_disp_smmu_stream_ids (nv_state_t *, NvU32 *, NvU32 *);
+
 /*
 * ---------------------------------------------------------------------------
 *
@@ -924,6 +950,7 @@ NV_STATUS  NV_API_CALL  rm_ioctl                 (nvidia_stack_t *, nv_state_t *
 NvBool     NV_API_CALL  rm_isr                   (nvidia_stack_t *, nv_state_t *, NvU32 *);
 void       NV_API_CALL  rm_isr_bh                (nvidia_stack_t *, nv_state_t *);
 void       NV_API_CALL  rm_isr_bh_unlocked       (nvidia_stack_t *, nv_state_t *);
+NvBool     NV_API_CALL  rm_is_msix_allowed       (nvidia_stack_t *, nv_state_t *);
 NV_STATUS  NV_API_CALL  rm_power_management      (nvidia_stack_t *, nv_state_t *, nv_pm_action_t);
 NV_STATUS  NV_API_CALL  rm_stop_user_channels    (nvidia_stack_t *, nv_state_t *);
 NV_STATUS  NV_API_CALL  rm_restart_user_channels (nvidia_stack_t *, nv_state_t *);
@@ -943,6 +970,7 @@ void       NV_API_CALL  rm_parse_option_string   (nvidia_stack_t *, const char *
 char*      NV_API_CALL  rm_remove_spaces         (const char *);
 char*      NV_API_CALL  rm_string_token          (char **, const char);
 void       NV_API_CALL  rm_vgpu_vfio_set_driver_vm(nvidia_stack_t *, NvBool);
+NV_STATUS  NV_API_CALL  rm_get_adapter_status_external(nvidia_stack_t *, nv_state_t *);

 NV_STATUS  NV_API_CALL  rm_run_rc_callback       (nvidia_stack_t *, nv_state_t *);
 void       NV_API_CALL  rm_execute_work_item     (nvidia_stack_t *, void *);
@@ -972,16 +1000,16 @@ NV_STATUS  NV_API_CALL  rm_p2p_init_mapping       (nvidia_stack_t *, NvU64, NvU6
 NV_STATUS  NV_API_CALL  rm_p2p_destroy_mapping    (nvidia_stack_t *, NvU64);
 NV_STATUS  NV_API_CALL  rm_p2p_get_pages          (nvidia_stack_t *, NvU64, NvU32, NvU64, NvU64, NvU64 *, NvU32 *, NvU32 *, NvU32 *, NvU8 **, void *);
 NV_STATUS  NV_API_CALL  rm_p2p_get_gpu_info       (nvidia_stack_t *, NvU64, NvU64, NvU8 **, void **);
-NV_STATUS  NV_API_CALL  rm_p2p_get_pages_persistent (nvidia_stack_t *,  NvU64, NvU64, void **, NvU64 *, NvU32 *, void *, void *);
+NV_STATUS  NV_API_CALL  rm_p2p_get_pages_persistent (nvidia_stack_t *,  NvU64, NvU64, void **, NvU64 *, NvU32 *, void *, void *, void **);
 NV_STATUS  NV_API_CALL  rm_p2p_register_callback  (nvidia_stack_t *, NvU64, NvU64, NvU64, void *, void (*)(void *), void *);
 NV_STATUS  NV_API_CALL  rm_p2p_put_pages          (nvidia_stack_t *, NvU64, NvU32, NvU64, void *);
-NV_STATUS  NV_API_CALL  rm_p2p_put_pages_persistent(nvidia_stack_t *, void *, void *);
+NV_STATUS  NV_API_CALL  rm_p2p_put_pages_persistent(nvidia_stack_t *, void *, void *, void *);
 NV_STATUS  NV_API_CALL  rm_p2p_dma_map_pages      (nvidia_stack_t *, nv_dma_device_t *, NvU8 *, NvU64, NvU32, NvU64 *, void **);
 NV_STATUS  NV_API_CALL  rm_dma_buf_dup_mem_handle (nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle, NvHandle, NvHandle, void *, NvHandle, NvU64, NvU64, NvHandle *, void **);
 void       NV_API_CALL  rm_dma_buf_undup_mem_handle(nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle);
 NV_STATUS  NV_API_CALL  rm_dma_buf_map_mem_handle (nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle, NvU64, NvU64, void *, nv_phys_addr_range_t **, NvU32 *);
 void       NV_API_CALL  rm_dma_buf_unmap_mem_handle(nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle, NvU64, nv_phys_addr_range_t **, NvU32);
-NV_STATUS  NV_API_CALL  rm_dma_buf_get_client_and_device(nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle *, NvHandle *, NvHandle *, void **, NvBool *);
+NV_STATUS  NV_API_CALL  rm_dma_buf_get_client_and_device(nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle, NvHandle *, NvHandle *, NvHandle *, void **, NvBool *);
 void       NV_API_CALL  rm_dma_buf_put_client_and_device(nvidia_stack_t *, nv_state_t *, NvHandle, NvHandle, NvHandle, void *);
 NV_STATUS  NV_API_CALL  rm_log_gpu_crash          (nv_stack_t *, nv_state_t *);

@@ -993,7 +1021,7 @@ NvBool     NV_API_CALL rm_gpu_need_4k_page_isolation(nv_state_t *);
 NvBool     NV_API_CALL rm_is_chipset_io_coherent(nv_stack_t *);
 NvBool     NV_API_CALL rm_init_event_locks(nvidia_stack_t *, nv_state_t *);
 void       NV_API_CALL rm_destroy_event_locks(nvidia_stack_t *, nv_state_t *);
-NV_STATUS  NV_API_CALL rm_get_gpu_numa_info(nvidia_stack_t *, nv_state_t *, NvS32 *, NvU64 *, NvU64 *, NvU64 *, NvU32 *);
+NV_STATUS  NV_API_CALL rm_get_gpu_numa_info(nvidia_stack_t *, nv_state_t *, nv_ioctl_numa_info_t *);
 NV_STATUS  NV_API_CALL rm_gpu_numa_online(nvidia_stack_t *, nv_state_t *);
 NV_STATUS  NV_API_CALL rm_gpu_numa_offline(nvidia_stack_t *, nv_state_t *);
 NvBool     NV_API_CALL rm_is_device_sequestered(nvidia_stack_t *, nv_state_t *);
@@ -1008,10 +1036,8 @@ void       NV_API_CALL rm_cleanup_dynamic_power_management(nvidia_stack_t *, nv_
 void       NV_API_CALL rm_enable_dynamic_power_management(nvidia_stack_t *, nv_state_t *);
 NV_STATUS  NV_API_CALL rm_ref_dynamic_power(nvidia_stack_t *, nv_state_t *, nv_dynamic_power_mode_t);
 void       NV_API_CALL rm_unref_dynamic_power(nvidia_stack_t *, nv_state_t *, nv_dynamic_power_mode_t);
-NV_STATUS  NV_API_CALL rm_transition_dynamic_power(nvidia_stack_t *, nv_state_t *, NvBool);
-const char* NV_API_CALL rm_get_vidmem_power_status(nvidia_stack_t *, nv_state_t *);
-const char* NV_API_CALL rm_get_dynamic_power_management_status(nvidia_stack_t *, nv_state_t *);
-const char* NV_API_CALL rm_get_gpu_gcx_support(nvidia_stack_t *, nv_state_t *, NvBool);
+NV_STATUS  NV_API_CALL rm_transition_dynamic_power(nvidia_stack_t *, nv_state_t *, NvBool, NvBool *);
+void       NV_API_CALL rm_get_power_info(nvidia_stack_t *, nv_state_t *, nv_power_info_t *);

 void       NV_API_CALL rm_acpi_notify(nvidia_stack_t *, nv_state_t *, NvU32);
 void       NV_API_CALL rm_acpi_nvpcf_notify(nvidia_stack_t *);
@@ -1023,12 +1049,12 @@ NV_STATUS  NV_API_CALL  nv_vgpu_create_request(nvidia_stack_t *, nv_state_t *, c
 NV_STATUS  NV_API_CALL  nv_vgpu_delete(nvidia_stack_t *, const NvU8 *, NvU16);
 NV_STATUS  NV_API_CALL  nv_vgpu_get_type_ids(nvidia_stack_t *, nv_state_t *, NvU32 *, NvU32 *, NvBool, NvU8, NvBool);
 NV_STATUS  NV_API_CALL  nv_vgpu_get_type_info(nvidia_stack_t *, nv_state_t *, NvU32, char *, int, NvU8);
-NV_STATUS  NV_API_CALL  nv_vgpu_get_bar_info(nvidia_stack_t *, nv_state_t *, const NvU8 *, NvU64 *, NvU32, void *);
-NV_STATUS  NV_API_CALL  nv_vgpu_start(nvidia_stack_t *, const NvU8 *, void *, NvS32 *, NvU8 *, NvU32);
-NV_STATUS  NV_API_CALL  nv_vgpu_get_sparse_mmap(nvidia_stack_t *, nv_state_t *, const NvU8 *, NvU64 **, NvU64 **, NvU32 *);
+NV_STATUS  NV_API_CALL  nv_vgpu_get_bar_info(nvidia_stack_t *, nv_state_t *, const NvU8 *, NvU64 *,
+                                             NvU64 *, NvU64 *, NvU32 *, NvBool *, NvU8 *);
+NV_STATUS  NV_API_CALL  nv_vgpu_get_hbm_info(nvidia_stack_t *, nv_state_t *, const NvU8 *, NvU64 *, NvU64 *);
 NV_STATUS  NV_API_CALL  nv_vgpu_process_vf_info(nvidia_stack_t *, nv_state_t *, NvU8, NvU32, NvU8, NvU8, NvU8, NvBool, void *);
-NV_STATUS  NV_API_CALL  nv_vgpu_update_request(nvidia_stack_t *, const NvU8 *, NvU32, NvU64 *, NvU64 *, const char *);
 NV_STATUS  NV_API_CALL  nv_gpu_bind_event(nvidia_stack_t *);
+NV_STATUS  NV_API_CALL  nv_gpu_unbind_event(nvidia_stack_t *, NvU32, NvBool *);

 NV_STATUS NV_API_CALL nv_get_usermap_access_params(nv_state_t*, nv_usermap_access_params_t*);
 nv_soc_irq_type_t NV_API_CALL nv_get_current_irq_type(nv_state_t*);
--- a/kernel-open/common/inc/nv_speculation_barrier.h
+++ b/kernel-open/common/inc/nv_speculation_barrier.h
@@ -86,7 +86,7 @@
 /* Not currently implemented for MSVC/ARM64. See bug 3366890. */
 #   define nv_speculation_barrier()
 #   define speculation_barrier() nv_speculation_barrier()
-#elif defined(NVCPU_NVRISCV64) && NVOS_IS_LIBOS
+#elif defined(NVCPU_IS_RISCV64)
 #   define nv_speculation_barrier()
 #else
 #error "Unknown compiler/chip family"
--- a/kernel-open/common/inc/nv_uvm_interface.h
+++ b/kernel-open/common/inc/nv_uvm_interface.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2013-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2013-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -62,10 +62,10 @@ typedef struct
 /*******************************************************************************
    nvUvmInterfaceRegisterGpu

-    Registers the GPU with the provided UUID for use. A GPU must be registered
-    before its UUID can be used with any other API. This call is ref-counted so
-    every nvUvmInterfaceRegisterGpu must be paired with a corresponding
-    nvUvmInterfaceUnregisterGpu.
+    Registers the GPU with the provided physical UUID for use. A GPU must be
+    registered before its UUID can be used with any other API. This call is
+    ref-counted so every nvUvmInterfaceRegisterGpu must be paired with a
+    corresponding nvUvmInterfaceUnregisterGpu.

    You don't need to call nvUvmInterfaceSessionCreate before calling this.

@@ -79,12 +79,13 @@ NV_STATUS nvUvmInterfaceRegisterGpu(const NvProcessorUuid *gpuUuid, UvmGpuPlatfo
 /*******************************************************************************
    nvUvmInterfaceUnregisterGpu

-    Unregisters the GPU with the provided UUID. This drops the ref count from
-    nvUvmInterfaceRegisterGpu. Once the reference count goes to 0 the device may
-    no longer be accessible until the next nvUvmInterfaceRegisterGpu call. No
-    automatic resource freeing is performed, so only make the last unregister
-    call after destroying all your allocations associated with that UUID (such
-    as those from nvUvmInterfaceAddressSpaceCreate).
+    Unregisters the GPU with the provided physical UUID. This drops the ref
+    count from nvUvmInterfaceRegisterGpu. Once the reference count goes to 0
+    the device may no longer be accessible until the next
+    nvUvmInterfaceRegisterGpu call. No automatic resource freeing is performed,
+    so only make the last unregister call after destroying all your allocations
+    associated with that UUID (such as those from
+    nvUvmInterfaceAddressSpaceCreate).

    If the UUID is not found, no operation is performed.
 */
@@ -121,10 +122,10 @@ NV_STATUS nvUvmInterfaceSessionDestroy(uvmGpuSessionHandle session);
    nvUvmInterfaceDeviceCreate

    Creates a device object under the given session for the GPU with the given
-    UUID. Also creates a partition object for the device iff bCreateSmcPartition
-    is true and pGpuInfo->smcEnabled is true. pGpuInfo->smcUserClientInfo will
-    be used to determine the SMC partition in this case. A device handle is
-    returned in the device output parameter.
+    physical UUID. Also creates a partition object for the device iff
+    bCreateSmcPartition is true and pGpuInfo->smcEnabled is true.
+    pGpuInfo->smcUserClientInfo will be used to determine the SMC partition in
+    this case. A device handle is returned in the device output parameter.

    Error codes:
      NV_ERR_GENERIC
@@ -161,6 +162,7 @@ void nvUvmInterfaceDeviceDestroy(uvmGpuDeviceHandle device);
 NV_STATUS nvUvmInterfaceAddressSpaceCreate(uvmGpuDeviceHandle device,
                                           unsigned long long vaBase,
                                           unsigned long long vaSize,
+                                           NvBool enableAts,
                                           uvmGpuAddressSpaceHandle *vaSpace,
                                           UvmGpuAddressSpaceInfo *vaSpaceInfo);

@@ -422,33 +424,6 @@ NV_STATUS nvUvmInterfacePmaPinPages(void *pPma,
                                    NvU64 pageSize,
                                    NvU32 flags);

-/*******************************************************************************
-    nvUvmInterfacePmaUnpinPages
-
-    This function will unpin the physical memory allocated using PMA. The pages
-    passed as input must be already pinned, else this function will return an
-    error and rollback any change if any page is not previously marked "pinned".
-    Behaviour is undefined if any blacklisted pages are unpinned.
-
-    Arguments:
-        pPma[IN]             - Pointer to PMA object.
-        pPages[IN]           - Array of pointers, containing the PA base
-                               address of each page to be unpinned.
-        pageCount [IN]       - Number of pages required to be unpinned.
-        pageSize [IN]        - Page size of each page to be unpinned.
-
-    Error codes:
-        NV_ERR_INVALID_ARGUMENT       - Invalid input arguments.
-        NV_ERR_GENERIC                - Unexpected error. We try hard to avoid
-                                        returning this error code as is not very
-                                        informative.
-        NV_ERR_NOT_SUPPORTED          - Operation not supported on broken FB
-*/
-NV_STATUS nvUvmInterfacePmaUnpinPages(void *pPma,
-                                      NvU64 *pPages,
-                                      NvLength pageCount,
-                                      NvU64 pageSize);
-
 /*******************************************************************************
    nvUvmInterfaceMemoryFree

@@ -638,6 +613,8 @@ NV_STATUS nvUvmInterfaceQueryCopyEnginesCaps(uvmGpuDeviceHandle device,
    nvUvmInterfaceGetGpuInfo

    Return various gpu info, refer to the UvmGpuInfo struct for details.
+    The input UUID is for the physical GPU and the pGpuClientInfo identifies
+    the SMC partition if SMC is enabled and the partition exists.
    If no gpu matching the uuid is found, an error will be returned.

    On Ampere+ GPUs, pGpuClientInfo contains SMC information provided by the
@@ -645,6 +622,9 @@ NV_STATUS nvUvmInterfaceQueryCopyEnginesCaps(uvmGpuDeviceHandle device,

    Error codes:
      NV_ERR_GENERIC
+      NV_ERR_NO_MEMORY
+      NV_ERR_GPU_UUID_NOT_FOUND
+      NV_ERR_INSUFFICIENT_PERMISSIONS
      NV_ERR_INSUFFICIENT_RESOURCES
 */
 NV_STATUS nvUvmInterfaceGetGpuInfo(const NvProcessorUuid *gpuUuid,
@@ -857,7 +837,7 @@ NV_STATUS nvUvmInterfaceGetEccInfo(uvmGpuDeviceHandle device,
        UVM GPU UNLOCK

    Arguments:
-        gpuUuid[IN]          - UUID of the GPU to operate on
+        device[IN]           - Device handle associated with the gpu
        bOwnInterrupts       - Set to NV_TRUE for UVM to take ownership of the
                               replayable page fault interrupts. Set to NV_FALSE
                               to return ownership of the page fault interrupts
@@ -973,14 +953,45 @@ NV_STATUS nvUvmInterfaceGetNonReplayableFaults(UvmGpuFaultInfo *pFaultInfo,
    NOTES:
    - This function DOES NOT acquire the RM API or GPU locks. That is because
    it is called during fault servicing, which could produce deadlocks.
+    - This function should not be called when interrupts are disabled.

    Arguments:
-        device[IN]        - Device handle associated with the gpu
+        pFaultInfo[IN]        - information provided by RM for fault handling.
+                                used for obtaining the device handle without locks.
+        bCopyAndFlush[IN]     - Instructs RM to perform the flush in the Copy+Flush mode.
+                                In this mode, RM will perform a copy of the packets from
+                                the HW buffer to UVM's SW buffer as part of performing
+                                the flush. This mode gives UVM the opportunity to observe
+                                the packets contained within the HW buffer at the time
+                                of issuing the call.

    Error codes:
      NV_ERR_INVALID_ARGUMENT
 */
-NV_STATUS nvUvmInterfaceFlushReplayableFaultBuffer(uvmGpuDeviceHandle device);
+NV_STATUS nvUvmInterfaceFlushReplayableFaultBuffer(UvmGpuFaultInfo *pFaultInfo,
+                                                   NvBool bCopyAndFlush);
+
+/*******************************************************************************
+    nvUvmInterfaceTogglePrefetchFaults
+
+    This function sends an RPC to GSP in order to toggle the prefetch fault PRI.
+
+    NOTES:
+    - This function DOES NOT acquire the RM API or GPU locks. That is because
+    it is called during fault servicing, which could produce deadlocks.
+    - This function should not be called when interrupts are disabled.
+
+    Arguments:
+        pFaultInfo[IN]        - Information provided by RM for fault handling.
+                                Used for obtaining the device handle without locks.
+        bEnable[IN]           - Instructs RM whether to toggle generating faults on
+                                prefetch on/off.
+
+    Error codes:
+      NV_ERR_INVALID_ARGUMENT
+*/
+NV_STATUS nvUvmInterfaceTogglePrefetchFaults(UvmGpuFaultInfo *pFaultInfo,
+                                             NvBool bEnable);

 /*******************************************************************************
    nvUvmInterfaceInitAccessCntrInfo
@@ -1087,7 +1098,8 @@ void nvUvmInterfaceDeRegisterUvmOps(void);

    Error codes:
      NV_ERR_INVALID_ARGUMENT
-      NV_ERR_OBJECT_NOT_FOUND : If device object associated with the uuids aren't found.
+      NV_ERR_OBJECT_NOT_FOUND : If device object associated with the device
+                                handles isn't found.
 */
 NV_STATUS nvUvmInterfaceP2pObjectCreate(uvmGpuDeviceHandle device1,
                                        uvmGpuDeviceHandle device2,
@@ -1140,6 +1152,8 @@ void nvUvmInterfaceP2pObjectDestroy(uvmGpuSessionHandle session,
        NV_ERR_NOT_READY                - Returned when querying the PTEs requires a deferred setup
                                          which has not yet completed. It is expected that the caller
                                          will reattempt the call until a different code is returned.
+                                          As an example, multi-node systems which require querying
+                                          PTEs from the Fabric Manager may return this code.
 */
 NV_STATUS nvUvmInterfaceGetExternalAllocPtes(uvmGpuAddressSpaceHandle vaSpace,
                                             NvHandle hMemory,
@@ -1449,18 +1463,30 @@ NV_STATUS nvUvmInterfacePagingChannelPushStream(UvmGpuPagingChannelHandle channe
                                                NvU32 methodStreamSize);

 /*******************************************************************************
-    CSL Interface and Locking
+    nvUvmInterfaceKeyRotationChannelDisable

-    The following functions do not acquire the RM API or GPU locks and must not be called
-    concurrently with the same UvmCslContext parameter in different threads. The caller must
-    guarantee this exclusion.
+    This function notifies RM that the given channels are idle.

-    * nvUvmInterfaceCslRotateIv
-    * nvUvmInterfaceCslEncrypt
-    * nvUvmInterfaceCslDecrypt
-    * nvUvmInterfaceCslSign
-    * nvUvmInterfaceCslQueryMessagePool
-    * nvUvmInterfaceCslIncrementIv
+    This function is called after RM has notified UVM that keys need to be rotated.
+    When called RM will disable the channels, rotate their keys, and then re-enable
+    the channels.
+
+    Locking: This function acquires an API and GPU lock.
+    Memory : This function dynamically allocates memory.
+
+    Arguments:
+        channelList[IN]      - An array of channel handles whose channels are idle.
+        channelListCount[IN] - Number of channels in channelList. Its value must be
+                               greater than 0.
+
+    Error codes:
+      NV_ERR_INVALID_ARGUMENT - channelList is NULL or channeListCount is 0.
+*/
+NV_STATUS nvUvmInterfaceKeyRotationChannelDisable(uvmGpuChannelHandle channelList[],
+                                                  NvU32 channeListCount);
+
+/*******************************************************************************
+    Cryptography Services Library (CSL) Interface
 */

 /*******************************************************************************
@@ -1471,8 +1497,11 @@ NV_STATUS nvUvmInterfacePagingChannelPushStream(UvmGpuPagingChannelHandle channe
    The lifetime of the context is the same as the lifetime of the secure channel
    it is paired with.

+    Locking: This function acquires an API lock.
+    Memory : This function dynamically allocates memory.
+
    Arguments:
-        uvmCslContext[IN/OUT] - The CSL context.
+        uvmCslContext[IN/OUT] - The CSL context associated with a channel.
        channel[IN]           - Handle to a secure channel.

    Error codes:
@@ -1490,11 +1519,35 @@ NV_STATUS nvUvmInterfaceCslInitContext(UvmCslContext *uvmCslContext,

    If context is already deinitialized then function returns immediately.

+    Locking: This function does not acquire an API or GPU lock.
+    Memory : This function may free memory.
+
    Arguments:
-        uvmCslContext[IN] - The CSL context.
+        uvmCslContext[IN] - The CSL context associated with a channel.
 */
 void nvUvmInterfaceDeinitCslContext(UvmCslContext *uvmCslContext);

+/*******************************************************************************
+    nvUvmInterfaceCslUpdateContext
+
+    Updates contexts after a key rotation event and can only be called once per
+    key rotation event. Following a key rotation event, and before
+    nvUvmInterfaceCslUpdateContext is called, data encrypted by the GPU with the
+    previous key can be decrypted with nvUvmInterfaceCslDecrypt.
+
+    Locking: This function acquires an API lock.
+    Memory : This function does not dynamically allocate memory.
+
+    Arguments:
+        contextList[IN/OUT]  - An array of pointers to CSL contexts.
+        contextListCount[IN] - Number of CSL contexts in contextList. Its value
+                               must be greater than 0.
+    Error codes:
+        NV_ERR_INVALID_ARGUMENT - contextList is NULL or contextListCount is 0.
+*/
+NV_STATUS nvUvmInterfaceCslUpdateContext(UvmCslContext *contextList[],
+                                         NvU32 contextListCount);
+
 /*******************************************************************************
    nvUvmInterfaceCslRotateIv

@@ -1509,11 +1562,13 @@ void nvUvmInterfaceDeinitCslContext(UvmCslContext *uvmCslContext);
    the channel must be idle before calling this function. This function can be
    called regardless of the value of the IV's message counter.

-    See "CSL Interface and Locking" for locking requirements.
-    This function does not perform dynamic memory allocation.
+    Locking: This function attempts to acquire the GPU lock.
+             In case of failure to acquire the return code
+             is NV_ERR_STATE_IN_USE.
+    Memory : This function does not dynamically allocate memory.

 Arguments:
-        uvmCslContext[IN/OUT] - The CSL context.
+        uvmCslContext[IN/OUT] - The CSL context associated with a channel.
        operation[IN]         - Either
                                - UVM_CSL_OPERATION_ENCRYPT
                                - UVM_CSL_OPERATION_DECRYPT
@@ -1521,7 +1576,11 @@ Arguments:
    Error codes:
      NV_ERR_INSUFFICIENT_RESOURCES - The rotate operation would cause a counter
                                      to overflow.
+      NV_ERR_STATE_IN_USE           - Unable to acquire lock / resource. Caller
+                                      can retry at a later time.
      NV_ERR_INVALID_ARGUMENT       - Invalid value for operation.
+      NV_ERR_GENERIC                - A failure other than _STATE_IN_USE occurred
+                                      when attempting to acquire a lock.
 */
 NV_STATUS nvUvmInterfaceCslRotateIv(UvmCslContext *uvmCslContext,
                                    UvmCslOperation operation);
@@ -1538,11 +1597,13 @@ NV_STATUS nvUvmInterfaceCslRotateIv(UvmCslContext *uvmCslContext,
    The encryptIV can be obtained from nvUvmInterfaceCslIncrementIv.
    However, it is optional. If it is NULL, the next IV in line will be used.

-    See "CSL Interface and Locking" for locking requirements.
-    This function does not perform dynamic memory allocation.
+    Locking: This function does not acquire an API or GPU lock.
+             If called concurrently in different threads with the same UvmCslContext
+             the caller must guarantee exclusion.
+    Memory : This function does not dynamically allocate memory.

 Arguments:
-        uvmCslContext[IN/OUT] - The CSL context.
+        uvmCslContext[IN/OUT] - The CSL context associated with a channel.
        bufferSize[IN]        - Size of the input and output buffers in
                                units of bytes. Value can range from 1 byte
                                to (2^32) - 1 bytes.
@@ -1553,8 +1614,9 @@ Arguments:
                                Its size is UVM_CSL_CRYPT_AUTH_TAG_SIZE_BYTES.

    Error codes:
-      NV_ERR_INVALID_ARGUMENT       - The size of the data is 0 bytes.
-                                    - The encryptIv has already been used.
+      NV_ERR_INVALID_ARGUMENT - The CSL context is not associated with a channel.
+                              - The size of the data is 0 bytes.
+                              - The encryptIv has already been used.
 */
 NV_STATUS nvUvmInterfaceCslEncrypt(UvmCslContext *uvmCslContext,
                                   NvU32 bufferSize,
@@ -1573,8 +1635,10 @@ NV_STATUS nvUvmInterfaceCslEncrypt(UvmCslContext *uvmCslContext,
    maximized when the input and output buffers are 16-byte aligned. This is
    natural alignment for AES block.

-    See "CSL Interface and Locking" for locking requirements.
-    This function does not perform dynamic memory allocation.
+    Locking: This function does not acquire an API or GPU lock.
+             If called concurrently in different threads with the same UvmCslContext
+             the caller must guarantee exclusion.
+    Memory : This function does not dynamically allocate memory.

    Arguments:
        uvmCslContext[IN/OUT] - The CSL context.
@@ -1616,11 +1680,13 @@ NV_STATUS nvUvmInterfaceCslDecrypt(UvmCslContext *uvmCslContext,
    Auth and input buffers must not overlap. If they do then calling this function produces
    undefined behavior.

-    See "CSL Interface and Locking" for locking requirements.
-    This function does not perform dynamic memory allocation.
+    Locking: This function does not acquire an API or GPU lock.
+             If called concurrently in different threads with the same UvmCslContext
+             the caller must guarantee exclusion.
+    Memory : This function does not dynamically allocate memory.

    Arguments:
-        uvmCslContext[IN/OUT] - The CSL context.
+        uvmCslContext[IN/OUT] - The CSL context associated with a channel.
        bufferSize[IN]        - Size of the input buffer in units of bytes.
                                Value can range from 1 byte to (2^32) - 1 bytes.
        inputBuffer[IN]       - Address of plaintext input buffer.
@@ -1629,7 +1695,8 @@ NV_STATUS nvUvmInterfaceCslDecrypt(UvmCslContext *uvmCslContext,

    Error codes:
      NV_ERR_INSUFFICIENT_RESOURCES - The signing operation would cause a counter overflow to occur.
-      NV_ERR_INVALID_ARGUMENT       - The size of the data is 0 bytes.
+      NV_ERR_INVALID_ARGUMENT       - The CSL context is not associated with a channel.
+                                    - The size of the data is 0 bytes.
 */
 NV_STATUS nvUvmInterfaceCslSign(UvmCslContext *uvmCslContext,
                                NvU32 bufferSize,
@@ -1641,8 +1708,10 @@ NV_STATUS nvUvmInterfaceCslSign(UvmCslContext *uvmCslContext,

    Returns the number of messages that can be encrypted before the message counter will overflow.

-    See "CSL Interface and Locking" for locking requirements.
-    This function does not perform dynamic memory allocation.
+    Locking: This function does not acquire an API or GPU lock.
+    Memory : This function does not dynamically allocate memory.
+             If called concurrently in different threads with the same UvmCslContext
+             the caller must guarantee exclusion.

    Arguments:
        uvmCslContext[IN/OUT] - The CSL context.
@@ -1666,8 +1735,10 @@ NV_STATUS nvUvmInterfaceCslQueryMessagePool(UvmCslContext *uvmCslContext,
    can be used in nvUvmInterfaceCslEncrypt. If operation is UVM_CSL_OPERATION_DECRYPT then
    the returned IV can be used in nvUvmInterfaceCslDecrypt.

-    See "CSL Interface and Locking" for locking requirements.
-    This function does not perform dynamic memory allocation.
+    Locking: This function does not acquire an API or GPU lock.
+             If called concurrently in different threads with the same UvmCslContext
+             the caller must guarantee exclusion.
+    Memory : This function does not dynamically allocate memory.

 Arguments:
        uvmCslContext[IN/OUT] - The CSL context.
@@ -1675,7 +1746,7 @@ Arguments:
                                - UVM_CSL_OPERATION_ENCRYPT
                                - UVM_CSL_OPERATION_DECRYPT
        increment[IN]         - The amount by which the IV is incremented. Can be 0.
-        iv[out]               - If non-NULL, a buffer to store the incremented IV.
+        iv[OUT]               - If non-NULL, a buffer to store the incremented IV.

    Error codes:
      NV_ERR_INVALID_ARGUMENT       - The value of the operation parameter is illegal.
@@ -1687,4 +1758,36 @@ NV_STATUS nvUvmInterfaceCslIncrementIv(UvmCslContext *uvmCslContext,
                                       NvU64 increment,
                                       UvmCslIv *iv);

+/*******************************************************************************
+    nvUvmInterfaceCslLogExternalEncryption
+
+    Checks and logs information about non-CSL encryptions, such as those that
+    originate from the GPU.
+
+    For contexts associated with channels, this function does not modify elements of
+    the UvmCslContext and must be called for each external encryption invocation.
+
+    For the context associated with fault buffers, bufferSize can encompass multiple
+    encryption invocations, and the UvmCslContext will be updated following a key
+    rotation event.
+
+    In either case the IV remains unmodified after this function is called.
+
+    Locking: This function does not acquire an API or GPU lock.
+    Memory : This function does not dynamically allocate memory.
+             If called concurrently in different threads with the same UvmCslContext
+             the caller must guarantee exclusion.
+
+    Arguments:
+        uvmCslContext[IN/OUT] - The CSL context.
+        bufferSize[OUT]       - The size of the buffer(s) encrypted by the
+                                external entity in units of bytes.
+
+    Error codes:
+      NV_ERR_INSUFFICIENT_RESOURCES - The device encryption would cause a counter
+                                      to overflow.
+*/
+NV_STATUS nvUvmInterfaceCslLogExternalEncryption(UvmCslContext *uvmCslContext,
+                                                 NvU32 bufferSize);
+
 #endif // _NV_UVM_INTERFACE_H_
--- a/kernel-open/common/inc/nv_uvm_types.h
+++ b/kernel-open/common/inc/nv_uvm_types.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2014-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2014-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -39,12 +39,12 @@
 // are multiple BIG page sizes in RM. These defines are used as flags to "0"
 // should be OK when user is not sure which pagesize allocation it wants
 //
-#define UVM_PAGE_SIZE_DEFAULT    0x0
-#define UVM_PAGE_SIZE_4K         0x1000
-#define UVM_PAGE_SIZE_64K        0x10000
-#define UVM_PAGE_SIZE_128K       0x20000
-#define UVM_PAGE_SIZE_2M         0x200000
-#define UVM_PAGE_SIZE_512M       0x20000000
+#define UVM_PAGE_SIZE_DEFAULT    0x0ULL
+#define UVM_PAGE_SIZE_4K         0x1000ULL
+#define UVM_PAGE_SIZE_64K        0x10000ULL
+#define UVM_PAGE_SIZE_128K       0x20000ULL
+#define UVM_PAGE_SIZE_2M         0x200000ULL
+#define UVM_PAGE_SIZE_512M       0x20000000ULL

 //
 // When modifying flags, make sure they are compatible with the mirrored
@@ -104,6 +104,10 @@ typedef struct UvmGpuMemoryInfo_tag
    // Out: Set to TRUE, if the allocation is in sysmem.
    NvBool sysmem;

+    // Out: Set to TRUE, if this allocation is treated as EGM.
+    //      sysmem is also TRUE when egm is TRUE.
+    NvBool egm;
+
    // Out: Set to TRUE, if the allocation is a constructed
    //      under a Device or Subdevice.
    //      All permutations of sysmem and deviceDescendant are valid.
@@ -125,6 +129,10 @@ typedef struct UvmGpuMemoryInfo_tag

    // Out: Uuid of the GPU to which the allocation belongs.
    //      This is only valid if deviceDescendant is NV_TRUE.
+    //      When egm is NV_TRUE, this is also the UUID of the GPU
+    //      for which EGM is local.
+    //      If the GPU has SMC enabled, the UUID is the GI UUID.
+    //      Otherwise, it is the UUID for the physical GPU.
    //      Note: If the allocation is owned by a device in
    //      an SLI group and the allocation is broadcast
    //      across the SLI group, this UUID will be any one
@@ -259,6 +267,7 @@ typedef struct UvmGpuChannelInfo_tag

    // The errorNotifier is filled out when the channel hits an RC error.
    NvNotification    *errorNotifier;
+    NvNotification    *keyRotationNotifier;

    NvU32              hwRunlistId;
    NvU32              hwChannelId;
@@ -284,13 +293,13 @@ typedef struct UvmGpuChannelInfo_tag

    // GPU VAs of both GPFIFO and GPPUT are needed in Confidential Computing
    // so a channel can be controlled via another channel (SEC2 or WLC/LCIC)
-    NvU64             gpFifoGpuVa;
-    NvU64             gpPutGpuVa;
-    NvU64             gpGetGpuVa;
+    NvU64              gpFifoGpuVa;
+    NvU64              gpPutGpuVa;
+    NvU64              gpGetGpuVa;
    // GPU VA of work submission offset is needed in Confidential Computing
    // so CE channels can ring doorbell of other channels as required for
    // WLC/LCIC work submission
-    NvU64             workSubmissionOffsetGpuVa;
+    NvU64              workSubmissionOffsetGpuVa;
 } UvmGpuChannelInfo;

 typedef enum
@@ -332,7 +341,7 @@ typedef struct UvmGpuPagingChannelAllocParams_tag

 // The max number of Copy Engines supported by a GPU.
 // The gpu ops build has a static assert that this is the correct number.
-#define UVM_COPY_ENGINE_COUNT_MAX 10
+#define UVM_COPY_ENGINE_COUNT_MAX 64

 typedef struct
 {
@@ -538,6 +547,10 @@ typedef struct UvmGpuP2PCapsParams_tag
    // the GPUs are direct peers.
    NvU32 peerIds[2];

+    // Out: peerId[i] contains gpu[i]'s EGM peer id of gpu[1 - i]. Only defined
+    // if the GPUs are direct peers and EGM enabled in the system.
+    NvU32 egmPeerIds[2];
+
    // Out: UVM_LINK_TYPE
    NvU32 p2pLink;

@@ -601,7 +614,8 @@ typedef struct UvmGpuInfo_tag
    // Printable gpu name
    char name[UVM_GPU_NAME_LENGTH];

-    // Uuid of this gpu
+    // Uuid of the physical GPU or GI UUID if nvUvmInterfaceGetGpuInfo()
+    // requested information for a valid SMC partition.
    NvProcessorUuid uuid;

    // Gpu architecture; NV2080_CTRL_MC_ARCH_INFO_ARCHITECTURE_*
@@ -683,6 +697,16 @@ typedef struct UvmGpuInfo_tag
    // to NVSwitch peers.
    NvBool connectedToSwitch;
    NvU64 nvswitchMemoryWindowStart;
+
+    // local EGM properties
+    // NV_TRUE if EGM is enabled
+    NvBool   egmEnabled;
+
+    // Peer ID to reach local EGM when EGM is enabled
+    NvU8     egmPeerId;
+
+    // EGM base address to offset in the GMMU PTE entry for EGM mappings
+    NvU64    egmBaseAddr;
 } UvmGpuInfo;

 typedef struct UvmGpuFbInfo_tag
@@ -691,9 +715,10 @@ typedef struct UvmGpuFbInfo_tag
    // RM regions that are not registered with PMA either.
    NvU64 maxAllocatableAddress;

-    NvU32 heapSize;         // RAM in KB available for user allocations
-    NvU32 reservedHeapSize; // RAM in KB reserved for internal RM allocation
-    NvBool bZeroFb;         // Zero FB mode enabled.
+    NvU32 heapSize;          // RAM in KB available for user allocations
+    NvU32 reservedHeapSize;  // RAM in KB reserved for internal RM allocation
+    NvBool bZeroFb;          // Zero FB mode enabled.
+    NvU64 maxVidmemPageSize; // Largest GPU page size to access vidmem.
 } UvmGpuFbInfo;

 typedef struct UvmGpuEccInfo_tag
@@ -771,14 +796,14 @@ typedef NV_STATUS (*uvmEventResume_t) (void);
 /*******************************************************************************
    uvmEventStartDevice
    This function will be called by the GPU driver once it has finished its
-    initialization to tell the UVM driver that this GPU has come up.
+    initialization to tell the UVM driver that this physical GPU has come up.
 */
 typedef NV_STATUS (*uvmEventStartDevice_t) (const NvProcessorUuid *pGpuUuidStruct);

 /*******************************************************************************
    uvmEventStopDevice
-    This function will be called by the GPU driver to let UVM know that a GPU
-    is going down.
+    This function will be called by the GPU driver to let UVM know that a
+    physical GPU is going down.
 */
 typedef NV_STATUS (*uvmEventStopDevice_t) (const NvProcessorUuid *pGpuUuidStruct);

@@ -809,7 +834,7 @@ typedef NV_STATUS (*uvmEventServiceInterrupt_t) (void *pDeviceObject,
 /*******************************************************************************
    uvmEventIsrTopHalf_t
    This function will be called by the GPU driver to let UVM know
-    that an interrupt has occurred.
+    that an interrupt has occurred on the given physical GPU.

    Returns:
        NV_OK if the UVM driver handled the interrupt
@@ -916,11 +941,6 @@ typedef struct UvmGpuFaultInfo_tag
        // CSL context used for performing decryption of replayable faults when
        // Confidential Computing is enabled.
        UvmCslContext cslCtx;
-
-        // Indicates whether UVM owns the replayable fault buffer.
-        // The value of this field is always NV_TRUE When Confidential Computing
-        // is disabled.
-        NvBool bUvmOwnsHwFaultBuffer;
    } replayable;
    struct
    {
@@ -1067,4 +1087,21 @@ typedef enum UvmCslOperation
    UVM_CSL_OPERATION_DECRYPT
 } UvmCslOperation;

+typedef enum UVM_KEY_ROTATION_STATUS {
+    // Key rotation complete/not in progress
+    UVM_KEY_ROTATION_STATUS_IDLE = 0,
+    // RM is waiting for clients to report their channels are idle for key rotation
+    UVM_KEY_ROTATION_STATUS_PENDING = 1,
+    // Key rotation is in progress
+    UVM_KEY_ROTATION_STATUS_IN_PROGRESS = 2,
+    // Key rotation timeout failure, RM will RC non-idle channels.
+    // UVM should never see this status value.
+    UVM_KEY_ROTATION_STATUS_FAILED_TIMEOUT = 3,
+    // Key rotation failed because upper threshold was crossed, RM will RC non-idle channels
+    UVM_KEY_ROTATION_STATUS_FAILED_THRESHOLD = 4,
+    // Internal RM failure while rotating keys for a certain channel, RM will RC the channel.
+    UVM_KEY_ROTATION_STATUS_FAILED_ROTATION = 5,
+    UVM_KEY_ROTATION_STATUS_MAX_COUNT = 6,
+} UVM_KEY_ROTATION_STATUS;
+
 #endif // _NV_UVM_TYPES_H_
--- a/kernel-open/common/inc/nvkms-api-types.h
+++ b/kernel-open/common/inc/nvkms-api-types.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2014-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2014-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -45,6 +45,11 @@

 #define NVKMS_DEVICE_ID_TEGRA                 0x0000ffff

+#define NVKMS_MAX_SUPERFRAME_VIEWS            4
+
+#define NVKMS_LOG2_LUT_ARRAY_SIZE             10
+#define NVKMS_LUT_ARRAY_SIZE                  (1 << NVKMS_LOG2_LUT_ARRAY_SIZE)
+
 typedef NvU32 NvKmsDeviceHandle;
 typedef NvU32 NvKmsDispHandle;
 typedef NvU32 NvKmsConnectorHandle;
@@ -53,6 +58,7 @@ typedef NvU32 NvKmsFrameLockHandle;
 typedef NvU32 NvKmsDeferredRequestFifoHandle;
 typedef NvU32 NvKmsSwapGroupHandle;
 typedef NvU32 NvKmsVblankSyncObjectHandle;
+typedef NvU32 NvKmsVblankSemControlHandle;

 struct NvKmsSize {
    NvU16 width;
@@ -179,6 +185,14 @@ enum NvKmsEventType {
    NVKMS_EVENT_TYPE_FLIP_OCCURRED,
 };

+enum NvKmsFlipResult {
+    NV_KMS_FLIP_RESULT_SUCCESS = 0,    /* Success */
+    NV_KMS_FLIP_RESULT_INVALID_PARAMS, /* Parameter validation failed */
+    NV_KMS_FLIP_RESULT_IN_PROGRESS,    /* Flip would fail because an outstanding
+                                          flip containing changes that cannot be
+                                          queued is in progress */
+};
+
 typedef enum {
    NV_EVO_SCALER_1TAP      = 0,
    NV_EVO_SCALER_2TAPS     = 1,
@@ -221,6 +235,16 @@ struct NvKmsUsageBounds {
    } layer[NVKMS_MAX_LAYERS_PER_HEAD];
 };

+/*!
+ * Per-component arrays of NvU16s describing the LUT; used for both the input
+ * LUT and output LUT.
+ */
+struct NvKmsLutRamps {
+    NvU16 red[NVKMS_LUT_ARRAY_SIZE];   /*! in */
+    NvU16 green[NVKMS_LUT_ARRAY_SIZE]; /*! in */
+    NvU16 blue[NVKMS_LUT_ARRAY_SIZE];  /*! in */
+};
+
 /*
 * A 3x4 row-major colorspace conversion matrix.
 *
@@ -531,6 +555,18 @@ typedef struct {
    NvBool noncoherent;
 } NvKmsDispIOCoherencyModes;

+enum NvKmsInputColorRange {
+    /*
+     * If DEFAULT is provided, driver will assume full range for RGB formats
+     * and limited range for YUV formats.
+     */
+    NVKMS_INPUT_COLORRANGE_DEFAULT = 0,
+
+    NVKMS_INPUT_COLORRANGE_LIMITED = 1,
+
+    NVKMS_INPUT_COLORRANGE_FULL = 2,
+};
+
 enum NvKmsInputColorSpace {
    /* Unknown colorspace; no de-gamma will be applied */
    NVKMS_INPUT_COLORSPACE_NONE = 0,
@@ -542,6 +578,12 @@ enum NvKmsInputColorSpace {
    NVKMS_INPUT_COLORSPACE_BT2100_PQ = 2,
 };

+enum NvKmsOutputColorimetry {
+    NVKMS_OUTPUT_COLORIMETRY_DEFAULT = 0,
+
+    NVKMS_OUTPUT_COLORIMETRY_BT2100 = 1,
+};
+
 enum NvKmsOutputTf {
    /*
     * NVKMS itself won't apply any OETF (clients are still
@@ -552,6 +594,17 @@ enum NvKmsOutputTf {
    NVKMS_OUTPUT_TF_PQ = 2,
 };

+/*!
+ * EOTF Data Byte 1 as per CTA-861-G spec.
+ * This is expected to match exactly with the spec.
+ */
+enum NvKmsInfoFrameEOTF {
+    NVKMS_INFOFRAME_EOTF_SDR_GAMMA = 0,
+    NVKMS_INFOFRAME_EOTF_HDR_GAMMA = 1,
+    NVKMS_INFOFRAME_EOTF_ST2084 = 2,
+    NVKMS_INFOFRAME_EOTF_HLG = 3,
+};
+
 /*!
 * HDR Static Metadata Type1 Descriptor as per CEA-861.3 spec.
 * This is expected to match exactly with the spec.
@@ -605,4 +658,29 @@ struct NvKmsHDRStaticMetadata {
    NvU16 maxFALL;
 };

+/*!
+ * A superframe is made of two or more video streams that are combined in
+ * a specific way. A DP serializer (an external device connected to a Tegra
+ * ARM SOC over DP or HDMI) can receive a video stream comprising multiple
+ * videos combined into a single frame and then split it into multiple
+ * video streams. The following structure describes the number of views
+ * and dimensions of each view inside a superframe.
+ */
+struct NvKmsSuperframeInfo {
+    NvU8 numViews;
+    struct {
+        /* x offset inside superframe at which this view starts */
+        NvU16 x;
+
+        /* y offset inside superframe at which this view starts */
+        NvU16 y;
+
+        /* Horizontal active width in pixels for this view */
+        NvU16 width;
+
+        /* Vertical active height in lines for this view */
+        NvU16 height;
+    } view[NVKMS_MAX_SUPERFRAME_VIEWS];
+};
+
 #endif /* NVKMS_API_TYPES_H */
--- a/kernel-open/common/inc/nvkms-kapi.h
+++ b/kernel-open/common/inc/nvkms-kapi.h
@@ -49,6 +49,8 @@ struct NvKmsKapiDevice;
 struct NvKmsKapiMemory;
 struct NvKmsKapiSurface;
 struct NvKmsKapiChannelEvent;
+struct NvKmsKapiSemaphoreSurface;
+struct NvKmsKapiSemaphoreSurfaceCallback;

 typedef NvU32 NvKmsKapiConnector;
 typedef NvU32 NvKmsKapiDisplay;
@@ -67,6 +69,14 @@ typedef NvU32 NvKmsKapiDisplay;
 */
 typedef void NvKmsChannelEventProc(void *dataPtr, NvU32 dataU32);

+/*
+ * Note: Same as above, this function must not call back into NVKMS-KAPI, nor
+ * directly into RM. Doing so could cause deadlocks given the notification
+ * function will most likely be called from within RM's interrupt handler
+ * callchain.
+ */
+typedef void NvKmsSemaphoreSurfaceCallbackProc(void *pData);
+
 /** @} */

 /**
@@ -126,6 +136,11 @@ struct NvKmsKapiDeviceResourcesInfo {
        NvU32 validCursorCompositionModes;
        NvU64 supportedCursorSurfaceMemoryFormats;

+        struct {
+            NvU64 maxSubmittedOffset;
+            NvU64 stride;
+        } semsurf;
+
        struct {
            NvU16 validRRTransforms;
            NvU32 validCompositionModes;
@@ -218,8 +233,10 @@ struct NvKmsKapiLayerConfig {
    struct NvKmsRRParams rrParams;
    struct NvKmsKapiSyncpt syncptParams;

-    struct NvKmsHDRStaticMetadata hdrMetadata;
-    NvBool hdrMetadataSpecified;
+    struct {
+        struct NvKmsHDRStaticMetadata val;
+        NvBool enabled;
+    } hdrMetadata;

    enum NvKmsOutputTf tf;

@@ -233,16 +250,21 @@ struct NvKmsKapiLayerConfig {
    NvU16 dstWidth, dstHeight;

    enum NvKmsInputColorSpace inputColorSpace;
+    struct NvKmsCscMatrix csc;
+    NvBool cscUseMain;
 };

 struct NvKmsKapiLayerRequestedConfig {
    struct NvKmsKapiLayerConfig config;
    struct {
-        NvBool surfaceChanged : 1;
-        NvBool srcXYChanged   : 1;
-        NvBool srcWHChanged   : 1;
-        NvBool dstXYChanged   : 1;
-        NvBool dstWHChanged   : 1;
+        NvBool surfaceChanged     : 1;
+        NvBool srcXYChanged       : 1;
+        NvBool srcWHChanged       : 1;
+        NvBool dstXYChanged       : 1;
+        NvBool dstWHChanged       : 1;
+        NvBool cscChanged         : 1;
+        NvBool tfChanged          : 1;
+        NvBool hdrMetadataChanged : 1;
    } flags;
 };

@@ -286,14 +308,41 @@ struct NvKmsKapiHeadModeSetConfig {
    struct NvKmsKapiDisplayMode mode;

    NvBool vrrEnabled;
+
+    struct {
+        NvBool enabled;
+        enum NvKmsInfoFrameEOTF eotf;
+        struct NvKmsHDRStaticMetadata staticMetadata;
+    } hdrInfoFrame;
+
+    enum NvKmsOutputColorimetry colorimetry;
+
+    struct {
+        struct {
+            NvBool specified;
+            NvU32 depth;
+            NvU32 start;
+            NvU32 end;
+            struct NvKmsLutRamps *pRamps;
+        } input;
+
+        struct {
+            NvBool specified;
+            NvBool enabled;
+            struct NvKmsLutRamps *pRamps;
+        } output;
+    } lut;
 };

 struct NvKmsKapiHeadRequestedConfig {
    struct NvKmsKapiHeadModeSetConfig modeSetConfig;
    struct {
-        NvBool activeChanged   : 1;
-        NvBool displaysChanged : 1;
-        NvBool modeChanged     : 1;
+        NvBool activeChanged       : 1;
+        NvBool displaysChanged     : 1;
+        NvBool modeChanged         : 1;
+        NvBool hdrInfoFrameChanged : 1;
+        NvBool colorimetryChanged  : 1;
+        NvBool lutChanged      : 1;
    } flags;

    struct NvKmsKapiCursorRequestedConfig cursorRequestedConfig;
@@ -318,6 +367,7 @@ struct NvKmsKapiHeadReplyConfig {
 };

 struct NvKmsKapiModeSetReplyConfig {
+    enum NvKmsFlipResult flipResult;
    struct NvKmsKapiHeadReplyConfig
        headReplyConfig[NVKMS_KAPI_MAX_HEADS];
 };
@@ -434,6 +484,14 @@ enum NvKmsKapiAllocationType {
    NVKMS_KAPI_ALLOCATION_TYPE_OFFSCREEN = 2,
 };

+typedef enum NvKmsKapiRegisterWaiterResultRec {
+    NVKMS_KAPI_REG_WAITER_FAILED,
+    NVKMS_KAPI_REG_WAITER_SUCCESS,
+    NVKMS_KAPI_REG_WAITER_ALREADY_SIGNALLED,
+} NvKmsKapiRegisterWaiterResult;
+
+typedef void NvKmsKapiSuspendResumeCallbackFunc(NvBool suspend);
+
 struct NvKmsKapiFunctionsTable {

    /*!
@@ -519,8 +577,8 @@ struct NvKmsKapiFunctionsTable {
    );

    /*!
-     * Revoke permissions previously granted. Only one (dispIndex, head,
-     * display) is currently supported.
+     * Revoke modeset permissions previously granted. Only one (dispIndex,
+     * head, display) is currently supported.
     *
     * \param [in]  device     A device returned by allocateDevice().
     *
@@ -537,6 +595,34 @@ struct NvKmsKapiFunctionsTable {
        NvKmsKapiDisplay display
    );

+    /*!
+     * Grant modeset sub-owner permissions to fd. This is used by clients to
+     * convert drm 'master' permissions into nvkms sub-owner permission.
+     *
+     * \param [in]  fd         fd from opening /dev/nvidia-modeset.
+     *
+     * \param [in]  device     A device returned by allocateDevice().
+     *
+     * \return NV_TRUE on success, NV_FALSE on failure.
+     */
+    NvBool (*grantSubOwnership)
+    (
+        NvS32 fd,
+        struct NvKmsKapiDevice *device
+    );
+
+    /*!
+     * Revoke sub-owner permissions previously granted.
+     *
+     * \param [in]  device     A device returned by allocateDevice().
+     *
+     * \return NV_TRUE on success, NV_FALSE on failure.
+     */
+    NvBool (*revokeSubOwnership)
+    (
+        struct NvKmsKapiDevice *device
+    );
+
    /*!
     * Registers for notification, via
     * NvKmsKapiAllocateDeviceParams::eventCallback, of the events specified
@@ -1122,6 +1208,208 @@ struct NvKmsKapiFunctionsTable {
                                       NvP64 dmaBuf,
                                       NvU32 limit);

+    /*!
+     * Import a semaphore surface allocated elsewhere to NVKMS and return a
+     * handle to the new object.
+     *
+     * \param [in] device            A device allocated using allocateDevice().
+     *
+     * \param [in] nvKmsParamsUser   Userspace pointer to driver-specific
+     *                               parameters describing the semaphore
+     *                               surface being imported.
+     *
+     * \param [in] nvKmsParamsSize   Size of the driver-specific parameter
+     *                               struct.
+     *
+     * \param [out] pSemaphoreMap    Returns a CPU mapping of the semaphore
+     *                               surface's semaphore memory to the client.
+     *
+     * \param [out] pMaxSubmittedMap Returns a CPU mapping of the semaphore
+     *                               surface's semaphore memory to the client.
+     *
+     * \return struct NvKmsKapiSemaphoreSurface* on success, NULL on failure.
+     */
+    struct NvKmsKapiSemaphoreSurface* (*importSemaphoreSurface)
+    (
+         struct NvKmsKapiDevice *device,
+         NvU64 nvKmsParamsUser,
+         NvU64 nvKmsParamsSize,
+         void **pSemaphoreMap,
+         void **pMaxSubmittedMap
+    );
+
+    /*!
+     * Free an imported semaphore surface.
+     *
+     * \param [in]  device              The device passed to
+     *                                  importSemaphoreSurface() when creating
+     *                                  semaphoreSurface.
+     *
+     * \param [in]  semaphoreSurface    A semaphore surface returned by
+     *                                  importSemaphoreSurface().
+     */
+    void (*freeSemaphoreSurface)
+    (
+        struct NvKmsKapiDevice *device,
+        struct NvKmsKapiSemaphoreSurface *semaphoreSurface
+    );
+
+    /*!
+     * Register a callback to be called when a semaphore reaches a value.
+     *
+     * The callback will be called when the semaphore at index in
+     * semaphoreSurface reaches the value wait_value.  The callback will
+     * be called at most once and is automatically unregistered when called.
+     * It may also be unregistered (i.e., cancelled) explicitly using the
+     * unregisterSemaphoreSurfaceCallback() function. To avoid leaking the
+     * memory used to track the registered callback, callers must ensure one
+     * of these methods of unregistration is used for every successful
+     * callback registration that returns a non-NULL pCallbackHandle.
+     *
+     * \param [in]  device              The device passed to
+     *                                  importSemaphoreSurface() when creating
+     *                                  semaphoreSurface.
+     *
+     * \param [in]  semaphoreSurface    A semaphore surface returned by
+     *                                  importSemaphoreSurface().
+     *
+     * \param [in]  pCallback           A pointer to the function to call when
+     *                                  the specified value is reached. NULL
+     *                                  means no callback.
+     *
+     * \param [in]  pData               Arbitrary data to be passed back to the
+     *                                  callback as its sole parameter.
+     *
+     * \param [in]  index               The index of the semaphore within
+     *                                  semaphoreSurface.
+     *
+     * \param [in]  wait_value          The value the semaphore must reach or
+     *                                  exceed before the callback is called.
+     *
+     * \param [in]  new_value           The value the semaphore will be set to
+     *                                  when it reaches or exceeds <wait_value>.
+     *                                  0 means do not update the value.
+     *
+     * \param [out] pCallbackHandle     On success, the value pointed to will
+     *                                  contain an opaque handle to the
+     *                                  registered callback that may be used to
+     *                                  cancel it if needed. Unused if pCallback
+     *                                  is NULL.
+     *
+     * \return NVKMS_KAPI_REG_WAITER_SUCCESS if the waiter was registered or if
+     *         no callback was requested and the semaphore at <index> has
+     *         already reached or exceeded <wait_value>
+     *
+     *         NVKMS_KAPI_REG_WAITER_ALREADY_SIGNALLED if a callback was
+     *         requested and the semaphore at <index> has already reached or
+     *         exceeded <wait_value>
+     *
+     *         NVKMS_KAPI_REG_WAITER_FAILED if waiter registration failed.
+     */
+    NvKmsKapiRegisterWaiterResult
+    (*registerSemaphoreSurfaceCallback)
+    (
+        struct NvKmsKapiDevice *device,
+        struct NvKmsKapiSemaphoreSurface *semaphoreSurface,
+        NvKmsSemaphoreSurfaceCallbackProc *pCallback,
+        void *pData,
+        NvU64 index,
+        NvU64 wait_value,
+        NvU64 new_value,
+        struct NvKmsKapiSemaphoreSurfaceCallback **pCallbackHandle
+    );
+
+    /*!
+     * Unregister a callback registered via registerSemaphoreSurfaceCallback()
+     *
+     * If the callback has not yet been called, this function will cancel the
+     * callback and free its associated resources.
+     *
+     * Note this function treats the callback handle as a pointer. While this
+     * function does not dereference that pointer itself, the underlying call
+     * to RM does within a properly guarded critical section that first ensures
+     * it is not in the process of being used within a callback. This means
+     * the callstack must take into consideration that pointers are not in
+     * general unique handles if they may have been freed, since a subsequent
+     * malloc could return the same pointer value at that point. This callchain
+     * avoids that by leveraging the behavior of the underlying RM APIs:
+     *
+     * 1) A callback handle is referenced relative to its corresponding
+     *    (semaphore surface, index, wait_value) tuple here and within RM. It
+     *    is not a valid handle outside of that scope.
+     *
+     * 2) A callback can not be registered against an already-reached value
+     *    for a given semaphore surface index.
+     *
+     * 3) A given callback handle can not be registered twice against the same
+     *    (semaphore surface, index, wait_value) tuple, so unregistration will
+     *    never race with registration at the RM level, and would only race at
+     *    a higher level if used incorrectly. Since this is kernel code, we
+     *    can safely assume there won't be malicious clients purposely misuing
+     *    the API, but the burden is placed on the caller to ensure its usage
+     *    does not lead to races at higher levels.
+     *
+     * These factors considered together ensure any valid registered handle is
+     * either still in the relevant waiter list and refers to the same event/
+     * callback as when it was registered, or has been removed from the list
+     * as part of a critical section that also destroys the list itself and
+     * makes future lookups in that list impossible, and hence eliminates the
+     * chance of comparing a stale handle with a new handle of the same value
+     * as part of a lookup.
+     *
+     * \param [in]  device              The device passed to
+     *                                  importSemaphoreSurface() when creating
+     *                                  semaphoreSurface.
+     *
+     * \param [in]  semaphoreSurface    The semaphore surface passed to
+     *                                  registerSemaphoreSurfaceCallback() when
+     *                                  registering the callback.
+     *
+     * \param [in]  index               The index passed to
+     *                                  registerSemaphoreSurfaceCallback() when
+     *                                  registering the callback.
+     *
+     * \param [in]  wait_value          The wait_value passed to
+     *                                  registerSemaphoreSurfaceCallback() when
+     *                                  registering the callback.
+     *
+     * \param [in]  callbackHandle      The callback handle returned by
+     *                                  registerSemaphoreSurfaceCallback().
+     */
+    NvBool
+    (*unregisterSemaphoreSurfaceCallback)
+    (
+        struct NvKmsKapiDevice *device,
+        struct NvKmsKapiSemaphoreSurface *semaphoreSurface,
+        NvU64 index,
+        NvU64 wait_value,
+        struct NvKmsKapiSemaphoreSurfaceCallback *callbackHandle
+    );
+
+    /*!
+     * Update the value of a semaphore surface from the CPU.
+     *
+     * Update the semaphore value at the specified index from the CPU, then
+     * wake up any pending CPU waiters associated with that index that are
+     * waiting on it reaching a value <= the new value.
+     */
+    NvBool
+    (*setSemaphoreSurfaceValue)
+    (
+        struct NvKmsKapiDevice *device,
+        struct NvKmsKapiSemaphoreSurface *semaphoreSurface,
+        NvU64 index,
+        NvU64 new_value
+    );
+
+    /*!
+     * Set the callback function for suspending and resuming the display system.
+     */
+    void
+    (*setSuspendResumeCallback)
+    (
+        NvKmsKapiSuspendResumeCallbackFunc *function
+    );
 };

 /** @} */
--- a/kernel-open/common/inc/nvmisc.h
+++ b/kernel-open/common/inc/nvmisc.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 1993-2020 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -494,6 +494,23 @@ do                                                      \
 //
 #define NV_TWO_N_MINUS_ONE(n) (((1ULL<<(n/2))<<((n+1)/2))-1)

+//
+// Create a 64b bitmask with n bits set
+// This is the same as ((1ULL<<n) - 1), but it doesn't overflow for n=64
+//
+// ...
+// n=-1, 0x0000000000000000
+// n=0,  0x0000000000000000
+// n=1,  0x0000000000000001
+// ...
+// n=63, 0x7FFFFFFFFFFFFFFF
+// n=64, 0xFFFFFFFFFFFFFFFF
+// n=65, 0xFFFFFFFFFFFFFFFF
+// n=66, 0xFFFFFFFFFFFFFFFF
+// ...
+//
+#define NV_BITMASK64(n) ((n<1) ? 0ULL : (NV_U64_MAX>>((n>64) ? 0 : (64-n))))
+
 #define DRF_READ_1WORD_BS(d,r,f,v) \
    ((DRF_EXTENT_MW(NV##d##r##f)<8)?DRF_READ_1BYTE_BS(NV##d##r##f,(v)): \
    ((DRF_EXTENT_MW(NV##d##r##f)<16)?DRF_READ_2BYTE_BS(NV##d##r##f,(v)): \
@@ -574,6 +591,13 @@ nvMaskPos32(const NvU32 mask, const NvU32 bitIdx)
    n32 = BIT_IDX_32(LOWESTBIT(n32));\
 }

+// Destructive operation on n64
+#define LOWESTBITIDX_64(n64)         \
+{                                    \
+    n64 = BIT_IDX_64(LOWESTBIT(n64));\
+}
+
+
 // Destructive operation on n32
 #define HIGHESTBITIDX_32(n32)   \
 {                               \
@@ -918,6 +942,14 @@ static NV_FORCEINLINE void *NV_NVUPTR_TO_PTR(NvUPtr address)
 // Use (lo) if (b) is less than 64, and (hi) if >= 64.
 //
 #define NV_BIT_SET_128(b, lo, hi)              { nvAssert( (b) < 128 ); if ( (b) < 64 ) (lo) |= NVBIT64(b); else (hi) |= NVBIT64( b & 0x3F ); }
+//
+// Clear the bit at pos (b) for U64 which is < 128.
+// Use (lo) if (b) is less than 64, and (hi) if >= 64.
+//
+#define NV_BIT_CLEAR_128(b, lo, hi)            { nvAssert( (b) < 128 ); if ( (b) < 64 ) (lo) &= ~NVBIT64(b); else (hi) &= ~NVBIT64( b & 0x3F ); }
+
+// Get the number of elements the specified fixed-size array
+#define NV_ARRAY_ELEMENTS(x)                   ((sizeof(x)/sizeof((x)[0])))

 #ifdef __cplusplus
 }
--- a/kernel-open/common/inc/nvstatuscodes.h
+++ b/kernel-open/common/inc/nvstatuscodes.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2014-2020 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2014-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -150,6 +150,9 @@ NV_STATUS_CODE(NV_ERR_NVLINK_CONFIGURATION_ERROR,      0x00000078, "Nvlink Confi
 NV_STATUS_CODE(NV_ERR_RISCV_ERROR,                     0x00000079, "Generic RISC-V assert or halt")
 NV_STATUS_CODE(NV_ERR_FABRIC_MANAGER_NOT_PRESENT,      0x0000007A, "Fabric Manager is not loaded")
 NV_STATUS_CODE(NV_ERR_ALREADY_SIGNALLED,               0x0000007B, "Semaphore Surface value already >= requested wait value")
+NV_STATUS_CODE(NV_ERR_QUEUE_TASK_SLOT_NOT_AVAILABLE,   0x0000007C, "PMU RPC error due to no queue slot available for this event")
+NV_STATUS_CODE(NV_ERR_KEY_ROTATION_IN_PROGRESS,        0x0000007D, "Operation not allowed as key rotation is in progress")
+NV_STATUS_CODE(NV_ERR_TEST_ONLY_CODE_NOT_ENABLED,      0x0000007E, "Test-only code path not enabled")

 // Warnings:
 NV_STATUS_CODE(NV_WARN_HOT_SWITCH,                     0x00010001, "WARNING Hot switch")
--- a/kernel-open/common/inc/nvtypes.h
+++ b/kernel-open/common/inc/nvtypes.h
@@ -145,7 +145,18 @@ typedef   signed short     NvS16; /* -32768 to 32767                         */
 #endif

 // Macro to build an NvU32 from four bytes, listed from msb to lsb
-#define NvU32_BUILD(a, b, c, d) (((a) << 24) | ((b) << 16) | ((c) << 8) | (d))
+#define NvU32_BUILD(a, b, c, d) \
+    ((NvU32)( \
+     (((NvU32)(a) & 0xff) << 24) | \
+     (((NvU32)(b) & 0xff) << 16) | \
+     (((NvU32)(c) & 0xff) << 8)  | \
+     (((NvU32)(d) & 0xff))))
+
+// Macro to build an NvU64 from two DWORDS, listed from msb to lsb
+#define NvU64_BUILD(a, b) \
+    ((NvU64)( \
+     (((NvU64)(a) & ~0U) << 32) | \
+     (((NvU64)(b) & ~0U))))

 #if NVTYPES_USE_STDINT
 typedef uint32_t           NvV32; /* "void": enumerated or multiple fields   */
--- a/kernel-open/common/inc/os-interface.h
+++ b/kernel-open/common/inc/os-interface.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 1999-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 1999-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -67,7 +67,6 @@ typedef struct os_wait_queue os_wait_queue;
 * ---------------------------------------------------------------------------
 */

-NvU64       NV_API_CALL  os_get_num_phys_pages       (void);
 NV_STATUS   NV_API_CALL  os_alloc_mem                (void **, NvU64);
 void        NV_API_CALL  os_free_mem                 (void *);
 NV_STATUS   NV_API_CALL  os_get_current_time         (NvU32 *, NvU32 *);
@@ -105,7 +104,6 @@ void*       NV_API_CALL  os_map_kernel_space         (NvU64, NvU64, NvU32);
 void        NV_API_CALL  os_unmap_kernel_space       (void *, NvU64);
 void*       NV_API_CALL  os_map_user_space           (NvU64, NvU64, NvU32, NvU32, void **);
 void        NV_API_CALL  os_unmap_user_space         (void *, NvU64, void *);
-NV_STATUS   NV_API_CALL  os_flush_cpu_cache          (void);
 NV_STATUS   NV_API_CALL  os_flush_cpu_cache_all      (void);
 NV_STATUS   NV_API_CALL  os_flush_user_cache         (void);
 void        NV_API_CALL  os_flush_cpu_write_combine_buffer(void);
@@ -162,10 +160,9 @@ NvBool      NV_API_CALL  os_is_vgx_hyper             (void);
 NV_STATUS   NV_API_CALL  os_inject_vgx_msi           (NvU16, NvU64, NvU32);
 NvBool      NV_API_CALL  os_is_grid_supported        (void);
 NvU32       NV_API_CALL  os_get_grid_csp_support     (void);
-void        NV_API_CALL  os_get_screen_info          (NvU64 *, NvU16 *, NvU16 *, NvU16 *, NvU16 *, NvU64, NvU64);
 void        NV_API_CALL  os_bug_check                (NvU32, const char *);
 NV_STATUS   NV_API_CALL  os_lock_user_pages          (void *, NvU64, void **, NvU32);
-NV_STATUS   NV_API_CALL  os_lookup_user_io_memory    (void *, NvU64, NvU64 **, void**);
+NV_STATUS   NV_API_CALL  os_lookup_user_io_memory    (void *, NvU64, NvU64 **);
 NV_STATUS   NV_API_CALL  os_unlock_user_pages        (NvU64, void *);
 NV_STATUS   NV_API_CALL  os_match_mmap_offset        (void *, NvU64, NvU64 *);
 NV_STATUS   NV_API_CALL  os_get_euid                 (NvU32 *);
@@ -200,6 +197,8 @@ nv_cap_t*   NV_API_CALL  os_nv_cap_create_file_entry  (nv_cap_t *, const char *,
 void        NV_API_CALL  os_nv_cap_destroy_entry      (nv_cap_t *);
 int         NV_API_CALL  os_nv_cap_validate_and_dup_fd(const nv_cap_t *, int);
 void        NV_API_CALL  os_nv_cap_close_fd           (int);
+NvS32       NV_API_CALL  os_imex_channel_get          (NvU64);
+NvS32       NV_API_CALL  os_imex_channel_count        (void);

 enum os_pci_req_atomics_type {
    OS_INTF_PCIE_REQ_ATOMICS_32BIT,
@@ -221,6 +220,7 @@ extern NvU8  os_page_shift;
 extern NvBool os_cc_enabled;
 extern NvBool os_cc_tdx_enabled;
 extern NvBool os_dma_buf_enabled;
+extern NvBool os_imex_channel_is_supported;

 /*
 * ---------------------------------------------------------------------------
--- a/kernel-open/common/inc/rm-gpu-ops.h
+++ b/kernel-open/common/inc/rm-gpu-ops.h
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 1999-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 1999-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -37,7 +37,7 @@ NV_STATUS  NV_API_CALL  rm_gpu_ops_create_session (nvidia_stack_t *, nvgpuSessio
 NV_STATUS  NV_API_CALL  rm_gpu_ops_destroy_session (nvidia_stack_t *, nvgpuSessionHandle_t);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_device_create (nvidia_stack_t *, nvgpuSessionHandle_t, const nvgpuInfo_t *, const NvProcessorUuid *, nvgpuDeviceHandle_t *, NvBool);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_device_destroy (nvidia_stack_t *, nvgpuDeviceHandle_t);
-NV_STATUS  NV_API_CALL  rm_gpu_ops_address_space_create(nvidia_stack_t *, nvgpuDeviceHandle_t, unsigned long long, unsigned long long, nvgpuAddressSpaceHandle_t *, nvgpuAddressSpaceInfo_t);
+NV_STATUS  NV_API_CALL  rm_gpu_ops_address_space_create(nvidia_stack_t *, nvgpuDeviceHandle_t, unsigned long long, unsigned long long, NvBool, nvgpuAddressSpaceHandle_t *, nvgpuAddressSpaceInfo_t);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_dup_address_space(nvidia_stack_t *, nvgpuDeviceHandle_t, NvHandle, NvHandle, nvgpuAddressSpaceHandle_t *, nvgpuAddressSpaceInfo_t);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_address_space_destroy(nvidia_stack_t *, nvgpuAddressSpaceHandle_t);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_memory_alloc_fb(nvidia_stack_t *, nvgpuAddressSpaceHandle_t, NvLength, NvU64 *, nvgpuAllocInfo_t);
@@ -45,7 +45,6 @@ NV_STATUS  NV_API_CALL  rm_gpu_ops_memory_alloc_fb(nvidia_stack_t *, nvgpuAddres
 NV_STATUS  NV_API_CALL  rm_gpu_ops_pma_alloc_pages(nvidia_stack_t *, void *, NvLength, NvU32 , nvgpuPmaAllocationOptions_t, NvU64 *);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_pma_free_pages(nvidia_stack_t *, void *, NvU64 *, NvLength , NvU32, NvU32);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_pma_pin_pages(nvidia_stack_t *, void *, NvU64 *, NvLength , NvU32, NvU32);
-NV_STATUS  NV_API_CALL  rm_gpu_ops_pma_unpin_pages(nvidia_stack_t *, void *, NvU64 *, NvLength , NvU32);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_get_pma_object(nvidia_stack_t *, nvgpuDeviceHandle_t, void **, const nvgpuPmaStatistics_t *);
 NV_STATUS  NV_API_CALL  rm_gpu_ops_pma_register_callbacks(nvidia_stack_t *sp, void *, nvPmaEvictPagesCallback, nvPmaEvictRangeCallback, void *);
 void       NV_API_CALL  rm_gpu_ops_pma_unregister_callbacks(nvidia_stack_t *sp, void *);
@@ -76,7 +75,8 @@ NV_STATUS NV_API_CALL rm_gpu_ops_own_page_fault_intr(nvidia_stack_t *, nvgpuDevi
 NV_STATUS  NV_API_CALL rm_gpu_ops_init_fault_info(nvidia_stack_t *, nvgpuDeviceHandle_t, nvgpuFaultInfo_t);
 NV_STATUS  NV_API_CALL rm_gpu_ops_destroy_fault_info(nvidia_stack_t *, nvgpuDeviceHandle_t, nvgpuFaultInfo_t);
 NV_STATUS  NV_API_CALL rm_gpu_ops_get_non_replayable_faults(nvidia_stack_t *, nvgpuFaultInfo_t, void *, NvU32 *);
-NV_STATUS  NV_API_CALL rm_gpu_ops_flush_replayable_fault_buffer(nvidia_stack_t *, nvgpuDeviceHandle_t);
+NV_STATUS  NV_API_CALL rm_gpu_ops_flush_replayable_fault_buffer(nvidia_stack_t *, nvgpuFaultInfo_t, NvBool);
+NV_STATUS  NV_API_CALL rm_gpu_ops_toggle_prefetch_faults(nvidia_stack_t *, nvgpuFaultInfo_t, NvBool);
 NV_STATUS  NV_API_CALL rm_gpu_ops_has_pending_non_replayable_faults(nvidia_stack_t *, nvgpuFaultInfo_t, NvBool *);
 NV_STATUS  NV_API_CALL rm_gpu_ops_init_access_cntr_info(nvidia_stack_t *, nvgpuDeviceHandle_t, nvgpuAccessCntrInfo_t, NvU32);
 NV_STATUS  NV_API_CALL rm_gpu_ops_destroy_access_cntr_info(nvidia_stack_t *, nvgpuDeviceHandle_t, nvgpuAccessCntrInfo_t);
@@ -101,8 +101,10 @@ NV_STATUS  NV_API_CALL rm_gpu_ops_paging_channels_map(nvidia_stack_t *, nvgpuAdd
 void       NV_API_CALL rm_gpu_ops_paging_channels_unmap(nvidia_stack_t *, nvgpuAddressSpaceHandle_t, NvU64, nvgpuDeviceHandle_t);
 NV_STATUS  NV_API_CALL rm_gpu_ops_paging_channel_push_stream(nvidia_stack_t *, nvgpuPagingChannelHandle_t, char *, NvU32);

+NV_STATUS  NV_API_CALL rm_gpu_ops_key_rotation_channel_disable(nvidia_stack_t *, nvgpuChannelHandle_t [], NvU32);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_context_init(nvidia_stack_t *, struct ccslContext_t **, nvgpuChannelHandle_t);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_context_clear(nvidia_stack_t *, struct ccslContext_t *);
+NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_context_update(nvidia_stack_t *, UvmCslContext *[], NvU32);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_rotate_iv(nvidia_stack_t *, struct ccslContext_t *, NvU8);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_encrypt(nvidia_stack_t *, struct ccslContext_t *, NvU32, NvU8 const *, NvU8 *, NvU8 *);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_encrypt_with_iv(nvidia_stack_t *, struct ccslContext_t *, NvU32, NvU8 const *, NvU8*, NvU8 *, NvU8 *);
@@ -110,5 +112,6 @@ NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_decrypt(nvidia_stack_t *, struct ccslCont
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_sign(nvidia_stack_t *, struct ccslContext_t *, NvU32, NvU8 const *, NvU8 *);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_query_message_pool(nvidia_stack_t *, struct ccslContext_t *, NvU8, NvU64 *);
 NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_increment_iv(nvidia_stack_t *, struct ccslContext_t *, NvU8, NvU64, NvU8 *);
+NV_STATUS  NV_API_CALL rm_gpu_ops_ccsl_log_device_encryption(nvidia_stack_t *, struct ccslContext_t *, NvU32);

 #endif
--- a/kernel-open/conftest.sh
+++ b/kernel-open/conftest.sh
--- a/kernel-open/header-presence-tests.mk
+++ b/kernel-open/header-presence-tests.mk
@@ -0,0 +1,102 @@
+# Each of these headers is checked for presence with a test #include; a
+# corresponding #define will be generated in conftest/headers.h.
+NV_HEADER_PRESENCE_TESTS = \
+  asm/system.h \
+  drm/drmP.h \
+  drm/drm_aperture.h \
+  drm/drm_auth.h \
+  drm/drm_gem.h \
+  drm/drm_crtc.h \
+  drm/drm_color_mgmt.h \
+  drm/drm_atomic.h \
+  drm/drm_atomic_helper.h \
+  drm/drm_atomic_state_helper.h \
+  drm/drm_encoder.h \
+  drm/drm_atomic_uapi.h \
+  drm/drm_drv.h \
+  drm/drm_fbdev_generic.h \
+  drm/drm_framebuffer.h \
+  drm/drm_connector.h \
+  drm/drm_probe_helper.h \
+  drm/drm_blend.h \
+  drm/drm_fourcc.h \
+  drm/drm_prime.h \
+  drm/drm_plane.h \
+  drm/drm_vblank.h \
+  drm/drm_file.h \
+  drm/drm_ioctl.h \
+  drm/drm_device.h \
+  drm/drm_mode_config.h \
+  drm/drm_modeset_lock.h \
+  dt-bindings/interconnect/tegra_icc_id.h \
+  generated/autoconf.h \
+  generated/compile.h \
+  generated/utsrelease.h \
+  linux/efi.h \
+  linux/kconfig.h \
+  linux/platform/tegra/mc_utils.h \
+  linux/printk.h \
+  linux/ratelimit.h \
+  linux/prio_tree.h \
+  linux/log2.h \
+  linux/of.h \
+  linux/bug.h \
+  linux/sched.h \
+  linux/sched/mm.h \
+  linux/sched/signal.h \
+  linux/sched/task.h \
+  linux/sched/task_stack.h \
+  xen/ioemu.h \
+  linux/fence.h \
+  linux/dma-fence.h \
+  linux/dma-resv.h \
+  soc/tegra/chip-id.h \
+  soc/tegra/fuse.h \
+  soc/tegra/fuse-helper.h \
+  soc/tegra/tegra_bpmp.h \
+  video/nv_internal.h \
+  linux/platform/tegra/dce/dce-client-ipc.h \
+  linux/nvhost.h \
+  linux/nvhost_t194.h \
+  linux/host1x-next.h \
+  asm/book3s/64/hash-64k.h \
+  asm/set_memory.h \
+  asm/prom.h \
+  asm/powernv.h \
+  linux/atomic.h \
+  asm/barrier.h \
+  asm/opal-api.h \
+  sound/hdaudio.h \
+  asm/pgtable_types.h \
+  asm/page.h \
+  linux/stringhash.h \
+  linux/dma-map-ops.h \
+  rdma/peer_mem.h \
+  sound/hda_codec.h \
+  linux/dma-buf.h \
+  linux/time.h \
+  linux/platform_device.h \
+  linux/mutex.h \
+  linux/reset.h \
+  linux/of_platform.h \
+  linux/of_device.h \
+  linux/of_gpio.h \
+  linux/gpio.h \
+  linux/gpio/consumer.h \
+  linux/interconnect.h \
+  linux/pm_runtime.h \
+  linux/clk.h \
+  linux/clk-provider.h \
+  linux/ioasid.h \
+  linux/stdarg.h \
+  linux/iosys-map.h \
+  asm/coco.h \
+  linux/vfio_pci_core.h \
+  linux/mdev.h \
+  soc/tegra/bpmp-abi.h \
+  soc/tegra/bpmp.h \
+  linux/sync_file.h \
+  linux/cc_platform.h \
+  asm/cpufeature.h \
+  linux/mpi.h
+
--- a/kernel-open/nvidia-drm/nv-kthread-q.c
+++ b/kernel-open/nvidia-drm/nv-kthread-q.c
@@ -0,0 +1,334 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2016 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#include "nv-kthread-q.h"
+#include "nv-list-helpers.h"
+
+#include <linux/kthread.h>
+#include <linux/interrupt.h>
+#include <linux/completion.h>
+#include <linux/module.h>
+#include <linux/mm.h>
+
+#if defined(NV_LINUX_BUG_H_PRESENT)
+    #include <linux/bug.h>
+#else
+    #include <asm/bug.h>
+#endif
+
+// Today's implementation is a little simpler and more limited than the
+// API description allows for in nv-kthread-q.h. Details include:
+//
+// 1. Each nv_kthread_q instance is a first-in, first-out queue.
+//
+// 2. Each nv_kthread_q instance is serviced by exactly one kthread.
+//
+// You can create any number of queues, each of which gets its own
+// named kernel thread (kthread). You can then insert arbitrary functions
+// into the queue, and those functions will be run in the context of the
+// queue's kthread.
+
+#ifndef WARN
+    // Only *really* old kernels (2.6.9) end up here. Just use a simple printk
+    // to implement this, because such kernels won't be supported much longer.
+    #define WARN(condition, format...) ({                    \
+        int __ret_warn_on = !!(condition);                   \
+        if (unlikely(__ret_warn_on))                         \
+            printk(KERN_ERR format);                         \
+        unlikely(__ret_warn_on);                             \
+    })
+#endif
+
+#define NVQ_WARN(fmt, ...)                                   \
+    do {                                                     \
+        if (in_interrupt()) {                                \
+            WARN(1, "nv_kthread_q: [in interrupt]: " fmt,    \
+            ##__VA_ARGS__);                                  \
+        }                                                    \
+        else {                                               \
+            WARN(1, "nv_kthread_q: task: %s: " fmt,          \
+                 current->comm,                              \
+                 ##__VA_ARGS__);                             \
+        }                                                    \
+    } while (0)
+
+static int _main_loop(void *args)
+{
+    nv_kthread_q_t *q = (nv_kthread_q_t *)args;
+    nv_kthread_q_item_t *q_item = NULL;
+    unsigned long flags;
+
+    while (1) {
+        // Normally this thread is never interrupted. However,
+        // down_interruptible (instead of down) is called here,
+        // in order to avoid being classified as a potentially
+        // hung task, by the kernel watchdog.
+        while (down_interruptible(&q->q_sem))
+            NVQ_WARN("Interrupted during semaphore wait\n");
+
+        if (atomic_read(&q->main_loop_should_exit))
+            break;
+
+        spin_lock_irqsave(&q->q_lock, flags);
+
+        // The q_sem semaphore prevents us from getting here unless there is
+        // at least one item in the list, so an empty list indicates a bug.
+        if (unlikely(list_empty(&q->q_list_head))) {
+            spin_unlock_irqrestore(&q->q_lock, flags);
+            NVQ_WARN("_main_loop: Empty queue: q: 0x%p\n", q);
+            continue;
+        }
+
+        // Consume one item from the queue
+        q_item = list_first_entry(&q->q_list_head,
+                                   nv_kthread_q_item_t,
+                                   q_list_node);
+
+        list_del_init(&q_item->q_list_node);
+
+        spin_unlock_irqrestore(&q->q_lock, flags);
+
+        // Run the item
+        q_item->function_to_run(q_item->function_args);
+
+        // Make debugging a little simpler by clearing this between runs:
+        q_item = NULL;
+    }
+
+    while (!kthread_should_stop())
+        schedule();
+
+    return 0;
+}
+
+void nv_kthread_q_stop(nv_kthread_q_t *q)
+{
+    // check if queue has been properly initialized
+    if (unlikely(!q->q_kthread))
+        return;
+
+    nv_kthread_q_flush(q);
+
+    // If this assertion fires, then a caller likely either broke the API rules,
+    // by adding items after calling nv_kthread_q_stop, or possibly messed up
+    // with inadequate flushing of self-rescheduling q_items.
+    if (unlikely(!list_empty(&q->q_list_head)))
+        NVQ_WARN("list not empty after flushing\n");
+
+    if (likely(!atomic_read(&q->main_loop_should_exit))) {
+
+        atomic_set(&q->main_loop_should_exit, 1);
+
+        // Wake up the kthread so that it can see that it needs to stop:
+        up(&q->q_sem);
+
+        kthread_stop(q->q_kthread);
+        q->q_kthread = NULL;
+    }
+}
+
+// When CONFIG_VMAP_STACK is defined, the kernel thread stack allocator used by
+// kthread_create_on_node relies on a 2 entry, per-core cache to minimize
+// vmalloc invocations. The cache is NUMA-unaware, so when there is a hit, the
+// stack location ends up being a function of the core assigned to the current
+// thread, instead of being a function of the specified NUMA node. The cache was
+// added to the kernel in commit ac496bf48d97f2503eaa353996a4dd5e4383eaf0
+// ("fork: Optimize task creation by caching two thread stacks per CPU if
+// CONFIG_VMAP_STACK=y")
+//
+// To work around the problematic cache, we create up to three kernel threads
+//   -If the first thread's stack is resident on the preferred node, return this
+//    thread.
+//   -Otherwise, create a second thread. If its stack is resident on the
+//    preferred node, stop the first thread and return this one.
+//   -Otherwise, create a third thread. The stack allocator does not find a
+//    cached stack, and so falls back to vmalloc, which takes the NUMA hint into
+//    consideration. The first two threads are then stopped.
+//
+// When CONFIG_VMAP_STACK is not defined, the first kernel thread is returned.
+//
+// This function is never invoked when there is no NUMA preference (preferred
+// node is NUMA_NO_NODE).
+static struct task_struct *thread_create_on_node(int (*threadfn)(void *data),
+                                                 nv_kthread_q_t *q,
+                                                 int preferred_node,
+                                                 const char *q_name)
+{
+
+    unsigned i, j;
+    const static unsigned attempts = 3;
+    struct task_struct *thread[3];
+
+    for (i = 0;; i++) {
+        struct page *stack;
+
+        thread[i] = kthread_create_on_node(threadfn, q, preferred_node, q_name);
+
+        if (unlikely(IS_ERR(thread[i]))) {
+
+            // Instead of failing, pick the previous thread, even if its
+            // stack is not allocated on the preferred node.
+            if (i > 0)
+                i--;
+
+            break;
+        }
+
+        // vmalloc is not used to allocate the stack, so simply return the
+        // thread, even if its stack may not be allocated on the preferred node
+        if (!is_vmalloc_addr(thread[i]->stack))
+            break;
+
+        // Ran out of attempts - return thread even if its stack may not be
+        // allocated on the preferred node
+        if ((i == (attempts - 1)))
+            break;
+
+        // Get the NUMA node where the first page of the stack is resident. If
+        // it is the preferred node, select this thread.
+        stack = vmalloc_to_page(thread[i]->stack);
+        if (page_to_nid(stack) == preferred_node)
+            break;
+    }
+
+    for (j = i; j > 0; j--)
+        kthread_stop(thread[j - 1]);
+
+    return thread[i];
+}
+
+int nv_kthread_q_init_on_node(nv_kthread_q_t *q, const char *q_name, int preferred_node)
+{
+    memset(q, 0, sizeof(*q));
+
+    INIT_LIST_HEAD(&q->q_list_head);
+    spin_lock_init(&q->q_lock);
+    sema_init(&q->q_sem, 0);
+
+    if (preferred_node == NV_KTHREAD_NO_NODE) {
+        q->q_kthread = kthread_create(_main_loop, q, q_name);
+    }
+    else {
+        q->q_kthread = thread_create_on_node(_main_loop, q, preferred_node, q_name);
+    }
+
+    if (IS_ERR(q->q_kthread)) {
+        int err = PTR_ERR(q->q_kthread);
+
+        // Clear q_kthread before returning so that nv_kthread_q_stop() can be
+        // safely called on it making error handling easier.
+        q->q_kthread = NULL;
+
+        return err;
+    }
+
+    wake_up_process(q->q_kthread);
+
+    return 0;
+}
+
+int nv_kthread_q_init(nv_kthread_q_t *q, const char *qname)
+{
+    return nv_kthread_q_init_on_node(q, qname, NV_KTHREAD_NO_NODE);
+}
+
+// Returns true (non-zero) if the item was actually scheduled, and false if the
+// item was already pending in a queue.
+static int _raw_q_schedule(nv_kthread_q_t *q, nv_kthread_q_item_t *q_item)
+{
+    unsigned long flags;
+    int ret = 1;
+
+    spin_lock_irqsave(&q->q_lock, flags);
+
+    if (likely(list_empty(&q_item->q_list_node)))
+        list_add_tail(&q_item->q_list_node, &q->q_list_head);
+    else
+        ret = 0;
+
+    spin_unlock_irqrestore(&q->q_lock, flags);
+
+    if (likely(ret))
+        up(&q->q_sem);
+
+    return ret;
+}
+
+void nv_kthread_q_item_init(nv_kthread_q_item_t *q_item,
+                            nv_q_func_t function_to_run,
+                            void *function_args)
+{
+    INIT_LIST_HEAD(&q_item->q_list_node);
+    q_item->function_to_run = function_to_run;
+    q_item->function_args   = function_args;
+}
+
+// Returns true (non-zero) if the q_item got scheduled, false otherwise.
+int nv_kthread_q_schedule_q_item(nv_kthread_q_t *q,
+                                 nv_kthread_q_item_t *q_item)
+{
+    if (unlikely(atomic_read(&q->main_loop_should_exit))) {
+        NVQ_WARN("Not allowed: nv_kthread_q_schedule_q_item was "
+                   "called with a non-alive q: 0x%p\n", q);
+        return 0;
+    }
+
+    return _raw_q_schedule(q, q_item);
+}
+
+static void _q_flush_function(void *args)
+{
+    struct completion *completion = (struct completion *)args;
+    complete(completion);
+}
+
+
+static void _raw_q_flush(nv_kthread_q_t *q)
+{
+    nv_kthread_q_item_t q_item;
+    DECLARE_COMPLETION_ONSTACK(completion);
+
+    nv_kthread_q_item_init(&q_item, _q_flush_function, &completion);
+
+    _raw_q_schedule(q, &q_item);
+
+    // Wait for the flush item to run. Once it has run, then all of the
+    // previously queued items in front of it will have run, so that means
+    // the flush is complete.
+    wait_for_completion(&completion);
+}
+
+void nv_kthread_q_flush(nv_kthread_q_t *q)
+{
+    if (unlikely(atomic_read(&q->main_loop_should_exit))) {
+        NVQ_WARN("Not allowed: nv_kthread_q_flush was called after "
+                   "nv_kthread_q_stop. q: 0x%p\n", q);
+        return;
+    }
+
+    // This 2x flush is not a typing mistake. The queue really does have to be
+    // flushed twice, in order to take care of the case of a q_item that
+    // reschedules itself.
+    _raw_q_flush(q);
+    _raw_q_flush(q);
+}
--- a/kernel-open/nvidia-drm/nv-pci-table.c
+++ b/kernel-open/nvidia-drm/nv-pci-table.c
@@ -25,6 +25,15 @@
 #include <linux/module.h>

 #include "nv-pci-table.h"
+#include "cpuopsys.h"
+
+#if defined(NV_BSD)
+/* Define PCI classes that FreeBSD's linuxkpi is missing */
+#define PCI_VENDOR_ID_NVIDIA 0x10de
+#define PCI_CLASS_DISPLAY_VGA 0x0300
+#define PCI_CLASS_DISPLAY_3D 0x0302
+#define PCI_CLASS_BRIDGE_OTHER 0x0680
+#endif

 /* Devices supported by RM */
 struct pci_device_id nv_pci_table[] = {
@@ -48,7 +57,7 @@ struct pci_device_id nv_pci_table[] = {
 };

 /* Devices supported by all drivers in nvidia.ko */
-struct pci_device_id nv_module_device_table[] = {
+struct pci_device_id nv_module_device_table[4] = {
    {
        .vendor      = PCI_VENDOR_ID_NVIDIA,
        .device      = PCI_ANY_ID,
@@ -76,4 +85,6 @@ struct pci_device_id nv_module_device_table[] = {
    { }
 };

+#if defined(NV_LINUX)
 MODULE_DEVICE_TABLE(pci, nv_module_device_table);
+#endif
--- a/kernel-open/nvidia-drm/nv-pci-table.h
+++ b/kernel-open/nvidia-drm/nv-pci-table.h
@@ -27,5 +27,6 @@
 #include <linux/pci.h>

 extern struct pci_device_id nv_pci_table[];
+extern struct pci_device_id nv_module_device_table[4];

 #endif /* _NV_PCI_TABLE_H_ */
--- a/kernel-open/nvidia-drm/nvidia-dma-fence-helper.h
+++ b/kernel-open/nvidia-drm/nvidia-dma-fence-helper.h
@@ -43,9 +43,13 @@
 #if defined(NV_LINUX_FENCE_H_PRESENT)
 typedef struct fence nv_dma_fence_t;
 typedef struct fence_ops nv_dma_fence_ops_t;
+typedef struct fence_cb nv_dma_fence_cb_t;
+typedef fence_func_t nv_dma_fence_func_t;
 #else
 typedef struct dma_fence nv_dma_fence_t;
 typedef struct dma_fence_ops nv_dma_fence_ops_t;
+typedef struct dma_fence_cb nv_dma_fence_cb_t;
+typedef dma_fence_func_t nv_dma_fence_func_t;
 #endif

 #if defined(NV_LINUX_FENCE_H_PRESENT)
@@ -97,6 +101,14 @@ static inline int nv_dma_fence_signal(nv_dma_fence_t *fence) {
 #endif
 }

+static inline int nv_dma_fence_signal_locked(nv_dma_fence_t *fence) {
+#if defined(NV_LINUX_FENCE_H_PRESENT)
+    return fence_signal_locked(fence);
+#else
+    return dma_fence_signal_locked(fence);
+#endif
+}
+
 static inline u64 nv_dma_fence_context_alloc(unsigned num) {
 #if defined(NV_LINUX_FENCE_H_PRESENT)
    return fence_context_alloc(num);
@@ -108,7 +120,7 @@ static inline u64 nv_dma_fence_context_alloc(unsigned num) {
 static inline void
 nv_dma_fence_init(nv_dma_fence_t *fence,
                  const nv_dma_fence_ops_t *ops,
-                  spinlock_t *lock, u64 context, unsigned seqno) {
+                  spinlock_t *lock, u64 context, uint64_t seqno) {
 #if defined(NV_LINUX_FENCE_H_PRESENT)
    fence_init(fence, ops, lock, context, seqno);
 #else
@@ -116,6 +128,29 @@ nv_dma_fence_init(nv_dma_fence_t *fence,
 #endif
 }

+static inline void
+nv_dma_fence_set_error(nv_dma_fence_t *fence,
+                       int error) {
+#if defined(NV_DMA_FENCE_SET_ERROR_PRESENT)
+    return dma_fence_set_error(fence, error);
+#elif defined(NV_FENCE_SET_ERROR_PRESENT)
+    return fence_set_error(fence, error);
+#else
+    fence->status = error;
+#endif
+}
+
+static inline int
+nv_dma_fence_add_callback(nv_dma_fence_t *fence,
+                          nv_dma_fence_cb_t *cb,
+                          nv_dma_fence_func_t func) {
+#if defined(NV_LINUX_FENCE_H_PRESENT)
+    return fence_add_callback(fence, cb, func);
+#else
+    return dma_fence_add_callback(fence, cb, func);
+#endif
+}
+
 #endif /* defined(NV_DRM_FENCE_AVAILABLE) */

 #endif /* __NVIDIA_DMA_FENCE_HELPER_H__ */
--- a/kernel-open/nvidia-drm/nvidia-dma-resv-helper.h
+++ b/kernel-open/nvidia-drm/nvidia-dma-resv-helper.h
@@ -121,6 +121,20 @@ static inline void nv_dma_resv_add_excl_fence(nv_dma_resv_t *obj,
 #endif
 }

+static inline void nv_dma_resv_add_shared_fence(nv_dma_resv_t *obj,
+                                                nv_dma_fence_t *fence)
+{
+#if defined(NV_LINUX_DMA_RESV_H_PRESENT)
+#if defined(NV_DMA_RESV_ADD_FENCE_PRESENT)
+    dma_resv_add_fence(obj, fence, DMA_RESV_USAGE_READ);
+#else
+    dma_resv_add_shared_fence(obj, fence);
+#endif
+#else
+    reservation_object_add_shared_fence(obj, fence);
+#endif
+}
+
 #endif /* defined(NV_DRM_FENCE_AVAILABLE) */

 #endif /* __NVIDIA_DMA_RESV_HELPER_H__ */
--- a/kernel-open/nvidia-drm/nvidia-drm-conftest.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-conftest.h
@@ -24,6 +24,7 @@
 #define __NVIDIA_DRM_CONFTEST_H__

 #include "conftest.h"
+#include "nvtypes.h"

 /*
 * NOTE: This file is expected to get included at the top before including any
@@ -61,4 +62,132 @@
 #undef NV_DRM_FENCE_AVAILABLE
 #endif

+/*
+ * We can support color management if either drm_helper_crtc_enable_color_mgmt()
+ * or drm_crtc_enable_color_mgmt() exist.
+ */
+#if defined(NV_DRM_HELPER_CRTC_ENABLE_COLOR_MGMT_PRESENT) || \
+    defined(NV_DRM_CRTC_ENABLE_COLOR_MGMT_PRESENT)
+#define NV_DRM_COLOR_MGMT_AVAILABLE
+#else
+#undef NV_DRM_COLOR_MGMT_AVAILABLE
+#endif
+
+/*
+ * Adapt to quirks in FreeBSD's Linux kernel compatibility layer.
+ */
+#if defined(NV_BSD)
+
+#include <linux/rwsem.h>
+#include <sys/param.h>
+#include <sys/lock.h>
+#include <sys/sx.h>
+
+/* For nv_drm_gem_prime_force_fence_signal */
+#ifndef spin_is_locked
+#define spin_is_locked(lock) mtx_owned(lock.m)
+#endif
+
+#ifndef rwsem_is_locked
+#define rwsem_is_locked(sem) (((sem)->sx.sx_lock & (SX_LOCK_SHARED)) \
+                              || ((sem)->sx.sx_lock & ~(SX_LOCK_FLAGMASK & ~SX_LOCK_SHARED)))
+#endif
+
+/*
+ * FreeBSD does not define vm_flags_t in its linuxkpi, since there is already
+ * a FreeBSD vm_flags_t (of a different size) and they don't want the names to
+ * collide. Temporarily redefine it when including nv-mm.h
+ */
+#define vm_flags_t unsigned long
+#include "nv-mm.h"
+#undef vm_flags_t
+
+/*
+ * sys/nv.h and nvidia/nv.h have the same header guard
+ * we need to clear it for nvlist_t to get loaded
+ */
+#undef _NV_H_
+#include <sys/nv.h>
+
+/*
+ * For now just use set_page_dirty as the lock variant
+ * is not ported for FreeBSD. (in progress). This calls
+ * vm_page_dirty. Used in nv-mm.h
+ */
+#define set_page_dirty_lock set_page_dirty
+
+/*
+ * FreeBSD does not implement drm_atomic_state_free, simply
+ * default to drm_atomic_state_put
+ */
+#define drm_atomic_state_free drm_atomic_state_put
+
+#if __FreeBSD_version < 1300000
+/* redefine LIST_HEAD_INIT to the linux version */
+#include <linux/list.h>
+#define LIST_HEAD_INIT(name) LINUX_LIST_HEAD_INIT(name)
+#endif
+
+/*
+ * FreeBSD currently has only vmf_insert_pfn_prot defined, and it has a
+ * static assert warning not to use it since all of DRM's usages are in
+ * loops with the vm obj lock(s) held. Instead we should use the lkpi
+ * function itself directly. For us none of this applies so we can just
+ * wrap it in our own definition of vmf_insert_pfn
+ */
+#ifndef NV_VMF_INSERT_PFN_PRESENT
+#define NV_VMF_INSERT_PFN_PRESENT 1
+
+#if __FreeBSD_version < 1300000
+#define VM_SHARED       (1 << 17)
+
+/* Not present in 12.2 */
+static inline vm_fault_t
+lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr,
+    unsigned long pfn, pgprot_t prot)
+{
+       vm_object_t vm_obj = vma->vm_obj;
+       vm_page_t page;
+       vm_pindex_t pindex;
+
+       VM_OBJECT_ASSERT_WLOCKED(vm_obj);
+       pindex = OFF_TO_IDX(addr - vma->vm_start);
+       if (vma->vm_pfn_count == 0)
+               vma->vm_pfn_first = pindex;
+       MPASS(pindex <= OFF_TO_IDX(vma->vm_end));
+
+       page = vm_page_grab(vm_obj, pindex, VM_ALLOC_NORMAL);
+       if (page == NULL) {
+               page = PHYS_TO_VM_PAGE(IDX_TO_OFF(pfn));
+               vm_page_xbusy(page);
+               if (vm_page_insert(page, vm_obj, pindex)) {
+                       vm_page_xunbusy(page);
+                       return (VM_FAULT_OOM);
+               }
+               page->valid = VM_PAGE_BITS_ALL;
+       }
+       pmap_page_set_memattr(page, pgprot2cachemode(prot));
+       vma->vm_pfn_count++;
+
+       return (VM_FAULT_NOPAGE);
+}
+#endif
+
+static inline vm_fault_t
+vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
+    unsigned long pfn)
+{
+       vm_fault_t ret;
+
+       VM_OBJECT_WLOCK(vma->vm_obj);
+       ret = lkpi_vmf_insert_pfn_prot_locked(vma, addr, pfn, vma->vm_page_prot);
+       VM_OBJECT_WUNLOCK(vma->vm_obj);
+
+       return (ret);
+}
+
+#endif
+
+#endif /* defined(NV_BSD) */
+
 #endif /* defined(__NVIDIA_DRM_CONFTEST_H__) */
--- a/kernel-open/nvidia-drm/nvidia-drm-connector.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-connector.c
@@ -349,10 +349,125 @@ nv_drm_connector_best_encoder(struct drm_connector *connector)
    return NULL;
 }

+#if defined(NV_DRM_MODE_CREATE_DP_COLORSPACE_PROPERTY_HAS_SUPPORTED_COLORSPACES_ARG)
+static const NvU32 __nv_drm_connector_supported_colorspaces =
+    BIT(DRM_MODE_COLORIMETRY_BT2020_RGB) |
+    BIT(DRM_MODE_COLORIMETRY_BT2020_YCC);
+#endif
+
+#if defined(NV_DRM_CONNECTOR_ATTACH_HDR_OUTPUT_METADATA_PROPERTY_PRESENT)
+static int
+__nv_drm_connector_atomic_check(struct drm_connector *connector,
+                                struct drm_atomic_state *state)
+{
+    struct drm_connector_state *new_connector_state =
+        drm_atomic_get_new_connector_state(state, connector);
+    struct drm_connector_state *old_connector_state =
+        drm_atomic_get_old_connector_state(state, connector);
+    struct nv_drm_device *nv_dev = to_nv_device(connector->dev);
+
+    struct drm_crtc *crtc = new_connector_state->crtc;
+    struct drm_crtc_state *crtc_state;
+    struct nv_drm_crtc_state *nv_crtc_state;
+    struct NvKmsKapiHeadRequestedConfig *req_config;
+
+    if (!crtc) {
+        return 0;
+    }
+
+    crtc_state = drm_atomic_get_new_crtc_state(state, crtc);
+    nv_crtc_state = to_nv_crtc_state(crtc_state);
+    req_config = &nv_crtc_state->req_config;
+
+    /*
+     * Override metadata for the entire head instead of allowing NVKMS to derive
+     * it from the layers' metadata.
+     *
+     * This is the metadata that will sent to the display, and if applicable,
+     * layers will be tone mapped to this metadata rather than that of the
+     * display.
+     */
+    req_config->flags.hdrInfoFrameChanged =
+        !drm_connector_atomic_hdr_metadata_equal(old_connector_state,
+                                                 new_connector_state);
+    if (new_connector_state->hdr_output_metadata &&
+        new_connector_state->hdr_output_metadata->data) {
+
+        /*
+         * Note that HDMI definitions are used here even though we might not
+         * be using HDMI. While that seems odd, it is consistent with
+         * upstream behavior.
+         */
+
+        struct hdr_output_metadata *hdr_metadata =
+            new_connector_state->hdr_output_metadata->data;
+        struct hdr_metadata_infoframe *info_frame =
+            &hdr_metadata->hdmi_metadata_type1;
+        unsigned int i;
+
+        if (hdr_metadata->metadata_type != HDMI_STATIC_METADATA_TYPE1) {
+            return -EINVAL;
+        }
+
+        for (i = 0; i < ARRAY_SIZE(info_frame->display_primaries); i++) {
+            req_config->modeSetConfig.hdrInfoFrame.staticMetadata.displayPrimaries[i].x =
+                info_frame->display_primaries[i].x;
+            req_config->modeSetConfig.hdrInfoFrame.staticMetadata.displayPrimaries[i].y =
+                info_frame->display_primaries[i].y;
+        }
+
+        req_config->modeSetConfig.hdrInfoFrame.staticMetadata.whitePoint.x =
+            info_frame->white_point.x;
+        req_config->modeSetConfig.hdrInfoFrame.staticMetadata.whitePoint.y =
+            info_frame->white_point.y;
+        req_config->modeSetConfig.hdrInfoFrame.staticMetadata.maxDisplayMasteringLuminance =
+            info_frame->max_display_mastering_luminance;
+        req_config->modeSetConfig.hdrInfoFrame.staticMetadata.minDisplayMasteringLuminance =
+            info_frame->min_display_mastering_luminance;
+        req_config->modeSetConfig.hdrInfoFrame.staticMetadata.maxCLL =
+            info_frame->max_cll;
+        req_config->modeSetConfig.hdrInfoFrame.staticMetadata.maxFALL =
+            info_frame->max_fall;
+
+        req_config->modeSetConfig.hdrInfoFrame.eotf = info_frame->eotf;
+
+        req_config->modeSetConfig.hdrInfoFrame.enabled = NV_TRUE;
+    } else {
+        req_config->modeSetConfig.hdrInfoFrame.enabled = NV_FALSE;
+    }
+
+    req_config->flags.colorimetryChanged =
+        (old_connector_state->colorspace != new_connector_state->colorspace);
+    // When adding a case here, also add to __nv_drm_connector_supported_colorspaces
+    switch (new_connector_state->colorspace) {
+        case DRM_MODE_COLORIMETRY_DEFAULT:
+            req_config->modeSetConfig.colorimetry =
+                NVKMS_OUTPUT_COLORIMETRY_DEFAULT;
+            break;
+        case DRM_MODE_COLORIMETRY_BT2020_RGB:
+        case DRM_MODE_COLORIMETRY_BT2020_YCC:
+            // Ignore RGB/YCC
+            // See https://patchwork.freedesktop.org/patch/525496/?series=111865&rev=4
+            req_config->modeSetConfig.colorimetry =
+                NVKMS_OUTPUT_COLORIMETRY_BT2100;
+            break;
+        default:
+            // XXX HDR TODO: Add support for more color spaces
+            NV_DRM_DEV_LOG_ERR(nv_dev, "Unsupported color space");
+            return -EINVAL;
+    }
+
+    return 0;
+}
+#endif /* defined(NV_DRM_CONNECTOR_ATTACH_HDR_OUTPUT_METADATA_PROPERTY_PRESENT) */
+
 static const struct drm_connector_helper_funcs nv_connector_helper_funcs = {
    .get_modes    = nv_drm_connector_get_modes,
    .mode_valid   = nv_drm_connector_mode_valid,
    .best_encoder = nv_drm_connector_best_encoder,
+#if defined(NV_DRM_CONNECTOR_ATTACH_HDR_OUTPUT_METADATA_PROPERTY_PRESENT)
+    .atomic_check = __nv_drm_connector_atomic_check,
+#endif
 };

 static struct drm_connector*
@@ -405,6 +520,32 @@ nv_drm_connector_new(struct drm_device *dev,
            DRM_CONNECTOR_POLL_CONNECT | DRM_CONNECTOR_POLL_DISCONNECT;
    }

+#if defined(NV_DRM_CONNECTOR_ATTACH_HDR_OUTPUT_METADATA_PROPERTY_PRESENT)
+    if (nv_connector->type == NVKMS_CONNECTOR_TYPE_HDMI) {
+#if defined(NV_DRM_MODE_CREATE_DP_COLORSPACE_PROPERTY_HAS_SUPPORTED_COLORSPACES_ARG)
+        if (drm_mode_create_hdmi_colorspace_property(
+                &nv_connector->base,
+                __nv_drm_connector_supported_colorspaces) == 0) {
+#else
+        if (drm_mode_create_hdmi_colorspace_property(&nv_connector->base) == 0) {
+#endif
+            drm_connector_attach_colorspace_property(&nv_connector->base);
+        }
+        drm_connector_attach_hdr_output_metadata_property(&nv_connector->base);
+    } else if (nv_connector->type == NVKMS_CONNECTOR_TYPE_DP) {
+#if defined(NV_DRM_MODE_CREATE_DP_COLORSPACE_PROPERTY_HAS_SUPPORTED_COLORSPACES_ARG)
+        if (drm_mode_create_dp_colorspace_property(
+                &nv_connector->base,
+                __nv_drm_connector_supported_colorspaces) == 0) {
+#else
+        if (drm_mode_create_dp_colorspace_property(&nv_connector->base) == 0) {
+#endif
+            drm_connector_attach_colorspace_property(&nv_connector->base);
+        }
+        drm_connector_attach_hdr_output_metadata_property(&nv_connector->base);
+    }
+#endif /* defined(NV_DRM_CONNECTOR_ATTACH_HDR_OUTPUT_METADATA_PROPERTY_PRESENT) */
+
    /* Register connector with DRM subsystem */

    ret = drm_connector_register(&nv_connector->base);
--- a/kernel-open/nvidia-drm/nvidia-drm-crtc.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-crtc.c
@@ -48,6 +48,11 @@
 #include <linux/host1x-next.h>
 #endif

+#if defined(NV_DRM_DRM_COLOR_MGMT_H_PRESENT)
+#include <drm/drm_color_mgmt.h>
+#endif
+
+
 #if defined(NV_DRM_HAS_HDR_OUTPUT_METADATA)
 static int
 nv_drm_atomic_replace_property_blob_from_id(struct drm_device *dev,
@@ -87,11 +92,22 @@ static void nv_drm_plane_destroy(struct drm_plane *plane)
    nv_drm_free(nv_plane);
 }

+static inline void
+plane_config_clear(struct NvKmsKapiLayerConfig *layerConfig)
+{
+    if (layerConfig == NULL) {
+        return;
+    }
+
+    memset(layerConfig, 0, sizeof(*layerConfig));
+    layerConfig->csc = NVKMS_IDENTITY_CSC_MATRIX;
+}
+
 static inline void
 plane_req_config_disable(struct NvKmsKapiLayerRequestedConfig *req_config)
 {
    /* Clear layer config */
-    memset(&req_config->config, 0, sizeof(req_config->config));
+    plane_config_clear(&req_config->config);

    /* Set flags to get cleared layer config applied */
    req_config->flags.surfaceChanged = NV_TRUE;
@@ -108,6 +124,45 @@ cursor_req_config_disable(struct NvKmsKapiCursorRequestedConfig *req_config)
    req_config->flags.surfaceChanged = NV_TRUE;
 }

+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+static void color_mgmt_config_ctm_to_csc(struct NvKmsCscMatrix *nvkms_csc,
+                                         struct drm_color_ctm  *drm_ctm)
+{
+    int y;
+
+    /* CTM is a 3x3 matrix while ours is 3x4. Zero out the last column. */
+    nvkms_csc->m[0][3] = nvkms_csc->m[1][3] = nvkms_csc->m[2][3] = 0;
+
+    for (y = 0; y < 3; y++) {
+        int x;
+
+        for (x = 0; x < 3; x++) {
+            /*
+             * Values in the CTM are encoded in S31.32 sign-magnitude fixed-
+             * point format, while NvKms CSC values are signed 2's-complement
+             * S15.16 (Ssign-extend12-3.16?) fixed-point format.
+             */
+            NvU64 ctmVal = drm_ctm->matrix[y*3 + x];
+            NvU64 signBit = ctmVal & (1ULL << 63);
+            NvU64 magnitude = ctmVal & ~signBit;
+
+            /*
+             * Drop the low 16 bits of the fractional part and the high 17 bits
+             * of the integral part. Drop 17 bits to avoid corner cases where
+             * the highest resulting bit is a 1, causing the `cscVal = -cscVal`
+             * line to result in a positive number.
+             */
+            NvS32 cscVal = (magnitude >> 16) & ((1ULL << 31) - 1);
+            if (signBit) {
+                cscVal = -cscVal;
+            }
+
+            nvkms_csc->m[y][x] = cscVal;
+        }
+    }
+}
+#endif /* NV_DRM_COLOR_MGMT_AVAILABLE */
+
 static void
 cursor_plane_req_config_update(struct drm_plane *plane,
                               struct drm_plane_state *plane_state,
@@ -121,12 +176,10 @@ cursor_plane_req_config_update(struct drm_plane *plane,
        return;
    }

-    *req_config = (struct NvKmsKapiCursorRequestedConfig) {
-        .surface = to_nv_framebuffer(plane_state->fb)->pSurface,
-
-        .dstX = plane_state->crtc_x,
-        .dstY = plane_state->crtc_y,
-    };
+    memset(req_config, 0, sizeof(*req_config));
+    req_config->surface = to_nv_framebuffer(plane_state->fb)->pSurface;
+    req_config->dstX = plane_state->crtc_x;
+    req_config->dstY = plane_state->crtc_y;

 #if defined(NV_DRM_ALPHA_BLENDING_AVAILABLE)
    if (plane->blend_mode_property != NULL && plane->alpha_property != NULL) {
@@ -220,22 +273,22 @@ plane_req_config_update(struct drm_plane *plane,
        return 0;
    }

-    *req_config = (struct NvKmsKapiLayerRequestedConfig) {
-        .config = {
-            .surface = to_nv_framebuffer(plane_state->fb)->pSurface,
+    memset(req_config, 0, sizeof(*req_config));

-            /* Source values are 16.16 fixed point */
-            .srcX = plane_state->src_x >> 16,
-            .srcY = plane_state->src_y >> 16,
-            .srcWidth  = plane_state->src_w >> 16,
-            .srcHeight = plane_state->src_h >> 16,
+    req_config->config.surface = to_nv_framebuffer(plane_state->fb)->pSurface;

-            .dstX = plane_state->crtc_x,
-            .dstY = plane_state->crtc_y,
-            .dstWidth  = plane_state->crtc_w,
-            .dstHeight = plane_state->crtc_h,
-        },
-    };
+    /* Source values are 16.16 fixed point */
+    req_config->config.srcX = plane_state->src_x >> 16;
+    req_config->config.srcY = plane_state->src_y >> 16;
+    req_config->config.srcWidth  = plane_state->src_w >> 16;
+    req_config->config.srcHeight = plane_state->src_h >> 16;
+
+    req_config->config.dstX = plane_state->crtc_x;
+    req_config->config.dstY = plane_state->crtc_y;
+    req_config->config.dstWidth  = plane_state->crtc_w;
+    req_config->config.dstHeight = plane_state->crtc_h;
+
+    req_config->config.csc = old_config.csc;

 #if defined(NV_DRM_ROTATION_AVAILABLE)
    /*
@@ -399,27 +452,25 @@ plane_req_config_update(struct drm_plane *plane,
        }

        for (i = 0; i < ARRAY_SIZE(info_frame->display_primaries); i ++) {
-            req_config->config.hdrMetadata.displayPrimaries[i].x =
+            req_config->config.hdrMetadata.val.displayPrimaries[i].x =
                info_frame->display_primaries[i].x;
-            req_config->config.hdrMetadata.displayPrimaries[i].y =
+            req_config->config.hdrMetadata.val.displayPrimaries[i].y =
                info_frame->display_primaries[i].y;
        }

-        req_config->config.hdrMetadata.whitePoint.x =
+        req_config->config.hdrMetadata.val.whitePoint.x =
            info_frame->white_point.x;
-        req_config->config.hdrMetadata.whitePoint.y =
+        req_config->config.hdrMetadata.val.whitePoint.y =
            info_frame->white_point.y;
-        req_config->config.hdrMetadata.maxDisplayMasteringLuminance =
+        req_config->config.hdrMetadata.val.maxDisplayMasteringLuminance =
            info_frame->max_display_mastering_luminance;
-        req_config->config.hdrMetadata.minDisplayMasteringLuminance =
+        req_config->config.hdrMetadata.val.minDisplayMasteringLuminance =
            info_frame->min_display_mastering_luminance;
-        req_config->config.hdrMetadata.maxCLL =
+        req_config->config.hdrMetadata.val.maxCLL =
            info_frame->max_cll;
-        req_config->config.hdrMetadata.maxFALL =
+        req_config->config.hdrMetadata.val.maxFALL =
            info_frame->max_fall;

-        req_config->config.hdrMetadataSpecified = true;
-
        switch (info_frame->eotf) {
            case HDMI_EOTF_SMPTE_ST2084:
                req_config->config.tf = NVKMS_OUTPUT_TF_PQ;
@@ -432,10 +483,21 @@ plane_req_config_update(struct drm_plane *plane,
                NV_DRM_DEV_LOG_ERR(nv_dev, "Unsupported EOTF");
                return -1;
        }
+
+        req_config->config.hdrMetadata.enabled = true;
    } else {
-        req_config->config.hdrMetadataSpecified = false;
+        req_config->config.hdrMetadata.enabled = false;
        req_config->config.tf = NVKMS_OUTPUT_TF_NONE;
    }
+
+    req_config->flags.hdrMetadataChanged =
+        ((old_config.hdrMetadata.enabled !=
+          req_config->config.hdrMetadata.enabled) ||
+         memcmp(&old_config.hdrMetadata.val,
+                &req_config->config.hdrMetadata.val,
+                sizeof(struct NvKmsHDRStaticMetadata)));
+
+    req_config->flags.tfChanged = (old_config.tf != req_config->config.tf);
 #endif

    /*
@@ -564,6 +626,24 @@ static int nv_drm_plane_atomic_check(struct drm_plane *plane,
                return ret;
            }

+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+            if (crtc_state->color_mgmt_changed) {
+                /*
+                 * According to the comment in the Linux kernel's
+                 * drivers/gpu/drm/drm_color_mgmt.c, if this property is NULL,
+                 * the CTM needs to be changed to the identity matrix
+                 */
+                if (crtc_state->ctm) {
+                    color_mgmt_config_ctm_to_csc(&plane_requested_config->config.csc,
+                                                 (struct drm_color_ctm *)crtc_state->ctm->data);
+                } else {
+                    plane_requested_config->config.csc = NVKMS_IDENTITY_CSC_MATRIX;
+                }
+                plane_requested_config->config.cscUseMain = NV_FALSE;
+                plane_requested_config->flags.cscChanged = NV_TRUE;
+            }
+#endif /* NV_DRM_COLOR_MGMT_AVAILABLE */
+
            if (__is_async_flip_requested(plane, crtc_state)) {
                /*
                 * Async flip requests that the flip happen 'as soon as
@@ -604,9 +684,7 @@ static int nv_drm_plane_atomic_set_property(
        to_nv_drm_plane_state(state);

    if (property == nv_dev->nv_out_fence_property) {
-#if defined(NV_LINUX_NVHOST_H_PRESENT) && defined(CONFIG_TEGRA_GRHOST)
-        nv_drm_plane_state->fd_user_ptr = u64_to_user_ptr(val);
-#endif
+        nv_drm_plane_state->fd_user_ptr = (void __user *)(uintptr_t)(val);
        return 0;
    } else if (property == nv_dev->nv_input_colorspace_property) {
        nv_drm_plane_state->input_colorspace = val;
@@ -654,6 +732,38 @@ static int nv_drm_plane_atomic_get_property(
    return -EINVAL;
 }

+/**
+ * nv_drm_plane_atomic_reset - plane state reset hook
+ * @plane: DRM plane
+ *
+ * Allocate an empty DRM plane state.
+ */
+static void nv_drm_plane_atomic_reset(struct drm_plane *plane)
+{
+    struct nv_drm_plane_state *nv_plane_state =
+        nv_drm_calloc(1, sizeof(*nv_plane_state));
+
+    if (!nv_plane_state) {
+        return;
+    }
+
+    drm_atomic_helper_plane_reset(plane);
+
+    /*
+     * The drm atomic helper function allocates a state object that is the wrong
+     * size. Copy its contents into the one we allocated above and replace the
+     * pointer.
+     */
+    if (plane->state) {
+        nv_plane_state->base = *plane->state;
+        kfree(plane->state);
+        plane->state = &nv_plane_state->base;
+    } else {
+        kfree(nv_plane_state);
+    }
+}
+
+
 static struct drm_plane_state *
 nv_drm_plane_atomic_duplicate_state(struct drm_plane *plane)
 {
@@ -692,9 +802,11 @@ static inline void __nv_drm_plane_atomic_destroy_state(
 #endif

 #if defined(NV_DRM_HAS_HDR_OUTPUT_METADATA)
-    struct nv_drm_plane_state *nv_drm_plane_state =
-        to_nv_drm_plane_state(state);
-    drm_property_blob_put(nv_drm_plane_state->hdr_output_metadata);
+    {
+        struct nv_drm_plane_state *nv_drm_plane_state =
+            to_nv_drm_plane_state(state);
+        drm_property_blob_put(nv_drm_plane_state->hdr_output_metadata);
+    }
 #endif
 }

@@ -711,7 +823,7 @@ static const struct drm_plane_funcs nv_plane_funcs = {
    .update_plane           = drm_atomic_helper_update_plane,
    .disable_plane          = drm_atomic_helper_disable_plane,
    .destroy                = nv_drm_plane_destroy,
-    .reset                  = drm_atomic_helper_plane_reset,
+    .reset                  = nv_drm_plane_atomic_reset,
    .atomic_get_property    = nv_drm_plane_atomic_get_property,
    .atomic_set_property    = nv_drm_plane_atomic_set_property,
    .atomic_duplicate_state = nv_drm_plane_atomic_duplicate_state,
@@ -757,14 +869,58 @@ static inline void nv_drm_crtc_duplicate_req_head_modeset_config(
     * there is no change in new configuration yet with respect
     * to older one!
     */
-    *new = (struct NvKmsKapiHeadRequestedConfig) {
-        .modeSetConfig = old->modeSetConfig,
-    };
+    memset(new, 0, sizeof(*new));
+    new->modeSetConfig = old->modeSetConfig;

    for (i = 0; i < ARRAY_SIZE(old->layerRequestedConfig); i++) {
-        new->layerRequestedConfig[i] = (struct NvKmsKapiLayerRequestedConfig) {
-            .config = old->layerRequestedConfig[i].config,
-        };
+        new->layerRequestedConfig[i].config =
+            old->layerRequestedConfig[i].config;
+    }
+}
+
+static inline struct nv_drm_crtc_state *nv_drm_crtc_state_alloc(void)
+{
+    struct nv_drm_crtc_state *nv_state = nv_drm_calloc(1, sizeof(*nv_state));
+    int i;
+
+    if (nv_state == NULL) {
+        return NULL;
+    }
+
+    for (i = 0; i < ARRAY_SIZE(nv_state->req_config.layerRequestedConfig); i++) {
+        plane_config_clear(&nv_state->req_config.layerRequestedConfig[i].config);
+    }
+    return nv_state;
+}
+
+
+/**
+ * nv_drm_atomic_crtc_reset - crtc state reset hook
+ * @crtc: DRM crtc
+ *
+ * Allocate an empty DRM crtc state.
+ */
+static void nv_drm_atomic_crtc_reset(struct drm_crtc *crtc)
+{
+    struct nv_drm_crtc_state *nv_state = nv_drm_crtc_state_alloc();
+
+    if (!nv_state) {
+        return;
+    }
+
+    drm_atomic_helper_crtc_reset(crtc);
+
+    /*
+     * The drm atomic helper function allocates a state object that is the wrong
+     * size. Copy its contents into the one we allocated above and replace the
+     * pointer.
+     */
+    if (crtc->state) {
+        nv_state->base = *crtc->state;
+        kfree(crtc->state);
+        crtc->state = &nv_state->base;
+    } else {
+        kfree(nv_state);
    }
 }

@@ -779,7 +935,7 @@ static inline void nv_drm_crtc_duplicate_req_head_modeset_config(
 static struct drm_crtc_state*
 nv_drm_atomic_crtc_duplicate_state(struct drm_crtc *crtc)
 {
-    struct nv_drm_crtc_state *nv_state = nv_drm_calloc(1, sizeof(*nv_state));
+    struct nv_drm_crtc_state *nv_state = nv_drm_crtc_state_alloc();

    if (nv_state == NULL) {
        return NULL;
@@ -800,6 +956,9 @@ nv_drm_atomic_crtc_duplicate_state(struct drm_crtc *crtc)
        &(to_nv_crtc_state(crtc->state)->req_config),
        &nv_state->req_config);

+    nv_state->ilut_ramps = NULL;
+    nv_state->olut_ramps = NULL;
+
    return &nv_state->base;
 }

@@ -823,16 +982,22 @@ static void nv_drm_atomic_crtc_destroy_state(struct drm_crtc *crtc,

    __nv_drm_atomic_helper_crtc_destroy_state(crtc, &nv_state->base);

+    nv_drm_free(nv_state->ilut_ramps);
+    nv_drm_free(nv_state->olut_ramps);
+
    nv_drm_free(nv_state);
 }

 static struct drm_crtc_funcs nv_crtc_funcs = {
    .set_config             = drm_atomic_helper_set_config,
    .page_flip              = drm_atomic_helper_page_flip,
-    .reset                  = drm_atomic_helper_crtc_reset,
+    .reset                  = nv_drm_atomic_crtc_reset,
    .destroy                = nv_drm_crtc_destroy,
    .atomic_duplicate_state = nv_drm_atomic_crtc_duplicate_state,
    .atomic_destroy_state   = nv_drm_atomic_crtc_destroy_state,
+#if defined(NV_DRM_ATOMIC_HELPER_LEGACY_GAMMA_SET_PRESENT)
+    .gamma_set = drm_atomic_helper_legacy_gamma_set,
+#endif
 };

 /*
@@ -866,6 +1031,132 @@ static int head_modeset_config_attach_connector(
    return 0;
 }

+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+static int color_mgmt_config_copy_lut(struct NvKmsLutRamps *nvkms_lut,
+                                      struct drm_color_lut *drm_lut,
+                                      uint64_t lut_len)
+{
+    uint64_t i = 0;
+    if (lut_len != NVKMS_LUT_ARRAY_SIZE) {
+        return -EINVAL;
+    }
+
+    /*
+     * Both NvKms and drm LUT values are 16-bit linear values. NvKms LUT ramps
+     * are in arrays in a single struct while drm LUT ramps are an array of
+     * structs.
+     */
+    for (i = 0; i < lut_len; i++) {
+        nvkms_lut->red[i]   = drm_lut[i].red;
+        nvkms_lut->green[i] = drm_lut[i].green;
+        nvkms_lut->blue[i]  = drm_lut[i].blue;
+    }
+    return 0;
+}
+
+static int color_mgmt_config_set_luts(struct nv_drm_crtc_state *nv_crtc_state,
+                                      struct NvKmsKapiHeadRequestedConfig *req_config)
+{
+    struct NvKmsKapiHeadModeSetConfig *modeset_config =
+        &req_config->modeSetConfig;
+    struct drm_crtc_state *crtc_state = &nv_crtc_state->base;
+    int ret = 0;
+
+    /*
+     * According to the comment in the Linux kernel's
+     * drivers/gpu/drm/drm_color_mgmt.c, if either property is NULL, that LUT
+     * needs to be changed to a linear LUT
+     */
+
+    req_config->flags.lutChanged = NV_TRUE;
+    if (crtc_state->degamma_lut) {
+        struct drm_color_lut *degamma_lut = NULL;
+        uint64_t degamma_len = 0;
+
+        nv_crtc_state->ilut_ramps = nv_drm_calloc(1, sizeof(*nv_crtc_state->ilut_ramps));
+        if (!nv_crtc_state->ilut_ramps) {
+            ret = -ENOMEM;
+            goto fail;
+        }
+
+        degamma_lut = (struct drm_color_lut *)crtc_state->degamma_lut->data;
+        degamma_len = crtc_state->degamma_lut->length /
+                      sizeof(struct drm_color_lut);
+
+        if ((ret = color_mgmt_config_copy_lut(nv_crtc_state->ilut_ramps,
+                                              degamma_lut,
+                                              degamma_len)) != 0) {
+            goto fail;
+        }
+
+        modeset_config->lut.input.specified = NV_TRUE;
+        modeset_config->lut.input.depth     = 30; /* specify the full LUT */
+        modeset_config->lut.input.start     = 0;
+        modeset_config->lut.input.end       = degamma_len - 1;
+        modeset_config->lut.input.pRamps    = nv_crtc_state->ilut_ramps;
+    } else {
+        /* setting input.end to 0 is equivalent to disabling the LUT, which
+         * should be equivalent to a linear LUT */
+        modeset_config->lut.input.specified = NV_TRUE;
+        modeset_config->lut.input.depth     = 30; /* specify the full LUT */
+        modeset_config->lut.input.start     = 0;
+        modeset_config->lut.input.end       = 0;
+        modeset_config->lut.input.pRamps    = NULL;
+
+    }
+
+    if (crtc_state->gamma_lut) {
+        struct drm_color_lut *gamma_lut = NULL;
+        uint64_t gamma_len = 0;
+
+        nv_crtc_state->olut_ramps = nv_drm_calloc(1, sizeof(*nv_crtc_state->olut_ramps));
+        if (!nv_crtc_state->olut_ramps) {
+            ret = -ENOMEM;
+            goto fail;
+        }
+
+        gamma_lut = (struct drm_color_lut *)crtc_state->gamma_lut->data;
+        gamma_len = crtc_state->gamma_lut->length /
+                    sizeof(struct drm_color_lut);
+
+        if ((ret = color_mgmt_config_copy_lut(nv_crtc_state->olut_ramps,
+                                              gamma_lut,
+                                              gamma_len)) != 0) {
+            goto fail;
+        }
+
+        modeset_config->lut.output.specified = NV_TRUE;
+        modeset_config->lut.output.enabled   = NV_TRUE;
+        modeset_config->lut.output.pRamps    = nv_crtc_state->olut_ramps;
+    } else {
+        /* disabling the output LUT should be equivalent to setting a linear
+         * LUT */
+        modeset_config->lut.output.specified = NV_TRUE;
+        modeset_config->lut.output.enabled   = NV_FALSE;
+        modeset_config->lut.output.pRamps    = NULL;
+    }
+
+    return 0;
+
+fail:
+    /* free allocated state */
+    nv_drm_free(nv_crtc_state->ilut_ramps);
+    nv_drm_free(nv_crtc_state->olut_ramps);
+
+    /* remove dangling pointers */
+    nv_crtc_state->ilut_ramps = NULL;
+    nv_crtc_state->olut_ramps = NULL;
+    modeset_config->lut.input.pRamps = NULL;
+    modeset_config->lut.output.pRamps = NULL;
+
+    /* prevent attempts at reading NULLs */
+    modeset_config->lut.input.specified = NV_FALSE;
+    modeset_config->lut.output.specified = NV_FALSE;
+
+    return ret;
+}
+#endif /* NV_DRM_COLOR_MGMT_AVAILABLE */
+
 /**
 * nv_drm_crtc_atomic_check() can fail after it has modified
 * the 'nv_drm_crtc_state::req_config', that is fine because 'nv_drm_crtc_state'
@@ -887,6 +1178,9 @@ static int nv_drm_crtc_atomic_check(struct drm_crtc *crtc,
    struct NvKmsKapiHeadRequestedConfig *req_config =
        &nv_crtc_state->req_config;
    int ret = 0;
+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+    struct nv_drm_device *nv_dev = to_nv_device(crtc_state->crtc->dev);
+#endif

    if (crtc_state->mode_changed) {
        drm_mode_to_nvkms_display_mode(&crtc_state->mode,
@@ -925,6 +1219,25 @@ static int nv_drm_crtc_atomic_check(struct drm_crtc *crtc,
        req_config->flags.activeChanged = NV_TRUE;
    }

+#if defined(NV_DRM_CRTC_STATE_HAS_VRR_ENABLED)
+    req_config->modeSetConfig.vrrEnabled = crtc_state->vrr_enabled;
+#endif
+
+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+    if (nv_dev->drmMasterChangedSinceLastAtomicCommit &&
+        (crtc_state->degamma_lut ||
+         crtc_state->ctm ||
+         crtc_state->gamma_lut)) {
+
+        crtc_state->color_mgmt_changed = NV_TRUE;
+    }
+    if (crtc_state->color_mgmt_changed) {
+        if ((ret = color_mgmt_config_set_luts(nv_crtc_state, req_config)) != 0) {
+            return ret;
+        }
+    }
+#endif
+
    return ret;
 }

@@ -1156,6 +1469,8 @@ nv_drm_plane_create(struct drm_device *dev,
            plane,
            validLayerRRTransforms);

+    nv_drm_free(formats);
+
    return plane;

 failed_plane_init:
@@ -1187,7 +1502,7 @@ static struct drm_crtc *__nv_drm_crtc_create(struct nv_drm_device *nv_dev,
        goto failed;
    }

-    nv_state = nv_drm_calloc(1, sizeof(*nv_state));
+    nv_state = nv_drm_crtc_state_alloc();
    if (nv_state == NULL) {
        goto failed_state_alloc;
    }
@@ -1220,6 +1535,22 @@ static struct drm_crtc *__nv_drm_crtc_create(struct nv_drm_device *nv_dev,

    drm_crtc_helper_add(&nv_crtc->base, &nv_crtc_helper_funcs);

+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+#if defined(NV_DRM_CRTC_ENABLE_COLOR_MGMT_PRESENT)
+    drm_crtc_enable_color_mgmt(&nv_crtc->base, NVKMS_LUT_ARRAY_SIZE, true,
+                               NVKMS_LUT_ARRAY_SIZE);
+#else
+    drm_helper_crtc_enable_color_mgmt(&nv_crtc->base, NVKMS_LUT_ARRAY_SIZE,
+                                      NVKMS_LUT_ARRAY_SIZE);
+#endif
+    ret = drm_mode_crtc_set_gamma_size(&nv_crtc->base, NVKMS_LUT_ARRAY_SIZE);
+    if (ret != 0) {
+        NV_DRM_DEV_LOG_WARN(
+            nv_dev,
+            "Failed to initialize legacy gamma support for head %u", head);
+    }
+#endif
+
    return &nv_crtc->base;

 failed_init_crtc:
@@ -1328,10 +1659,16 @@ static void NvKmsKapiCrcsToDrm(const struct NvKmsKapiCrcs *crcs,
 {
    drmCrcs->outputCrc32.value = crcs->outputCrc32.value;
    drmCrcs->outputCrc32.supported = crcs->outputCrc32.supported;
+    drmCrcs->outputCrc32.__pad0 = 0;
+    drmCrcs->outputCrc32.__pad1 = 0;
    drmCrcs->rasterGeneratorCrc32.value = crcs->rasterGeneratorCrc32.value;
    drmCrcs->rasterGeneratorCrc32.supported = crcs->rasterGeneratorCrc32.supported;
+    drmCrcs->rasterGeneratorCrc32.__pad0 = 0;
+    drmCrcs->rasterGeneratorCrc32.__pad1 = 0;
    drmCrcs->compositorCrc32.value = crcs->compositorCrc32.value;
    drmCrcs->compositorCrc32.supported = crcs->compositorCrc32.supported;
+    drmCrcs->compositorCrc32.__pad0 = 0;
+    drmCrcs->compositorCrc32.__pad1 = 0;
 }

 int nv_drm_get_crtc_crc32_v2_ioctl(struct drm_device *dev,
--- a/kernel-open/nvidia-drm/nvidia-drm-crtc.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-crtc.h
@@ -129,6 +129,9 @@ struct nv_drm_crtc_state {
     */
    struct NvKmsKapiHeadRequestedConfig req_config;

+    struct NvKmsLutRamps *ilut_ramps;
+    struct NvKmsLutRamps *olut_ramps;
+
    /**
     * @nv_flip:
     *
--- a/kernel-open/nvidia-drm/nvidia-drm-drv.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-drv.c
@@ -44,6 +44,10 @@
 #include <drm/drmP.h>
 #endif

+#if defined(NV_DRM_DRM_ATOMIC_UAPI_H_PRESENT)
+#include <drm/drm_atomic_uapi.h>
+#endif
+
 #if defined(NV_DRM_DRM_VBLANK_H_PRESENT)
 #include <drm/drm_vblank.h>
 #endif
@@ -60,7 +64,17 @@
 #include <drm/drm_ioctl.h>
 #endif

+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+#include <drm/drm_aperture.h>
+#include <drm/drm_fb_helper.h>
+#endif
+
+#if defined(NV_DRM_DRM_FBDEV_GENERIC_H_PRESENT)
+#include <drm/drm_fbdev_generic.h>
+#endif
+
 #include <linux/pci.h>
+#include <linux/workqueue.h>

 /*
 * Commit fcd70cd36b9b ("drm: Split out drm_probe_helper.h")
@@ -84,6 +98,11 @@
 #include <drm/drm_atomic_helper.h>
 #endif

+static int nv_drm_revoke_modeset_permission(struct drm_device *dev,
+                                            struct drm_file *filep,
+                                            NvU32 dpyId);
+static int nv_drm_revoke_sub_ownership(struct drm_device *dev);
+
 static struct nv_drm_device *dev_list = NULL;

 static const char* nv_get_input_colorspace_name(
@@ -354,19 +373,15 @@ static int nv_drm_create_properties(struct nv_drm_device *nv_dev)
        len++;
    }

-#if defined(NV_LINUX_NVHOST_H_PRESENT) && defined(CONFIG_TEGRA_GRHOST)
-    if (!nv_dev->supportsSyncpts) {
-        return 0;
+    if (nv_dev->supportsSyncpts) {
+        nv_dev->nv_out_fence_property =
+            drm_property_create_range(nv_dev->dev, DRM_MODE_PROP_ATOMIC,
+                    "NV_DRM_OUT_FENCE_PTR", 0, U64_MAX);
+        if (nv_dev->nv_out_fence_property == NULL) {
+            return -ENOMEM;
+        }
    }

-    nv_dev->nv_out_fence_property =
-        drm_property_create_range(nv_dev->dev, DRM_MODE_PROP_ATOMIC,
-            "NV_DRM_OUT_FENCE_PTR", 0, U64_MAX);
-    if (nv_dev->nv_out_fence_property == NULL) {
-        return -ENOMEM;
-    }
-#endif
-
    nv_dev->nv_input_colorspace_property =
        drm_property_create_enum(nv_dev->dev, 0, "NV_INPUT_COLORSPACE",
                                 enum_list, len);
@@ -387,6 +402,27 @@ static int nv_drm_create_properties(struct nv_drm_device *nv_dev)
    return 0;
 }

+#if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
+/*
+ * We can't just call drm_kms_helper_hotplug_event directly because
+ * fbdev_generic may attempt to set a mode from inside the hotplug event
+ * handler. Because kapi event handling runs on nvkms_kthread_q, this blocks
+ * other event processing including the flip completion notifier expected by
+ * nv_drm_atomic_commit.
+ *
+ * Defer hotplug event handling to a work item so that nvkms_kthread_q can
+ * continue processing events while a DRM modeset is in progress.
+ */
+static void nv_drm_handle_hotplug_event(struct work_struct *work)
+{
+    struct delayed_work *dwork = to_delayed_work(work);
+    struct nv_drm_device *nv_dev =
+        container_of(dwork, struct nv_drm_device, hotplug_event_work);
+
+    drm_kms_helper_hotplug_event(nv_dev->dev);
+}
+#endif
+
 static int nv_drm_load(struct drm_device *dev, unsigned long flags)
 {
 #if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
@@ -440,6 +476,22 @@ static int nv_drm_load(struct drm_device *dev, unsigned long flags)
        return -ENODEV;
    }

+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+    /*
+     * If fbdev is enabled, take modeset ownership now before other DRM clients
+     * can take master (and thus NVKMS ownership).
+     */
+    if (nv_drm_fbdev_module_param) {
+        if (!nvKms->grabOwnership(pDevice)) {
+            nvKms->freeDevice(pDevice);
+            NV_DRM_DEV_LOG_ERR(nv_dev, "Failed to grab NVKMS modeset ownership");
+            return -EBUSY;
+        }
+
+        nv_dev->hasFramebufferConsole = NV_TRUE;
+    }
+#endif
+
    mutex_lock(&nv_dev->lock);

    /* Set NvKmsKapiDevice */
@@ -460,6 +512,11 @@ static int nv_drm_load(struct drm_device *dev, unsigned long flags)

    nv_dev->supportsSyncpts = resInfo.caps.supportsSyncpts;

+    nv_dev->semsurf_stride = resInfo.caps.semsurf.stride;
+
+    nv_dev->semsurf_max_submitted_offset =
+        resInfo.caps.semsurf.maxSubmittedOffset;
+
 #if defined(NV_DRM_FORMAT_MODIFIERS_PRESENT)
    gen = nv_dev->pageKindGeneration;
    kind = nv_dev->genericPageKind;
@@ -517,6 +574,7 @@ static int nv_drm_load(struct drm_device *dev, unsigned long flags)

    /* Enable event handling */

+    INIT_DELAYED_WORK(&nv_dev->hotplug_event_work, nv_drm_handle_hotplug_event);
    atomic_set(&nv_dev->enable_event_handling, true);

    init_waitqueue_head(&nv_dev->flip_event_wq);
@@ -544,8 +602,20 @@ static void __nv_drm_unload(struct drm_device *dev)
        return;
    }

+    /* Release modeset ownership if fbdev is enabled */
+
+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+    if (nv_dev->hasFramebufferConsole) {
+        drm_atomic_helper_shutdown(dev);
+        nvKms->releaseOwnership(nv_dev->pDevice);
+    }
+#endif
+
+    cancel_delayed_work_sync(&nv_dev->hotplug_event_work);
    mutex_lock(&nv_dev->lock);

+    WARN_ON(nv_dev->subOwnershipGranted);
+
    /* Disable event handling */

    atomic_set(&nv_dev->enable_event_handling, false);
@@ -595,9 +665,15 @@ static int __nv_drm_master_set(struct drm_device *dev,
 {
    struct nv_drm_device *nv_dev = to_nv_device(dev);

-    if (!nvKms->grabOwnership(nv_dev->pDevice)) {
+    /*
+     * If this device is driving a framebuffer, then nvidia-drm already has
+     * modeset ownership. Otherwise, grab ownership now.
+     */
+    if (!nv_dev->hasFramebufferConsole &&
+        !nvKms->grabOwnership(nv_dev->pDevice)) {
        return -EINVAL;
    }
+    nv_dev->drmMasterChangedSinceLastAtomicCommit = NV_TRUE;

    return 0;
 }
@@ -631,6 +707,9 @@ void nv_drm_master_drop(struct drm_device *dev, struct drm_file *file_priv)
    struct nv_drm_device *nv_dev = to_nv_device(dev);
    int err;

+    nv_drm_revoke_modeset_permission(dev, file_priv, 0);
+    nv_drm_revoke_sub_ownership(dev);
+
    /*
     * After dropping nvkms modeset onwership, it is not guaranteed that
     * drm and nvkms modeset state will remain in sync.  Therefore, disable
@@ -655,7 +734,9 @@ void nv_drm_master_drop(struct drm_device *dev, struct drm_file *file_priv)

    drm_modeset_unlock_all(dev);

-    nvKms->releaseOwnership(nv_dev->pDevice);
+    if (!nv_dev->hasFramebufferConsole) {
+        nvKms->releaseOwnership(nv_dev->pDevice);
+    }
 }
 #endif /* NV_DRM_ATOMIC_MODESET_AVAILABLE */

@@ -693,16 +774,39 @@ static int nv_drm_get_dev_info_ioctl(struct drm_device *dev,

    params->gpu_id = nv_dev->gpu_info.gpu_id;
    params->primary_index = dev->primary->index;
-#if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
-    params->generic_page_kind = nv_dev->genericPageKind;
-    params->page_kind_generation = nv_dev->pageKindGeneration;
-    params->sector_layout = nv_dev->sectorLayout;
-#else
+    params->supports_alloc = false;
    params->generic_page_kind = 0;
    params->page_kind_generation = 0;
    params->sector_layout = 0;
-#endif
+    params->supports_sync_fd = false;
+    params->supports_semsurf = false;

+#if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
+    /* Memory allocation and semaphore surfaces are only supported
+     * if the modeset = 1 parameter is set */
+    if (nv_dev->pDevice != NULL) {
+        params->supports_alloc = true;
+        params->generic_page_kind = nv_dev->genericPageKind;
+        params->page_kind_generation = nv_dev->pageKindGeneration;
+        params->sector_layout = nv_dev->sectorLayout;
+
+        if (nv_dev->semsurf_stride != 0) {
+            params->supports_semsurf = true;
+#if defined(NV_SYNC_FILE_GET_FENCE_PRESENT)
+            params->supports_sync_fd = true;
+#endif /* defined(NV_SYNC_FILE_GET_FENCE_PRESENT) */
+        }
+    }
+#endif /* defined(NV_DRM_ATOMIC_MODESET_AVAILABLE) */
+
+    return 0;
+}
+
+static int nv_drm_get_drm_file_unique_id_ioctl(struct drm_device *dev,
+                                               void *data, struct drm_file *filep)
+{
+    struct drm_nvidia_get_drm_file_unique_id_params *params = data;
+    params->id = (u64)(filep->driver_priv);
    return 0;
 }

@@ -833,10 +937,10 @@ static NvU32 nv_drm_get_head_bit_from_connector(struct drm_connector *connector)
    return 0;
 }

-static int nv_drm_grant_permission_ioctl(struct drm_device *dev, void *data,
-                                         struct drm_file *filep)
+static int nv_drm_grant_modeset_permission(struct drm_device *dev,
+                                           struct drm_nvidia_grant_permissions_params *params,
+                                           struct drm_file *filep)
 {
-    struct drm_nvidia_grant_permissions_params *params = data;
    struct nv_drm_device *nv_dev = to_nv_device(dev);
    struct nv_drm_connector *target_nv_connector = NULL;
    struct nv_drm_crtc *target_nv_crtc = NULL;
@@ -958,26 +1062,102 @@ done:
    return ret;
 }

-static bool nv_drm_revoke_connector(struct nv_drm_device *nv_dev,
-                                    struct nv_drm_connector *nv_connector)
+static int nv_drm_grant_sub_ownership(struct drm_device *dev,
+                                      struct drm_nvidia_grant_permissions_params *params)
 {
-    bool ret = true;
-    if (nv_connector->modeset_permission_crtc) {
-        if (nv_connector->nv_detected_encoder) {
-            ret = nvKms->revokePermissions(
-                nv_dev->pDevice, nv_connector->modeset_permission_crtc->head,
-                nv_connector->nv_detected_encoder->hDisplay);
-        }
-        nv_connector->modeset_permission_crtc->modeset_permission_filep = NULL;
-        nv_connector->modeset_permission_crtc = NULL;
+    int ret = -EINVAL;
+    struct nv_drm_device *nv_dev = to_nv_device(dev);
+    struct drm_modeset_acquire_ctx *pctx;
+#if NV_DRM_MODESET_LOCK_ALL_END_ARGUMENT_COUNT == 3
+    struct drm_modeset_acquire_ctx ctx;
+    DRM_MODESET_LOCK_ALL_BEGIN(dev, ctx, DRM_MODESET_ACQUIRE_INTERRUPTIBLE,
+                                ret);
+    pctx = &ctx;
+#else
+    mutex_lock(&dev->mode_config.mutex);
+    pctx = dev->mode_config.acquire_ctx;
+#endif
+
+    if (nv_dev->subOwnershipGranted ||
+        !nvKms->grantSubOwnership(params->fd, nv_dev->pDevice)) {
+        goto done;
    }
-    nv_connector->modeset_permission_filep = NULL;
-    return ret;
+
+    /*
+     * When creating an ownership grant, shut down all heads and disable flip
+     * notifications.
+     */
+    ret = nv_drm_atomic_helper_disable_all(dev, pctx);
+    if (ret != 0) {
+        NV_DRM_DEV_LOG_ERR(
+            nv_dev,
+            "nv_drm_atomic_helper_disable_all failed with error code %d!",
+            ret);
+    }
+
+    atomic_set(&nv_dev->enable_event_handling, false);
+    nv_dev->subOwnershipGranted = NV_TRUE;
+
+    ret = 0;
+
+done:
+#if NV_DRM_MODESET_LOCK_ALL_END_ARGUMENT_COUNT == 3
+    DRM_MODESET_LOCK_ALL_END(dev, ctx, ret);
+#else
+    mutex_unlock(&dev->mode_config.mutex);
+#endif
+    return 0;
 }

-static int nv_drm_revoke_permission(struct drm_device *dev,
-                                    struct drm_file *filep, NvU32 dpyId)
+static int nv_drm_grant_permission_ioctl(struct drm_device *dev, void *data,
+                                         struct drm_file *filep)
 {
+    struct drm_nvidia_grant_permissions_params *params = data;
+
+    if (params->type == NV_DRM_PERMISSIONS_TYPE_MODESET) {
+        return nv_drm_grant_modeset_permission(dev, params, filep);
+    } else if (params->type == NV_DRM_PERMISSIONS_TYPE_SUB_OWNER) {
+        return nv_drm_grant_sub_ownership(dev, params);
+    }
+
+    return -EINVAL;
+}
+
+static int
+nv_drm_atomic_disable_connector(struct drm_atomic_state *state,
+                                struct nv_drm_connector *nv_connector)
+{
+    struct drm_crtc_state *crtc_state;
+    struct drm_connector_state *connector_state;
+    int ret = 0;
+
+    if (nv_connector->modeset_permission_crtc) {
+        crtc_state = drm_atomic_get_crtc_state(
+            state, &nv_connector->modeset_permission_crtc->base);
+        if (!crtc_state) {
+            return -EINVAL;
+        }
+
+        crtc_state->active = false;
+        ret = drm_atomic_set_mode_prop_for_crtc(crtc_state, NULL);
+        if (ret < 0) {
+            return ret;
+        }
+    }
+
+    connector_state = drm_atomic_get_connector_state(state, &nv_connector->base);
+    if (!connector_state) {
+        return -EINVAL;
+    }
+
+    return drm_atomic_set_crtc_for_connector(connector_state, NULL);
+}
+
+static int nv_drm_revoke_modeset_permission(struct drm_device *dev,
+                                            struct drm_file *filep, NvU32 dpyId)
+{
+    struct drm_modeset_acquire_ctx *pctx;
+    struct drm_atomic_state *state;
    struct drm_connector *connector;
    struct drm_crtc *crtc;
    int ret = 0;
@@ -988,10 +1168,19 @@ static int nv_drm_revoke_permission(struct drm_device *dev,
    struct drm_modeset_acquire_ctx ctx;
    DRM_MODESET_LOCK_ALL_BEGIN(dev, ctx, DRM_MODESET_ACQUIRE_INTERRUPTIBLE,
                               ret);
+    pctx = &ctx;
 #else
    mutex_lock(&dev->mode_config.mutex);
+    pctx = dev->mode_config.acquire_ctx;
 #endif

+    state = drm_atomic_state_alloc(dev);
+    if (!state) {
+        ret = -ENOMEM;
+        goto done;
+    }
+    state->acquire_ctx = pctx;
+
    /*
     * If dpyId is set, only revoke those specific resources. Otherwise,
     * it is from closing the file so revoke all resources for that filep.
@@ -1003,10 +1192,13 @@ static int nv_drm_revoke_permission(struct drm_device *dev,
        struct nv_drm_connector *nv_connector = to_nv_connector(connector);
        if (nv_connector->modeset_permission_filep == filep &&
            (!dpyId || nv_drm_connector_is_dpy_id(connector, dpyId))) {
-            if (!nv_drm_connector_revoke_permissions(dev, nv_connector)) {
-                ret = -EINVAL;
-                // Continue trying to revoke as much as possible.
+            ret = nv_drm_atomic_disable_connector(state, nv_connector);
+            if (ret < 0) {
+                goto done;
            }
+
+            // Continue trying to revoke as much as possible.
+            nv_drm_connector_revoke_permissions(dev, nv_connector);
        }
    }
 #if defined(NV_DRM_CONNECTOR_LIST_ITER_PRESENT)
@@ -1020,6 +1212,25 @@ static int nv_drm_revoke_permission(struct drm_device *dev,
        }
    }

+    ret = drm_atomic_commit(state);
+done:
+#if defined(NV_DRM_ATOMIC_STATE_REF_COUNTING_PRESENT)
+    drm_atomic_state_put(state);
+#else
+    if (ret != 0) {
+        drm_atomic_state_free(state);
+    } else {
+        /*
+         * In case of success, drm_atomic_commit() takes care to cleanup and
+         * free @state.
+         *
+         * Comment placed above drm_atomic_commit() says: The caller must not
+         * free or in any other way access @state. If the function fails then
+         * the caller must clean up @state itself.
+         */
+    }
+#endif
+
 #if NV_DRM_MODESET_LOCK_ALL_END_ARGUMENT_COUNT == 3
    DRM_MODESET_LOCK_ALL_END(dev, ctx, ret);
 #else
@@ -1029,14 +1240,55 @@ static int nv_drm_revoke_permission(struct drm_device *dev,
    return ret;
 }

+static int nv_drm_revoke_sub_ownership(struct drm_device *dev)
+{
+    int ret = -EINVAL;
+    struct nv_drm_device *nv_dev = to_nv_device(dev);
+#if NV_DRM_MODESET_LOCK_ALL_END_ARGUMENT_COUNT == 3
+    struct drm_modeset_acquire_ctx ctx;
+    DRM_MODESET_LOCK_ALL_BEGIN(dev, ctx, DRM_MODESET_ACQUIRE_INTERRUPTIBLE,
+                               ret);
+#else
+    mutex_lock(&dev->mode_config.mutex);
+#endif
+
+    if (!nv_dev->subOwnershipGranted) {
+        goto done;
+    }
+
+    if (!nvKms->revokeSubOwnership(nv_dev->pDevice)) {
+        NV_DRM_DEV_LOG_ERR(nv_dev, "Failed to revoke sub-ownership from NVKMS");
+        goto done;
+    }
+
+    nv_dev->subOwnershipGranted = NV_FALSE;
+    atomic_set(&nv_dev->enable_event_handling, true);
+    ret = 0;
+
+done:
+#if NV_DRM_MODESET_LOCK_ALL_END_ARGUMENT_COUNT == 3
+    DRM_MODESET_LOCK_ALL_END(dev, ctx, ret);
+#else
+    mutex_unlock(&dev->mode_config.mutex);
+#endif
+    return ret;
+}
+
 static int nv_drm_revoke_permission_ioctl(struct drm_device *dev, void *data,
                                          struct drm_file *filep)
 {
    struct drm_nvidia_revoke_permissions_params *params = data;
-    if (!params->dpyId) {
-        return -EINVAL;
+
+    if (params->type == NV_DRM_PERMISSIONS_TYPE_MODESET) {
+        if (!params->dpyId) {
+            return -EINVAL;
+        }
+        return nv_drm_revoke_modeset_permission(dev, filep, params->dpyId);
+    } else if (params->type == NV_DRM_PERMISSIONS_TYPE_SUB_OWNER) {
+        return nv_drm_revoke_sub_ownership(dev);
    }
-    return nv_drm_revoke_permission(dev, filep, params->dpyId);
+
+    return -EINVAL;
 }

 static void nv_drm_postclose(struct drm_device *dev, struct drm_file *filep)
@@ -1051,11 +1303,22 @@ static void nv_drm_postclose(struct drm_device *dev, struct drm_file *filep)
        dev->mode_config.num_connector > 0 &&
        dev->mode_config.connector_list.next != NULL &&
        dev->mode_config.connector_list.prev != NULL) {
-        nv_drm_revoke_permission(dev, filep, 0);
+        nv_drm_revoke_modeset_permission(dev, filep, 0);
    }
 }
 #endif /* NV_DRM_ATOMIC_MODESET_AVAILABLE */

+static int nv_drm_open(struct drm_device *dev, struct drm_file *filep)
+{
+    _Static_assert(sizeof(filep->driver_priv) >= sizeof(u64),
+                   "filep->driver_priv can not hold an u64");
+    static atomic64_t id = ATOMIC_INIT(0);
+
+    filep->driver_priv = (void *)atomic64_inc_return(&id);
+
+    return 0;
+}
+
 #if defined(NV_DRM_MASTER_HAS_LEASES)
 static struct drm_master *nv_drm_find_lessee(struct drm_master *master,
                                             int lessee_id)
@@ -1299,6 +1562,9 @@ static const struct drm_ioctl_desc nv_drm_ioctls[] = {
    DRM_IOCTL_DEF_DRV(NVIDIA_GET_DEV_INFO,
                      nv_drm_get_dev_info_ioctl,
                      DRM_RENDER_ALLOW|DRM_UNLOCKED),
+    DRM_IOCTL_DEF_DRV(NVIDIA_GET_DRM_FILE_UNIQUE_ID,
+                      nv_drm_get_drm_file_unique_id_ioctl,
+                      DRM_RENDER_ALLOW|DRM_UNLOCKED),

 #if defined(NV_DRM_FENCE_AVAILABLE)
    DRM_IOCTL_DEF_DRV(NVIDIA_FENCE_SUPPORTED,
@@ -1310,11 +1576,35 @@ static const struct drm_ioctl_desc nv_drm_ioctls[] = {
    DRM_IOCTL_DEF_DRV(NVIDIA_GEM_PRIME_FENCE_ATTACH,
                      nv_drm_gem_prime_fence_attach_ioctl,
                      DRM_RENDER_ALLOW|DRM_UNLOCKED),
+    DRM_IOCTL_DEF_DRV(NVIDIA_SEMSURF_FENCE_CTX_CREATE,
+                      nv_drm_semsurf_fence_ctx_create_ioctl,
+                      DRM_RENDER_ALLOW|DRM_UNLOCKED),
+    DRM_IOCTL_DEF_DRV(NVIDIA_SEMSURF_FENCE_CREATE,
+                      nv_drm_semsurf_fence_create_ioctl,
+                      DRM_RENDER_ALLOW|DRM_UNLOCKED),
+    DRM_IOCTL_DEF_DRV(NVIDIA_SEMSURF_FENCE_WAIT,
+                      nv_drm_semsurf_fence_wait_ioctl,
+                      DRM_RENDER_ALLOW|DRM_UNLOCKED),
+    DRM_IOCTL_DEF_DRV(NVIDIA_SEMSURF_FENCE_ATTACH,
+                      nv_drm_semsurf_fence_attach_ioctl,
+                      DRM_RENDER_ALLOW|DRM_UNLOCKED),
 #endif

+    /*
+     * DRM_UNLOCKED is implicit for all non-legacy DRM driver IOCTLs since Linux
+     * v4.10 commit fa5386459f06 "drm: Used DRM_LEGACY for all legacy functions"
+     * (Linux v4.4 commit ea487835e887 "drm: Enforce unlocked ioctl operation
+     * for kms driver ioctls" previously did it only for drivers that set the
+     * DRM_MODESET flag), so this will race with SET_CLIENT_CAP. Linux v4.11
+     * commit dcf727ab5d17 "drm: setclientcap doesn't need the drm BKL" also
+     * removed locking from SET_CLIENT_CAP so there is no use attempting to lock
+     * manually. The latter commit acknowledges that this can expose userspace
+     * to inconsistent behavior when racing with itself, but accepts that risk.
+     */
    DRM_IOCTL_DEF_DRV(NVIDIA_GET_CLIENT_CAPABILITY,
                      nv_drm_get_client_capability_ioctl,
                      0),
+
 #if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
    DRM_IOCTL_DEF_DRV(NVIDIA_GET_CRTC_CRC32,
                      nv_drm_get_crtc_crc32_ioctl,
@@ -1357,6 +1647,9 @@ static struct drm_driver nv_drm_driver = {
    .driver_features        =
 #if defined(NV_DRM_DRIVER_PRIME_FLAG_PRESENT)
                               DRIVER_PRIME |
+#endif
+#if defined(NV_DRM_SYNCOBJ_FEATURES_PRESENT)
+                               DRIVER_SYNCOBJ | DRIVER_SYNCOBJ_TIMELINE |
 #endif
                               DRIVER_GEM  | DRIVER_RENDER,

@@ -1368,14 +1661,14 @@ static struct drm_driver nv_drm_driver = {
    .num_ioctls             = ARRAY_SIZE(nv_drm_ioctls),

 /*
- * linux-next commit 71a7974ac701 ("drm/prime: Unexport helpers for fd/handle
- * conversion") unexports drm_gem_prime_handle_to_fd() and
+ * Linux kernel v6.6 commit 71a7974ac701 ("drm/prime: Unexport helpers
+ * for fd/handle conversion") unexports drm_gem_prime_handle_to_fd() and
 * drm_gem_prime_fd_to_handle().
 *
- * Prior linux-next commit 6b85aa68d9d5 ("drm: Enable PRIME import/export for
- * all drivers") made these helpers the default when .prime_handle_to_fd /
- * .prime_fd_to_handle are unspecified, so it's fine to just skip specifying
- * them if the helpers aren't present.
+ * Prior Linux kernel v6.6 commit 6b85aa68d9d5 ("drm: Enable PRIME
+ * import/export for all drivers") made these helpers the default when
+ * .prime_handle_to_fd / .prime_fd_to_handle are unspecified, so it's fine
+ * to just skip specifying them if the helpers aren't present.
 */
 #if NV_IS_EXPORT_SYMBOL_PRESENT_drm_gem_prime_handle_to_fd
    .prime_handle_to_fd     = drm_gem_prime_handle_to_fd,
@@ -1409,6 +1702,7 @@ static struct drm_driver nv_drm_driver = {
 #if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
    .postclose              = nv_drm_postclose,
 #endif
+    .open                   = nv_drm_open,

    .fops                   = &nv_drm_fops,

@@ -1436,7 +1730,7 @@ static struct drm_driver nv_drm_driver = {
 * kernel supports atomic modeset and the 'modeset' kernel module
 * parameter is true.
 */
-static void nv_drm_update_drm_driver_features(void)
+void nv_drm_update_drm_driver_features(void)
 {
 #if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)

@@ -1462,11 +1756,12 @@ static void nv_drm_update_drm_driver_features(void)
 /*
 * Helper function for allocate/register DRM device for given NVIDIA GPU ID.
 */
-static void nv_drm_register_drm_device(const nv_gpu_info_t *gpu_info)
+void nv_drm_register_drm_device(const nv_gpu_info_t *gpu_info)
 {
    struct nv_drm_device *nv_dev = NULL;
    struct drm_device *dev = NULL;
    struct device *device = gpu_info->os_device_ptr;
+    bool bus_is_pci;

    DRM_DEBUG(
        "Registering device for NVIDIA GPU ID 0x08%x",
@@ -1500,8 +1795,15 @@ static void nv_drm_register_drm_device(const nv_gpu_info_t *gpu_info)
    dev->dev_private = nv_dev;
    nv_dev->dev = dev;

+    bus_is_pci =
+#if defined(NV_LINUX)
+        device->bus == &pci_bus_type;
+#elif defined(NV_BSD)
+        devclass_find("pci");
+#endif
+
 #if defined(NV_DRM_DEVICE_HAS_PDEV)
-    if (device->bus == &pci_bus_type) {
+    if (bus_is_pci) {
        dev->pdev = to_pci_dev(device);
    }
 #endif
@@ -1513,6 +1815,23 @@ static void nv_drm_register_drm_device(const nv_gpu_info_t *gpu_info)
        goto failed_drm_register;
    }

+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+    if (nv_drm_fbdev_module_param &&
+        drm_core_check_feature(dev, DRIVER_MODESET)) {
+
+        if (bus_is_pci) {
+            struct pci_dev *pdev = to_pci_dev(device);
+
+#if defined(NV_DRM_APERTURE_REMOVE_CONFLICTING_PCI_FRAMEBUFFERS_HAS_DRIVER_ARG)
+            drm_aperture_remove_conflicting_pci_framebuffers(pdev, &nv_drm_driver);
+#else
+            drm_aperture_remove_conflicting_pci_framebuffers(pdev, nv_drm_driver.name);
+#endif
+        }
+        drm_fbdev_generic_setup(dev, 32);
+    }
+#endif /* defined(NV_DRM_FBDEV_GENERIC_AVAILABLE) */
+
    /* Add NVIDIA-DRM device into list */

    nv_dev->next = dev_list;
@@ -1532,6 +1851,7 @@ failed_drm_alloc:
 /*
 * Enumerate NVIDIA GPUs and allocate/register DRM device for each of them.
 */
+#if defined(NV_LINUX)
 int nv_drm_probe_devices(void)
 {
    nv_gpu_info_t *gpu_info = NULL;
@@ -1574,6 +1894,7 @@ done:

    return ret;
 }
+#endif

 /*
 * Unregister all NVIDIA DRM devices.
@@ -1582,9 +1903,10 @@ void nv_drm_remove_devices(void)
 {
    while (dev_list != NULL) {
        struct nv_drm_device *next = dev_list->next;
+        struct drm_device *dev = dev_list->dev;

-        drm_dev_unregister(dev_list->dev);
-        nv_drm_dev_free(dev_list->dev);
+        drm_dev_unregister(dev);
+        nv_drm_dev_free(dev);

        nv_drm_free(dev_list);

@@ -1592,4 +1914,79 @@ void nv_drm_remove_devices(void)
    }
 }

+/*
+ * Handle system suspend and resume.
+ *
+ * Normally, a DRM driver would use drm_mode_config_helper_suspend() to save the
+ * current state on suspend and drm_mode_config_helper_resume() to restore it
+ * after resume. This works for upstream drivers because user-mode tasks are
+ * frozen before the suspend hook is called.
+ *
+ * In the case of nvidia-drm, the suspend hook is also called when 'suspend' is
+ * written to /proc/driver/nvidia/suspend, before user-mode tasks are frozen.
+ * However, we don't actually need to save and restore the display state because
+ * the driver requires a VT switch to an unused VT before suspending and a
+ * switch back to the application (or fbdev console) on resume. The DRM client
+ * (or fbdev helper functions) will restore the appropriate mode on resume.
+ *
+ */
+void nv_drm_suspend_resume(NvBool suspend)
+{
+    static DEFINE_MUTEX(nv_drm_suspend_mutex);
+    static NvU32 nv_drm_suspend_count = 0;
+    struct nv_drm_device *nv_dev;
+
+    mutex_lock(&nv_drm_suspend_mutex);
+
+    /*
+     * Count the number of times the driver is asked to suspend. Suspend all DRM
+     * devices on the first suspend call and resume them on the last resume
+     * call.  This is necessary because the kernel may call nvkms_suspend()
+     * simultaneously for each GPU, but NVKMS itself also suspends all GPUs on
+     * the first call.
+     */
+    if (suspend) {
+        if (nv_drm_suspend_count++ > 0) {
+            goto done;
+        }
+    } else {
+        BUG_ON(nv_drm_suspend_count == 0);
+
+        if (--nv_drm_suspend_count > 0) {
+            goto done;
+        }
+    }
+
+#if defined(NV_DRM_ATOMIC_MODESET_AVAILABLE)
+    nv_dev = dev_list;
+
+    /*
+     * NVKMS shuts down all heads on suspend. Update DRM state accordingly.
+     */
+    for (nv_dev = dev_list; nv_dev; nv_dev = nv_dev->next) {
+        struct drm_device *dev = nv_dev->dev;
+
+        if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
+            continue;
+        }
+
+        if (suspend) {
+            drm_kms_helper_poll_disable(dev);
+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+            drm_fb_helper_set_suspend_unlocked(dev->fb_helper, 1);
+#endif
+            drm_mode_config_reset(dev);
+        } else {
+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+            drm_fb_helper_set_suspend_unlocked(dev->fb_helper, 0);
+#endif
+            drm_kms_helper_poll_enable(dev);
+        }
+    }
+#endif /* NV_DRM_ATOMIC_MODESET_AVAILABLE */
+
+done:
+    mutex_unlock(&nv_drm_suspend_mutex);
+}
+
 #endif /* NV_DRM_AVAILABLE */
--- a/kernel-open/nvidia-drm/nvidia-drm-drv.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-drv.h
@@ -31,6 +31,12 @@ int nv_drm_probe_devices(void);

 void nv_drm_remove_devices(void);

+void nv_drm_suspend_resume(NvBool suspend);
+
+void nv_drm_register_drm_device(const nv_gpu_info_t *);
+
+void nv_drm_update_drm_driver_features(void);
+
 #endif /* defined(NV_DRM_AVAILABLE) */

 #endif /* __NVIDIA_DRM_DRV_H__ */
--- a/kernel-open/nvidia-drm/nvidia-drm-encoder.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-encoder.c
@@ -300,7 +300,7 @@ void nv_drm_handle_display_change(struct nv_drm_device *nv_dev,

    nv_drm_connector_mark_connection_status_dirty(nv_encoder->nv_connector);

-    drm_kms_helper_hotplug_event(dev);
+    schedule_delayed_work(&nv_dev->hotplug_event_work, 0);
 }

 void nv_drm_handle_dynamic_display_connected(struct nv_drm_device *nv_dev,
@@ -347,6 +347,6 @@ void nv_drm_handle_dynamic_display_connected(struct nv_drm_device *nv_dev,
    drm_reinit_primary_mode_group(dev);
 #endif

-    drm_kms_helper_hotplug_event(dev);
+    schedule_delayed_work(&nv_dev->hotplug_event_work, 0);
 }
 #endif
--- a/kernel-open/nvidia-drm/nvidia-drm-fb.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-fb.c
@@ -240,7 +240,7 @@ struct drm_framebuffer *nv_drm_internal_framebuffer_create(
        if (nv_dev->modifiers[i] == DRM_FORMAT_MOD_INVALID) {
            NV_DRM_DEV_DEBUG_DRIVER(
                nv_dev,
-                "Invalid format modifier for framebuffer object: 0x%016llx",
+                "Invalid format modifier for framebuffer object: 0x%016" NvU64_fmtx,
                modifier);
            return ERR_PTR(-EINVAL);
        }
--- a/kernel-open/nvidia-drm/nvidia-drm-fence.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-fence.c
--- a/kernel-open/nvidia-drm/nvidia-drm-fence.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-fence.h
@@ -41,6 +41,22 @@ int nv_drm_prime_fence_context_create_ioctl(struct drm_device *dev,
 int nv_drm_gem_prime_fence_attach_ioctl(struct drm_device *dev,
                                        void *data, struct drm_file *filep);

+int nv_drm_semsurf_fence_ctx_create_ioctl(struct drm_device *dev,
+                                          void *data,
+                                          struct drm_file *filep);
+
+int nv_drm_semsurf_fence_create_ioctl(struct drm_device *dev,
+                                      void *data,
+                                      struct drm_file *filep);
+
+int nv_drm_semsurf_fence_wait_ioctl(struct drm_device *dev,
+                                    void *data,
+                                    struct drm_file *filep);
+
+int nv_drm_semsurf_fence_attach_ioctl(struct drm_device *dev,
+                                      void *data,
+                                      struct drm_file *filep);
+
 #endif /* NV_DRM_FENCE_AVAILABLE */

 #endif /* NV_DRM_AVAILABLE */
--- a/kernel-open/nvidia-drm/nvidia-drm-gem-dma-buf.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-gem-dma-buf.c
@@ -71,12 +71,42 @@ static int __nv_drm_gem_dma_buf_create_mmap_offset(
 static int __nv_drm_gem_dma_buf_mmap(struct nv_drm_gem_object *nv_gem,
                                     struct vm_area_struct *vma)
 {
+#if defined(NV_LINUX)
    struct dma_buf_attachment *attach = nv_gem->base.import_attach;
    struct dma_buf *dma_buf = attach->dmabuf;
+#endif
    struct file *old_file;
    int ret;

    /* check if buffer supports mmap */
+#if defined(NV_BSD)
+    /*
+     * Most of the FreeBSD DRM code refers to struct file*, which is actually
+     * a struct linux_file*. The dmabuf code in FreeBSD is not actually plumbed
+     * through the same linuxkpi bits it seems (probably so it can be used
+     * elsewhere), so dma_buf->file really is a native FreeBSD struct file...
+     */
+    if (!nv_gem->base.filp->f_op->mmap)
+        return -EINVAL;
+
+    /* readjust the vma */
+    get_file(nv_gem->base.filp);
+    old_file = vma->vm_file;
+    vma->vm_file = nv_gem->base.filp;
+    vma->vm_pgoff -= drm_vma_node_start(&nv_gem->base.vma_node);
+
+    ret = nv_gem->base.filp->f_op->mmap(nv_gem->base.filp, vma);
+
+    if (ret) {
+        /* restore old parameters on failure */
+        vma->vm_file = old_file;
+        vma->vm_pgoff += drm_vma_node_start(&nv_gem->base.vma_node);
+        fput(nv_gem->base.filp);
+    } else {
+        if (old_file)
+            fput(old_file);
+    }
+#else
    if (!dma_buf->file->f_op->mmap)
        return -EINVAL;

@@ -84,18 +114,20 @@ static int __nv_drm_gem_dma_buf_mmap(struct nv_drm_gem_object *nv_gem,
    get_file(dma_buf->file);
    old_file = vma->vm_file;
    vma->vm_file = dma_buf->file;
-    vma->vm_pgoff -= drm_vma_node_start(&nv_gem->base.vma_node);;
+    vma->vm_pgoff -= drm_vma_node_start(&nv_gem->base.vma_node);

    ret = dma_buf->file->f_op->mmap(dma_buf->file, vma);

    if (ret) {
        /* restore old parameters on failure */
        vma->vm_file = old_file;
+        vma->vm_pgoff += drm_vma_node_start(&nv_gem->base.vma_node);
        fput(dma_buf->file);
    } else {
        if (old_file)
            fput(old_file);
    }
+#endif

    return ret;
 }
--- a/kernel-open/nvidia-drm/nvidia-drm-gem-nvkms-memory.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-gem-nvkms-memory.c
@@ -37,6 +37,9 @@
 #endif

 #include <linux/io.h>
+#if defined(NV_BSD)
+#include <vm/vm_pageout.h>
+#endif

 #include "nv-mm.h"

@@ -93,7 +96,17 @@ static vm_fault_t __nv_drm_gem_nvkms_handle_vma_fault(
    if (nv_nvkms_memory->pages_count == 0) {
        pfn = (unsigned long)(uintptr_t)nv_nvkms_memory->pPhysicalAddress;
        pfn >>= PAGE_SHIFT;
+#if defined(NV_LINUX)
+        /*
+         * FreeBSD doesn't set pgoff. We instead have pfn be the base physical
+         * address, and we will calculate the index pidx from the virtual address.
+         *
+         * This only works because linux_cdev_pager_populate passes the pidx as
+         * vmf->virtual_address. Then we turn the virtual address
+         * into a physical page number.
+         */
        pfn += page_offset;
+#endif
    } else {
        BUG_ON(page_offset >= nv_nvkms_memory->pages_count);
        pfn = page_to_pfn(nv_nvkms_memory->pages[page_offset]);
@@ -243,6 +256,15 @@ static int __nv_drm_nvkms_gem_obj_init(
    NvU64 *pages = NULL;
    NvU32 numPages = 0;

+    if ((size % PAGE_SIZE) != 0) {
+        NV_DRM_DEV_LOG_ERR(
+            nv_dev,
+            "NvKmsKapiMemory 0x%p size should be in a multiple of page size to "
+            "create a gem object",
+            pMemory);
+        return -EINVAL;
+    }
+
    nv_nvkms_memory->pPhysicalAddress = NULL;
    nv_nvkms_memory->pWriteCombinedIORemapAddress = NULL;
    nv_nvkms_memory->physically_mapped = false;
@@ -314,7 +336,7 @@ int nv_drm_dumb_create(
        ret = -ENOMEM;
        NV_DRM_DEV_LOG_ERR(
            nv_dev,
-            "Failed to allocate NvKmsKapiMemory for dumb object of size %llu",
+            "Failed to allocate NvKmsKapiMemory for dumb object of size %" NvU64_fmtu,
            args->size);
        goto nvkms_alloc_memory_failed;
    }
@@ -465,7 +487,7 @@ int nv_drm_gem_alloc_nvkms_memory_ioctl(struct drm_device *dev,
        goto failed;
    }

-    if (p->__pad != 0) {
+    if ((p->__pad0 != 0) || (p->__pad1 != 0)) {
        ret = -EINVAL;
        NV_DRM_DEV_LOG_ERR(nv_dev, "non-zero value in padding field");
        goto failed;
@@ -529,14 +551,12 @@ static struct drm_gem_object *__nv_drm_gem_nvkms_prime_dup(
 {
    struct nv_drm_device *nv_dev = to_nv_device(dev);
    const struct nv_drm_device *nv_dev_src;
-    const struct nv_drm_gem_nvkms_memory *nv_nvkms_memory_src;
    struct nv_drm_gem_nvkms_memory *nv_nvkms_memory;
    struct NvKmsKapiMemory *pMemory;

    BUG_ON(nv_gem_src == NULL || nv_gem_src->ops != &nv_gem_nvkms_memory_ops);

    nv_dev_src = to_nv_device(nv_gem_src->base.dev);
-    nv_nvkms_memory_src = to_nv_nvkms_memory_const(nv_gem_src);

    if ((nv_nvkms_memory =
            nv_drm_calloc(1, sizeof(*nv_nvkms_memory))) == NULL) {
--- a/kernel-open/nvidia-drm/nvidia-drm-gem-user-memory.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-gem-user-memory.c
@@ -36,6 +36,10 @@
 #include "linux/mm.h"
 #include "nv-mm.h"

+#if defined(NV_BSD)
+#include <vm/vm_pageout.h>
+#endif
+
 static inline
 void __nv_drm_gem_user_memory_free(struct nv_drm_gem_object *nv_gem)
 {
@@ -113,6 +117,10 @@ static vm_fault_t __nv_drm_gem_user_memory_handle_vma_fault(
    page_offset = vmf->pgoff - drm_vma_node_start(&gem->vma_node);

    BUG_ON(page_offset >= nv_user_memory->pages_count);
+
+#if !defined(NV_LINUX)
+    ret = vmf_insert_pfn(vma, address, page_to_pfn(nv_user_memory->pages[page_offset]));
+#else /* !defined(NV_LINUX) */
    ret = vm_insert_page(vma, address, nv_user_memory->pages[page_offset]);
    switch (ret) {
        case 0:
@@ -131,6 +139,7 @@ static vm_fault_t __nv_drm_gem_user_memory_handle_vma_fault(
            ret = VM_FAULT_SIGBUS;
            break;
    }
+#endif /* !defined(NV_LINUX) */

    return ret;
 }
@@ -170,7 +179,7 @@ int nv_drm_gem_import_userspace_memory_ioctl(struct drm_device *dev,
    if ((params->size % PAGE_SIZE) != 0) {
        NV_DRM_DEV_LOG_ERR(
            nv_dev,
-            "Userspace memory 0x%llx size should be in a multiple of page "
+            "Userspace memory 0x%" NvU64_fmtx " size should be in a multiple of page "
            "size to create a gem object",
            params->address);
        return -EINVAL;
@@ -183,7 +192,7 @@ int nv_drm_gem_import_userspace_memory_ioctl(struct drm_device *dev,
    if (ret != 0) {
        NV_DRM_DEV_LOG_ERR(
            nv_dev,
-            "Failed to lock user pages for address 0x%llx: %d",
+            "Failed to lock user pages for address 0x%" NvU64_fmtx ": %d",
            params->address, ret);
        return ret;
    }
--- a/kernel-open/nvidia-drm/nvidia-drm-gem.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-gem.h
@@ -95,6 +95,16 @@ static inline struct nv_drm_gem_object *to_nv_gem_object(
 * 3e70fd160cf0b1945225eaa08dd2cb8544f21cb8 (2018-11-15).
 */

+static inline void
+nv_drm_gem_object_reference(struct nv_drm_gem_object *nv_gem)
+{
+#if defined(NV_DRM_GEM_OBJECT_GET_PRESENT)
+    drm_gem_object_get(&nv_gem->base);
+#else
+    drm_gem_object_reference(&nv_gem->base);
+#endif
+}
+
 static inline void
 nv_drm_gem_object_unreference_unlocked(struct nv_drm_gem_object *nv_gem)
 {
--- a/kernel-open/nvidia-drm/nvidia-drm-helper.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-helper.c
@@ -45,8 +45,7 @@

 /*
 * The inclusion of drm_framebuffer.h was removed from drm_crtc.h by commit
- * 720cf96d8fecde29b72e1101f8a567a0ce99594f ("drm: Drop drm_framebuffer.h from
- * drm_crtc.h") in linux-next, expected in v5.19-rc7.
+ * 720cf96d8fec ("drm: Drop drm_framebuffer.h from drm_crtc.h") in v6.0.
 *
 * We only need drm_framebuffer.h for drm_framebuffer_put(), and it is always
 * present (v4.9+) when drm_framebuffer_{put,get}() is present (v4.12+), so it
--- a/kernel-open/nvidia-drm/nvidia-drm-helper.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-helper.h
@@ -306,6 +306,36 @@ int nv_drm_atomic_helper_disable_all(struct drm_device *dev,
    for_each_plane_in_state(__state, plane, plane_state, __i)
 #endif

+/*
+ * for_each_new_plane_in_state() was added by kernel commit
+ * 581e49fe6b411f407102a7f2377648849e0fa37f which was Signed-off-by:
+ *      Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
+ *      Daniel Vetter <daniel.vetter@ffwll.ch>
+ *
+ * This commit also added the old_state and new_state pointers to
+ * __drm_planes_state. Because of this, the best that can be done on kernel
+ * versions without this macro is for_each_plane_in_state.
+ */
+
+/**
+ * nv_drm_for_each_new_plane_in_state - iterate over all planes in an atomic update
+ * @__state: &struct drm_atomic_state pointer
+ * @plane: &struct drm_plane iteration cursor
+ * @new_plane_state: &struct drm_plane_state iteration cursor for the new state
+ * @__i: int iteration cursor, for macro-internal use
+ *
+ * This iterates over all planes in an atomic update, tracking only the new
+ * state. This is useful in enable functions, where we need the new state the
+ * hardware should be in when the atomic commit operation has completed.
+ */
+#if !defined(for_each_new_plane_in_state)
+#define nv_drm_for_each_new_plane_in_state(__state, plane, new_plane_state, __i) \
+    nv_drm_for_each_plane_in_state(__state, plane, new_plane_state, __i)
+#else
+#define nv_drm_for_each_new_plane_in_state(__state, plane, new_plane_state, __i) \
+    for_each_new_plane_in_state(__state, plane, new_plane_state, __i)
+#endif
+
 static inline struct drm_connector *
 nv_drm_connector_lookup(struct drm_device *dev, struct drm_file *filep,
                        uint32_t id)
@@ -582,6 +612,19 @@ static inline int nv_drm_format_num_planes(uint32_t format)

 #endif /* defined(NV_DRM_FORMAT_MODIFIERS_PRESENT) */

+/*
+ * DRM_UNLOCKED was removed with commit 2798ffcc1d6a ("drm: Remove locking for
+ * legacy ioctls and DRM_UNLOCKED") in v6.8, but it was previously made
+ * implicit for all non-legacy DRM driver IOCTLs since Linux v4.10 commit
+ * fa5386459f06 "drm: Used DRM_LEGACY for all legacy functions" (Linux v4.4
+ * commit ea487835e887 "drm: Enforce unlocked ioctl operation for kms driver
+ * ioctls" previously did it only for drivers that set the DRM_MODESET flag), so
+ * it was effectively a no-op anyway.
+ */
+#if !defined(NV_DRM_UNLOCKED_IOCTL_FLAG_PRESENT)
+#define DRM_UNLOCKED 0
+#endif
+
 /*
 * drm_vma_offset_exact_lookup_locked() were added
 * by kernel commit 2225cfe46bcc which was Signed-off-by:
--- a/kernel-open/nvidia-drm/nvidia-drm-ioctl.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-ioctl.h
@@ -48,6 +48,11 @@
 #define DRM_NVIDIA_GET_CONNECTOR_ID_FOR_DPY_ID      0x11
 #define DRM_NVIDIA_GRANT_PERMISSIONS                0x12
 #define DRM_NVIDIA_REVOKE_PERMISSIONS               0x13
+#define DRM_NVIDIA_SEMSURF_FENCE_CTX_CREATE         0x14
+#define DRM_NVIDIA_SEMSURF_FENCE_CREATE             0x15
+#define DRM_NVIDIA_SEMSURF_FENCE_WAIT               0x16
+#define DRM_NVIDIA_SEMSURF_FENCE_ATTACH             0x17
+#define DRM_NVIDIA_GET_DRM_FILE_UNIQUE_ID           0x18

 #define DRM_IOCTL_NVIDIA_GEM_IMPORT_NVKMS_MEMORY                           \
    DRM_IOWR((DRM_COMMAND_BASE + DRM_NVIDIA_GEM_IMPORT_NVKMS_MEMORY),      \
@@ -67,7 +72,7 @@
 *
 * 'warning: suggest parentheses around arithmetic in operand of |'
 */
-#if defined(NV_LINUX)
+#if defined(NV_LINUX) || defined(NV_BSD)
 #define DRM_IOCTL_NVIDIA_FENCE_SUPPORTED                         \
    DRM_IO(DRM_COMMAND_BASE + DRM_NVIDIA_FENCE_SUPPORTED)
 #define DRM_IOCTL_NVIDIA_DMABUF_SUPPORTED                        \
@@ -133,6 +138,31 @@
    DRM_IOWR((DRM_COMMAND_BASE + DRM_NVIDIA_REVOKE_PERMISSIONS),        \
             struct drm_nvidia_revoke_permissions_params)

+#define DRM_IOCTL_NVIDIA_SEMSURF_FENCE_CTX_CREATE                       \
+    DRM_IOWR((DRM_COMMAND_BASE +                                        \
+              DRM_NVIDIA_SEMSURF_FENCE_CTX_CREATE),                     \
+              struct drm_nvidia_semsurf_fence_ctx_create_params)
+
+#define DRM_IOCTL_NVIDIA_SEMSURF_FENCE_CREATE                           \
+    DRM_IOWR((DRM_COMMAND_BASE +                                        \
+              DRM_NVIDIA_SEMSURF_FENCE_CREATE),                         \
+              struct drm_nvidia_semsurf_fence_create_params)
+
+#define DRM_IOCTL_NVIDIA_SEMSURF_FENCE_WAIT                             \
+    DRM_IOW((DRM_COMMAND_BASE +                                         \
+              DRM_NVIDIA_SEMSURF_FENCE_WAIT),                           \
+              struct drm_nvidia_semsurf_fence_wait_params)
+
+#define DRM_IOCTL_NVIDIA_SEMSURF_FENCE_ATTACH                           \
+    DRM_IOW((DRM_COMMAND_BASE +                                         \
+              DRM_NVIDIA_SEMSURF_FENCE_ATTACH),                         \
+              struct drm_nvidia_semsurf_fence_attach_params)
+
+#define DRM_IOCTL_NVIDIA_GET_DRM_FILE_UNIQUE_ID                         \
+    DRM_IOWR((DRM_COMMAND_BASE +                                        \
+              DRM_NVIDIA_GET_DRM_FILE_UNIQUE_ID),                       \
+              struct drm_nvidia_get_drm_file_unique_id_params)
+
 struct drm_nvidia_gem_import_nvkms_memory_params {
    uint64_t mem_size;           /* IN */

@@ -154,10 +184,15 @@ struct drm_nvidia_get_dev_info_params {
    uint32_t gpu_id;             /* OUT */
    uint32_t primary_index;      /* OUT; the "card%d" value */

-    /* See DRM_FORMAT_MOD_NVIDIA_BLOCK_LINEAR_2D definitions of these */
+    uint32_t supports_alloc;     /* OUT */
+    /* The generic_page_kind, page_kind_generation, and sector_layout
+     * fields are only valid if supports_alloc is true.
+     * See DRM_FORMAT_MOD_NVIDIA_BLOCK_LINEAR_2D definitions of these. */
    uint32_t generic_page_kind;    /* OUT */
    uint32_t page_kind_generation; /* OUT */
    uint32_t sector_layout;        /* OUT */
+    uint32_t supports_sync_fd;     /* OUT */
+    uint32_t supports_semsurf;     /* OUT */
 };

 struct drm_nvidia_prime_fence_context_create_params {
@@ -179,6 +214,7 @@ struct drm_nvidia_gem_prime_fence_attach_params {
    uint32_t handle;                /* IN GEM handle to attach fence to */
    uint32_t fence_context_handle;  /* IN GEM handle to fence context on which fence is run on */
    uint32_t sem_thresh;            /* IN Semaphore value to reach before signal */
+    uint32_t __pad;
 };

 struct drm_nvidia_get_client_capability_params {
@@ -190,6 +226,8 @@ struct drm_nvidia_get_client_capability_params {
 struct drm_nvidia_crtc_crc32 {
    uint32_t value; /* Read value, undefined if supported is false */
    uint8_t supported; /* Supported boolean, true if readable by hardware */
+    uint8_t __pad0;
+    uint16_t __pad1;
 };

 struct drm_nvidia_crtc_crc32_v2_out {
@@ -229,10 +267,11 @@ struct drm_nvidia_gem_alloc_nvkms_memory_params {
    uint32_t handle;              /* OUT */
    uint8_t  block_linear;        /* IN */
    uint8_t  compressible;        /* IN/OUT */
-    uint16_t __pad;
+    uint16_t __pad0;

    uint64_t memory_size;         /* IN */
    uint32_t flags;               /* IN */
+    uint32_t __pad1;
 };

 struct drm_nvidia_gem_export_dmabuf_memory_params {
@@ -266,13 +305,94 @@ struct drm_nvidia_get_connector_id_for_dpy_id_params {
    uint32_t connectorId; /* OUT */
 };

+enum drm_nvidia_permissions_type {
+    NV_DRM_PERMISSIONS_TYPE_MODESET = 2,
+    NV_DRM_PERMISSIONS_TYPE_SUB_OWNER = 3
+};
+
 struct drm_nvidia_grant_permissions_params {
    int32_t fd;           /* IN */
    uint32_t dpyId;       /* IN */
+    uint32_t type;        /* IN */
 };

 struct drm_nvidia_revoke_permissions_params {
    uint32_t dpyId;       /* IN */
+    uint32_t type;        /* IN */
+};
+
+struct drm_nvidia_semsurf_fence_ctx_create_params {
+    uint64_t index;             /* IN Index of the desired semaphore in the
+                                 * fence context's semaphore surface */
+
+    /* Params for importing userspace semaphore surface */
+    uint64_t nvkms_params_ptr;  /* IN */
+    uint64_t nvkms_params_size; /* IN */
+
+    uint32_t handle;            /* OUT GEM handle to fence context */
+    uint32_t __pad;
+};
+
+struct drm_nvidia_semsurf_fence_create_params {
+    uint32_t fence_context_handle;  /* IN GEM handle to fence context on which
+                                     * fence is run on */
+
+    uint32_t timeout_value_ms;      /* IN Timeout value in ms for the fence
+                                     * after which the fence will be signaled
+                                     * with its error status set to -ETIMEDOUT.
+                                     * Default timeout value is 5000ms */
+
+    uint64_t wait_value;            /* IN Semaphore value to reach before signal */
+
+    int32_t  fd;                    /* OUT sync FD object representing the
+                                     * semaphore at the specified index reaching
+                                     * a value >= wait_value */
+    uint32_t __pad;
+};
+
+/*
+ * Note there is no provision for timeouts in this ioctl. The kernel
+ * documentation asserts timeouts should be handled by fence producers, and
+ * that waiters should not second-guess their logic, as it is producers rather
+ * than consumers that have better information when it comes to determining a
+ * reasonable timeout for a given workload.
+ */
+struct drm_nvidia_semsurf_fence_wait_params {
+    uint32_t fence_context_handle;  /* IN GEM handle to fence context which will
+                                     * be used to wait on the sync FD.  Need not
+                                     * be the fence context used to create the
+                                     * sync FD. */
+
+    int32_t  fd;                    /* IN sync FD object to wait on */
+
+    uint64_t pre_wait_value;        /* IN Wait for the semaphore represented by
+                                     * fence_context to reach this value before
+                                     * waiting for the sync file. */
+
+    uint64_t post_wait_value;       /* IN Signal the semaphore represented by
+                                     * fence_context to this value after waiting
+                                     * for the sync file */
+};
+
+struct drm_nvidia_semsurf_fence_attach_params {
+    uint32_t handle;                /* IN GEM handle of buffer */
+
+    uint32_t fence_context_handle;  /* IN GEM handle of fence context */
+
+    uint32_t timeout_value_ms;      /* IN Timeout value in ms for the fence
+                                     * after which the fence will be signaled
+                                     * with its error status set to -ETIMEDOUT.
+                                     * Default timeout value is 5000ms */
+
+    uint32_t shared;                /* IN If true, fence will reserve shared
+                                     * access to the buffer, otherwise it will
+                                     * reserve exclusive access */
+
+    uint64_t wait_value;            /* IN Semaphore value to reach before signal */
+};
+
+struct drm_nvidia_get_drm_file_unique_id_params {
+    uint64_t id;                    /* OUT Unique ID of the DRM file */
 };

 #endif /* _UAPI_NVIDIA_DRM_IOCTL_H_ */
--- a/kernel-open/nvidia-drm/nvidia-drm-linux.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-linux.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015, NVIDIA CORPORATION. All rights reserved.
+ * Copyright (c) 2015-2023, NVIDIA CORPORATION. All rights reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
@@ -21,8 +21,6 @@
 */

 #include <linux/module.h>
-#include <linux/slab.h>
-#include <linux/err.h>

 #include "nvidia-drm-os-interface.h"
 #include "nvidia-drm.h"
@@ -31,135 +29,18 @@

 #if defined(NV_DRM_AVAILABLE)

-#if defined(NV_DRM_DRMP_H_PRESENT)
-#include <drm/drmP.h>
-#endif
-
-#include <linux/vmalloc.h>
-
-#include "nv-mm.h"
-
 MODULE_PARM_DESC(
    modeset,
    "Enable atomic kernel modesetting (1 = enable, 0 = disable (default))");
-bool nv_drm_modeset_module_param = false;
 module_param_named(modeset, nv_drm_modeset_module_param, bool, 0400);

-void *nv_drm_calloc(size_t nmemb, size_t size)
-{
-    size_t total_size = nmemb * size;
-    //
-    // Check for overflow.
-    //
-    if ((nmemb != 0) && ((total_size / nmemb) != size))
-    {
-        return NULL;
-    }
-    return kzalloc(nmemb * size, GFP_KERNEL);
-}
-
-void nv_drm_free(void *ptr)
-{
-    if (IS_ERR(ptr)) {
-        return;
-    }
-
-    kfree(ptr);
-}
-
-char *nv_drm_asprintf(const char *fmt, ...)
-{
-    va_list ap;
-    char *p;
-
-    va_start(ap, fmt);
-    p = kvasprintf(GFP_KERNEL, fmt, ap);
-    va_end(ap);
-
-    return p;
-}
-
-#if defined(NVCPU_X86) || defined(NVCPU_X86_64)
-  #define WRITE_COMBINE_FLUSH()    asm volatile("sfence":::"memory")
-#elif defined(NVCPU_FAMILY_ARM)
-  #if defined(NVCPU_ARM)
-    #define WRITE_COMBINE_FLUSH()  { dsb(); outer_sync(); }
-  #elif defined(NVCPU_AARCH64)
-    #define WRITE_COMBINE_FLUSH()  mb()
-  #endif
-#elif defined(NVCPU_PPC64LE)
-  #define WRITE_COMBINE_FLUSH()    asm volatile("sync":::"memory")
+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+MODULE_PARM_DESC(
+    fbdev,
+    "Create a framebuffer device (1 = enable, 0 = disable (default)) (EXPERIMENTAL)");
+module_param_named(fbdev, nv_drm_fbdev_module_param, bool, 0400);
 #endif

-void nv_drm_write_combine_flush(void)
-{
-    WRITE_COMBINE_FLUSH();
-}
-
-int nv_drm_lock_user_pages(unsigned long address,
-                           unsigned long pages_count, struct page ***pages)
-{
-    struct mm_struct *mm = current->mm;
-    struct page **user_pages;
-    int pages_pinned;
-
-    user_pages = nv_drm_calloc(pages_count, sizeof(*user_pages));
-
-    if (user_pages == NULL) {
-        return -ENOMEM;
-    }
-
-    nv_mmap_read_lock(mm);
-
-    pages_pinned = NV_PIN_USER_PAGES(address, pages_count, FOLL_WRITE,
-                                     user_pages, NULL);
-    nv_mmap_read_unlock(mm);
-
-    if (pages_pinned < 0 || (unsigned)pages_pinned < pages_count) {
-        goto failed;
-    }
-
-    *pages = user_pages;
-
-    return 0;
-
-failed:
-
-    if (pages_pinned > 0) {
-        int i;
-
-        for (i = 0; i < pages_pinned; i++) {
-           NV_UNPIN_USER_PAGE(user_pages[i]);
-        }
-    }
-
-    nv_drm_free(user_pages);
-
-    return (pages_pinned < 0) ? pages_pinned : -EINVAL;
-}
-
-void nv_drm_unlock_user_pages(unsigned long  pages_count, struct page **pages)
-{
-    unsigned long i;
-
-    for (i = 0; i < pages_count; i++) {
-        set_page_dirty_lock(pages[i]);
-        NV_UNPIN_USER_PAGE(pages[i]);
-    }
-
-    nv_drm_free(pages);
-}
-
-void *nv_drm_vmap(struct page **pages, unsigned long pages_count)
-{
-    return vmap(pages, pages_count, VM_USERMAP, PAGE_KERNEL);
-}
-
-void nv_drm_vunmap(void *address)
-{
-    vunmap(address);
-}
-
 #endif /* NV_DRM_AVAILABLE */

 /*************************************************************************
--- a/kernel-open/nvidia-drm/nvidia-drm-modeset.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-modeset.c
@@ -237,6 +237,14 @@ nv_drm_atomic_apply_modeset_config(struct drm_device *dev,
    int i;
    int ret;

+    /*
+     * If sub-owner permission was granted to another NVKMS client, disallow
+     * modesets through the DRM interface.
+     */
+    if (nv_dev->subOwnershipGranted) {
+        return -EINVAL;
+    }
+
    memset(requested_config, 0, sizeof(*requested_config));

    /* Loop over affected crtcs and construct NvKmsKapiRequestedModeSetConfig */
@@ -274,9 +282,6 @@ nv_drm_atomic_apply_modeset_config(struct drm_device *dev,

                nv_new_crtc_state->nv_flip = NULL;
            }
-#if defined(NV_DRM_CRTC_STATE_HAS_VRR_ENABLED)
-            requested_config->headRequestedConfig[nv_crtc->head].modeSetConfig.vrrEnabled = new_crtc_state->vrr_enabled;
-#endif
        }
    }

@@ -292,7 +297,9 @@ nv_drm_atomic_apply_modeset_config(struct drm_device *dev,
                                   requested_config,
                                   &reply_config,
                                   commit)) {
-        return -EINVAL;
+        if (commit || reply_config.flipResult != NV_KMS_FLIP_RESULT_IN_PROGRESS) {
+            return -EINVAL;
+        }
    }

    if (commit && nv_dev->supportsSyncpts) {
@@ -314,6 +321,24 @@ int nv_drm_atomic_check(struct drm_device *dev,
 {
    int ret = 0;

+#if defined(NV_DRM_COLOR_MGMT_AVAILABLE)
+    struct drm_crtc *crtc;
+    struct drm_crtc_state *crtc_state;
+    int i;
+
+    nv_drm_for_each_crtc_in_state(state, crtc, crtc_state, i) {
+        /*
+         * if the color management changed on the crtc, we need to update the
+         * crtc's plane's CSC matrices, so add the crtc's planes to the commit
+         */
+        if (crtc_state->color_mgmt_changed) {
+            if ((ret = drm_atomic_add_affected_planes(state, crtc)) != 0) {
+                goto done;
+            }
+        }
+    }
+#endif /* NV_DRM_COLOR_MGMT_AVAILABLE */
+
    if ((ret = drm_atomic_helper_check(dev, state)) != 0) {
        goto done;
    }
@@ -388,42 +413,56 @@ int nv_drm_atomic_commit(struct drm_device *dev,
    struct nv_drm_device *nv_dev = to_nv_device(dev);

    /*
-     * drm_mode_config_funcs::atomic_commit() mandates to return -EBUSY
-     * for nonblocking commit if previous updates (commit tasks/flip event) are
-     * pending. In case of blocking commits it mandates to wait for previous
-     * updates to complete.
+     * XXX: drm_mode_config_funcs::atomic_commit() mandates to return -EBUSY
+     * for nonblocking commit if the commit would need to wait for previous
+     * updates (commit tasks/flip event) to complete. In case of blocking
+     * commits it mandates to wait for previous updates to complete. However,
+     * the kernel DRM-KMS documentation does explicitly allow maintaining a
+     * queue of outstanding commits.
+     *
+     * Our system already implements such a queue, but due to
+     * bug 4054608, it is currently not used.
     */
-    if (nonblock) {
-        nv_drm_for_each_crtc_in_state(state, crtc, crtc_state, i) {
-            struct nv_drm_crtc *nv_crtc = to_nv_crtc(crtc);
+    nv_drm_for_each_crtc_in_state(state, crtc, crtc_state, i) {
+        struct nv_drm_crtc *nv_crtc = to_nv_crtc(crtc);

-            /*
-             * Here you aren't required to hold nv_drm_crtc::flip_list_lock
-             * because:
-             *
-             * The core DRM driver acquires lock for all affected crtcs before
-             * calling into ->commit() hook, therefore it is not possible for
-             * other threads to call into ->commit() hook affecting same crtcs
-             * and enqueue flip objects into flip_list -
-             *
-             *   nv_drm_atomic_commit_internal()
-             *     |-> nv_drm_atomic_apply_modeset_config(commit=true)
-             *           |-> nv_drm_crtc_enqueue_flip()
-             *
-             * Only possibility is list_empty check races with code path
-             * dequeuing flip object -
-             *
-             *   __nv_drm_handle_flip_event()
-             *     |-> nv_drm_crtc_dequeue_flip()
-             *
-             * But this race condition can't lead list_empty() to return
-             * incorrect result. nv_drm_crtc_dequeue_flip() in the middle of
-             * updating the list could not trick us into thinking the list is
-             * empty when it isn't.
-             */
+        /*
+         * Here you aren't required to hold nv_drm_crtc::flip_list_lock
+         * because:
+         *
+         * The core DRM driver acquires lock for all affected crtcs before
+         * calling into ->commit() hook, therefore it is not possible for
+         * other threads to call into ->commit() hook affecting same crtcs
+         * and enqueue flip objects into flip_list -
+         *
+         *   nv_drm_atomic_commit_internal()
+         *     |-> nv_drm_atomic_apply_modeset_config(commit=true)
+         *           |-> nv_drm_crtc_enqueue_flip()
+         *
+         * Only possibility is list_empty check races with code path
+         * dequeuing flip object -
+         *
+         *   __nv_drm_handle_flip_event()
+         *     |-> nv_drm_crtc_dequeue_flip()
+         *
+         * But this race condition can't lead list_empty() to return
+         * incorrect result. nv_drm_crtc_dequeue_flip() in the middle of
+         * updating the list could not trick us into thinking the list is
+         * empty when it isn't.
+         */
+        if (nonblock) {
            if (!list_empty(&nv_crtc->flip_list)) {
                return -EBUSY;
            }
+        } else {
+            if (wait_event_timeout(
+                    nv_dev->flip_event_wq,
+                    list_empty(&nv_crtc->flip_list),
+                    3 * HZ /* 3 second */) == 0) {
+                NV_DRM_DEV_LOG_ERR(
+                    nv_dev,
+                    "Flip event timeout on head %u", nv_crtc->head);
+            }
        }
    }

@@ -467,6 +506,7 @@ int nv_drm_atomic_commit(struct drm_device *dev,

        goto done;
    }
+    nv_dev->drmMasterChangedSinceLastAtomicCommit = NV_FALSE;

    nv_drm_for_each_crtc_in_state(state, crtc, crtc_state, i) {
        struct nv_drm_crtc *nv_crtc = to_nv_crtc(crtc);
@@ -547,6 +587,9 @@ int nv_drm_atomic_commit(struct drm_device *dev,
                NV_DRM_DEV_LOG_ERR(
                    nv_dev,
                    "Flip event timeout on head %u", nv_crtc->head);
+                while (!list_empty(&nv_crtc->flip_list)) {
+                    __nv_drm_handle_flip_event(nv_crtc);
+                }
            }
        }
    }
--- a/kernel-open/nvidia-drm/nvidia-drm-os-interface.c
+++ b/kernel-open/nvidia-drm/nvidia-drm-os-interface.c
@@ -0,0 +1,285 @@
+/*
+ * Copyright (c) 2015-2023, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/slab.h>
+
+#include "nvidia-drm-os-interface.h"
+
+#if defined(NV_DRM_AVAILABLE)
+
+#if defined(NV_LINUX_SYNC_FILE_H_PRESENT)
+#include <linux/file.h>
+#include <linux/sync_file.h>
+#endif
+
+#include <linux/vmalloc.h>
+#include <linux/sched.h>
+#include <linux/device.h>
+
+#include "nv-mm.h"
+
+#if defined(NV_DRM_DRMP_H_PRESENT)
+#include <drm/drmP.h>
+#endif
+
+bool nv_drm_modeset_module_param = false;
+bool nv_drm_fbdev_module_param = false;
+
+void *nv_drm_calloc(size_t nmemb, size_t size)
+{
+    size_t total_size = nmemb * size;
+    //
+    // Check for overflow.
+    //
+    if ((nmemb != 0) && ((total_size / nmemb) != size))
+    {
+        return NULL;
+    }
+    return kzalloc(nmemb * size, GFP_KERNEL);
+}
+
+void nv_drm_free(void *ptr)
+{
+    if (IS_ERR(ptr)) {
+        return;
+    }
+
+    kfree(ptr);
+}
+
+char *nv_drm_asprintf(const char *fmt, ...)
+{
+    va_list ap;
+    char *p;
+
+    va_start(ap, fmt);
+    p = kvasprintf(GFP_KERNEL, fmt, ap);
+    va_end(ap);
+
+    return p;
+}
+
+#if defined(NVCPU_X86) || defined(NVCPU_X86_64)
+  #define WRITE_COMBINE_FLUSH()    asm volatile("sfence":::"memory")
+#elif defined(NVCPU_PPC64LE)
+  #define WRITE_COMBINE_FLUSH()    asm volatile("sync":::"memory")
+#else
+  #define WRITE_COMBINE_FLUSH()    mb()
+#endif
+
+void nv_drm_write_combine_flush(void)
+{
+    WRITE_COMBINE_FLUSH();
+}
+
+int nv_drm_lock_user_pages(unsigned long address,
+                           unsigned long pages_count, struct page ***pages)
+{
+    struct mm_struct *mm = current->mm;
+    struct page **user_pages;
+    int pages_pinned;
+
+    user_pages = nv_drm_calloc(pages_count, sizeof(*user_pages));
+
+    if (user_pages == NULL) {
+        return -ENOMEM;
+    }
+
+    nv_mmap_read_lock(mm);
+
+    pages_pinned = NV_PIN_USER_PAGES(address, pages_count, FOLL_WRITE,
+                                     user_pages);
+    nv_mmap_read_unlock(mm);
+
+    if (pages_pinned < 0 || (unsigned)pages_pinned < pages_count) {
+        goto failed;
+    }
+
+    *pages = user_pages;
+
+    return 0;
+
+failed:
+
+    if (pages_pinned > 0) {
+        int i;
+
+        for (i = 0; i < pages_pinned; i++) {
+           NV_UNPIN_USER_PAGE(user_pages[i]);
+        }
+    }
+
+    nv_drm_free(user_pages);
+
+    return (pages_pinned < 0) ? pages_pinned : -EINVAL;
+}
+
+void nv_drm_unlock_user_pages(unsigned long  pages_count, struct page **pages)
+{
+    unsigned long i;
+
+    for (i = 0; i < pages_count; i++) {
+        set_page_dirty_lock(pages[i]);
+        NV_UNPIN_USER_PAGE(pages[i]);
+    }
+
+    nv_drm_free(pages);
+}
+
+/*
+ * linuxkpi vmap doesn't use the flags argument as it
+ * doesn't seem to be needed. Define VM_USERMAP to 0
+ * to make errors go away
+ *
+ * vmap: sys/compat/linuxkpi/common/src/linux_compat.c
+ */
+#if defined(NV_BSD)
+#define VM_USERMAP 0
+#endif
+
+void *nv_drm_vmap(struct page **pages, unsigned long pages_count)
+{
+    return vmap(pages, pages_count, VM_USERMAP, PAGE_KERNEL);
+}
+
+void nv_drm_vunmap(void *address)
+{
+    vunmap(address);
+}
+
+bool nv_drm_workthread_init(nv_drm_workthread *worker, const char *name)
+{
+    worker->shutting_down = false;
+    if (nv_kthread_q_init(&worker->q, name)) {
+        return false;
+    }
+
+    spin_lock_init(&worker->lock);
+
+    return true;
+}
+
+void nv_drm_workthread_shutdown(nv_drm_workthread *worker)
+{
+    unsigned long flags;
+
+    spin_lock_irqsave(&worker->lock, flags);
+    worker->shutting_down = true;
+    spin_unlock_irqrestore(&worker->lock, flags);
+
+    nv_kthread_q_stop(&worker->q);
+}
+
+void nv_drm_workthread_work_init(nv_drm_work *work,
+                                 void (*callback)(void *),
+                                 void *arg)
+{
+    nv_kthread_q_item_init(work, callback, arg);
+}
+
+int nv_drm_workthread_add_work(nv_drm_workthread *worker, nv_drm_work *work)
+{
+    unsigned long flags;
+    int ret = 0;
+
+    spin_lock_irqsave(&worker->lock, flags);
+    if (!worker->shutting_down) {
+        ret = nv_kthread_q_schedule_q_item(&worker->q, work);
+    }
+    spin_unlock_irqrestore(&worker->lock, flags);
+
+    return ret;
+}
+
+void nv_drm_timer_setup(nv_drm_timer *timer, void (*callback)(nv_drm_timer *nv_drm_timer))
+{
+    nv_timer_setup(timer, callback);
+}
+
+void nv_drm_mod_timer(nv_drm_timer *timer, unsigned long timeout_native)
+{
+    mod_timer(&timer->kernel_timer, timeout_native);
+}
+
+unsigned long nv_drm_timer_now(void)
+{
+    return jiffies;
+}
+
+unsigned long nv_drm_timeout_from_ms(NvU64 relative_timeout_ms)
+{
+    return jiffies + msecs_to_jiffies(relative_timeout_ms);
+}
+
+bool nv_drm_del_timer_sync(nv_drm_timer *timer)
+{
+    if (del_timer_sync(&timer->kernel_timer)) {
+        return true;
+    } else {
+        return false;
+    }
+}
+
+#if defined(NV_DRM_FENCE_AVAILABLE)
+int nv_drm_create_sync_file(nv_dma_fence_t *fence)
+{
+#if defined(NV_LINUX_SYNC_FILE_H_PRESENT)
+    struct sync_file *sync;
+    int fd = get_unused_fd_flags(O_CLOEXEC);
+
+    if (fd < 0) {
+        return fd;
+    }
+
+    /* sync_file_create() generates its own reference to the fence */
+    sync = sync_file_create(fence);
+
+    if (IS_ERR(sync)) {
+        put_unused_fd(fd);
+        return PTR_ERR(sync);
+    }
+
+    fd_install(fd, sync->file);
+
+    return fd;
+#else /* defined(NV_LINUX_SYNC_FILE_H_PRESENT) */
+    return -EINVAL;
+#endif  /* defined(NV_LINUX_SYNC_FILE_H_PRESENT) */
+}
+
+nv_dma_fence_t *nv_drm_sync_file_get_fence(int fd)
+{
+#if defined(NV_SYNC_FILE_GET_FENCE_PRESENT)
+    return sync_file_get_fence(fd);
+#else /* defined(NV_SYNC_FILE_GET_FENCE_PRESENT) */
+    return NULL;
+#endif  /* defined(NV_SYNC_FILE_GET_FENCE_PRESENT) */
+}
+#endif /* defined(NV_DRM_FENCE_AVAILABLE) */
+
+void nv_drm_yield(void)
+{
+    set_current_state(TASK_INTERRUPTIBLE);
+    schedule_timeout(1);
+}
+
+#endif /* NV_DRM_AVAILABLE */
--- a/kernel-open/nvidia-drm/nvidia-drm-os-interface.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-os-interface.h
@@ -29,10 +29,47 @@

 #if defined(NV_DRM_AVAILABLE)

+#if defined(NV_DRM_FENCE_AVAILABLE)
+#include "nvidia-dma-fence-helper.h"
+#endif
+
+#if defined(NV_LINUX) || defined(NV_BSD)
+#include "nv-kthread-q.h"
+#include "linux/spinlock.h"
+
+typedef struct nv_drm_workthread {
+    spinlock_t lock;
+    struct nv_kthread_q q;
+    bool shutting_down;
+} nv_drm_workthread;
+
+typedef nv_kthread_q_item_t nv_drm_work;
+
+#else
+#error "Need to define deferred work primitives for this OS"
+#endif
+
+#if defined(NV_LINUX) || defined(NV_BSD)
+#include "nv-timer.h"
+
+typedef struct nv_timer nv_drm_timer;
+
+#else
+#error "Need to define kernel timer callback primitives for this OS"
+#endif
+
+#if defined(NV_DRM_FBDEV_GENERIC_SETUP_PRESENT) && defined(NV_DRM_APERTURE_REMOVE_CONFLICTING_PCI_FRAMEBUFFERS_PRESENT)
+#define NV_DRM_FBDEV_GENERIC_AVAILABLE
+#endif
+
 struct page;

 /* Set to true when the atomic modeset feature is enabled. */
 extern bool nv_drm_modeset_module_param;
+#if defined(NV_DRM_FBDEV_GENERIC_AVAILABLE)
+/* Set to true when the nvidia-drm driver should install a framebuffer device */
+extern bool nv_drm_fbdev_module_param;
+#endif

 void *nv_drm_calloc(size_t nmemb, size_t size);

@@ -51,6 +88,37 @@ void *nv_drm_vmap(struct page **pages, unsigned long pages_count);

 void nv_drm_vunmap(void *address);

-#endif
+bool nv_drm_workthread_init(nv_drm_workthread *worker, const char *name);
+
+/* Can be called concurrently with nv_drm_workthread_add_work() */
+void nv_drm_workthread_shutdown(nv_drm_workthread *worker);
+
+void nv_drm_workthread_work_init(nv_drm_work *work,
+                                 void (*callback)(void *),
+                                 void *arg);
+
+/* Can be called concurrently with nv_drm_workthread_shutdown() */
+int nv_drm_workthread_add_work(nv_drm_workthread *worker, nv_drm_work *work);
+
+void nv_drm_timer_setup(nv_drm_timer *timer,
+                        void (*callback)(nv_drm_timer *nv_drm_timer));
+
+void nv_drm_mod_timer(nv_drm_timer *timer, unsigned long relative_timeout_ms);
+
+bool nv_drm_del_timer_sync(nv_drm_timer *timer);
+
+unsigned long nv_drm_timer_now(void);
+
+unsigned long nv_drm_timeout_from_ms(NvU64 relative_timeout_ms);
+
+#if defined(NV_DRM_FENCE_AVAILABLE)
+int nv_drm_create_sync_file(nv_dma_fence_t *fence);
+
+nv_dma_fence_t *nv_drm_sync_file_get_fence(int fd);
+#endif /* defined(NV_DRM_FENCE_AVAILABLE) */
+
+void nv_drm_yield(void);
+
+#endif /* defined(NV_DRM_AVAILABLE) */

 #endif /* __NVIDIA_DRM_OS_INTERFACE_H__ */
--- a/kernel-open/nvidia-drm/nvidia-drm-priv.h
+++ b/kernel-open/nvidia-drm/nvidia-drm-priv.h
@@ -46,12 +46,33 @@
 #define NV_DRM_LOG_ERR(__fmt, ...) \
    DRM_ERROR("[nvidia-drm] " __fmt "\n", ##__VA_ARGS__)

+/*
+ * DRM_WARN() was added in v4.9 by kernel commit
+ * 30b0da8d556e65ff935a56cd82c05ba0516d3e4a
+ *
+ * Before this commit, only DRM_INFO and DRM_ERROR were defined and
+ * DRM_INFO(fmt, ...) was defined as
+ * printk(KERN_INFO "[" DRM_NAME "] " fmt, ##__VA_ARGS__). So, if
+ * DRM_WARN is undefined this defines NV_DRM_LOG_WARN following the
+ * same pattern as DRM_INFO.
+ */
+#ifdef DRM_WARN
+#define NV_DRM_LOG_WARN(__fmt, ...) \
+    DRM_WARN("[nvidia-drm] " __fmt "\n", ##__VA_ARGS__)
+#else
+#define NV_DRM_LOG_WARN(__fmt, ...) \
+    printk(KERN_WARNING "[" DRM_NAME "] [nvidia-drm] " __fmt "\n", ##__VA_ARGS__)
+#endif
+
 #define NV_DRM_LOG_INFO(__fmt, ...) \
    DRM_INFO("[nvidia-drm] " __fmt "\n", ##__VA_ARGS__)

 #define NV_DRM_DEV_LOG_INFO(__dev, __fmt, ...) \
    NV_DRM_LOG_INFO("[GPU ID 0x%08x] " __fmt, __dev->gpu_info.gpu_id, ##__VA_ARGS__)

+#define NV_DRM_DEV_LOG_WARN(__dev, __fmt, ...) \
+    NV_DRM_LOG_WARN("[GPU ID 0x%08x] " __fmt, __dev->gpu_info.gpu_id, ##__VA_ARGS__)
+
 #define NV_DRM_DEV_LOG_ERR(__dev, __fmt, ...) \
    NV_DRM_LOG_ERR("[GPU ID 0x%08x] " __fmt, __dev->gpu_info.gpu_id, ##__VA_ARGS__)

@@ -105,6 +126,7 @@ struct nv_drm_device {
    NvU64 modifiers[6 /* block linear */ + 1 /* linear */ + 1 /* terminator */];
 #endif

+    struct delayed_work hotplug_event_work;
    atomic_t enable_event_handling;

    /**
@@ -117,9 +139,26 @@ struct nv_drm_device {

 #endif

+#if defined(NV_DRM_FENCE_AVAILABLE)
+    NvU64 semsurf_stride;
+    NvU64 semsurf_max_submitted_offset;
+#endif
+
    NvBool hasVideoMemory;

    NvBool supportsSyncpts;
+    NvBool subOwnershipGranted;
+    NvBool hasFramebufferConsole;
+
+    /**
+     * @drmMasterChangedSinceLastAtomicCommit:
+     *
+     * This flag is set in nv_drm_master_set and reset after a completed atomic
+     * commit. It is used to restore or recommit state that is lost by the
+     * NvKms modeset owner change, such as the CRTC color management
+     * properties.
+     */
+    NvBool drmMasterChangedSinceLastAtomicCommit;

    struct drm_property *nv_out_fence_property;
    struct drm_property *nv_input_colorspace_property;
--- a/kernel-open/nvidia-drm/nvidia-drm-sources.mk
+++ b/kernel-open/nvidia-drm/nvidia-drm-sources.mk
@@ -0,0 +1,132 @@
+###########################################################################
+# Kbuild fragment for nvidia-drm.ko
+###########################################################################
+
+#
+# Define NVIDIA_DRM_SOURCES
+#
+
+NVIDIA_DRM_SOURCES =
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-drv.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-utils.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-crtc.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-encoder.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-connector.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-fb.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-modeset.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-fence.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-helper.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nv-kthread-q.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nv-pci-table.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem-nvkms-memory.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem-user-memory.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem-dma-buf.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-format.c
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-os-interface.c
+
+#
+# Register the conftests needed by nvidia-drm.ko
+#
+
+NV_CONFTEST_GENERIC_COMPILE_TESTS += drm_available
+NV_CONFTEST_GENERIC_COMPILE_TESTS += drm_atomic_available
+NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_gpl_refcount_inc
+NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_gpl_refcount_dec_and_test
+NV_CONFTEST_GENERIC_COMPILE_TESTS += drm_alpha_blending_available
+NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_present_drm_gem_prime_fd_to_handle
+NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_present_drm_gem_prime_handle_to_fd
+
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_dev_unref
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_reinit_primary_mode_group
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += get_user_pages_remote
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += get_user_pages
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += pin_user_pages_remote
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += pin_user_pages
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_gem_object_lookup
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_atomic_state_ref_counting
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_driver_has_gem_prime_res_obj
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_atomic_helper_connector_dpms
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_funcs_have_mode_in_name
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_has_vrr_capable_property
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += vmf_insert_pfn
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_framebuffer_get
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_gem_object_get
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_dev_put
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_format_num_planes
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_for_each_possible_encoder
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_rotation_available
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_vma_offset_exact_lookup_locked
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_gem_object_put_unlocked
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += nvhost_dma_fence_unpack
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += list_is_first
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += timer_setup
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += dma_fence_set_error
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += fence_set_error
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += sync_file_get_fence
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_aperture_remove_conflicting_pci_framebuffers
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_fbdev_generic_setup
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_attach_hdr_output_metadata_property
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_helper_crtc_enable_color_mgmt
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_crtc_enable_color_mgmt
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_atomic_helper_legacy_gamma_set
+
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_present
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_has_bus_type
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_has_get_irq
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_has_get_name
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_device_list
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_legacy_dev_list
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_set_busid
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_connectors_changed
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_init_function_args
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_helper_mode_fill_fb_struct
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_master_drop_has_from_release_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_unload_has_int_return_type
+NV_CONFTEST_TYPE_COMPILE_TESTS += vm_fault_has_address
+NV_CONFTEST_TYPE_COMPILE_TESTS += vm_ops_fault_removed_vma_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_atomic_helper_crtc_destroy_state_has_crtc_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_atomic_helper_plane_destroy_state_has_plane_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_mode_object_find_has_file_priv_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += dma_buf_owner
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_list_iter
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_atomic_helper_swap_state_has_stall_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_prime_flag_present
+NV_CONFTEST_TYPE_COMPILE_TESTS += vm_fault_t
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_gem_object_has_resv
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_async_flip
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_pageflip_flags
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_vrr_enabled
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_format_modifiers_present
+NV_CONFTEST_TYPE_COMPILE_TESTS += mm_has_mmap_lock
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_vma_node_is_allowed_has_tag_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_vma_offset_node_has_readonly
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_display_mode_has_vrefresh
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_master_set_has_int_return_type
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_gem_free_object
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_prime_pages_to_sg_has_drm_device_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_gem_prime_callbacks
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_atomic_check_has_atomic_state_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_gem_object_vmap_has_map_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_plane_atomic_check_has_atomic_state_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_device_has_pdev
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_no_vblank
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_mode_config_has_allow_fb_modifiers
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_has_hdr_output_metadata
+NV_CONFTEST_TYPE_COMPILE_TESTS += dma_resv_add_fence
+NV_CONFTEST_TYPE_COMPILE_TESTS += dma_resv_reserve_fences
+NV_CONFTEST_TYPE_COMPILE_TESTS += reservation_object_reserve_shared_has_num_fences_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_has_override_edid
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_master_has_leases
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_file_get_master
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_modeset_lock_all_end
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_lookup
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_put
+NV_CONFTEST_TYPE_COMPILE_TESTS += vm_area_struct_has_const_vm_flags
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_dumb_destroy
+NV_CONFTEST_TYPE_COMPILE_TESTS += fence_ops_use_64bit_seqno
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_aperture_remove_conflicting_pci_framebuffers_has_driver_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_mode_create_dp_colorspace_property_has_supported_colorspaces_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_syncobj_features_present
+NV_CONFTEST_TYPE_COMPILE_TESTS += drm_unlocked_ioctl_flag_present
--- a/kernel-open/nvidia-drm/nvidia-drm.Kbuild
+++ b/kernel-open/nvidia-drm/nvidia-drm.Kbuild
@@ -2,29 +2,16 @@
 # Kbuild fragment for nvidia-drm.ko
 ###########################################################################

+# Get our source file list and conftest list from the common file
+include $(src)/nvidia-drm/nvidia-drm-sources.mk
+
+# Linux-specific sources
+NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-linux.c
+
 #
 # Define NVIDIA_DRM_{SOURCES,OBJECTS}
 #

-NVIDIA_DRM_SOURCES =
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-drv.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-utils.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-crtc.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-encoder.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-connector.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-fb.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-modeset.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-fence.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-linux.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-helper.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nv-pci-table.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem-nvkms-memory.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem-user-memory.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-gem-dma-buf.c
-NVIDIA_DRM_SOURCES += nvidia-drm/nvidia-drm-format.c
-
 NVIDIA_DRM_OBJECTS = $(patsubst %.c,%.o,$(NVIDIA_DRM_SOURCES))

 obj-m += nvidia-drm.o
@@ -43,93 +30,4 @@ NVIDIA_DRM_CFLAGS += -UDEBUG -U_DEBUG -DNDEBUG -DNV_BUILD_MODULE_INSTANCES=0

 $(call ASSIGN_PER_OBJ_CFLAGS, $(NVIDIA_DRM_OBJECTS), $(NVIDIA_DRM_CFLAGS))

-#
-# Register the conftests needed by nvidia-drm.ko
-#
-
 NV_OBJECTS_DEPEND_ON_CONFTEST += $(NVIDIA_DRM_OBJECTS)
-
-NV_CONFTEST_GENERIC_COMPILE_TESTS += drm_available
-NV_CONFTEST_GENERIC_COMPILE_TESTS += drm_atomic_available
-NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_gpl_refcount_inc
-NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_gpl_refcount_dec_and_test
-NV_CONFTEST_GENERIC_COMPILE_TESTS += drm_alpha_blending_available
-NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_present_drm_gem_prime_fd_to_handle
-NV_CONFTEST_GENERIC_COMPILE_TESTS += is_export_symbol_present_drm_gem_prime_handle_to_fd
-
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_dev_unref
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_reinit_primary_mode_group
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += get_user_pages_remote
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += get_user_pages
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += pin_user_pages_remote
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += pin_user_pages
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_gem_object_lookup
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_atomic_state_ref_counting
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_driver_has_gem_prime_res_obj
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_atomic_helper_connector_dpms
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_funcs_have_mode_in_name
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_has_vrr_capable_property
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += vmf_insert_pfn
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_framebuffer_get
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_gem_object_get
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_dev_put
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_format_num_planes
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_connector_for_each_possible_encoder
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_rotation_available
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_vma_offset_exact_lookup_locked
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += drm_gem_object_put_unlocked
-NV_CONFTEST_FUNCTION_COMPILE_TESTS += nvhost_dma_fence_unpack
-
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_present
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_has_bus_type
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_has_get_irq
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_bus_has_get_name
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_device_list
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_legacy_dev_list
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_set_busid
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_connectors_changed
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_init_function_args
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_helper_mode_fill_fb_struct
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_master_drop_has_from_release_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_unload_has_int_return_type
-NV_CONFTEST_TYPE_COMPILE_TESTS += vm_fault_has_address
-NV_CONFTEST_TYPE_COMPILE_TESTS += vm_ops_fault_removed_vma_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_atomic_helper_crtc_destroy_state_has_crtc_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_atomic_helper_plane_destroy_state_has_plane_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_mode_object_find_has_file_priv_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += dma_buf_owner
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_list_iter
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_atomic_helper_swap_state_has_stall_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_prime_flag_present
-NV_CONFTEST_TYPE_COMPILE_TESTS += vm_fault_t
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_gem_object_has_resv
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_async_flip
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_pageflip_flags
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_vrr_enabled
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_format_modifiers_present
-NV_CONFTEST_TYPE_COMPILE_TESTS += mm_has_mmap_lock
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_vma_node_is_allowed_has_tag_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_vma_offset_node_has_readonly
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_display_mode_has_vrefresh
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_master_set_has_int_return_type
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_gem_free_object
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_prime_pages_to_sg_has_drm_device_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_gem_prime_callbacks
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_atomic_check_has_atomic_state_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_gem_object_vmap_has_map_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_plane_atomic_check_has_atomic_state_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_device_has_pdev
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_crtc_state_has_no_vblank
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_mode_config_has_allow_fb_modifiers
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_has_hdr_output_metadata
-NV_CONFTEST_TYPE_COMPILE_TESTS += dma_resv_add_fence
-NV_CONFTEST_TYPE_COMPILE_TESTS += dma_resv_reserve_fences
-NV_CONFTEST_TYPE_COMPILE_TESTS += reservation_object_reserve_shared_has_num_fences_arg
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_has_override_edid
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_master_has_leases
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_file_get_master
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_modeset_lock_all_end
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_lookup
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_connector_put
-NV_CONFTEST_TYPE_COMPILE_TESTS += vm_area_struct_has_const_vm_flags
-NV_CONFTEST_TYPE_COMPILE_TESTS += drm_driver_has_dumb_destroy
--- a/kernel-open/nvidia-drm/nvidia-drm.c
+++ b/kernel-open/nvidia-drm/nvidia-drm.c
@@ -45,6 +45,7 @@ int nv_drm_init(void)
        return -EINVAL;
    }

+    nvKms->setSuspendResumeCallback(nv_drm_suspend_resume);
    return nv_drm_probe_devices();
 #else
    return 0;
@@ -54,6 +55,7 @@ int nv_drm_init(void)
 void nv_drm_exit(void)
 {
 #if defined(NV_DRM_AVAILABLE)
+    nvKms->setSuspendResumeCallback(NULL);
    nv_drm_remove_devices();
 #endif
 }
--- a/kernel-open/nvidia-modeset/nv-kthread-q.c
+++ b/kernel-open/nvidia-modeset/nv-kthread-q.c
@@ -247,6 +247,11 @@ int nv_kthread_q_init_on_node(nv_kthread_q_t *q, const char *q_name, int preferr
    return 0;
 }

+int nv_kthread_q_init(nv_kthread_q_t *q, const char *qname)
+{
+    return nv_kthread_q_init_on_node(q, qname, NV_KTHREAD_NO_NODE);
+}
+
 // Returns true (non-zero) if the item was actually scheduled, and false if the
 // item was already pending in a queue.
 static int _raw_q_schedule(nv_kthread_q_t *q, nv_kthread_q_item_t *q_item)
--- a/kernel-open/nvidia-modeset/nvidia-modeset-linux.c
+++ b/kernel-open/nvidia-modeset/nvidia-modeset-linux.c
@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2015-21 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-FileCopyrightText: Copyright (c) 2015-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
@@ -35,12 +35,13 @@
 #include <linux/list.h>
 #include <linux/rwsem.h>
 #include <linux/freezer.h>
+#include <linux/poll.h>
+#include <linux/cdev.h>

 #include <acpi/video.h>

 #include "nvstatus.h"

-#include "nv-register-module.h"
 #include "nv-modeset-interface.h"
 #include "nv-kref.h"

@@ -53,8 +54,13 @@
 #include "nv-kthread-q.h"
 #include "nv-time.h"
 #include "nv-lock.h"
+#include "nv-chardev-numbers.h"

-#if !defined(CONFIG_RETPOLINE)
+/*
+ * Commit aefb2f2e619b ("x86/bugs: Rename CONFIG_RETPOLINE =>
+ * CONFIG_MITIGATION_RETPOLINE) in v6.8 renamed CONFIG_RETPOLINE.
+ */
+#if !defined(CONFIG_RETPOLINE) && !defined(CONFIG_MITIGATION_RETPOLINE)
 #include "nv-retpoline.h"
 #endif

@@ -65,9 +71,21 @@
 static bool output_rounding_fix = true;
 module_param_named(output_rounding_fix, output_rounding_fix, bool, 0400);

+static bool disable_hdmi_frl = false;
+module_param_named(disable_hdmi_frl, disable_hdmi_frl, bool, 0400);
+
 static bool disable_vrr_memclk_switch = false;
 module_param_named(disable_vrr_memclk_switch, disable_vrr_memclk_switch, bool, 0400);

+static bool hdmi_deepcolor = true;
+module_param_named(hdmi_deepcolor, hdmi_deepcolor, bool, 0400);
+
+static bool vblank_sem_control = true;
+module_param_named(vblank_sem_control, vblank_sem_control, bool, 0400);
+
+static bool opportunistic_display_sync = true;
+module_param_named(opportunistic_display_sync, opportunistic_display_sync, bool, 0400);
+
 /* These parameters are used for fault injection tests.  Normally the defaults
 * should be used. */
 MODULE_PARM_DESC(fail_malloc, "Fail the Nth call to nvkms_alloc");
@@ -78,6 +96,7 @@ MODULE_PARM_DESC(malloc_verbose, "Report information about malloc calls on modul
 static bool malloc_verbose = false;
 module_param_named(malloc_verbose, malloc_verbose, bool, 0400);

+#if NVKMS_CONFIG_FILE_SUPPORTED
 /* This parameter is used to find the dpy override conf file */
 #define NVKMS_CONF_FILE_SPECIFIED (nvkms_conf != NULL)

@@ -86,6 +105,7 @@ MODULE_PARM_DESC(config_file,
                 "(default: disabled)");
 static char *nvkms_conf = NULL;
 module_param_named(config_file, nvkms_conf, charp, 0400);
+#endif

 static atomic_t nvkms_alloc_called_count;

@@ -94,11 +114,45 @@ NvBool nvkms_output_rounding_fix(void)
    return output_rounding_fix;
 }

+NvBool nvkms_disable_hdmi_frl(void)
+{
+    return disable_hdmi_frl;
+}
+
 NvBool nvkms_disable_vrr_memclk_switch(void)
 {
    return disable_vrr_memclk_switch;
 }

+NvBool nvkms_hdmi_deepcolor(void)
+{
+    return hdmi_deepcolor;
+}
+
+NvBool nvkms_vblank_sem_control(void)
+{
+    return vblank_sem_control;
+}
+
+NvBool nvkms_opportunistic_display_sync(void)
+{
+    return opportunistic_display_sync;
+}
+
+NvBool nvkms_kernel_supports_syncpts(void)
+{
+/*
+ * Note this only checks that the kernel has the prerequisite
+ * support for syncpts; callers must also check that the hardware
+ * supports syncpts.
+ */
+#if (defined(CONFIG_TEGRA_GRHOST) || defined(NV_LINUX_HOST1X_NEXT_H_PRESENT))
+    return NV_TRUE;
+#else
+    return NV_FALSE;
+#endif
+}
+
 #define NVKMS_SYNCPT_STUBS_NEEDED

 /*************************************************************************
@@ -200,9 +254,23 @@ static inline int nvkms_read_trylock_pm_lock(void)

 static inline void nvkms_read_lock_pm_lock(void)
 {
-    while (!down_read_trylock(&nvkms_pm_lock)) {
-        try_to_freeze();
-        cond_resched();
+    if ((current->flags & PF_NOFREEZE)) {
+        /*
+         * Non-freezable tasks (i.e. kthreads in this case) don't have to worry
+         * about being frozen during system suspend, but do need to block so
+         * that the CPU can go idle during s2idle. Do a normal uninterruptible
+         * blocking wait for the PM lock.
+         */
+        down_read(&nvkms_pm_lock);
+    } else {
+        /*
+         * For freezable tasks, make sure we give the kernel an opportunity to
+         * freeze if taking the PM lock fails.
+         */
+        while (!down_read_trylock(&nvkms_pm_lock)) {
+            try_to_freeze();
+            cond_resched();
+        }
    }
 }

@@ -335,7 +403,7 @@ NvU64 nvkms_get_usec(void)
    struct timespec64 ts;
    NvU64 ns;

-    ktime_get_real_ts64(&ts);
+    ktime_get_raw_ts64(&ts);

    ns = timespec64_to_ns(&ts);
    return ns / 1000;
@@ -449,6 +517,8 @@ nvkms_event_queue_changed(nvkms_per_open_handle_t *pOpenKernel,

 static void nvkms_suspend(NvU32 gpuId)
 {
+    nvKmsKapiSuspendResume(NV_TRUE /* suspend */);
+
    if (gpuId == 0) {
        nvkms_write_lock_pm_lock();
    }
@@ -467,6 +537,8 @@ static void nvkms_resume(NvU32 gpuId)
    if (gpuId == 0) {
        nvkms_write_unlock_pm_lock();
    }
+
+    nvKmsKapiSuspendResume(NV_FALSE /* suspend */);
 }


@@ -795,49 +867,6 @@ void nvkms_free_timer(nvkms_timer_handle_t *handle)
    timer->cancel = NV_TRUE;
 }

-void* nvkms_get_per_open_data(int fd)
-{
-    struct file *filp = fget(fd);
-    struct nvkms_per_open *popen = NULL;
-    dev_t rdev = 0;
-    void *data = NULL;
-
-    if (filp == NULL) {
-        return NULL;
-    }
-
-    if (filp->f_inode == NULL) {
-        goto done;
-    }
-    rdev = filp->f_inode->i_rdev;
-
-    if ((MAJOR(rdev) != NVKMS_MAJOR_DEVICE_NUMBER) ||
-        (MINOR(rdev) != NVKMS_MINOR_DEVICE_NUMBER)) {
-        goto done;
-    }
-
-    popen = filp->private_data;
-    if (popen == NULL) {
-        goto done;
-    }
-
-    data = popen->data;
-
-done:
-    /*
-     * fget() incremented the struct file's reference count, which
-     * needs to be balanced with a call to fput().  It is safe to
-     * decrement the reference count before returning
-     * filp->private_data because core NVKMS is currently holding the
-     * nvkms_lock, which prevents the nvkms_close() => nvKmsClose()
-     * call chain from freeing the file out from under the caller of
-     * nvkms_get_per_open_data().
-     */
-    fput(filp);
-
-    return data;
-}
-
 NvBool nvkms_fd_is_nvidia_chardev(int fd)
 {
    struct file *filp = fget(fd);
@@ -1219,6 +1248,26 @@ void nvkms_close_from_kapi(struct nvkms_per_open *popen)
    nvkms_close_pm_unlocked(popen);
 }

+NvBool nvkms_ioctl_from_kapi_try_pmlock
+(
+    struct nvkms_per_open *popen,
+    NvU32 cmd, void *params_address, const size_t param_size
+)
+{
+    NvBool ret;
+
+    if (nvkms_read_trylock_pm_lock()) {
+        return NV_FALSE;
+    }
+
+    ret = nvkms_ioctl_common(popen,
+                             cmd,
+                             (NvU64)(NvUPtr)params_address, param_size) == 0;
+    nvkms_read_unlock_pm_lock();
+
+    return ret;
+}
+
 NvBool nvkms_ioctl_from_kapi
 (
    struct nvkms_per_open *popen,
@@ -1382,6 +1431,7 @@ static void nvkms_proc_exit(void)
 /*************************************************************************
 * NVKMS Config File Read
 ************************************************************************/
+#if NVKMS_CONFIG_FILE_SUPPORTED
 static NvBool nvkms_fs_mounted(void)
 {
    return current->fs != NULL;
@@ -1489,6 +1539,11 @@ static void nvkms_read_config_file_locked(void)

    nvkms_free(buffer, buf_size);
 }
+#else
+static void nvkms_read_config_file_locked(void)
+{
+}
+#endif

 /*************************************************************************
 * NVKMS KAPI functions
@@ -1583,6 +1638,12 @@ static int nvkms_ioctl(struct inode *inode, struct file *filp,
    return status;
 }

+static long nvkms_unlocked_ioctl(struct file *filp, unsigned int cmd,
+                                 unsigned long arg)
+{
+    return nvkms_ioctl(filp->f_inode, filp, cmd, arg);
+}
+
 static unsigned int nvkms_poll(struct file *filp, poll_table *wait)
 {
    unsigned int mask = 0;
@@ -1610,17 +1671,73 @@ static unsigned int nvkms_poll(struct file *filp, poll_table *wait)
 * Module loading support code.
 *************************************************************************/

-static nvidia_module_t nvidia_modeset_module = {
+#define NVKMS_RDEV  (MKDEV(NV_MAJOR_DEVICE_NUMBER, \
+                           NV_MINOR_DEVICE_NUMBER_MODESET_DEVICE))
+
+static struct file_operations nvkms_fops = {
    .owner       = THIS_MODULE,
-    .module_name = "nvidia-modeset",
-    .instance    = 1, /* minor number: 255-1=254 */
-    .open        = nvkms_open,
-    .close       = nvkms_close,
-    .mmap        = nvkms_mmap,
-    .ioctl       = nvkms_ioctl,
    .poll        = nvkms_poll,
+    .unlocked_ioctl = nvkms_unlocked_ioctl,
+#if NVCPU_IS_X86_64 || NVCPU_IS_AARCH64
+    .compat_ioctl = nvkms_unlocked_ioctl,
+#endif
+    .mmap        = nvkms_mmap,
+    .open        = nvkms_open,
+    .release     = nvkms_close,
 };

+static struct cdev nvkms_device_cdev;
+
+static int __init nvkms_register_chrdev(void)
+{
+    int ret;
+
+    ret = register_chrdev_region(NVKMS_RDEV, 1, "nvidia-modeset");
+    if (ret < 0) {
+        return ret;
+    }
+
+    cdev_init(&nvkms_device_cdev, &nvkms_fops);
+    ret = cdev_add(&nvkms_device_cdev, NVKMS_RDEV, 1);
+    if (ret < 0) {
+        unregister_chrdev_region(NVKMS_RDEV, 1);
+        return ret;
+    }
+
+    return ret;
+}
+
+static void nvkms_unregister_chrdev(void)
+{
+    cdev_del(&nvkms_device_cdev);
+    unregister_chrdev_region(NVKMS_RDEV, 1);
+}
+
+void* nvkms_get_per_open_data(int fd)
+{
+    struct file *filp = fget(fd);
+    void *data = NULL;
+
+    if (filp) {
+        if (filp->f_op == &nvkms_fops && filp->private_data) {
+            struct nvkms_per_open *popen = filp->private_data;
+            data = popen->data;
+        }
+
+        /*
+         * fget() incremented the struct file's reference count, which needs to
+         * be balanced with a call to fput().  It is safe to decrement the
+         * reference count before returning filp->private_data because core
+         * NVKMS is currently holding the nvkms_lock, which prevents the
+         * nvkms_close() => nvKmsClose() call chain from freeing the file out
+         * from under the caller of nvkms_get_per_open_data().
+         */
+        fput(filp);
+    }
+
+    return data;
+}
+
 static int __init nvkms_init(void)
 {
    int ret;
@@ -1651,10 +1768,9 @@ static int __init nvkms_init(void)
    INIT_LIST_HEAD(&nvkms_timers.list);
    spin_lock_init(&nvkms_timers.lock);

-    ret = nvidia_register_module(&nvidia_modeset_module);
-
+    ret = nvkms_register_chrdev();
    if (ret != 0) {
-        goto fail_register_module;
+        goto fail_register_chrdev;
    }

    down(&nvkms_lock);
@@ -1673,8 +1789,8 @@ static int __init nvkms_init(void)
    return 0;

 fail_module_load:
-    nvidia_unregister_module(&nvidia_modeset_module);
-fail_register_module:
+    nvkms_unregister_chrdev();
+fail_register_chrdev:
    nv_kthread_q_stop(&nvkms_deferred_close_kthread_q);
 fail_deferred_close_kthread:
    nv_kthread_q_stop(&nvkms_kthread_q);
@@ -1738,7 +1854,7 @@ restart:
    nv_kthread_q_stop(&nvkms_deferred_close_kthread_q);
    nv_kthread_q_stop(&nvkms_kthread_q);

-    nvidia_unregister_module(&nvidia_modeset_module);
+    nvkms_unregister_chrdev();
    nvkms_free_rm();

    if (malloc_verbose) {
--- a/kernel-open/nvidia-modeset/nvidia-modeset-os-interface.h
+++ b/kernel-open/nvidia-modeset/nvidia-modeset-os-interface.h
@@ -97,8 +97,11 @@ typedef struct {
 } NvKmsSyncPtOpParams;

 NvBool nvkms_output_rounding_fix(void);
-
+NvBool nvkms_disable_hdmi_frl(void);
 NvBool nvkms_disable_vrr_memclk_switch(void);
+NvBool nvkms_hdmi_deepcolor(void);
+NvBool nvkms_vblank_sem_control(void);
+NvBool nvkms_opportunistic_display_sync(void);

 void   nvkms_call_rm    (void *ops);
 void*  nvkms_alloc      (size_t size,
@@ -301,6 +304,11 @@ NvU32 nvkms_enumerate_gpus(nv_gpu_info_t *gpu_info);

 NvBool nvkms_allow_write_combining(void);

+/*!
+ * Check if OS supports syncpoints.
+ */
+NvBool nvkms_kernel_supports_syncpts(void);
+
 /*!
 * Checks whether the fd is associated with an nvidia character device.
 */
@@ -325,6 +333,16 @@ NvBool nvkms_ioctl_from_kapi
    NvU32 cmd, void *params_address, const size_t params_size
 );

+/*!
+ * Like nvkms_ioctl_from_kapi, but return NV_FALSE instead of waiting if the
+ * power management read lock cannot be acquired.
+ */
+NvBool nvkms_ioctl_from_kapi_try_pmlock
+(
+    struct nvkms_per_open *popen,
+    NvU32 cmd, void *params_address, const size_t params_size
+);
+
 /*!
 * APIs for locking.
 */
--- a/kernel-open/nvidia-modeset/nvidia-modeset.Kbuild
+++ b/kernel-open/nvidia-modeset/nvidia-modeset.Kbuild
@@ -58,6 +58,18 @@ nvidia-modeset-y += $(NVIDIA_MODESET_BINARY_OBJECT_O)
 NVIDIA_MODESET_CFLAGS += -I$(src)/nvidia-modeset
 NVIDIA_MODESET_CFLAGS += -UDEBUG -U_DEBUG -DNDEBUG -DNV_BUILD_MODULE_INSTANCES=0

+# Some Android kernels prohibit driver use of filesystem functions like
+# filp_open() and kernel_read().  Disable the NVKMS_CONFIG_FILE_SUPPORTED
+# functionality that uses those functions when building for Android.
+
+PLATFORM_IS_ANDROID ?= 0
+
+ifeq ($(PLATFORM_IS_ANDROID),1)
+  NVIDIA_MODESET_CFLAGS += -DNVKMS_CONFIG_FILE_SUPPORTED=0
+else
+  NVIDIA_MODESET_CFLAGS += -DNVKMS_CONFIG_FILE_SUPPORTED=1
+endif
+
 $(call ASSIGN_PER_OBJ_CFLAGS, $(NVIDIA_MODESET_OBJECTS), $(NVIDIA_MODESET_CFLAGS))


@@ -93,3 +105,4 @@ NV_CONFTEST_FUNCTION_COMPILE_TESTS += list_is_first
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += ktime_get_real_ts64
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += ktime_get_raw_ts64
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += acpi_video_backlight_use_native
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += kernel_read_has_pointer_pos_arg
--- a/kernel-open/nvidia-modeset/nvkms.h
+++ b/kernel-open/nvidia-modeset/nvkms.h
@@ -66,6 +66,8 @@ enum NvKmsClientType {
    NVKMS_CLIENT_KERNEL_SPACE,
 };

+struct NvKmsPerOpenDev;
+
 NvBool nvKmsIoctl(
    void *pOpenVoid,
    NvU32 cmd,
@@ -101,7 +103,11 @@ NvBool nvKmsKapiGetFunctionsTableInternal
    struct NvKmsKapiFunctionsTable *funcsTable
 );

+void nvKmsKapiSuspendResume(NvBool suspend);
+
 NvBool nvKmsGetBacklight(NvU32 display_id, void *drv_priv, NvU32 *brightness);
 NvBool nvKmsSetBacklight(NvU32 display_id, void *drv_priv, NvU32 brightness);

+NvBool nvKmsOpenDevHasSubOwnerPermissionOrBetter(const struct NvKmsPerOpenDev *pOpenDev);
+
 #endif /* __NV_KMS_H__ */
--- a/kernel-open/nvidia-peermem/nvidia-peermem.c
+++ b/kernel-open/nvidia-peermem/nvidia-peermem.c
@@ -1,20 +1,25 @@
-/* SPDX-License-Identifier: Linux-OpenIB */
 /*
 * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
 *
- * Redistribution and use in source and binary forms, with or
- * without modification, are permitted provided that the following
- * conditions are met:
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
 *
- *  - Redistributions of source code must retain the above
- *    copyright notice, this list of conditions and the following
- *    disclaimer.
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
 *
- *  - Redistributions in binary form must reproduce the above
- *    copyright notice, this list of conditions and the following
- *    disclaimer in the documentation and/or other materials
- *    provided with the distribution.
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
@@ -43,7 +48,9 @@

 MODULE_AUTHOR("Yishai Hadas");
 MODULE_DESCRIPTION("NVIDIA GPU memory plug-in");
-MODULE_LICENSE("Linux-OpenIB");
+
+MODULE_LICENSE("Dual BSD/GPL");
+
 MODULE_VERSION(DRV_VERSION);
 enum {
        NV_MEM_PEERDIRECT_SUPPORT_DEFAULT = 0,
@@ -53,7 +60,13 @@ static int peerdirect_support = NV_MEM_PEERDIRECT_SUPPORT_DEFAULT;
 module_param(peerdirect_support, int, S_IRUGO);
 MODULE_PARM_DESC(peerdirect_support, "Set level of support for Peer-direct, 0 [default] or 1 [legacy, for example MLNX_OFED 4.9 LTS]");

-#define peer_err(FMT, ARGS...) printk(KERN_ERR "nvidia-peermem" " %s:%d " FMT, __FUNCTION__, __LINE__, ## ARGS)
+
+#define peer_err(FMT, ARGS...) printk(KERN_ERR "nvidia-peermem" " %s:%d ERROR " FMT, __FUNCTION__, __LINE__, ## ARGS)
+#ifdef NV_MEM_DEBUG
+#define peer_trace(FMT, ARGS...) printk(KERN_DEBUG "nvidia-peermem" " %s:%d TRACE " FMT, __FUNCTION__, __LINE__, ## ARGS)
+#else
+#define peer_trace(FMT, ARGS...) do {} while (0)
+#endif

 #if defined(NV_MLNX_IB_PEER_MEM_SYMBOLS_PRESENT)

@@ -74,7 +87,10 @@ invalidate_peer_memory mem_invalidate_callback;
 static void *reg_handle = NULL;
 static void *reg_handle_nc = NULL;

+#define NV_MEM_CONTEXT_MAGIC ((u64)0xF1F4F1D0FEF0DAD0ULL)
+
 struct nv_mem_context {
+    u64 pad1;
    struct nvidia_p2p_page_table *page_table;
    struct nvidia_p2p_dma_mapping *dma_mapping;
    u64 core_context;
@@ -86,8 +102,22 @@ struct nv_mem_context {
    struct task_struct *callback_task;
    int sg_allocated;
    struct sg_table sg_head;
+    u64 pad2;
 };

+#define NV_MEM_CONTEXT_CHECK_OK(MC) ({                                  \
+    struct nv_mem_context *mc = (MC);                                   \
+    int rc = ((0 != mc) &&                                              \
+              (READ_ONCE(mc->pad1) == NV_MEM_CONTEXT_MAGIC) &&          \
+              (READ_ONCE(mc->pad2) == NV_MEM_CONTEXT_MAGIC));           \
+    if (!rc) {                                                          \
+        peer_trace("invalid nv_mem_context=%px pad1=%016llx pad2=%016llx\n", \
+                   mc,                                                  \
+                   mc?mc->pad1:0,                                       \
+                   mc?mc->pad2:0);                                      \
+    }                                                                   \
+    rc;                                                                 \
+})

 static void nv_get_p2p_free_callback(void *data)
 {
@@ -97,8 +127,9 @@ static void nv_get_p2p_free_callback(void *data)
    struct nvidia_p2p_dma_mapping *dma_mapping = NULL;

    __module_get(THIS_MODULE);
-    if (!nv_mem_context) {
-        peer_err("nv_get_p2p_free_callback -- invalid nv_mem_context\n");
+
+    if (!NV_MEM_CONTEXT_CHECK_OK(nv_mem_context)) {
+        peer_err("detected invalid context, skipping further processing\n");
        goto out;
    }

@@ -169,9 +200,11 @@ static int nv_mem_acquire(unsigned long addr, size_t size, void *peer_mem_privat
        /* Error case handled as not mine */
        return 0;

+    nv_mem_context->pad1 = NV_MEM_CONTEXT_MAGIC;
    nv_mem_context->page_virt_start = addr & GPU_PAGE_MASK;
    nv_mem_context->page_virt_end   = (addr + size + GPU_PAGE_SIZE - 1) & GPU_PAGE_MASK;
    nv_mem_context->mapped_size  = nv_mem_context->page_virt_end - nv_mem_context->page_virt_start;
+    nv_mem_context->pad2 = NV_MEM_CONTEXT_MAGIC;

    ret = nvidia_p2p_get_pages(0, 0, nv_mem_context->page_virt_start, nv_mem_context->mapped_size,
                               &nv_mem_context->page_table, nv_mem_dummy_callback, nv_mem_context);
@@ -195,6 +228,7 @@ static int nv_mem_acquire(unsigned long addr, size_t size, void *peer_mem_privat
    return 1;

 err:
+    memset(nv_mem_context, 0, sizeof(*nv_mem_context));
    kfree(nv_mem_context);

    /* Error case handled as not mine */
@@ -249,8 +283,8 @@ static int nv_dma_map(struct sg_table *sg_head, void *context,
    nv_mem_context->sg_allocated = 1;
    for_each_sg(sg_head->sgl, sg, nv_mem_context->npages, i) {
        sg_set_page(sg, NULL, nv_mem_context->page_size, 0);
-        sg->dma_address = dma_mapping->dma_addresses[i];
-        sg->dma_length = nv_mem_context->page_size;
+        sg_dma_address(sg) = dma_mapping->dma_addresses[i];
+        sg_dma_len(sg) = nv_mem_context->page_size;
    }
    nv_mem_context->sg_head = *sg_head;
    *nmap = nv_mem_context->npages;
@@ -304,8 +338,13 @@ static void nv_mem_put_pages_common(int nc,
        return;

    if (nc) {
+#ifdef NVIDIA_P2P_CAP_GET_PAGES_PERSISTENT_API
        ret = nvidia_p2p_put_pages_persistent(nv_mem_context->page_virt_start,
                                              nv_mem_context->page_table, 0);
+#else
+        ret = nvidia_p2p_put_pages(0, 0, nv_mem_context->page_virt_start,
+                                   nv_mem_context->page_table);
+#endif
    } else {
        ret = nvidia_p2p_put_pages(0, 0, nv_mem_context->page_virt_start,
                                   nv_mem_context->page_table);
@@ -342,6 +381,7 @@ static void nv_mem_release(void *context)
        sg_free_table(&nv_mem_context->sg_head);
        nv_mem_context->sg_allocated = 0;
    }
+    memset(nv_mem_context, 0, sizeof(*nv_mem_context));
    kfree(nv_mem_context);
    module_put(THIS_MODULE);
    return;
@@ -412,9 +452,15 @@ static int nv_mem_get_pages_nc(unsigned long addr,
    nv_mem_context->core_context = core_context;
    nv_mem_context->page_size = GPU_PAGE_SIZE;

+#ifdef NVIDIA_P2P_CAP_GET_PAGES_PERSISTENT_API
    ret = nvidia_p2p_get_pages_persistent(nv_mem_context->page_virt_start,
                                          nv_mem_context->mapped_size,
                                          &nv_mem_context->page_table, 0);
+#else
+    ret = nvidia_p2p_get_pages(0, 0, nv_mem_context->page_virt_start, nv_mem_context->mapped_size,
+                               &nv_mem_context->page_table, NULL, NULL);
+#endif
+
    if (ret < 0) {
        peer_err("error %d while calling nvidia_p2p_get_pages() with NULL callback\n", ret);
        return ret;
@@ -459,8 +505,6 @@ static int __init nv_mem_client_init(void)
    }

 #if defined (NV_MLNX_IB_PEER_MEM_SYMBOLS_PRESENT)
-    int status = 0;
-
    // off by one, to leave space for the trailing '1' which is flagging
    // the new client type
    BUG_ON(strlen(DRV_NAME) > IB_PEER_MEMORY_NAME_MAX-1);
@@ -489,7 +533,7 @@ static int __init nv_mem_client_init(void)
                         &mem_invalidate_callback);
    if (!reg_handle) {
        peer_err("nv_mem_client_init -- error while registering traditional client\n");
-        status = -EINVAL;
+        rc = -EINVAL;
        goto out;
    }

@@ -499,12 +543,12 @@ static int __init nv_mem_client_init(void)
    reg_handle_nc = ib_register_peer_memory_client(&nv_mem_client_nc, NULL);
    if (!reg_handle_nc) {
        peer_err("nv_mem_client_init -- error while registering nc client\n");
-        status = -EINVAL;
+        rc = -EINVAL;
        goto out;
    }

 out:
-    if (status) {
+    if (rc) {
        if (reg_handle) {
            ib_unregister_peer_memory_client(reg_handle);
            reg_handle = NULL;
@@ -516,7 +560,7 @@ out:
        }
    }

-    return status;
+    return rc;
 #else
    return -EINVAL;
 #endif
--- a/kernel-open/nvidia-uvm/clc365.h
+++ b/kernel-open/nvidia-uvm/clc365.h
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2022 NVIDIA Corporation
+    Copyright (c) 2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
--- a/kernel-open/nvidia-uvm/clc369.h
+++ b/kernel-open/nvidia-uvm/clc369.h
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2022 NVIDIA Corporation
+    Copyright (c) 2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
--- a/kernel-open/nvidia-uvm/ctrl2080mc.h
+++ b/kernel-open/nvidia-uvm/ctrl2080mc.h
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2013-2022 NVIDIA Corporation
+    Copyright (c) 2013-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
--- a/kernel-open/nvidia-uvm/nv-kthread-q.c
+++ b/kernel-open/nvidia-uvm/nv-kthread-q.c
@@ -247,6 +247,11 @@ int nv_kthread_q_init_on_node(nv_kthread_q_t *q, const char *q_name, int preferr
    return 0;
 }

+int nv_kthread_q_init(nv_kthread_q_t *q, const char *qname)
+{
+    return nv_kthread_q_init_on_node(q, qname, NV_KTHREAD_NO_NODE);
+}
+
 // Returns true (non-zero) if the item was actually scheduled, and false if the
 // item was already pending in a queue.
 static int _raw_q_schedule(nv_kthread_q_t *q, nv_kthread_q_item_t *q_item)
--- a/kernel-open/nvidia-uvm/nvidia-uvm-sources.Kbuild
+++ b/kernel-open/nvidia-uvm/nvidia-uvm-sources.Kbuild
@@ -27,6 +27,7 @@ NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_rm_mem.c
 NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_channel.c
 NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_lock.c
 NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_hal.c
+NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_processors.c
 NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_range_tree.c
 NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_rb_tree.c
 NVIDIA_UVM_SOURCES += nvidia-uvm/uvm_range_allocator.c
--- a/kernel-open/nvidia-uvm/nvidia-uvm.Kbuild
+++ b/kernel-open/nvidia-uvm/nvidia-uvm.Kbuild
@@ -86,6 +86,8 @@ NV_CONFTEST_FUNCTION_COMPILE_TESTS += mmget_not_zero
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += mmgrab
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += iommu_sva_bind_device_has_drvdata_arg
 NV_CONFTEST_FUNCTION_COMPILE_TESTS += vm_fault_to_errno
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += find_next_bit_wrap
+NV_CONFTEST_FUNCTION_COMPILE_TESTS += iommu_is_dma_domain

 NV_CONFTEST_TYPE_COMPILE_TESTS += backing_dev_info
 NV_CONFTEST_TYPE_COMPILE_TESTS += mm_context_t
@@ -99,6 +101,7 @@ NV_CONFTEST_TYPE_COMPILE_TESTS += kmem_cache_has_kobj_remove_work
 NV_CONFTEST_TYPE_COMPILE_TESTS += sysfs_slab_unlink
 NV_CONFTEST_TYPE_COMPILE_TESTS += vm_fault_t
 NV_CONFTEST_TYPE_COMPILE_TESTS += mmu_notifier_ops_invalidate_range
+NV_CONFTEST_TYPE_COMPILE_TESTS += mmu_notifier_ops_arch_invalidate_secondary_tlbs
 NV_CONFTEST_TYPE_COMPILE_TESTS += proc_ops
 NV_CONFTEST_TYPE_COMPILE_TESTS += timespec64
 NV_CONFTEST_TYPE_COMPILE_TESTS += mm_has_mmap_lock
--- a/kernel-open/nvidia-uvm/nvstatus.c
+++ b/kernel-open/nvidia-uvm/nvstatus.c
@@ -24,16 +24,17 @@
 #include "nvstatus.h"

 #if !defined(NV_PRINTF_STRING_SECTION)
-#if defined(NVRM) && NVCPU_IS_RISCV64
-#define NV_PRINTF_STRING_SECTION         __attribute__ ((section (".logging")))
-#else // defined(NVRM) && NVCPU_IS_RISCV64
+#if defined(NVRM) && NVOS_IS_LIBOS
+#include "libos_log.h"
+#define NV_PRINTF_STRING_SECTION LIBOS_SECTION_LOGGING
+#else // defined(NVRM) && NVOS_IS_LIBOS
 #define NV_PRINTF_STRING_SECTION
-#endif // defined(NVRM) && NVCPU_IS_RISCV64
+#endif // defined(NVRM) && NVOS_IS_LIBOS
 #endif // !defined(NV_PRINTF_STRING_SECTION)

 /*
 * Include nvstatuscodes.h twice.  Once for creating constant strings in the
- * the NV_PRINTF_STRING_SECTION section of the ececutable, and once to build
+ * the NV_PRINTF_STRING_SECTION section of the executable, and once to build
 * the g_StatusCodeList table.
 */
 #undef NV_STATUS_CODE
--- a/kernel-open/nvidia-uvm/uvm.c
+++ b/kernel-open/nvidia-uvm/uvm.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2015-2022 NVIDIA Corporation
+    Copyright (c) 2015-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -571,7 +571,6 @@ static void uvm_vm_open_managed_entry(struct vm_area_struct *vma)
 static void uvm_vm_close_managed(struct vm_area_struct *vma)
 {
    uvm_va_space_t *va_space = uvm_va_space_get(vma->vm_file);
-    uvm_processor_id_t gpu_id;
    bool make_zombie = false;

    if (current->mm != NULL)
@@ -606,12 +605,6 @@ static void uvm_vm_close_managed(struct vm_area_struct *vma)

    uvm_destroy_vma_managed(vma, make_zombie);

-    // Notify GPU address spaces that the fault buffer needs to be flushed to
-    // avoid finding stale entries that can be attributed to new VA ranges
-    // reallocated at the same address.
-    for_each_gpu_id_in_mask(gpu_id, &va_space->registered_gpu_va_spaces) {
-        uvm_processor_mask_set_atomic(&va_space->needs_fault_buffer_flush, gpu_id);
-    }
    uvm_va_space_up_write(va_space);

    if (current->mm != NULL)
@@ -1060,7 +1053,7 @@ NV_STATUS uvm_test_register_unload_state_buffer(UVM_TEST_REGISTER_UNLOAD_STATE_B
    // are not used because unload_state_buf may be a managed memory pointer and
    // therefore a locking assertion from the CPU fault handler could be fired.
    nv_mmap_read_lock(current->mm);
-    ret = NV_PIN_USER_PAGES(params->unload_state_buf, 1, FOLL_WRITE, &page, NULL);
+    ret = NV_PIN_USER_PAGES(params->unload_state_buf, 1, FOLL_WRITE, &page);
    nv_mmap_read_unlock(current->mm);

    if (ret < 0)
--- a/kernel-open/nvidia-uvm/uvm.h
+++ b/kernel-open/nvidia-uvm/uvm.h
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2013-2022 NVIDIA Corporation
+    Copyright (c) 2013-2024 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -45,16 +45,20 @@
 //     #endif
 // 3) Do the same thing for the function definition, and for any structs that
 //    are taken as arguments to these functions.
-// 4) Let this change propagate over to cuda_a, so that the CUDA driver can
-//    start using the new API by bumping up the API version number its using.
-//    This can be found in gpgpu/cuda/cuda.nvmk.
-// 5) Once the cuda_a changes have made it back into chips_a, remove the old API
-//    declaration, definition, and any old structs that were in use.
+// 4) Let this change propagate over to cuda_a and dev_a, so that the CUDA and
+//    nvidia-cfg libraries can start using the new API by bumping up the API
+//    version number it's using.
+//    Places where UVM_API_REVISION is defined are:
+//      drivers/gpgpu/cuda/cuda.nvmk (cuda_a)
+//      drivers/setup/linux/nvidia-cfg/makefile.nvmk (dev_a)
+// 5) Once the dev_a and cuda_a changes have made it back into chips_a,
+//    remove the old API declaration, definition, and any old structs that were
+//    in use.

 #ifndef _UVM_H_
 #define _UVM_H_

-#define UVM_API_LATEST_REVISION 8
+#define UVM_API_LATEST_REVISION 11

 #if !defined(UVM_API_REVISION)
 #error "please define UVM_API_REVISION macro to a desired version number or UVM_API_LATEST_REVISION macro"
@@ -180,12 +184,8 @@ NV_STATUS UvmSetDriverVersion(NvU32 major, NvU32 changelist);
 //         because it is not very informative.
 //
 //------------------------------------------------------------------------------
-#if UVM_API_REV_IS_AT_MOST(4)
-NV_STATUS UvmInitialize(UvmFileDescriptor fd);
-#else
 NV_STATUS UvmInitialize(UvmFileDescriptor fd,
                        NvU64             flags);
-#endif

 //------------------------------------------------------------------------------
 // UvmDeinitialize
@@ -216,6 +216,10 @@ NV_STATUS UvmDeinitialize(void);
 // Note that it is not required to release VA ranges that were reserved with
 // UvmReserveVa().
 //
+// This is useful for per-process checkpoint and restore, where kernel-mode
+// state needs to be reconfigured to match the expectations of a pre-existing
+// user-mode process.
+//
 // UvmReopen() closes the open file returned by UvmGetFileDescriptor() and
 // replaces it with a new open file with the same name.
 //
@@ -293,7 +297,9 @@ NV_STATUS UvmIsPageableMemoryAccessSupported(NvBool *pageableMemAccess);
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU for which pageable memory access support is queried.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition for which
+//         pageable memory access support is queried.
 //
 //     pageableMemAccess: (OUTPUT)
 //         Returns true (non-zero) if the GPU represented by gpuUuid supports
@@ -323,9 +329,19 @@ NV_STATUS UvmIsPageableMemoryAccessSupportedOnGpu(const NvProcessorUuid *gpuUuid
 // usage. Calling UvmRegisterGpu multiple times on the same GPU from the same
 // process results in an error.
 //
+// After successfully registering a GPU partition, all subsequent API calls
+// which take a NvProcessorUuid argument (including UvmGpuMappingAttributes),
+// must use the GI partition UUID which can be obtained with
+// NvRmControl(NVC637_CTRL_CMD_GET_UUID). Otherwise, if the GPU is not SMC
+// capable or SMC enabled, the physical GPU UUID must be used.
+//
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to register.
+//         UUID of the physical GPU to register.
+//
+//     platformParams: (INPUT)
+//         User handles identifying the GPU partition to register.
+//         This should be NULL if the GPU is not SMC capable or SMC enabled.
 //
 // Error codes:
 //     NV_ERR_NO_MEMORY:
@@ -360,27 +376,31 @@ NV_STATUS UvmIsPageableMemoryAccessSupportedOnGpu(const NvProcessorUuid *gpuUuid
 //         OS state required to register the GPU is not present.
 //
 //     NV_ERR_INVALID_STATE:
-//         OS state required to register the GPU is malformed.
+//         OS state required to register the GPU is malformed, or the partition
+//         identified by the user handles or its configuration changed.
 //
 //     NV_ERR_GENERIC:
 //         Unexpected error. We try hard to avoid returning this error code,
 //         because it is not very informative.
 //
 //------------------------------------------------------------------------------
+#if UVM_API_REV_IS_AT_MOST(8)
 NV_STATUS UvmRegisterGpu(const NvProcessorUuid *gpuUuid);
+#else
+NV_STATUS UvmRegisterGpu(const NvProcessorUuid *gpuUuid,
+                         const UvmGpuPlatformParams *platformParams);
+#endif

+#if UVM_API_REV_IS_AT_MOST(8)
 //------------------------------------------------------------------------------
 // UvmRegisterGpuSmc
 //
 // The same as UvmRegisterGpu, but takes additional parameters to specify the
 // GPU partition being registered if SMC is enabled.
 //
-// TODO: Bug 2844714: Merge UvmRegisterGpuSmc() with UvmRegisterGpu() once
-//       the initial SMC support is in place.
-//
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the parent GPU of the SMC partition to register.
+//         UUID of the physical GPU of the SMC partition to register.
 //
 //     platformParams: (INPUT)
 //         User handles identifying the partition to register.
@@ -393,6 +413,7 @@ NV_STATUS UvmRegisterGpu(const NvProcessorUuid *gpuUuid);
 //
 NV_STATUS UvmRegisterGpuSmc(const NvProcessorUuid *gpuUuid,
                            const UvmGpuPlatformParams *platformParams);
+#endif

 //------------------------------------------------------------------------------
 // UvmUnregisterGpu
@@ -418,7 +439,8 @@ NV_STATUS UvmRegisterGpuSmc(const NvProcessorUuid *gpuUuid,
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to unregister.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to unregister.
 //
 // Error codes:
 //     NV_ERR_INVALID_DEVICE:
@@ -476,7 +498,8 @@ NV_STATUS UvmUnregisterGpu(const NvProcessorUuid *gpuUuid);
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to register.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to register.
 //
 //     platformParams: (INPUT)
 //         On Linux: RM ctrl fd, hClient and hVaSpace.
@@ -547,7 +570,9 @@ NV_STATUS UvmRegisterGpuVaSpace(const NvProcessorUuid             *gpuUuid,
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU whose VA space should be unregistered.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition whose VA space
+//         should be unregistered.
 //
 // Error codes:
 //     NV_ERR_INVALID_DEVICE:
@@ -577,7 +602,7 @@ NV_STATUS UvmUnregisterGpuVaSpace(const NvProcessorUuid *gpuUuid);
 //
 // The two GPUs must be connected via PCIe. An error is returned if the GPUs are
 // not connected or are connected over an interconnect different than PCIe
-// (NVLink, for example).
+// (NVLink or SMC partitions, for example).
 //
 // If both GPUs have GPU VA spaces registered for them, the two GPU VA spaces
 // must support the same set of page sizes for GPU mappings.
@@ -590,10 +615,12 @@ NV_STATUS UvmUnregisterGpuVaSpace(const NvProcessorUuid *gpuUuid);
 //
 // Arguments:
 //     gpuUuidA: (INPUT)
-//         UUID of GPU A.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition A.
 //
 //     gpuUuidB: (INPUT)
-//         UUID of GPU B.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition B.
 //
 // Error codes:
 //     NV_ERR_NO_MEMORY:
@@ -639,10 +666,12 @@ NV_STATUS UvmEnablePeerAccess(const NvProcessorUuid *gpuUuidA,
 //
 // Arguments:
 //     gpuUuidA: (INPUT)
-//         UUID of GPU A.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition A.
 //
 //     gpuUuidB: (INPUT)
-//         UUID of GPU B.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition B.
 //
 // Error codes:
 //     NV_ERR_INVALID_DEVICE:
@@ -687,7 +716,9 @@ NV_STATUS UvmDisablePeerAccess(const NvProcessorUuid *gpuUuidA,
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//        UUID of the GPU that the channel is associated with.
+//        UUID of the physical GPU if the GPU is not SMC capable or SMC
+//        enabled, or the GPU instance UUID of the partition that the channel is
+//        associated with.
 //
 //     platformParams: (INPUT)
 //         On Linux: RM ctrl fd, hClient and hChannel.
@@ -1126,11 +1157,14 @@ NV_STATUS UvmAllowMigrationRangeGroups(const NvU64 *rangeGroupIds,
 //         Length, in bytes, of the range.
 //
 //     preferredLocationUuid: (INPUT)
-//         UUID of the preferred location for this VA range.
+//         UUID of the CPU, UUID of the physical GPU if the GPU is not SMC
+//         capable or SMC enabled, or the GPU instance UUID of the partition of
+//         the preferred location for this VA range.
 //
 //     accessedByUuids: (INPUT)
-//         UUIDs of all processors that should have persistent mappings to this
-//         VA range.
+//         UUID of the CPU, UUID of the physical GPUs if the GPUs are not SMC
+//         capable or SMC enabled, or the GPU instance UUID of the partitions
+//         that should have persistent mappings to this VA range.
 //
 //     accessedByCount: (INPUT)
 //         Number of elements in the accessedByUuids array.
@@ -1408,12 +1442,13 @@ NV_STATUS UvmAllocSemaphorePool(void                          *base,
 //         Length, in bytes, of the range.
 //
 //     destinationUuid: (INPUT)
-//         UUID of the destination processor to migrate pages to.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID to
+//         migrate pages to.
 //
 //     preferredCpuMemoryNode: (INPUT)
 //         Preferred CPU NUMA memory node used if the destination processor is
-//         the CPU. This argument is ignored if the given virtual address range
-//         corresponds to managed memory.
+//         the CPU.
 //
 // Error codes:
 //     NV_ERR_INVALID_ADDRESS:
@@ -1452,16 +1487,10 @@ NV_STATUS UvmAllocSemaphorePool(void                          *base,
 //         pages were associated with a non-migratable range group.
 //
 //------------------------------------------------------------------------------
-#if UVM_API_REV_IS_AT_MOST(5)
-NV_STATUS UvmMigrate(void                  *base,
-                     NvLength               length,
-                     const NvProcessorUuid *destinationUuid);
-#else
 NV_STATUS UvmMigrate(void                  *base,
                     NvLength               length,
                     const NvProcessorUuid *destinationUuid,
                     NvS32                  preferredCpuMemoryNode);
-#endif

 //------------------------------------------------------------------------------
 // UvmMigrateAsync
@@ -1493,7 +1522,9 @@ NV_STATUS UvmMigrate(void                  *base,
 //         Length, in bytes, of the range.
 //
 //     destinationUuid: (INPUT)
-//         UUID of the destination processor to migrate pages to.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID to
+//         migrate pages to.
 //
 //     preferredCpuMemoryNode: (INPUT)
 //         Preferred CPU NUMA memory node used if the destination processor is
@@ -1543,20 +1574,12 @@ NV_STATUS UvmMigrate(void                  *base,
 //         pages were associated with a non-migratable range group.
 //
 //------------------------------------------------------------------------------
-#if UVM_API_REV_IS_AT_MOST(5)
-NV_STATUS UvmMigrateAsync(void                  *base,
-                          NvLength               length,
-                          const NvProcessorUuid *destinationUuid,
-                          void                  *semaphoreAddress,
-                          NvU32                  semaphorePayload);
-#else
 NV_STATUS UvmMigrateAsync(void                  *base,
                          NvLength               length,
                          const NvProcessorUuid *destinationUuid,
                          NvS32                  preferredCpuMemoryNode,
                          void                  *semaphoreAddress,
                          NvU32                  semaphorePayload);
-#endif

 //------------------------------------------------------------------------------
 // UvmMigrateRangeGroup
@@ -1564,9 +1587,7 @@ NV_STATUS UvmMigrateAsync(void                  *base,
 // Migrates the backing of all virtual address ranges associated with the given
 // range group to the specified destination processor. The behavior of this API
 // is equivalent to calling UvmMigrate on each VA range associated with this
-// range group. The value for the preferredCpuMemoryNode is irrelevant in this
-// case as it only applies to migrations of pageable address, which cannot be
-// used to create range groups.
+// range group.
 //
 // Any errors encountered during migration are returned immediately. No attempt
 // is made to migrate the remaining unmigrated ranges and the ranges that are
@@ -1580,7 +1601,9 @@ NV_STATUS UvmMigrateAsync(void                  *base,
 //         Id of the range group whose associated VA ranges have to be migrated.
 //
 //     destinationUuid: (INPUT)
-//         UUID of the destination processor to migrate pages to.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID to
+//         migrate pages to.
 //
 // Error codes:
 //     NV_ERR_OBJECT_NOT_FOUND:
@@ -1942,7 +1965,9 @@ NV_STATUS UvmMapExternalAllocation(void                              *base,
 //
 //
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to map the sparse region on.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to map the sparse
+//         region on.
 //
 // Errors:
 //     NV_ERR_INVALID_ADDRESS:
@@ -1999,7 +2024,9 @@ NV_STATUS UvmMapExternalSparse(void                  *base,
 //         The length of the virtual address range.
 //
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to unmap the VA range from.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to unmap the VA
+//         range from.
 //
 // Errors:
 //     NV_ERR_INVALID_ADDRESS:
@@ -2066,7 +2093,9 @@ NV_STATUS UvmUnmapExternalAllocation(void                  *base,
 //         supported by the GPU.
 //
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to map the dynamic parallelism region on.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to map the
+//         dynamic parallelism region on.
 //
 // Errors:
 //     NV_ERR_UVM_ADDRESS_IN_USE:
@@ -2297,15 +2326,14 @@ NV_STATUS UvmDisableReadDuplication(void     *base,
 //         Length, in bytes, of the range.
 //
 //     preferredLocationUuid: (INPUT)
-//         UUID of the preferred location.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID
+//         preferred location.
 //
-//     preferredCpuNumaNode: (INPUT)
+//     preferredCpuMemoryNode: (INPUT)
 //         Preferred CPU NUMA memory node used if preferredLocationUuid is the
 //         UUID of the CPU. -1 is a special value which indicates all CPU nodes
-//         allowed by the global and thread memory policies. This argument is
-//         ignored if preferredLocationUuid refers to a GPU or the given virtual
-//         address range corresponds to managed memory. If NUMA is not enabled,
-//         only 0 or -1 is allowed.
+//         allowed by the global and thread memory policies.
 //
 // Errors:
 //     NV_ERR_INVALID_ADDRESS:
@@ -2335,10 +2363,11 @@ NV_STATUS UvmDisableReadDuplication(void     *base,
 //
 //      NV_ERR_INVALID_ARGUMENT:
 //         One of the following occured:
-//         - preferredLocationUuid is the UUID of a CPU and preferredCpuNumaNode
-//           refers to a registered GPU.
-//         - preferredCpuNumaNode is invalid and preferredLocationUuid is the
-//           UUID of the CPU.
+//         - preferredLocationUuid is the UUID of the CPU and
+//           preferredCpuMemoryNode is either:
+//              - not a valid NUMA node,
+//              - not a possible NUMA node, or
+//              - a NUMA node ID corresponding to a registered GPU.
 //
 //     NV_ERR_NOT_SUPPORTED:
 //         The UVM file descriptor is associated with another process and the
@@ -2349,16 +2378,10 @@ NV_STATUS UvmDisableReadDuplication(void     *base,
 //         because it is not very informative.
 //
 //------------------------------------------------------------------------------
-#if UVM_API_REV_IS_AT_MOST(7)
-NV_STATUS UvmSetPreferredLocation(void                  *base,
-                                  NvLength               length,
-                                  const NvProcessorUuid *preferredLocationUuid);
-#else
 NV_STATUS UvmSetPreferredLocation(void                  *base,
                                  NvLength               length,
                                  const NvProcessorUuid *preferredLocationUuid,
-                                  NvS32                  preferredCpuNumaNode);
-#endif
+                                  NvS32                  preferredCpuMemoryNode);

 //------------------------------------------------------------------------------
 // UvmUnsetPreferredLocation
@@ -2481,8 +2504,9 @@ NV_STATUS UvmUnsetPreferredLocation(void     *base,
 //         Length, in bytes, of the range.
 //
 //     accessedByUuid: (INPUT)
-//         UUID of the processor that should have pages in the the VA range
-//         mapped when possible.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID
+//         that should have pages in the VA range mapped when possible.
 //
 // Errors:
 //     NV_ERR_INVALID_ADDRESS:
@@ -2550,8 +2574,10 @@ NV_STATUS UvmSetAccessedBy(void                  *base,
 //         Length, in bytes, of the range.
 //
 //     accessedByUuid: (INPUT)
-//         UUID of the processor from which any policies set by
-//         UvmSetAccessedBy should be revoked for the given VA range.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID
+//         from which any policies set by UvmSetAccessedBy should be revoked
+//         for the given VA range.
 //
 // Errors:
 //     NV_ERR_INVALID_ADDRESS:
@@ -2609,7 +2635,9 @@ NV_STATUS UvmUnsetAccessedBy(void                  *base,
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to enable software-assisted system-wide atomics on.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to enable
+//         software-assisted system-wide atomics on.
 //
 // Error codes:
 //     NV_ERR_NO_MEMORY:
@@ -2645,7 +2673,9 @@ NV_STATUS UvmEnableSystemWideAtomics(const NvProcessorUuid *gpuUuid);
 //
 // Arguments:
 //     gpuUuid: (INPUT)
-//         UUID of the GPU to disable software-assisted system-wide atomics on.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition to disable
+//         software-assisted system-wide atomics on.
 //
 // Error codes:
 //     NV_ERR_INVALID_DEVICE:
@@ -2874,7 +2904,9 @@ NV_STATUS UvmDebugCountersEnable(UvmDebugSession   session,
 //         Name of the counter in that scope.
 //
 //     gpu: (INPUT)
-//         Gpuid of the scoped GPU. This parameter is ignored in AllGpu scopes.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, or the GPU instance UUID of the partition of the scoped GPU.
+//         This parameter is ignored in AllGpu scopes.
 //
 //     pCounterHandle: (OUTPUT)
 //         Handle to the counter address.
@@ -2928,7 +2960,7 @@ NV_STATUS UvmDebugGetCounterVal(UvmDebugSession     session,
 // UvmEventQueueCreate
 //
 // This call creates an event queue of the given size.
-// No events are added in the queue till they are enabled by the user.
+// No events are added in the queue until they are enabled by the user.
 // Event queue data is visible to the user even after the target process dies
 // if the session is active and queue is not freed.
 //
@@ -2979,7 +3011,7 @@ NV_STATUS UvmEventQueueCreate(UvmDebugSession        sessionHandle,
 // UvmEventQueueDestroy
 //
 // This call frees all interal resources associated with the queue, including
-// upinning of the memory associated with that queue. Freeing user buffer is
+// unpinning of the memory associated with that queue. Freeing user buffer is
 // responsibility of a caller. Event queue might be also destroyed as a side
 // effect of destroying a session associated with this queue.
 //
@@ -3163,9 +3195,9 @@ NV_STATUS UvmEventGetNotificationHandles(UvmEventQueueHandle  *queueHandleArray,
 // UvmEventGetGpuUuidTable
 //
 // Each migration event entry contains the gpu index to/from where data is
-// migrated. This index maps to a corresponding gpu UUID in the gpuUuidTable.
-// Using indices saves on the size of each event entry. This API provides the
-// gpuIndex to gpuUuid relation to the user.
+// migrated. This index maps to a corresponding physical gpu UUID in the
+// gpuUuidTable. Using indices saves on the size of each event entry. This API
+// provides the gpuIndex to gpuUuid relation to the user.
 //
 // This API does not access the queue state maintained in the user
 // library and so the user doesn't need to acquire a lock to protect the
@@ -3173,9 +3205,9 @@ NV_STATUS UvmEventGetNotificationHandles(UvmEventQueueHandle  *queueHandleArray,
 //
 // Arguments:
 //     gpuUuidTable: (OUTPUT)
-//         The return value is an array of UUIDs. The array index is the
-//         corresponding gpuIndex. There can be at max 32 gpus associated with
-//         UVM, so array size is 32.
+//         The return value is an array of physical GPU UUIDs. The array index
+//         is the corresponding gpuIndex. There can be at max 32 GPUs
+//         associated with UVM, so array size is 32.
 //
 //     validCount: (OUTPUT)
 //         The system doesn't normally contain 32 GPUs. This field gives the
@@ -3234,7 +3266,7 @@ NV_STATUS UvmEventGetGpuUuidTable(NvProcessorUuid *gpuUuidTable,
 //------------------------------------------------------------------------------
 NV_STATUS UvmEventFetch(UvmDebugSession      sessionHandle,
                        UvmEventQueueHandle  queueHandle,
-                        UvmEventEntry       *pBuffer,
+                        UvmEventEntry_V1    *pBuffer,
                        NvU64               *nEntries);

 //------------------------------------------------------------------------------
@@ -3430,10 +3462,14 @@ NV_STATUS UvmToolsDestroySession(UvmToolsSessionHandle session);
 // 4. Destroy event Queue using UvmToolsDestroyEventQueue
 //

-
+#if UVM_API_REV_IS_AT_MOST(10)
+// This is deprecated and replaced by sizeof(UvmToolsEventControlData).
 NvLength UvmToolsGetEventControlSize(void);

+// This is deprecated and replaced by sizeof(UvmEventEntry_V1) or
+// sizeof(UvmEventEntry_V2).
 NvLength UvmToolsGetEventEntrySize(void);
+#endif

 NvLength UvmToolsGetNumberOfCounters(void);

@@ -3448,6 +3484,10 @@ NvLength UvmToolsGetNumberOfCounters(void);
 //     session: (INPUT)
 //         Handle to the tools session.
 //
+//     version: (INPUT)
+//         Requested version for events or counters.
+//         See UvmEventEntry_V1 and UvmEventEntry_V2.
+//
 //     event_buffer: (INPUT)
 //         User allocated buffer. Must be page-aligned. Must be large enough to
 //         hold at least event_buffer_size events. Gets pinned until queue is
@@ -3460,9 +3500,7 @@ NvLength UvmToolsGetNumberOfCounters(void);
 //     event_control (INPUT)
 //         User allocated buffer. Must be page-aligned. Must be large enough to
 //         hold UvmToolsEventControlData (although single page-size allocation
-//         should be more than enough). One could call
-//         UvmToolsGetEventControlSize() function to find out current size of
-//         UvmToolsEventControlData. Gets pinned until queue is destroyed.
+//         should be more than enough). Gets pinned until queue is destroyed.
 //
 //     queue: (OUTPUT)
 //         Handle to the created queue.
@@ -3472,22 +3510,32 @@ NvLength UvmToolsGetNumberOfCounters(void);
 //         Session handle does not refer to a valid session
 //
 //     NV_ERR_INVALID_ARGUMENT:
+//         The version is not UvmEventEntry_V1 or UvmEventEntry_V2.
 //         One of the parameters: event_buffer, event_buffer_size, event_control
 //         is not valid
 //
 //     NV_ERR_INSUFFICIENT_RESOURCES:
-//         There could be multiple reasons for this error. One would be that it's
-//         not possible to allocate a queue of requested size. Another would be
-//         that either event_buffer or event_control memory couldn't be pinned
-//         (e.g. because of OS limitation of pinnable memory). Also it could not
-//         have been possible to create UvmToolsEventQueueDescriptor.
+//         There could be multiple reasons for this error. One would be that
+//         it's not possible to allocate a queue of requested size. Another
+//         would be either event_buffer or event_control memory couldn't be
+//         pinned (e.g. because of OS limitation of pinnable memory). Also it
+//         could not have been possible to create UvmToolsEventQueueDescriptor.
 //
 //------------------------------------------------------------------------------
+#if UVM_API_REV_IS_AT_MOST(10)
 NV_STATUS UvmToolsCreateEventQueue(UvmToolsSessionHandle     session,
                                   void                     *event_buffer,
                                   NvLength                  event_buffer_size,
                                   void                     *event_control,
                                   UvmToolsEventQueueHandle *queue);
+#else
+NV_STATUS UvmToolsCreateEventQueue(UvmToolsSessionHandle        session,
+                                   UvmToolsEventQueueVersion    version,
+                                   void                        *event_buffer,
+                                   NvLength                     event_buffer_size,
+                                   void                        *event_control,
+                                   UvmToolsEventQueueHandle    *queue);
+#endif

 UvmToolsEventQueueDescriptor UvmToolsGetEventQueueDescriptor(UvmToolsEventQueueHandle queue);

@@ -3524,7 +3572,7 @@ NV_STATUS UvmToolsSetNotificationThreshold(UvmToolsEventQueueHandle queue,
 //------------------------------------------------------------------------------
 // UvmToolsDestroyEventQueue
 //
-// Destroys all internal resources associated with the queue. It unpinns the
+// Destroys all internal resources associated with the queue. It unpins the
 // buffers provided in UvmToolsCreateEventQueue. Event Queue is also auto
 // destroyed when corresponding session gets destroyed.
 //
@@ -3546,7 +3594,7 @@ NV_STATUS UvmToolsDestroyEventQueue(UvmToolsEventQueueHandle queue);
 // UvmEventQueueEnableEvents
 //
 // This call enables a particular event type in the event queue. All events are
-// disabled by default. Any event type is considered listed if and only if it's
+// disabled by default. Any event type is considered listed if and only if its
 // corresponding value is equal to 1 (in other words, bit is set). Disabled
 // events listed in eventTypeFlags are going to be enabled. Enabled events and
 // events not listed in eventTypeFlags are not affected by this call.
@@ -3579,7 +3627,7 @@ NV_STATUS UvmToolsEventQueueEnableEvents(UvmToolsEventQueueHandle queue,
 // UvmToolsEventQueueDisableEvents
 //
 // This call disables a particular event type in the event queue. Any event type
-// is considered listed if and only if it's corresponding value is equal to 1
+// is considered listed if and only if its corresponding value is equal to 1
 // (in other words, bit is set). Enabled events listed in eventTypeFlags are
 // going to be disabled. Disabled events and events not listed in eventTypeFlags
 // are not affected by this call.
@@ -3617,7 +3665,7 @@ NV_STATUS UvmToolsEventQueueDisableEvents(UvmToolsEventQueueHandle queue,
 //
 // Counters position follows the layout of the memory that UVM driver decides to
 // use. To obtain particular counter value, user should perform consecutive
-// atomic reads at a a given buffer + offset address.
+// atomic reads at a given buffer + offset address.
 //
 // It is not defined what is the initial value of a counter. User should rely on
 // a difference between each snapshot.
@@ -3640,9 +3688,9 @@ NV_STATUS UvmToolsEventQueueDisableEvents(UvmToolsEventQueueHandle queue,
 //         Provided session is not valid
 //
 //     NV_ERR_INSUFFICIENT_RESOURCES
-//         There could be multiple reasons for this error. One would be that it's
-//         not possible to allocate counters structure. Another would be that
-//         either event_buffer or event_control memory couldn't be pinned
+//         There could be multiple reasons for this error. One would be that
+//         it's not possible to allocate counters structure. Another would be
+//         that either event_buffer or event_control memory couldn't be pinned
 //         (e.g. because of OS limitation of pinnable memory)
 //
 //------------------------------------------------------------------------------
@@ -3653,12 +3701,12 @@ NV_STATUS UvmToolsCreateProcessAggregateCounters(UvmToolsSessionHandle   session
 //------------------------------------------------------------------------------
 // UvmToolsCreateProcessorCounters
 //
-// Creates the counters structure for tracking per-process counters.
+// Creates the counters structure for tracking per-processor counters.
 // These counters are disabled by default.
 //
 // Counters position follows the layout of the memory that UVM driver decides to
 // use. To obtain particular counter value, user should perform consecutive
-// atomic reads at a a given buffer + offset address.
+// atomic reads at a given buffer + offset address.
 //
 // It is not defined what is the initial value of a counter. User should rely on
 // a difference between each snapshot.
@@ -3674,7 +3722,9 @@ NV_STATUS UvmToolsCreateProcessAggregateCounters(UvmToolsSessionHandle   session
 //         counters are destroyed.
 //
 //     processorUuid: (INPUT)
-//        UUID of the resource, for which counters will provide statistic data.
+//         UUID of the physical GPU if the GPU is not SMC capable or SMC
+//         enabled, the GPU instance UUID of the partition, or the CPU UUID of
+//         the resource, for which counters will provide statistic data.
 //
 //     counters: (OUTPUT)
 //         Handle to the created counters.
@@ -3684,9 +3734,9 @@ NV_STATUS UvmToolsCreateProcessAggregateCounters(UvmToolsSessionHandle   session
 //         session handle does not refer to a valid tools session
 //
 //     NV_ERR_INSUFFICIENT_RESOURCES
-//         There could be multiple reasons for this error. One would be that it's
-//         not possible to allocate counters structure. Another would be that
-//         either event_buffer or event_control memory couldn't be pinned
+//         There could be multiple reasons for this error. One would be that
+//         it's not possible to allocate counters structure. Another would be
+//         that either event_buffer or event_control memory couldn't be pinned
 //         (e.g. because of OS limitation of pinnable memory)
 //
 //     NV_ERR_INVALID_ARGUMENT
@@ -3702,7 +3752,7 @@ NV_STATUS UvmToolsCreateProcessorCounters(UvmToolsSessionHandle   session,
 // UvmToolsDestroyCounters
 //
 // Destroys all internal resources associated with this counters structure.
-// It unpinns the buffer provided in UvmToolsCreate*Counters. Counters structure
+// It unpins the buffer provided in UvmToolsCreate*Counters. Counters structure
 // also gest destroyed when corresponding session is destroyed.
 //
 // Arguments:
@@ -3723,7 +3773,7 @@ NV_STATUS UvmToolsDestroyCounters(UvmToolsCountersHandle counters);
 // UvmToolsEnableCounters
 //
 // This call enables certain counter types in the counters structure. Any
-// counter type is considered listed if and only if it's corresponding value is
+// counter type is considered listed if and only if its corresponding value is
 // equal to 1 (in other words, bit is set). Disabled counter types listed in
 // counterTypeFlags are going to be enabled. Already enabled counter types and
 // counter types not listed in counterTypeFlags are not affected by this call.
@@ -3757,7 +3807,7 @@ NV_STATUS UvmToolsEnableCounters(UvmToolsCountersHandle counters,
 // UvmToolsDisableCounters
 //
 // This call disables certain counter types in the counters structure. Any
-// counter type is considered listed if and only if it's corresponding value is
+// counter type is considered listed if and only if its corresponding value is
 // equal to 1 (in other words, bit is set). Enabled counter types listed in
 // counterTypeFlags are going to be disabled. Already disabled counter types and
 // counter types not listed in counterTypeFlags are not affected by this call.
@@ -3902,32 +3952,72 @@ NV_STATUS UvmToolsWriteProcessMemory(UvmToolsSessionHandle  session,
 // UvmToolsGetProcessorUuidTable
 //
 // Populate a table with the UUIDs of all the currently registered processors
-// in the target process.  When a GPU is registered, it is added to the table.
-// When a GPU is unregistered, it is removed.  As long as a GPU remains registered,
-// its index in the table does not change.  New registrations obtain the first
-// unused index.
+// in the target process. When a GPU is registered, it is added to the table.
+// When a GPU is unregistered, it is removed. As long as a GPU remains
+// registered, its index in the table does not change.
+// Note that the index in the table corresponds to the processor ID reported
+// in UvmEventEntry event records and that the table is not contiguously packed
+// with non-zero UUIDs even with no GPU unregistrations.
 //
 // Arguments:
 //     session: (INPUT)
 //         Handle to the tools session.
 //
+//     version: (INPUT)
+//         Requested version for the UUID table returned. The version must
+//         match the requested version of the event queue created with
+//         UvmToolsCreateEventQueue().
+//         See UvmEventEntry_V1 and UvmEventEntry_V2.
+//
 //     table: (OUTPUT)
 //         Array of processor UUIDs, including the CPU's UUID which is always
 //         at index zero.  The srcIndex and dstIndex fields of the
 //         UvmEventMigrationInfo struct index this array.  Unused indices will
-//         have a UUID of zero.
+//         have a UUID of zero. Version UvmEventEntry_V1 only uses GPU UUIDs
+//         for the UUID of the physical GPU and only supports a single SMC
+//         partition registered per process. Version UvmEventEntry_V2 supports
+//         multiple SMC partitions registered per process and uses physical GPU
+//         UUIDs if the GPU is not SMC capable or SMC enabled and GPU instance
+//         UUIDs for SMC partitions.
+//         The table pointer can be NULL in which case, the size of the table
+//         needed to hold all the UUIDs is returned in 'count'.
+//
+//     table_size: (INPUT)
+//         The size of the table in number of array elements. This can be
+//         zero if the table pointer is NULL.
 //
 //     count: (OUTPUT)
-//         Set by UVM to the number of UUIDs written, including any gaps in
-//         the table due to unregistered GPUs.
+//         On output, it is set by UVM to the number of UUIDs needed to hold
+//         all the UUIDs, including any gaps in the table due to unregistered
+//         GPUs.
 //
 // Error codes:
 //     NV_ERR_INVALID_ADDRESS:
-//         writing to table failed.
+//         writing to table failed or the count pointer was invalid.
+//
+//     NV_ERR_INVALID_ARGUMENT:
+//         The version is not UvmEventEntry_V1 or UvmEventEntry_V2.
+//         The count pointer is NULL.
+//         See UvmToolsEventQueueVersion.
+//
+//     NV_WARN_MISMATCHED_TARGET:
+//         The kernel returned a table suitable for UvmEventEntry_V1 events.
+//         (i.e., the kernel is older and doesn't support UvmEventEntry_V2).
+//
+//     NV_ERR_NO_MEMORY:
+//         Internal memory allocation failed.
 //------------------------------------------------------------------------------
+#if UVM_API_REV_IS_AT_MOST(10)
 NV_STATUS UvmToolsGetProcessorUuidTable(UvmToolsSessionHandle  session,
                                        NvProcessorUuid       *table,
                                        NvLength              *count);
+#else
+NV_STATUS UvmToolsGetProcessorUuidTable(UvmToolsSessionHandle      session,
+                                        UvmToolsEventQueueVersion  version,
+                                        NvProcessorUuid           *table,
+                                        NvLength                   table_size,
+                                        NvLength                  *count);
+#endif

 //------------------------------------------------------------------------------
 // UvmToolsFlushEvents
--- a/kernel-open/nvidia-uvm/uvm_ada.c
+++ b/kernel-open/nvidia-uvm/uvm_ada.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2021 NVIDIA Corporation
+    Copyright (c) 2021-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -79,6 +79,8 @@ void uvm_hal_ada_arch_init_properties(uvm_parent_gpu_t *parent_gpu)

    parent_gpu->access_counters_supported = true;

+    parent_gpu->access_counters_can_use_physical_addresses = false;
+
    parent_gpu->fault_cancel_va_supported = true;

    parent_gpu->scoped_atomics_supported = true;
@@ -94,4 +96,6 @@ void uvm_hal_ada_arch_init_properties(uvm_parent_gpu_t *parent_gpu)
    parent_gpu->map_remap_larger_page_promotion = false;

    parent_gpu->plc_supported = true;
+
+    parent_gpu->no_ats_range_required = false;
 }
--- a/kernel-open/nvidia-uvm/uvm_ampere.c
+++ b/kernel-open/nvidia-uvm/uvm_ampere.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2018-20221 NVIDIA Corporation
+    Copyright (c) 2018-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -38,10 +38,12 @@ void uvm_hal_ampere_arch_init_properties(uvm_parent_gpu_t *parent_gpu)

    parent_gpu->utlb_per_gpc_count = uvm_ampere_get_utlbs_per_gpc(parent_gpu);

-    parent_gpu->fault_buffer_info.replayable.utlb_count = parent_gpu->rm_info.maxGpcCount * parent_gpu->utlb_per_gpc_count;
+    parent_gpu->fault_buffer_info.replayable.utlb_count = parent_gpu->rm_info.maxGpcCount *
+                                                          parent_gpu->utlb_per_gpc_count;
    {
        uvm_fault_buffer_entry_t *dummy;
-        UVM_ASSERT(parent_gpu->fault_buffer_info.replayable.utlb_count <= (1 << (sizeof(dummy->fault_source.utlb_id) * 8)));
+        UVM_ASSERT(parent_gpu->fault_buffer_info.replayable.utlb_count <= (1 <<
+                                                                           (sizeof(dummy->fault_source.utlb_id) * 8)));
    }

    // A single top level PDE on Ampere covers 128 TB and that's the minimum
@@ -53,7 +55,7 @@ void uvm_hal_ampere_arch_init_properties(uvm_parent_gpu_t *parent_gpu)
    parent_gpu->uvm_mem_va_size = UVM_MEM_VA_SIZE;

    // See uvm_mmu.h for mapping placement
-    parent_gpu->flat_vidmem_va_base = 136 * UVM_SIZE_1TB;
+    parent_gpu->flat_vidmem_va_base = 160 * UVM_SIZE_1TB;
    parent_gpu->flat_sysmem_va_base = 256 * UVM_SIZE_1TB;

    parent_gpu->ce_phys_vidmem_write_supported = true;
@@ -81,6 +83,8 @@ void uvm_hal_ampere_arch_init_properties(uvm_parent_gpu_t *parent_gpu)

    parent_gpu->access_counters_supported = true;

+    parent_gpu->access_counters_can_use_physical_addresses = false;
+
    parent_gpu->fault_cancel_va_supported = true;

    parent_gpu->scoped_atomics_supported = true;
@@ -101,4 +105,6 @@ void uvm_hal_ampere_arch_init_properties(uvm_parent_gpu_t *parent_gpu)
        parent_gpu->map_remap_larger_page_promotion = false;

    parent_gpu->plc_supported = true;
+
+    parent_gpu->no_ats_range_required = false;
 }
--- a/kernel-open/nvidia-uvm/uvm_ampere_ce.c
+++ b/kernel-open/nvidia-uvm/uvm_ampere_ce.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2018-2022 NVIDIA Corporation
+    Copyright (c) 2018-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -117,7 +117,7 @@ bool uvm_hal_ampere_ce_memcopy_is_valid_c6b5(uvm_push_t *push, uvm_gpu_address_t
    NvU64 push_begin_gpu_va;
    uvm_gpu_t *gpu = uvm_push_get_gpu(push);

-    if (!uvm_gpu_is_virt_mode_sriov_heavy(gpu))
+    if (!uvm_parent_gpu_is_virt_mode_sriov_heavy(gpu->parent))
        return true;

    if (uvm_channel_is_proxy(push->channel)) {
@@ -196,7 +196,7 @@ bool uvm_hal_ampere_ce_memset_is_valid_c6b5(uvm_push_t *push,
 {
    uvm_gpu_t *gpu = uvm_push_get_gpu(push);

-    if (!uvm_gpu_is_virt_mode_sriov_heavy(gpu))
+    if (!uvm_parent_gpu_is_virt_mode_sriov_heavy(gpu->parent))
        return true;

    if (uvm_channel_is_proxy(push->channel)) {
--- a/kernel-open/nvidia-uvm/uvm_ampere_host.c
+++ b/kernel-open/nvidia-uvm/uvm_ampere_host.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2018-2022 NVIDIA Corporation
+    Copyright (c) 2018-2024 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -33,7 +33,7 @@ bool uvm_hal_ampere_host_method_is_valid(uvm_push_t *push, NvU32 method_address,
 {
    uvm_gpu_t *gpu = uvm_push_get_gpu(push);

-    if (!uvm_gpu_is_virt_mode_sriov_heavy(gpu))
+    if (!uvm_parent_gpu_is_virt_mode_sriov_heavy(gpu->parent))
        return true;

    if (uvm_channel_is_privileged(push->channel)) {
@@ -205,17 +205,18 @@ void uvm_hal_ampere_host_clear_faulted_channel_sw_method(uvm_push_t *push,
                     CLEAR_FAULTED_B, HWVALUE(C076, CLEAR_FAULTED_B, INST_HI, instance_ptr_hi));
 }

-// Copy from Pascal, this version sets TLB_INVALIDATE_INVAL_SCOPE.
+// Copy from Turing, this version sets TLB_INVALIDATE_INVAL_SCOPE.
 void uvm_hal_ampere_host_tlb_invalidate_all(uvm_push_t *push,
-                                            uvm_gpu_phys_address_t pdb,
-                                            NvU32 depth,
-                                            uvm_membar_t membar)
+                                           uvm_gpu_phys_address_t pdb,
+                                           NvU32 depth,
+                                           uvm_membar_t membar)
 {
    NvU32 aperture_value;
    NvU32 page_table_level;
    NvU32 pdb_lo;
    NvU32 pdb_hi;
    NvU32 ack_value = 0;
+    NvU32 sysmembar_value = 0;

    UVM_ASSERT_MSG(pdb.aperture == UVM_APERTURE_VID || pdb.aperture == UVM_APERTURE_SYS, "aperture: %u", pdb.aperture);

@@ -230,8 +231,8 @@ void uvm_hal_ampere_host_tlb_invalidate_all(uvm_push_t *push,
    pdb_lo = pdb.address & HWMASK(C56F, MEM_OP_C, TLB_INVALIDATE_PDB_ADDR_LO);
    pdb_hi = pdb.address >> HWSIZE(C56F, MEM_OP_C, TLB_INVALIDATE_PDB_ADDR_LO);

-    // PDE3 is the highest level on Pascal, see the comment in uvm_pascal_mmu.c
-    // for details.
+    // PDE3 is the highest level on Pascal-Ampere, see the comment in
+    // uvm_pascal_mmu.c for details.
    UVM_ASSERT_MSG(depth < NVC56F_MEM_OP_C_TLB_INVALIDATE_PAGE_TABLE_LEVEL_UP_TO_PDE3, "depth %u", depth);
    page_table_level = NVC56F_MEM_OP_C_TLB_INVALIDATE_PAGE_TABLE_LEVEL_UP_TO_PDE3 - depth;

@@ -242,7 +243,12 @@ void uvm_hal_ampere_host_tlb_invalidate_all(uvm_push_t *push,
        ack_value = HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_ACK_TYPE, GLOBALLY);
    }

-    NV_PUSH_4U(C56F, MEM_OP_A, HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS) |
+    if (membar == UVM_MEMBAR_SYS)
+        sysmembar_value = HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, EN);
+    else
+        sysmembar_value = HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS);
+
+    NV_PUSH_4U(C56F, MEM_OP_A, sysmembar_value |
                               HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_INVAL_SCOPE, NON_LINK_TLBS),
                     MEM_OP_B, 0,
                     MEM_OP_C, HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_PDB, ONE) |
@@ -255,16 +261,18 @@ void uvm_hal_ampere_host_tlb_invalidate_all(uvm_push_t *push,
                     MEM_OP_D, HWCONST(C56F, MEM_OP_D, OPERATION, MMU_TLB_INVALIDATE) |
                               HWVALUE(C56F, MEM_OP_D, TLB_INVALIDATE_PDB_ADDR_HI, pdb_hi));

-    uvm_hal_tlb_invalidate_membar(push, membar);
+    // GPU membar still requires an explicit membar method.
+    if (membar == UVM_MEMBAR_GPU)
+        uvm_push_get_gpu(push)->parent->host_hal->membar_gpu(push);
 }

-// Copy from Volta, this version sets TLB_INVALIDATE_INVAL_SCOPE.
+// Copy from Turing, this version sets TLB_INVALIDATE_INVAL_SCOPE.
 void uvm_hal_ampere_host_tlb_invalidate_va(uvm_push_t *push,
                                           uvm_gpu_phys_address_t pdb,
                                           NvU32 depth,
                                           NvU64 base,
                                           NvU64 size,
-                                           NvU32 page_size,
+                                           NvU64 page_size,
                                           uvm_membar_t membar)
 {
    NvU32 aperture_value;
@@ -272,6 +280,7 @@ void uvm_hal_ampere_host_tlb_invalidate_va(uvm_push_t *push,
    NvU32 pdb_lo;
    NvU32 pdb_hi;
    NvU32 ack_value = 0;
+    NvU32 sysmembar_value = 0;
    NvU32 va_lo;
    NvU32 va_hi;
    NvU64 end;
@@ -281,9 +290,9 @@ void uvm_hal_ampere_host_tlb_invalidate_va(uvm_push_t *push,
    NvU32 log2_invalidation_size;
    uvm_gpu_t *gpu = uvm_push_get_gpu(push);

-    UVM_ASSERT_MSG(IS_ALIGNED(page_size, 1 << 12), "page_size 0x%x\n", page_size);
-    UVM_ASSERT_MSG(IS_ALIGNED(base, page_size), "base 0x%llx page_size 0x%x\n", base, page_size);
-    UVM_ASSERT_MSG(IS_ALIGNED(size, page_size), "size 0x%llx page_size 0x%x\n", size, page_size);
+    UVM_ASSERT_MSG(IS_ALIGNED(page_size, 1 << 12), "page_size 0x%llx\n", page_size);
+    UVM_ASSERT_MSG(IS_ALIGNED(base, page_size), "base 0x%llx page_size 0x%llx\n", base, page_size);
+    UVM_ASSERT_MSG(IS_ALIGNED(size, page_size), "size 0x%llx page_size 0x%llx\n", size, page_size);
    UVM_ASSERT_MSG(size > 0, "size 0x%llx\n", size);

    // The invalidation size must be a power-of-two number of pages containing
@@ -325,7 +334,7 @@ void uvm_hal_ampere_host_tlb_invalidate_va(uvm_push_t *push,
    pdb_lo = pdb.address & HWMASK(C56F, MEM_OP_C, TLB_INVALIDATE_PDB_ADDR_LO);
    pdb_hi = pdb.address >> HWSIZE(C56F, MEM_OP_C, TLB_INVALIDATE_PDB_ADDR_LO);

-    // PDE3 is the highest level on Pascal-Ampere , see the comment in
+    // PDE3 is the highest level on Pascal-Ampere, see the comment in
    // uvm_pascal_mmu.c for details.
    UVM_ASSERT_MSG(depth < NVC56F_MEM_OP_C_TLB_INVALIDATE_PAGE_TABLE_LEVEL_UP_TO_PDE3, "depth %u", depth);
    page_table_level = NVC56F_MEM_OP_C_TLB_INVALIDATE_PAGE_TABLE_LEVEL_UP_TO_PDE3 - depth;
@@ -337,10 +346,15 @@ void uvm_hal_ampere_host_tlb_invalidate_va(uvm_push_t *push,
        ack_value = HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_ACK_TYPE, GLOBALLY);
    }

+    if (membar == UVM_MEMBAR_SYS)
+        sysmembar_value = HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, EN);
+    else
+        sysmembar_value = HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS);
+
    NV_PUSH_4U(C56F, MEM_OP_A, HWVALUE(C56F, MEM_OP_A, TLB_INVALIDATE_INVALIDATION_SIZE, log2_invalidation_size) |
-                               HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS) |
-                               HWVALUE(C56F, MEM_OP_A, TLB_INVALIDATE_TARGET_ADDR_LO, va_lo) |
-                               HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_INVAL_SCOPE, NON_LINK_TLBS),
+                               HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_INVAL_SCOPE, NON_LINK_TLBS) |
+                               sysmembar_value |
+                               HWVALUE(C56F, MEM_OP_A, TLB_INVALIDATE_TARGET_ADDR_LO, va_lo),
                     MEM_OP_B, HWVALUE(C56F, MEM_OP_B, TLB_INVALIDATE_TARGET_ADDR_HI, va_hi),
                     MEM_OP_C, HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_PDB, ONE) |
                               HWVALUE(C56F, MEM_OP_C, TLB_INVALIDATE_PDB_ADDR_LO, pdb_lo) |
@@ -352,21 +366,23 @@ void uvm_hal_ampere_host_tlb_invalidate_va(uvm_push_t *push,
                     MEM_OP_D, HWCONST(C56F, MEM_OP_D, OPERATION, MMU_TLB_INVALIDATE_TARGETED) |
                               HWVALUE(C56F, MEM_OP_D, TLB_INVALIDATE_PDB_ADDR_HI, pdb_hi));

-    uvm_hal_tlb_invalidate_membar(push, membar);
+    // GPU membar still requires an explicit membar method.
+    if (membar == UVM_MEMBAR_GPU)
+        gpu->parent->host_hal->membar_gpu(push);
 }

-// Copy from Pascal, this version sets TLB_INVALIDATE_INVAL_SCOPE.
+// Copy from Turing, this version sets TLB_INVALIDATE_INVAL_SCOPE.
 void uvm_hal_ampere_host_tlb_invalidate_test(uvm_push_t *push,
                                             uvm_gpu_phys_address_t pdb,
                                             UVM_TEST_INVALIDATE_TLB_PARAMS *params)
 {
    NvU32 ack_value = 0;
+    NvU32 sysmembar_value = 0;
    NvU32 invalidate_gpc_value = 0;
    NvU32 aperture_value = 0;
    NvU32 pdb_lo = 0;
    NvU32 pdb_hi = 0;
    NvU32 page_table_level = 0;
-    uvm_membar_t membar;

    UVM_ASSERT_MSG(pdb.aperture == UVM_APERTURE_VID || pdb.aperture == UVM_APERTURE_SYS, "aperture: %u", pdb.aperture);
    if (pdb.aperture == UVM_APERTURE_VID)
@@ -381,7 +397,7 @@ void uvm_hal_ampere_host_tlb_invalidate_test(uvm_push_t *push,
    pdb_hi = pdb.address >> HWSIZE(C56F, MEM_OP_C, TLB_INVALIDATE_PDB_ADDR_LO);

    if (params->page_table_level != UvmInvalidatePageTableLevelAll) {
-        // PDE3 is the highest level on Pascal, see the comment in
+        // PDE3 is the highest level on Pascal-Ampere, see the comment in
        // uvm_pascal_mmu.c for details.
        page_table_level = min((NvU32)UvmInvalidatePageTableLevelPde3, params->page_table_level) - 1;
    }
@@ -393,6 +409,11 @@ void uvm_hal_ampere_host_tlb_invalidate_test(uvm_push_t *push,
        ack_value = HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_ACK_TYPE, GLOBALLY);
    }

+    if (params->membar == UvmInvalidateTlbMemBarSys)
+        sysmembar_value = HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, EN);
+    else
+        sysmembar_value = HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS);
+
    if (params->disable_gpc_invalidate)
        invalidate_gpc_value = HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_GPC, DISABLE);
    else
@@ -403,9 +424,9 @@ void uvm_hal_ampere_host_tlb_invalidate_test(uvm_push_t *push,

        NvU32 va_lo = va & HWMASK(C56F, MEM_OP_A, TLB_INVALIDATE_TARGET_ADDR_LO);
        NvU32 va_hi = va >> HWSIZE(C56F, MEM_OP_A, TLB_INVALIDATE_TARGET_ADDR_LO);
-        NV_PUSH_4U(C56F, MEM_OP_A, HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS) |
-                                   HWVALUE(C56F, MEM_OP_A, TLB_INVALIDATE_TARGET_ADDR_LO, va_lo) |
-                                   HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_INVAL_SCOPE, NON_LINK_TLBS),
+        NV_PUSH_4U(C56F, MEM_OP_A, sysmembar_value |
+                                   HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_INVAL_SCOPE, NON_LINK_TLBS) |
+                                   HWVALUE(C56F, MEM_OP_A, TLB_INVALIDATE_TARGET_ADDR_LO, va_lo),
                         MEM_OP_B, HWVALUE(C56F, MEM_OP_B, TLB_INVALIDATE_TARGET_ADDR_HI, va_hi),
                         MEM_OP_C, HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_REPLAY, NONE) |
                                   HWVALUE(C56F, MEM_OP_C, TLB_INVALIDATE_PAGE_TABLE_LEVEL, page_table_level) |
@@ -418,7 +439,7 @@ void uvm_hal_ampere_host_tlb_invalidate_test(uvm_push_t *push,
                                   HWVALUE(C56F, MEM_OP_D, TLB_INVALIDATE_PDB_ADDR_HI, pdb_hi));
    }
    else {
-        NV_PUSH_4U(C56F, MEM_OP_A, HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_SYSMEMBAR, DIS) |
+        NV_PUSH_4U(C56F, MEM_OP_A, sysmembar_value |
                                   HWCONST(C56F, MEM_OP_A, TLB_INVALIDATE_INVAL_SCOPE, NON_LINK_TLBS),
                         MEM_OP_B, 0,
                         MEM_OP_C, HWCONST(C56F, MEM_OP_C, TLB_INVALIDATE_REPLAY, NONE) |
@@ -432,12 +453,7 @@ void uvm_hal_ampere_host_tlb_invalidate_test(uvm_push_t *push,
                                   HWVALUE(C56F, MEM_OP_D, TLB_INVALIDATE_PDB_ADDR_HI, pdb_hi));
    }

-    if (params->membar == UvmInvalidateTlbMemBarSys)
-        membar = UVM_MEMBAR_SYS;
-    else if (params->membar == UvmInvalidateTlbMemBarLocal)
-        membar = UVM_MEMBAR_GPU;
-    else
-        membar = UVM_MEMBAR_NONE;
-
-    uvm_hal_tlb_invalidate_membar(push, membar);
+    // GPU membar still requires an explicit membar method.
+    if (params->membar == UvmInvalidateTlbMemBarLocal)
+        uvm_push_get_gpu(push)->parent->host_hal->membar_gpu(push);
 }
--- a/kernel-open/nvidia-uvm/uvm_ampere_mmu.c
+++ b/kernel-open/nvidia-uvm/uvm_ampere_mmu.c
@@ -51,7 +51,7 @@ uvm_mmu_engine_type_t uvm_hal_ampere_mmu_engine_id_to_type(NvU16 mmu_engine_id)
    return UVM_MMU_ENGINE_TYPE_GRAPHICS;
 }

-static NvU32 page_table_depth_ampere(NvU32 page_size)
+static NvU32 page_table_depth_ampere(NvU64 page_size)
 {
    // The common-case is page_size == UVM_PAGE_SIZE_2M, hence the first check
    if (page_size == UVM_PAGE_SIZE_2M)
@@ -62,14 +62,14 @@ static NvU32 page_table_depth_ampere(NvU32 page_size)
        return 4;
 }

-static NvU32 page_sizes_ampere(void)
+static NvU64 page_sizes_ampere(void)
 {
    return UVM_PAGE_SIZE_512M | UVM_PAGE_SIZE_2M | UVM_PAGE_SIZE_64K | UVM_PAGE_SIZE_4K;
 }

 static uvm_mmu_mode_hal_t ampere_mmu_mode_hal;

-uvm_mmu_mode_hal_t *uvm_hal_mmu_mode_ampere(NvU32 big_page_size)
+uvm_mmu_mode_hal_t *uvm_hal_mmu_mode_ampere(NvU64 big_page_size)
 {
    static bool initialized = false;

--- a/kernel-open/nvidia-uvm/uvm_ats.c
+++ b/kernel-open/nvidia-uvm/uvm_ats.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2018-2021 NVIDIA Corporation
+    Copyright (c) 2018-2024 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
--- a/kernel-open/nvidia-uvm/uvm_ats.h
+++ b/kernel-open/nvidia-uvm/uvm_ats.h
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2018-2021 NVIDIA Corporation
+    Copyright (c) 2018-2024 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -29,20 +29,9 @@
 #include "uvm_ats_ibm.h"
 #include "nv_uvm_types.h"
 #include "uvm_lock.h"
+#include "uvm_ats_sva.h"

-    #include "uvm_ats_sva.h"
-
-    #define UVM_ATS_SUPPORTED() (UVM_ATS_IBM_SUPPORTED() || UVM_ATS_SVA_SUPPORTED())
-
-// ATS prefetcher uses hmm_range_fault() to query residency information.
-// hmm_range_fault() needs CONFIG_HMM_MIRROR. To detect racing CPU invalidates
-// of memory regions while hmm_range_fault() is being called, MMU interval
-// notifiers are needed.
-    #if defined(CONFIG_HMM_MIRROR) && defined(NV_MMU_INTERVAL_NOTIFIER)
-        #define UVM_ATS_PREFETCH_SUPPORTED() 1
-    #else
-        #define UVM_ATS_PREFETCH_SUPPORTED() 0
-    #endif
+#define UVM_ATS_SUPPORTED() (UVM_ATS_IBM_SUPPORTED() || UVM_ATS_SVA_SUPPORTED())

 typedef struct
 {
--- a/kernel-open/nvidia-uvm/uvm_ats_faults.c
+++ b/kernel-open/nvidia-uvm/uvm_ats_faults.c
@@ -30,23 +30,36 @@
 #include <linux/mempolicy.h>
 #include <linux/mmu_notifier.h>

-#if UVM_ATS_PREFETCH_SUPPORTED()
+#if UVM_HMM_RANGE_FAULT_SUPPORTED()
 #include <linux/hmm.h>
 #endif

-static NV_STATUS service_ats_faults(uvm_gpu_va_space_t *gpu_va_space,
-                                    struct vm_area_struct *vma,
-                                    NvU64 start,
-                                    size_t length,
-                                    uvm_fault_access_type_t access_type,
-                                    uvm_ats_fault_context_t *ats_context)
+typedef enum
+{
+    UVM_ATS_SERVICE_TYPE_FAULTS = 0,
+    UVM_ATS_SERVICE_TYPE_ACCESS_COUNTERS,
+    UVM_ATS_SERVICE_TYPE_COUNT
+} uvm_ats_service_type_t;
+
+static NV_STATUS service_ats_requests(uvm_gpu_va_space_t *gpu_va_space,
+                                      struct vm_area_struct *vma,
+                                      NvU64 start,
+                                      size_t length,
+                                      uvm_fault_access_type_t access_type,
+                                      uvm_ats_service_type_t service_type,
+                                      uvm_ats_fault_context_t *ats_context)
 {
    uvm_va_space_t *va_space = gpu_va_space->va_space;
    struct mm_struct *mm = va_space->va_space_mm.mm;
-    bool write = (access_type >= UVM_FAULT_ACCESS_TYPE_WRITE);
    NV_STATUS status;
    NvU64 user_space_start;
    NvU64 user_space_length;
+    bool write = (access_type >= UVM_FAULT_ACCESS_TYPE_WRITE);
+    bool fault_service_type = (service_type == UVM_ATS_SERVICE_TYPE_FAULTS);
+    uvm_populate_permissions_t populate_permissions = fault_service_type ?
+                                            (write ? UVM_POPULATE_PERMISSIONS_WRITE : UVM_POPULATE_PERMISSIONS_ANY) :
+                                            UVM_POPULATE_PERMISSIONS_INHERIT;
+

    // Request uvm_migrate_pageable() to touch the corresponding page after
    // population.
@@ -83,10 +96,10 @@ static NV_STATUS service_ats_faults(uvm_gpu_va_space_t *gpu_va_space,
        .dst_node_id                    = ats_context->residency_node,
        .start                          = start,
        .length                         = length,
-        .populate_permissions           = write ? UVM_POPULATE_PERMISSIONS_WRITE : UVM_POPULATE_PERMISSIONS_ANY,
-        .touch                          = true,
-        .skip_mapped                    = true,
-        .populate_on_cpu_alloc_failures = true,
+        .populate_permissions           = populate_permissions,
+        .touch                          = fault_service_type,
+        .skip_mapped                    = fault_service_type,
+        .populate_on_cpu_alloc_failures = fault_service_type,
        .user_space_start               = &user_space_start,
        .user_space_length              = &user_space_length,
    };
@@ -107,10 +120,10 @@ static NV_STATUS service_ats_faults(uvm_gpu_va_space_t *gpu_va_space,
    return status;
 }

-static void flush_tlb_write_faults(uvm_gpu_va_space_t *gpu_va_space,
-                                   NvU64 addr,
-                                   size_t size,
-                                   uvm_fault_client_type_t client_type)
+static void flush_tlb_va_region(uvm_gpu_va_space_t *gpu_va_space,
+                                NvU64 addr,
+                                size_t size,
+                                uvm_fault_client_type_t client_type)
 {
    uvm_ats_fault_invalidate_t *ats_invalidate;

@@ -119,12 +132,12 @@ static void flush_tlb_write_faults(uvm_gpu_va_space_t *gpu_va_space,
    else
        ats_invalidate = &gpu_va_space->gpu->parent->fault_buffer_info.non_replayable.ats_invalidate;

-    if (!ats_invalidate->write_faults_in_batch) {
-        uvm_tlb_batch_begin(&gpu_va_space->page_tables, &ats_invalidate->write_faults_tlb_batch);
-        ats_invalidate->write_faults_in_batch = true;
+    if (!ats_invalidate->tlb_batch_pending) {
+        uvm_tlb_batch_begin(&gpu_va_space->page_tables, &ats_invalidate->tlb_batch);
+        ats_invalidate->tlb_batch_pending = true;
    }

-    uvm_tlb_batch_invalidate(&ats_invalidate->write_faults_tlb_batch, addr, size, PAGE_SIZE, UVM_MEMBAR_NONE);
+    uvm_tlb_batch_invalidate(&ats_invalidate->tlb_batch, addr, size, PAGE_SIZE, UVM_MEMBAR_NONE);
 }

 static void ats_batch_select_residency(uvm_gpu_va_space_t *gpu_va_space,
@@ -190,7 +203,7 @@ done:
    ats_context->prefetch_state.has_preferred_location = false;
 #endif

-    ats_context->residency_id = gpu ? gpu->parent->id : UVM_ID_CPU;
+    ats_context->residency_id = gpu ? gpu->id : UVM_ID_CPU;
    ats_context->residency_node = residency;
 }

@@ -233,7 +246,7 @@ static uvm_va_block_region_t uvm_ats_region_from_vma(struct vm_area_struct *vma,
    return uvm_ats_region_from_start_end(start, end);
 }

-#if UVM_ATS_PREFETCH_SUPPORTED()
+#if UVM_HMM_RANGE_FAULT_SUPPORTED()

 static bool uvm_ats_invalidate_notifier(struct mmu_interval_notifier *mni, unsigned long cur_seq)
 {
@@ -271,12 +284,12 @@ static NV_STATUS ats_compute_residency_mask(uvm_gpu_va_space_t *gpu_va_space,
                                            uvm_ats_fault_context_t *ats_context)
 {
    NV_STATUS status = NV_OK;
+    uvm_page_mask_t *residency_mask = &ats_context->prefetch_state.residency_mask;

-#if UVM_ATS_PREFETCH_SUPPORTED()
+#if UVM_HMM_RANGE_FAULT_SUPPORTED()
    int ret;
    NvU64 start;
    NvU64 end;
-    uvm_page_mask_t *residency_mask = &ats_context->prefetch_state.residency_mask;
    struct hmm_range range;
    uvm_page_index_t page_index;
    uvm_va_block_region_t vma_region;
@@ -357,78 +370,83 @@ static NV_STATUS ats_compute_residency_mask(uvm_gpu_va_space_t *gpu_va_space,

    mmu_interval_notifier_remove(range.notifier);

+#else
+    uvm_page_mask_zero(residency_mask);
 #endif

    return status;
 }

-static void ats_expand_fault_region(uvm_gpu_va_space_t *gpu_va_space,
-                                    struct vm_area_struct *vma,
-                                    uvm_ats_fault_context_t *ats_context,
-                                    uvm_va_block_region_t max_prefetch_region,
-                                    uvm_page_mask_t *faulted_mask)
+static void ats_compute_prefetch_mask(uvm_gpu_va_space_t *gpu_va_space,
+                                      struct vm_area_struct *vma,
+                                      uvm_ats_fault_context_t *ats_context,
+                                      uvm_va_block_region_t max_prefetch_region)
 {
-    uvm_page_mask_t *read_fault_mask = &ats_context->read_fault_mask;
-    uvm_page_mask_t *write_fault_mask = &ats_context->write_fault_mask;
+    uvm_page_mask_t *accessed_mask = &ats_context->accessed_mask;
    uvm_page_mask_t *residency_mask = &ats_context->prefetch_state.residency_mask;
    uvm_page_mask_t *prefetch_mask = &ats_context->prefetch_state.prefetch_pages_mask;
    uvm_perf_prefetch_bitmap_tree_t *bitmap_tree = &ats_context->prefetch_state.bitmap_tree;

-    if (uvm_page_mask_empty(faulted_mask))
+    if (uvm_page_mask_empty(accessed_mask))
        return;

    uvm_perf_prefetch_compute_ats(gpu_va_space->va_space,
-                                  faulted_mask,
-                                  uvm_va_block_region_from_mask(NULL, faulted_mask),
+                                  accessed_mask,
+                                  uvm_va_block_region_from_mask(NULL, accessed_mask),
                                  max_prefetch_region,
                                  residency_mask,
                                  bitmap_tree,
                                  prefetch_mask);
-
-    uvm_page_mask_or(read_fault_mask, read_fault_mask, prefetch_mask);
-
-    if (vma->vm_flags & VM_WRITE)
-        uvm_page_mask_or(write_fault_mask, write_fault_mask, prefetch_mask);
 }

-static NV_STATUS ats_fault_prefetch(uvm_gpu_va_space_t *gpu_va_space,
-                                    struct vm_area_struct *vma,
-                                    NvU64 base,
-                                    uvm_ats_fault_context_t *ats_context)
+static NV_STATUS ats_compute_prefetch(uvm_gpu_va_space_t *gpu_va_space,
+                                      struct vm_area_struct *vma,
+                                      NvU64 base,
+                                      uvm_ats_service_type_t service_type,
+                                      uvm_ats_fault_context_t *ats_context)
 {
-    NV_STATUS status = NV_OK;
-    uvm_page_mask_t *read_fault_mask = &ats_context->read_fault_mask;
-    uvm_page_mask_t *write_fault_mask = &ats_context->write_fault_mask;
-    uvm_page_mask_t *faulted_mask = &ats_context->faulted_mask;
+    NV_STATUS status;
+    uvm_page_mask_t *accessed_mask = &ats_context->accessed_mask;
    uvm_page_mask_t *prefetch_mask = &ats_context->prefetch_state.prefetch_pages_mask;
    uvm_va_block_region_t max_prefetch_region = uvm_ats_region_from_vma(vma, base);

+    // Residency mask needs to be computed even if prefetching is disabled since
+    // the residency information is also needed by access counters servicing in
+    // uvm_ats_service_access_counters()
+    status = ats_compute_residency_mask(gpu_va_space, vma, base, ats_context);
+    if (status != NV_OK)
+        return status;
+
    if (!uvm_perf_prefetch_enabled(gpu_va_space->va_space))
        return status;

-    if (uvm_page_mask_empty(faulted_mask))
-        return status;
-
-    status = ats_compute_residency_mask(gpu_va_space, vma, base, ats_context);
-    if (status != NV_OK)
+    if (uvm_page_mask_empty(accessed_mask))
        return status;

    // Prefetch the entire region if none of the pages are resident on any node
    // and if preferred_location is the faulting GPU.
    if (ats_context->prefetch_state.has_preferred_location &&
-        ats_context->prefetch_state.first_touch &&
-        uvm_id_equal(ats_context->residency_id, gpu_va_space->gpu->parent->id)) {
+        (ats_context->prefetch_state.first_touch || (service_type == UVM_ATS_SERVICE_TYPE_ACCESS_COUNTERS)) &&
+        uvm_id_equal(ats_context->residency_id, gpu_va_space->gpu->id)) {

        uvm_page_mask_init_from_region(prefetch_mask, max_prefetch_region, NULL);
+    }
+    else {
+        ats_compute_prefetch_mask(gpu_va_space, vma, ats_context, max_prefetch_region);
+    }
+
+    if (service_type == UVM_ATS_SERVICE_TYPE_FAULTS) {
+        uvm_page_mask_t *read_fault_mask = &ats_context->read_fault_mask;
+        uvm_page_mask_t *write_fault_mask = &ats_context->write_fault_mask;
+
        uvm_page_mask_or(read_fault_mask, read_fault_mask, prefetch_mask);

        if (vma->vm_flags & VM_WRITE)
            uvm_page_mask_or(write_fault_mask, write_fault_mask, prefetch_mask);
-
-        return status;
    }
-
-    ats_expand_fault_region(gpu_va_space, vma, ats_context, max_prefetch_region, faulted_mask);
+    else {
+        uvm_page_mask_or(accessed_mask, accessed_mask, prefetch_mask);
+    }

    return status;
 }
@@ -446,6 +464,7 @@ NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,
    uvm_page_mask_t *faults_serviced_mask = &ats_context->faults_serviced_mask;
    uvm_page_mask_t *reads_serviced_mask = &ats_context->reads_serviced_mask;
    uvm_fault_client_type_t client_type = ats_context->client_type;
+    uvm_ats_service_type_t service_type = UVM_ATS_SERVICE_TYPE_FAULTS;

    UVM_ASSERT(vma);
    UVM_ASSERT(IS_ALIGNED(base, UVM_VA_BLOCK_SIZE));
@@ -454,6 +473,9 @@ NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,
    UVM_ASSERT(gpu_va_space->ats.enabled);
    UVM_ASSERT(uvm_gpu_va_space_state(gpu_va_space) == UVM_GPU_VA_SPACE_STATE_ACTIVE);

+    uvm_assert_mmap_lock_locked(vma->vm_mm);
+    uvm_assert_rwsem_locked(&gpu_va_space->va_space->lock);
+
    uvm_page_mask_zero(faults_serviced_mask);
    uvm_page_mask_zero(reads_serviced_mask);

@@ -479,7 +501,7 @@ NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,

    ats_batch_select_residency(gpu_va_space, vma, ats_context);

-    ats_fault_prefetch(gpu_va_space, vma, base, ats_context);
+    ats_compute_prefetch(gpu_va_space, vma, base, service_type, ats_context);

    for_each_va_block_subregion_in_mask(subregion, write_fault_mask, region) {
        NvU64 start = base + (subregion.first * PAGE_SIZE);
@@ -491,12 +513,13 @@ NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,
        UVM_ASSERT(start >= vma->vm_start);
        UVM_ASSERT((start + length) <= vma->vm_end);

-        status = service_ats_faults(gpu_va_space, vma, start, length, access_type, ats_context);
+        status = service_ats_requests(gpu_va_space, vma, start, length, access_type, service_type, ats_context);
        if (status != NV_OK)
            return status;

        if (vma->vm_flags & VM_WRITE) {
            uvm_page_mask_region_fill(faults_serviced_mask, subregion);
+            uvm_ats_smmu_invalidate_tlbs(gpu_va_space, start, length);

            // The Linux kernel never invalidates TLB entries on mapping
            // permission upgrade. This is a problem if the GPU has cached
@@ -507,7 +530,7 @@ NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,
            // infinite loop because we just forward the fault to the Linux
            // kernel and it will see that the permissions in the page table are
            // correct. Therefore, we flush TLB entries on ATS write faults.
-            flush_tlb_write_faults(gpu_va_space, start, length, client_type);
+            flush_tlb_va_region(gpu_va_space, start, length, client_type);
        }
        else {
            uvm_page_mask_region_fill(reads_serviced_mask, subregion);
@@ -525,11 +548,20 @@ NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,
        UVM_ASSERT(start >= vma->vm_start);
        UVM_ASSERT((start + length) <= vma->vm_end);

-        status = service_ats_faults(gpu_va_space, vma, start, length, access_type, ats_context);
+        status = service_ats_requests(gpu_va_space, vma, start, length, access_type, service_type, ats_context);
        if (status != NV_OK)
            return status;

        uvm_page_mask_region_fill(faults_serviced_mask, subregion);
+
+        // Similarly to permission upgrade scenario, discussed above, GPU
+        // will not re-fetch the entry if the PTE is invalid and page size
+        // is 4K. To avoid infinite faulting loop, invalidate TLB for every
+        // new translation written explicitly like in the case of permission
+        // upgrade.
+        if (PAGE_SIZE == UVM_PAGE_SIZE_4K)
+            flush_tlb_va_region(gpu_va_space, start, length, client_type);
+
    }

    return status;
@@ -564,7 +596,7 @@ NV_STATUS uvm_ats_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space,
    NV_STATUS status;
    uvm_push_t push;

-    if (!ats_invalidate->write_faults_in_batch)
+    if (!ats_invalidate->tlb_batch_pending)
        return NV_OK;

    UVM_ASSERT(gpu_va_space);
@@ -576,7 +608,7 @@ NV_STATUS uvm_ats_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space,
                            "Invalidate ATS entries");

    if (status == NV_OK) {
-        uvm_tlb_batch_end(&ats_invalidate->write_faults_tlb_batch, &push, UVM_MEMBAR_NONE);
+        uvm_tlb_batch_end(&ats_invalidate->tlb_batch, &push, UVM_MEMBAR_NONE);
        uvm_push_end(&push);

        // Add this push to the GPU's tracker so that fault replays/clears can
@@ -584,8 +616,57 @@ NV_STATUS uvm_ats_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space,
        status = uvm_tracker_add_push_safe(out_tracker, &push);
    }

-    ats_invalidate->write_faults_in_batch = false;
+    ats_invalidate->tlb_batch_pending = false;

    return status;
 }

+NV_STATUS uvm_ats_service_access_counters(uvm_gpu_va_space_t *gpu_va_space,
+                                          struct vm_area_struct *vma,
+                                          NvU64 base,
+                                          uvm_ats_fault_context_t *ats_context)
+{
+    uvm_va_block_region_t subregion;
+    uvm_va_block_region_t region = uvm_va_block_region(0, PAGES_PER_UVM_VA_BLOCK);
+    uvm_ats_service_type_t service_type = UVM_ATS_SERVICE_TYPE_ACCESS_COUNTERS;
+
+    UVM_ASSERT(vma);
+    UVM_ASSERT(IS_ALIGNED(base, UVM_VA_BLOCK_SIZE));
+    UVM_ASSERT(g_uvm_global.ats.enabled);
+    UVM_ASSERT(gpu_va_space);
+    UVM_ASSERT(gpu_va_space->ats.enabled);
+    UVM_ASSERT(uvm_gpu_va_space_state(gpu_va_space) == UVM_GPU_VA_SPACE_STATE_ACTIVE);
+
+    uvm_assert_mmap_lock_locked(vma->vm_mm);
+    uvm_assert_rwsem_locked(&gpu_va_space->va_space->lock);
+
+    ats_batch_select_residency(gpu_va_space, vma, ats_context);
+
+    // Ignoring the return value of ats_compute_prefetch is ok since prefetching
+    // is just an optimization and servicing access counter migrations is still
+    // worthwhile even without any prefetching added. So, let servicing continue
+    // instead of returning early even if the prefetch computation fails.
+    ats_compute_prefetch(gpu_va_space, vma, base, service_type, ats_context);
+
+    // Remove pages which are already resident at the intended destination from
+    // the accessed_mask.
+    uvm_page_mask_andnot(&ats_context->accessed_mask,
+                         &ats_context->accessed_mask,
+                         &ats_context->prefetch_state.residency_mask);
+
+    for_each_va_block_subregion_in_mask(subregion, &ats_context->accessed_mask, region) {
+        NV_STATUS status;
+        NvU64 start = base + (subregion.first * PAGE_SIZE);
+        size_t length = uvm_va_block_region_num_pages(subregion) * PAGE_SIZE;
+        uvm_fault_access_type_t access_type = UVM_FAULT_ACCESS_TYPE_COUNT;
+
+        UVM_ASSERT(start >= vma->vm_start);
+        UVM_ASSERT((start + length) <= vma->vm_end);
+
+        status = service_ats_requests(gpu_va_space, vma, start, length, access_type, service_type, ats_context);
+        if (status != NV_OK)
+            return status;
+    }
+
+    return NV_OK;
+}
--- a/kernel-open/nvidia-uvm/uvm_ats_faults.h
+++ b/kernel-open/nvidia-uvm/uvm_ats_faults.h
@@ -42,17 +42,37 @@
 // corresponding bit in read_fault_mask. These returned masks are only valid if
 // the return status is NV_OK. Status other than NV_OK indicate system global
 // fault servicing failures.
+//
+// LOCKING: The caller must retain and hold the mmap_lock and hold the va_space
+// lock.
 NV_STATUS uvm_ats_service_faults(uvm_gpu_va_space_t *gpu_va_space,
                                 struct vm_area_struct *vma,
                                 NvU64 base,
                                 uvm_ats_fault_context_t *ats_context);

+// Service access counter notifications on ATS regions in the range (base, base
+// + UVM_VA_BLOCK_SIZE) for individual pages in the range requested by page_mask
+// set in ats_context->accessed_mask. base must be aligned to UVM_VA_BLOCK_SIZE.
+// The caller is responsible for ensuring that the addresses in the
+// accessed_mask is completely covered by the VMA. The caller is also
+// responsible for handling any errors returned by this function.
+//
+// Returns NV_OK if servicing was successful. Any other error indicates an error
+// while servicing the range.
+//
+// LOCKING: The caller must retain and hold the mmap_lock and hold the va_space
+// lock.
+NV_STATUS uvm_ats_service_access_counters(uvm_gpu_va_space_t *gpu_va_space,
+                                          struct vm_area_struct *vma,
+                                          NvU64 base,
+                                          uvm_ats_fault_context_t *ats_context);
+
 // Return whether there are any VA ranges (and thus GMMU mappings) within the
 // UVM_GMMU_ATS_GRANULARITY-aligned region containing address.
 bool uvm_ats_check_in_gmmu_region(uvm_va_space_t *va_space, NvU64 address, uvm_va_range_t *next);

 // This function performs pending TLB invalidations for ATS and clears the
-// ats_invalidate->write_faults_in_batch flag
+// ats_invalidate->tlb_batch_pending flag
 NV_STATUS uvm_ats_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space,
                                  uvm_ats_fault_invalidate_t *ats_invalidate,
                                  uvm_tracker_t *out_tracker);
--- a/kernel-open/nvidia-uvm/uvm_ats_sva.c
+++ b/kernel-open/nvidia-uvm/uvm_ats_sva.c
@@ -29,8 +29,13 @@
 #include "uvm_va_space.h"
 #include "uvm_va_space_mm.h"

+#include <asm/io.h>
+#include <linux/log2.h>
 #include <linux/iommu.h>
 #include <linux/mm_types.h>
+#include <linux/acpi.h>
+#include <linux/device.h>
+#include <linux/mmu_context.h>

 // linux/sched/mm.h is needed for mmget_not_zero and mmput to get the mm
 // reference required for the iommu_sva_bind_device() call. This header is not
@@ -46,17 +51,276 @@
 #define UVM_IOMMU_SVA_BIND_DEVICE(dev, mm) iommu_sva_bind_device(dev, mm)
 #endif

+// Type to represent a 128-bit SMMU command queue command.
+struct smmu_cmd {
+    NvU64 low;
+    NvU64 high;
+};
+
+// Base address of SMMU CMDQ-V for GSMMU0.
+#define SMMU_CMDQV_BASE_ADDR(smmu_base) (smmu_base + 0x200000)
+#define SMMU_CMDQV_BASE_LEN 0x00830000
+
+// CMDQV configuration is done by firmware but we check status here.
+#define SMMU_CMDQV_CONFIG 0x0
+#define SMMU_CMDQV_CONFIG_CMDQV_EN BIT(0)
+
+// Used to map a particular VCMDQ to a VINTF.
+#define SMMU_CMDQV_CMDQ_ALLOC_MAP(vcmdq_id) (0x200 + 0x4 * (vcmdq_id))
+#define SMMU_CMDQV_CMDQ_ALLOC_MAP_ALLOC BIT(0)
+
+// Shift for the field containing the index of the virtual interface
+// owning the VCMDQ.
+#define SMMU_CMDQV_CMDQ_ALLOC_MAP_VIRT_INTF_INDX_SHIFT 15
+
+// Base address for the VINTF registers.
+#define SMMU_VINTF_BASE_ADDR(cmdqv_base_addr, vintf_id) (cmdqv_base_addr + 0x1000 + 0x100 * (vintf_id))
+
+// Virtual interface (VINTF) configuration registers. The WAR only
+// works on baremetal so we need to configure ourselves as the
+// hypervisor owner.
+#define SMMU_VINTF_CONFIG 0x0
+#define SMMU_VINTF_CONFIG_ENABLE BIT(0)
+#define SMMU_VINTF_CONFIG_HYP_OWN BIT(17)
+
+#define SMMU_VINTF_STATUS 0x0
+#define SMMU_VINTF_STATUS_ENABLED BIT(0)
+
+// Caclulates the base address for a particular VCMDQ instance.
+#define SMMU_VCMDQ_BASE_ADDR(cmdqv_base_addr, vcmdq_id) (cmdqv_base_addr + 0x10000 + 0x80 * (vcmdq_id))
+
+// SMMU command queue consumer index register. Updated by SMMU
+// when commands are consumed.
+#define SMMU_VCMDQ_CONS 0x0
+
+// SMMU command queue producer index register. Updated by UVM when
+// commands are added to the queue.
+#define SMMU_VCMDQ_PROD 0x4
+
+// Configuration register used to enable a VCMDQ.
+#define SMMU_VCMDQ_CONFIG 0x8
+#define SMMU_VCMDQ_CONFIG_ENABLE BIT(0)
+
+// Status register used to check the VCMDQ is enabled.
+#define SMMU_VCMDQ_STATUS 0xc
+#define SMMU_VCMDQ_STATUS_ENABLED BIT(0)
+
+// Base address offset for the VCMDQ registers.
+#define SMMU_VCMDQ_CMDQ_BASE 0x10000
+
+// Size of the command queue. Each command is 16 bytes and we can't
+// have a command queue greater than one page in size.
+#define SMMU_VCMDQ_CMDQ_BASE_LOG2SIZE (PAGE_SHIFT - ilog2(sizeof(struct smmu_cmd)))
+#define SMMU_VCMDQ_CMDQ_ENTRIES (1UL << SMMU_VCMDQ_CMDQ_BASE_LOG2SIZE)
+
+// We always use VINTF63 for the WAR
+#define VINTF 63
+static void smmu_vintf_write32(void __iomem *smmu_cmdqv_base, int reg, NvU32 val)
+{
+    iowrite32(val, SMMU_VINTF_BASE_ADDR(smmu_cmdqv_base, VINTF) + reg);
+}
+
+static NvU32 smmu_vintf_read32(void __iomem *smmu_cmdqv_base, int reg)
+{
+    return ioread32(SMMU_VINTF_BASE_ADDR(smmu_cmdqv_base, VINTF) + reg);
+}
+
+// We always use VCMDQ127 for the WAR
+#define VCMDQ 127
+void smmu_vcmdq_write32(void __iomem *smmu_cmdqv_base, int reg, NvU32 val)
+{
+    iowrite32(val, SMMU_VCMDQ_BASE_ADDR(smmu_cmdqv_base, VCMDQ) + reg);
+}
+
+NvU32 smmu_vcmdq_read32(void __iomem *smmu_cmdqv_base, int reg)
+{
+    return ioread32(SMMU_VCMDQ_BASE_ADDR(smmu_cmdqv_base, VCMDQ) + reg);
+}
+
+static void smmu_vcmdq_write64(void __iomem *smmu_cmdqv_base, int reg, NvU64 val)
+{
+    iowrite64(val, SMMU_VCMDQ_BASE_ADDR(smmu_cmdqv_base, VCMDQ) + reg);
+}
+
+// Fix for Bug 4130089: [GH180][r535] WAR for kernel not issuing SMMU
+// TLB invalidates on read-only to read-write upgrades
+static NV_STATUS uvm_ats_smmu_war_init(uvm_parent_gpu_t *parent_gpu)
+{
+    uvm_spin_loop_t spin;
+    NV_STATUS status;
+    unsigned long cmdqv_config;
+    void __iomem *smmu_cmdqv_base;
+    struct acpi_iort_node *node;
+    struct acpi_iort_smmu_v3 *iort_smmu;
+
+    node = *(struct acpi_iort_node **) dev_get_platdata(parent_gpu->pci_dev->dev.iommu->iommu_dev->dev->parent);
+    iort_smmu = (struct acpi_iort_smmu_v3 *) node->node_data;
+
+    smmu_cmdqv_base = ioremap(SMMU_CMDQV_BASE_ADDR(iort_smmu->base_address), SMMU_CMDQV_BASE_LEN);
+    if (!smmu_cmdqv_base)
+        return NV_ERR_NO_MEMORY;
+
+    parent_gpu->smmu_war.smmu_cmdqv_base = smmu_cmdqv_base;
+    cmdqv_config = ioread32(smmu_cmdqv_base + SMMU_CMDQV_CONFIG);
+    if (!(cmdqv_config & SMMU_CMDQV_CONFIG_CMDQV_EN)) {
+        status = NV_ERR_OBJECT_NOT_FOUND;
+        goto out;
+    }
+
+    // Allocate SMMU CMDQ pages for WAR
+    parent_gpu->smmu_war.smmu_cmdq = alloc_page(NV_UVM_GFP_FLAGS | __GFP_ZERO);
+    if (!parent_gpu->smmu_war.smmu_cmdq) {
+        status = NV_ERR_NO_MEMORY;
+        goto out;
+    }
+
+    // Initialise VINTF for the WAR
+    smmu_vintf_write32(smmu_cmdqv_base, SMMU_VINTF_CONFIG, SMMU_VINTF_CONFIG_ENABLE | SMMU_VINTF_CONFIG_HYP_OWN);
+    UVM_SPIN_WHILE(!(smmu_vintf_read32(smmu_cmdqv_base, SMMU_VINTF_STATUS) & SMMU_VINTF_STATUS_ENABLED), &spin);
+
+    // Allocate VCMDQ to VINTF
+    iowrite32((VINTF << SMMU_CMDQV_CMDQ_ALLOC_MAP_VIRT_INTF_INDX_SHIFT) | SMMU_CMDQV_CMDQ_ALLOC_MAP_ALLOC,
+              smmu_cmdqv_base + SMMU_CMDQV_CMDQ_ALLOC_MAP(VCMDQ));
+
+    smmu_vcmdq_write64(smmu_cmdqv_base, SMMU_VCMDQ_CMDQ_BASE,
+                       page_to_phys(parent_gpu->smmu_war.smmu_cmdq) | SMMU_VCMDQ_CMDQ_BASE_LOG2SIZE);
+    smmu_vcmdq_write32(smmu_cmdqv_base, SMMU_VCMDQ_CONS, 0);
+    smmu_vcmdq_write32(smmu_cmdqv_base, SMMU_VCMDQ_PROD, 0);
+    smmu_vcmdq_write32(smmu_cmdqv_base, SMMU_VCMDQ_CONFIG, SMMU_VCMDQ_CONFIG_ENABLE);
+    UVM_SPIN_WHILE(!(smmu_vcmdq_read32(smmu_cmdqv_base, SMMU_VCMDQ_STATUS) & SMMU_VCMDQ_STATUS_ENABLED), &spin);
+
+    uvm_mutex_init(&parent_gpu->smmu_war.smmu_lock, UVM_LOCK_ORDER_LEAF);
+    parent_gpu->smmu_war.smmu_prod = 0;
+    parent_gpu->smmu_war.smmu_cons = 0;
+
+    return NV_OK;
+
+out:
+    iounmap(parent_gpu->smmu_war.smmu_cmdqv_base);
+    parent_gpu->smmu_war.smmu_cmdqv_base = NULL;
+
+    return status;
+}
+
+static void uvm_ats_smmu_war_deinit(uvm_parent_gpu_t *parent_gpu)
+{
+    void __iomem *smmu_cmdqv_base = parent_gpu->smmu_war.smmu_cmdqv_base;
+    NvU32 cmdq_alloc_map;
+
+    if (parent_gpu->smmu_war.smmu_cmdqv_base) {
+        smmu_vcmdq_write32(smmu_cmdqv_base, SMMU_VCMDQ_CONFIG, 0);
+        cmdq_alloc_map = ioread32(smmu_cmdqv_base + SMMU_CMDQV_CMDQ_ALLOC_MAP(VCMDQ));
+        iowrite32(cmdq_alloc_map & SMMU_CMDQV_CMDQ_ALLOC_MAP_ALLOC, smmu_cmdqv_base + SMMU_CMDQV_CMDQ_ALLOC_MAP(VCMDQ));
+        smmu_vintf_write32(smmu_cmdqv_base, SMMU_VINTF_CONFIG, 0);
+    }
+
+    if (parent_gpu->smmu_war.smmu_cmdq)
+        __free_page(parent_gpu->smmu_war.smmu_cmdq);
+
+    if (parent_gpu->smmu_war.smmu_cmdqv_base)
+        iounmap(parent_gpu->smmu_war.smmu_cmdqv_base);
+}
+
+// The SMMU on ARM64 can run under different translation regimes depending on
+// what features the OS and CPU variant support. The CPU for GH180 supports
+// virtualisation extensions and starts the kernel at EL2 meaning SMMU operates
+// under the NS-EL2-E2H translation regime. Therefore we need to use the
+// TLBI_EL2_* commands which invalidate TLB entries created under this
+// translation regime.
+#define CMDQ_OP_TLBI_EL2_ASID 0x21;
+#define CMDQ_OP_TLBI_EL2_VA 0x22;
+#define CMDQ_OP_CMD_SYNC 0x46
+
+// Use the same maximum as used for MAX_TLBI_OPS in the upstream
+// kernel.
+#define UVM_MAX_TLBI_OPS (1UL << (PAGE_SHIFT - 3))
+
+#if UVM_ATS_SMMU_WAR_REQUIRED()
+void uvm_ats_smmu_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space, NvU64 addr, size_t size)
+{
+    struct mm_struct *mm = gpu_va_space->va_space->va_space_mm.mm;
+    uvm_parent_gpu_t *parent_gpu = gpu_va_space->gpu->parent;
+    struct {
+        NvU64 low;
+        NvU64 high;
+    } *vcmdq;
+    unsigned long vcmdq_prod;
+    NvU64 end;
+    uvm_spin_loop_t spin;
+    NvU16 asid;
+
+    if (!parent_gpu->smmu_war.smmu_cmdqv_base)
+        return;
+
+    asid = arm64_mm_context_get(mm);
+    vcmdq = kmap(parent_gpu->smmu_war.smmu_cmdq);
+    uvm_mutex_lock(&parent_gpu->smmu_war.smmu_lock);
+    vcmdq_prod = parent_gpu->smmu_war.smmu_prod;
+
+    // Our queue management is very simple. The mutex prevents multiple
+    // producers writing to the queue and all our commands require waiting for
+    // the queue to drain so we know it's empty. If we can't fit enough commands
+    // in the queue we just invalidate the whole ASID.
+    //
+    // The command queue is a cirular buffer with the MSB representing a wrap
+    // bit that must toggle on each wrap. See the SMMU architecture
+    // specification for more details.
+    //
+    // SMMU_VCMDQ_CMDQ_ENTRIES - 1 because we need to leave space for the
+    // CMD_SYNC.
+    if ((size >> PAGE_SHIFT) > min(UVM_MAX_TLBI_OPS, SMMU_VCMDQ_CMDQ_ENTRIES - 1)) {
+        vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].low = CMDQ_OP_TLBI_EL2_ASID;
+        vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].low |= (NvU64) asid << 48;
+        vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].high = 0;
+        vcmdq_prod++;
+    }
+    else {
+        for (end = addr + size; addr < end; addr += PAGE_SIZE) {
+            vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].low = CMDQ_OP_TLBI_EL2_VA;
+            vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].low |= (NvU64) asid << 48;
+            vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].high = addr & ~((1UL << 12) - 1);
+            vcmdq_prod++;
+        }
+    }
+
+    vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].low = CMDQ_OP_CMD_SYNC;
+    vcmdq[vcmdq_prod % SMMU_VCMDQ_CMDQ_ENTRIES].high = 0x0;
+    vcmdq_prod++;
+
+    // MSB is the wrap bit
+    vcmdq_prod &= (1UL << (SMMU_VCMDQ_CMDQ_BASE_LOG2SIZE + 1)) - 1;
+    parent_gpu->smmu_war.smmu_prod = vcmdq_prod;
+    smmu_vcmdq_write32(parent_gpu->smmu_war.smmu_cmdqv_base, SMMU_VCMDQ_PROD, parent_gpu->smmu_war.smmu_prod);
+
+    UVM_SPIN_WHILE(
+        (smmu_vcmdq_read32(parent_gpu->smmu_war.smmu_cmdqv_base, SMMU_VCMDQ_CONS) & GENMASK(19, 0)) != vcmdq_prod,
+        &spin);
+
+    uvm_mutex_unlock(&parent_gpu->smmu_war.smmu_lock);
+    kunmap(parent_gpu->smmu_war.smmu_cmdq);
+    arm64_mm_context_put(mm);
+}
+#endif
+
 NV_STATUS uvm_ats_sva_add_gpu(uvm_parent_gpu_t *parent_gpu)
 {
    int ret;

    ret = iommu_dev_enable_feature(&parent_gpu->pci_dev->dev, IOMMU_DEV_FEAT_SVA);
+    if (ret)
+        return errno_to_nv_status(ret);

-    return errno_to_nv_status(ret);
+    if (UVM_ATS_SMMU_WAR_REQUIRED())
+        return uvm_ats_smmu_war_init(parent_gpu);
+    else
+        return NV_OK;
 }

 void uvm_ats_sva_remove_gpu(uvm_parent_gpu_t *parent_gpu)
 {
+    if (UVM_ATS_SMMU_WAR_REQUIRED())
+        uvm_ats_smmu_war_deinit(parent_gpu);
+
    iommu_dev_disable_feature(&parent_gpu->pci_dev->dev, IOMMU_DEV_FEAT_SVA);
 }

--- a/kernel-open/nvidia-uvm/uvm_ats_sva.h
+++ b/kernel-open/nvidia-uvm/uvm_ats_sva.h
@@ -53,6 +53,18 @@
        #define UVM_ATS_SVA_SUPPORTED() 0
    #endif

+// If NV_MMU_NOTIFIER_OPS_HAS_ARCH_INVALIDATE_SECONDARY_TLBS is defined it
+// means the upstream fix is in place so no need for the WAR from
+// Bug 4130089: [GH180][r535] WAR for  kernel not issuing SMMU TLB
+// invalidates on read-only
+#if defined(NV_MMU_NOTIFIER_OPS_HAS_ARCH_INVALIDATE_SECONDARY_TLBS)
+    #define UVM_ATS_SMMU_WAR_REQUIRED() 0
+#elif NVCPU_IS_AARCH64
+    #define UVM_ATS_SMMU_WAR_REQUIRED() 1
+#else
+    #define UVM_ATS_SMMU_WAR_REQUIRED() 0
+#endif
+
 typedef struct
 {
    int placeholder;
@@ -81,6 +93,17 @@ typedef struct

    // LOCKING: None
    void uvm_ats_sva_unregister_gpu_va_space(uvm_gpu_va_space_t *gpu_va_space);
+
+    // Fix for Bug 4130089: [GH180][r535] WAR for kernel not issuing SMMU
+    // TLB invalidates on read-only to read-write upgrades
+    #if UVM_ATS_SMMU_WAR_REQUIRED()
+        void uvm_ats_smmu_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space, NvU64 addr, size_t size);
+    #else
+        static void uvm_ats_smmu_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space, NvU64 addr, size_t size)
+        {
+
+        }
+    #endif
 #else
    static NV_STATUS uvm_ats_sva_add_gpu(uvm_parent_gpu_t *parent_gpu)
    {
@@ -111,6 +134,11 @@ typedef struct
    {

    }
+
+    static void uvm_ats_smmu_invalidate_tlbs(uvm_gpu_va_space_t *gpu_va_space, NvU64 addr, size_t size)
+    {
+
+    }
 #endif // UVM_ATS_SVA_SUPPORTED

 #endif // __UVM_ATS_SVA_H__
--- a/kernel-open/nvidia-uvm/uvm_ce_test.c
+++ b/kernel-open/nvidia-uvm/uvm_ce_test.c
@@ -56,7 +56,7 @@ static NV_STATUS test_non_pipelined(uvm_gpu_t *gpu)

    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = uvm_rm_mem_alloc_and_map_cpu(gpu, UVM_RM_MEM_TYPE_SYS, CE_TEST_MEM_SIZE, 0, &host_mem);
@@ -176,7 +176,7 @@ static NV_STATUS test_membar(uvm_gpu_t *gpu)

    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = uvm_rm_mem_alloc_and_map_cpu(gpu, UVM_RM_MEM_TYPE_SYS, sizeof(NvU32), 0, &host_mem);
@@ -411,10 +411,11 @@ static NV_STATUS test_memcpy_and_memset(uvm_gpu_t *gpu)
    size_t i, j, k, s;
    uvm_mem_alloc_params_t mem_params = {0};

-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        TEST_NV_CHECK_GOTO(uvm_mem_alloc_sysmem_dma_and_map_cpu_kernel(size, gpu, current->mm, &verif_mem), done);
    else
        TEST_NV_CHECK_GOTO(uvm_mem_alloc_sysmem_and_map_cpu_kernel(size, current->mm, &verif_mem), done);
+
    TEST_NV_CHECK_GOTO(uvm_mem_map_gpu_kernel(verif_mem, gpu), done);

    gpu_verif_addr = uvm_mem_gpu_address_virtual_kernel(verif_mem, gpu);
@@ -436,7 +437,7 @@ static NV_STATUS test_memcpy_and_memset(uvm_gpu_t *gpu)
    TEST_NV_CHECK_GOTO(uvm_rm_mem_alloc(gpu, UVM_RM_MEM_TYPE_SYS, size, 0, &sys_rm_mem), done);
    gpu_addresses[0] = uvm_rm_mem_get_gpu_va(sys_rm_mem, gpu, is_proxy_va_space);

-    if (uvm_conf_computing_mode_enabled(gpu)) {
+    if (g_uvm_global.conf_computing_enabled) {
        for (i = 0; i < iterations; ++i) {
            for (s = 0; s < ARRAY_SIZE(element_sizes); s++) {
                TEST_NV_CHECK_GOTO(test_memcpy_and_memset_inner(gpu,
@@ -559,7 +560,7 @@ static NV_STATUS test_semaphore_reduction_inc(uvm_gpu_t *gpu)

    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = test_semaphore_alloc_sem(gpu, size, &mem);
@@ -611,7 +612,7 @@ static NV_STATUS test_semaphore_release(uvm_gpu_t *gpu)

    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = test_semaphore_alloc_sem(gpu, size, &mem);
@@ -665,7 +666,7 @@ static NV_STATUS test_semaphore_timestamp(uvm_gpu_t *gpu)

    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = test_semaphore_alloc_sem(gpu, size, &mem);
@@ -1153,7 +1154,7 @@ static NV_STATUS test_encryption_decryption(uvm_gpu_t *gpu,
    } small_sizes[] = {{1, 1}, {3, 1}, {8, 1}, {2, 2}, {8, 4}, {UVM_PAGE_SIZE_4K - 8, 8}, {UVM_PAGE_SIZE_4K + 8, 8}};

    // Only Confidential Computing uses CE encryption/decryption
-    if (!uvm_conf_computing_mode_enabled(gpu))
+    if (!g_uvm_global.conf_computing_enabled)
        return NV_OK;

    // Use a size, and copy size, that are not a multiple of common page sizes.
--- a/kernel-open/nvidia-uvm/uvm_channel.c
+++ b/kernel-open/nvidia-uvm/uvm_channel.c
@@ -83,7 +83,7 @@ bool uvm_channel_pool_uses_mutex(uvm_channel_pool_t *pool)
    // submission uses UVM_SPIN_LOOP, which can call 'schedule', to wait for
    // LCIC completion. Indirect submission is synchronous, calling
    // uvm_push_wait which again uses UVM_SPIN_LOOP.
-    if (uvm_conf_computing_mode_enabled(pool->manager->gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return true;

    // Unless the mutex is required, the spinlock is preferred when work
@@ -95,7 +95,7 @@ static void channel_pool_lock_init(uvm_channel_pool_t *pool)
 {
    uvm_lock_order_t order = UVM_LOCK_ORDER_CHANNEL;

-    if (uvm_conf_computing_mode_enabled(pool->manager->gpu) && uvm_channel_pool_is_wlc(pool))
+    if (g_uvm_global.conf_computing_enabled && uvm_channel_pool_is_wlc(pool))
        order = UVM_LOCK_ORDER_WLC_CHANNEL;

    if (uvm_channel_pool_uses_mutex(pool))
@@ -137,7 +137,7 @@ static NvU32 uvm_channel_update_progress_with_max(uvm_channel_t *channel,
    // Completed value should never exceed the queued value
    UVM_ASSERT_MSG_RELEASE(completed_value <= channel->tracking_sem.queued_value,
                           "GPU %s channel %s unexpected completed_value 0x%llx > queued_value 0x%llx\n",
-                           channel->pool->manager->gpu->parent->name,
+                           uvm_gpu_name(uvm_channel_get_gpu(channel)),
                           channel->name,
                           completed_value,
                           channel->tracking_sem.queued_value);
@@ -273,9 +273,8 @@ static bool try_claim_channel(uvm_channel_t *channel, NvU32 num_gpfifo_entries)
 static void unlock_channel_for_push(uvm_channel_t *channel)
 {
    NvU32 index;
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);

-    if (!uvm_conf_computing_mode_enabled(gpu))
+    if (!g_uvm_global.conf_computing_enabled)
        return;

    index = uvm_channel_index_in_pool(channel);
@@ -287,25 +286,22 @@ static void unlock_channel_for_push(uvm_channel_t *channel)
    uvm_up_out_of_order(&channel->pool->push_sem);
 }

-static bool is_channel_locked_for_push(uvm_channel_t *channel)
+bool uvm_channel_is_locked_for_push(uvm_channel_t *channel)
 {
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
-
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return test_bit(uvm_channel_index_in_pool(channel), channel->pool->push_locks);

    // For CE and proxy channels, we always return that the channel is locked,
    // which has no functional impact in the UVM channel code-flow, this is only
-    // used on UVM_ASSERTs.
+    // used in UVM_ASSERTs.
    return true;
 }

 static void lock_channel_for_push(uvm_channel_t *channel)
 {
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
    NvU32 index = uvm_channel_index_in_pool(channel);

-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
    uvm_channel_pool_assert_locked(channel->pool);
    UVM_ASSERT(!test_bit(index, channel->pool->push_locks));

@@ -314,10 +310,9 @@ static void lock_channel_for_push(uvm_channel_t *channel)

 static bool test_claim_and_lock_channel(uvm_channel_t *channel, NvU32 num_gpfifo_entries)
 {
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
    NvU32 index = uvm_channel_index_in_pool(channel);

-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
    uvm_channel_pool_assert_locked(channel->pool);

    if (!test_bit(index, channel->pool->push_locks) && try_claim_channel_locked(channel, num_gpfifo_entries)) {
@@ -337,7 +332,7 @@ static NV_STATUS channel_reserve_and_lock_in_pool(uvm_channel_pool_t *pool, uvm_
    NvU32 index;

    UVM_ASSERT(pool);
-    UVM_ASSERT(uvm_conf_computing_mode_enabled(pool->manager->gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);

    // This semaphore is uvm_up() in unlock_channel_for_push() as part of the
    // uvm_channel_end_push() routine.
@@ -399,7 +394,7 @@ static NV_STATUS channel_reserve_in_pool(uvm_channel_pool_t *pool, uvm_channel_t

    UVM_ASSERT(pool);

-    if (uvm_conf_computing_mode_enabled(pool->manager->gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return channel_reserve_and_lock_in_pool(pool, channel_out);

    uvm_for_each_channel_in_pool(channel, pool) {
@@ -509,7 +504,7 @@ static void channel_semaphore_gpu_encrypt_payload(uvm_push_t *push, NvU64 semaph
    NvU32 payload_size = sizeof(*semaphore->payload);
    NvU32 *last_pushed_notifier = &semaphore->conf_computing.last_pushed_notifier;

-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
    UVM_ASSERT(uvm_channel_is_ce(channel));

    encrypted_payload_gpu_va = uvm_rm_mem_get_gpu_va(semaphore->conf_computing.encrypted_payload, gpu, false);
@@ -540,29 +535,97 @@ static void push_reserve_csl_sign_buf(uvm_push_t *push)
    UVM_ASSERT((buf - UVM_METHOD_SIZE / sizeof(*buf)) == push->begin);
 }

+static uvm_channel_t *get_paired_channel(uvm_channel_t *channel)
+{
+    unsigned index;
+    uvm_channel_pool_t *paired_pool;
+    uvm_channel_type_t paired_channel_type;
+
+    UVM_ASSERT(channel);
+    UVM_ASSERT(uvm_channel_is_wlc(channel) || uvm_channel_is_lcic(channel));
+
+    index = uvm_channel_index_in_pool(channel);
+    paired_channel_type = uvm_channel_is_wlc(channel) ? UVM_CHANNEL_TYPE_LCIC : UVM_CHANNEL_TYPE_WLC;
+    paired_pool = channel->pool->manager->pool_to_use.default_for_type[paired_channel_type];
+    return paired_pool->channels + index;
+}
+
+uvm_channel_t *uvm_channel_lcic_get_paired_wlc(uvm_channel_t *lcic_channel)
+{
+    UVM_ASSERT(lcic_channel);
+    UVM_ASSERT(uvm_channel_is_lcic(lcic_channel));
+
+    return get_paired_channel(lcic_channel);
+}
+
+uvm_channel_t *uvm_channel_wlc_get_paired_lcic(uvm_channel_t *wlc_channel)
+{
+    UVM_ASSERT(wlc_channel);
+    UVM_ASSERT(uvm_channel_is_wlc(wlc_channel));
+
+    return get_paired_channel(wlc_channel);
+}
+
+static NV_STATUS channel_rotate_and_reserve_launch_channel(uvm_channel_t *channel, uvm_channel_t **launch_channel)
+{
+    uvm_channel_manager_t *manager = channel->pool->manager;
+    NV_STATUS status;
+
+    status = uvm_conf_computing_maybe_rotate_channel_ivs(channel);
+    if (status != NV_OK)
+        return status;
+
+    // CE channels, other than WLC fix launch schedule setup, need a launch
+    // channel that needs to be reserved
+    if (uvm_channel_is_ce(channel) &&
+        !(uvm_channel_is_wlc(channel) && uvm_channel_manager_is_wlc_ready(manager))) {
+        uvm_channel_t *local_launch_channel = NULL;
+        uvm_channel_type_t indirect_channel_type = uvm_channel_manager_is_wlc_ready(manager) ?
+                                                   UVM_CHANNEL_TYPE_WLC :
+                                                   UVM_CHANNEL_TYPE_SEC2;
+        status = uvm_channel_reserve_type(manager, indirect_channel_type, &local_launch_channel);
+        if (status != NV_OK)
+            return status;
+
+        // Indirect launch relies on pre-allocated resources to avoid failure
+        // paths. This includes pre-allocating IV space. There's no way to
+        // undo the launch channel reservation, so just return an error.
+        status = uvm_conf_computing_maybe_rotate_channel_ivs(local_launch_channel);
+        if (status != NV_OK) {
+            uvm_channel_release(local_launch_channel, 1);
+            return status;
+        }
+
+        if (uvm_channel_is_wlc(local_launch_channel)) {
+            status = uvm_conf_computing_maybe_rotate_channel_ivs(uvm_channel_wlc_get_paired_lcic(local_launch_channel));
+            if (status != NV_OK) {
+                uvm_channel_release(local_launch_channel, 1);
+                return status;
+            }
+        }
+        *launch_channel = local_launch_channel;
+    }
+
+
+    return NV_OK;
+}
+
 NV_STATUS uvm_channel_begin_push(uvm_channel_t *channel, uvm_push_t *push)
 {
-    NV_STATUS status;
+    NV_STATUS status = NV_OK;
    uvm_channel_manager_t *manager;
-    uvm_gpu_t *gpu;

    UVM_ASSERT(channel);
    UVM_ASSERT(push);

    manager = channel->pool->manager;

-    gpu = uvm_channel_get_gpu(channel);
-
    // Only SEC2 and WLC with set up fixed schedule can use direct push
    // submission. All other cases (including WLC pre-schedule) need to
    // reserve a launch channel that will be used to submit this push
    // indirectly.
-    if (uvm_conf_computing_mode_enabled(gpu) && uvm_channel_is_ce(channel) &&
-        !(uvm_channel_is_wlc(channel) && uvm_channel_manager_is_wlc_ready(manager))) {
-        uvm_channel_type_t indirect_channel_type = uvm_channel_manager_is_wlc_ready(manager) ?
-                                                   UVM_CHANNEL_TYPE_WLC :
-                                                   UVM_CHANNEL_TYPE_SEC2;
-        status = uvm_channel_reserve_type(manager, indirect_channel_type, &push->launch_channel);
+    if (g_uvm_global.conf_computing_enabled) {
+        status = channel_rotate_and_reserve_launch_channel(channel, &push->launch_channel);
        if (status != NV_OK)
            return status;
    }
@@ -570,7 +633,7 @@ NV_STATUS uvm_channel_begin_push(uvm_channel_t *channel, uvm_push_t *push)
    // When the Confidential Computing feature is enabled, the channel's lock
    // should have already been acquired in uvm_channel_reserve() or
    // channel_reserve_and_lock_in_pool().
-    UVM_ASSERT(is_channel_locked_for_push(channel));
+    UVM_ASSERT(uvm_channel_is_locked_for_push(channel));

    push->channel = channel;
    push->channel_tracking_value = 0;
@@ -603,7 +666,7 @@ static void internal_channel_submit_work(uvm_push_t *push, NvU32 push_size, NvU3
    gpfifo_entry = (NvU64*)channel->channel_info.gpFifoEntries + channel->cpu_put;
    pushbuffer_va = uvm_pushbuffer_get_gpu_va_for_push(pushbuffer, push);

-    if (uvm_conf_computing_mode_enabled(gpu)) {
+    if (g_uvm_global.conf_computing_enabled) {
        void *unprotected_pb = uvm_pushbuffer_get_unprotected_cpu_va_for_push(pushbuffer, push);
        UVM_ASSERT(uvm_channel_is_sec2(channel));

@@ -674,45 +737,14 @@ static void uvm_channel_tracking_semaphore_release(uvm_push_t *push, NvU64 semap
    // needs to be scheduled to get an encrypted shadow copy in unprotected
    // sysmem. This allows UVM to later decrypt it and observe the new
    // semaphore value.
-    if (uvm_conf_computing_mode_enabled(push->gpu) && uvm_channel_is_ce(push->channel))
+    if (g_uvm_global.conf_computing_enabled && uvm_channel_is_ce(push->channel))
        channel_semaphore_gpu_encrypt_payload(push, semaphore_va);
 }

-static uvm_channel_t *get_paired_channel(uvm_channel_t *channel)
-{
-    unsigned index;
-    uvm_channel_pool_t *paired_pool;
-    uvm_channel_type_t paired_channel_type;
-
-    UVM_ASSERT(channel);
-    UVM_ASSERT(uvm_channel_is_wlc(channel) || uvm_channel_is_lcic(channel));
-
-    index = uvm_channel_index_in_pool(channel);
-    paired_channel_type = uvm_channel_is_wlc(channel) ? UVM_CHANNEL_TYPE_LCIC : UVM_CHANNEL_TYPE_WLC;
-    paired_pool = channel->pool->manager->pool_to_use.default_for_type[paired_channel_type];
-    return paired_pool->channels + index;
-}
-
-static uvm_channel_t *wlc_get_paired_lcic(uvm_channel_t *wlc_channel)
-{
-    UVM_ASSERT(wlc_channel);
-    UVM_ASSERT(uvm_channel_is_wlc(wlc_channel));
-
-    return get_paired_channel(wlc_channel);
-}
-
-static uvm_channel_t *lcic_get_paired_wlc(uvm_channel_t *lcic_channel)
-{
-    UVM_ASSERT(lcic_channel);
-    UVM_ASSERT(uvm_channel_is_lcic(lcic_channel));
-
-    return get_paired_channel(lcic_channel);
-}
-
 static void internal_channel_submit_work_wlc(uvm_push_t *push)
 {
    uvm_channel_t *wlc_channel = push->channel;
-    uvm_channel_t *lcic_channel = wlc_get_paired_lcic(wlc_channel);
+    uvm_channel_t *lcic_channel = uvm_channel_wlc_get_paired_lcic(wlc_channel);
    UvmCslIv *iv_cpu_addr = lcic_channel->tracking_sem.semaphore.conf_computing.ivs;
    NvU32 *last_pushed_notifier;
    NvU32 iv_index;
@@ -926,7 +958,7 @@ static void set_gpfifo_via_sec2(uvm_push_t *sec2_push, uvm_channel_t *channel, N
            uvm_pushbuffer_t *pushbuffer = uvm_channel_get_pushbuffer(channel);
            NvU64 prev_pb_va = uvm_pushbuffer_get_gpu_va_base(pushbuffer) + previous_gpfifo->pushbuffer_offset;

-            // Reconstruct the previous gpfifo entry. UVM_GPFIFO_SYNC_WAIT is
+            // Reconstruct the previous GPFIFO entry. UVM_GPFIFO_SYNC_WAIT is
            // used only in static WLC schedule.
            // Overwriting the previous entry with the same value doesn't hurt,
            // whether the previous entry has been processed or not
@@ -1053,7 +1085,7 @@ static void encrypt_push(uvm_push_t *push)
    uvm_pushbuffer_t *pushbuffer = uvm_channel_get_pushbuffer(channel);
    unsigned auth_tag_offset = UVM_CONF_COMPUTING_AUTH_TAG_SIZE * push->push_info_index;

-    if (!uvm_conf_computing_mode_enabled(gpu))
+    if (!g_uvm_global.conf_computing_enabled)
        return;

    if (!push_info->on_complete)
@@ -1111,7 +1143,7 @@ void uvm_channel_end_push(uvm_push_t *push)
    uvm_channel_tracking_semaphore_release(push, semaphore_va, new_payload);

    if (uvm_channel_is_wlc(channel) && uvm_channel_manager_is_wlc_ready(channel_manager)) {
-        uvm_channel_t *paired_lcic = wlc_get_paired_lcic(channel);
+        uvm_channel_t *paired_lcic = uvm_channel_wlc_get_paired_lcic(channel);

        gpu->parent->ce_hal->semaphore_reduction_inc(push,
                                                     paired_lcic->channel_info.gpPutGpuVa,
@@ -1125,7 +1157,7 @@ void uvm_channel_end_push(uvm_push_t *push)
            // The UVM_MAX_WLC_PUSH_SIZE is set to fit indirect work launch
            // pushes. However, direct pushes to WLC can be smaller than this
            // size. This is used e.g. by indirect submission of control
-            // gpfifo entries.
+            // GPFIFO entries.
            gpu->parent->host_hal->noop(push, UVM_MAX_WLC_PUSH_SIZE - uvm_push_get_size(push));
        }
    }
@@ -1144,8 +1176,9 @@ void uvm_channel_end_push(uvm_push_t *push)
    // Indirect submission via SEC2/WLC needs pushes to be aligned for
    // encryption/decryption. The pushbuffer_size of this push
    // influences starting address of the next push.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        entry->pushbuffer_size = UVM_ALIGN_UP(push_size, UVM_CONF_COMPUTING_BUF_ALIGNMENT);
+
    entry->push_info = &channel->push_infos[push->push_info_index];
    entry->type = UVM_GPFIFO_ENTRY_TYPE_NORMAL;

@@ -1158,7 +1191,7 @@ void uvm_channel_end_push(uvm_push_t *push)
    else if (uvm_channel_is_wlc(channel) && uvm_channel_manager_is_wlc_ready(channel_manager)) {
        internal_channel_submit_work_wlc(push);
    }
-    else if (uvm_conf_computing_mode_enabled(gpu) && uvm_channel_is_ce(channel)) {
+    else if (g_uvm_global.conf_computing_enabled && uvm_channel_is_ce(channel)) {
        if (uvm_channel_manager_is_wlc_ready(channel_manager)) {
            internal_channel_submit_work_indirect_wlc(push, cpu_put, new_cpu_put);
        }
@@ -1209,7 +1242,7 @@ static void submit_ctrl_gpfifo(uvm_channel_t *channel, uvm_gpfifo_entry_t *entry

    UVM_ASSERT(entry == &channel->gpfifo_entries[cpu_put]);

-    if (uvm_conf_computing_mode_enabled(gpu) && uvm_channel_is_ce(channel))
+    if (g_uvm_global.conf_computing_enabled && uvm_channel_is_ce(channel))
        return;

    gpfifo_entry = (NvU64*)channel->channel_info.gpFifoEntries + cpu_put;
@@ -1291,8 +1324,6 @@ static void write_ctrl_gpfifo(uvm_channel_t *channel, NvU64 ctrl_fifo_entry_valu
    uvm_gpfifo_entry_t *entry;
    NvU32 cpu_put;
    NvU32 new_cpu_put;
-    bool needs_indirect_submit = false;
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);

    channel_pool_lock(channel->pool);

@@ -1315,8 +1346,6 @@ static void write_ctrl_gpfifo(uvm_channel_t *channel, NvU64 ctrl_fifo_entry_valu
    --channel->current_gpfifo_count;

    submit_ctrl_gpfifo(channel, entry, new_cpu_put);
-    if (uvm_conf_computing_mode_enabled(gpu) && uvm_channel_is_ce(channel))
-        needs_indirect_submit = true;

    channel->cpu_put = new_cpu_put;

@@ -1327,7 +1356,8 @@ static void write_ctrl_gpfifo(uvm_channel_t *channel, NvU64 ctrl_fifo_entry_valu
    // semaphore release, where the channel is unlocked.
    channel_pool_unlock(channel->pool);

-    if (needs_indirect_submit) {
+    // Trigger indirect submission when needed.
+    if (g_uvm_global.conf_computing_enabled && uvm_channel_is_ce(channel)) {
        NV_STATUS status = submit_ctrl_gpfifo_indirect(channel, entry, cpu_put, new_cpu_put);

        // All failures are globally fatal. There's nothing we do to recover.
@@ -1344,12 +1374,11 @@ static void write_ctrl_gpfifo(uvm_channel_t *channel, NvU64 ctrl_fifo_entry_valu
 NV_STATUS uvm_channel_write_ctrl_gpfifo(uvm_channel_t *channel, NvU64 ctrl_fifo_entry_value)
 {
    NV_STATUS status;
-    uvm_gpu_t *gpu = channel->pool->manager->gpu;
    uvm_push_t push;

    UVM_ASSERT(!uvm_channel_is_proxy(channel));

-    // WLC/LCIC channels can only process custom gpfifo entries before
+    // WLC/LCIC channels can only process custom GPFIFO entries before
    // their schedule is set up.
    UVM_ASSERT(!uvm_channel_is_lcic(channel) || !uvm_channel_manager_is_wlc_ready(channel->pool->manager));
    UVM_ASSERT(!uvm_channel_is_wlc(channel) || !uvm_channel_manager_is_wlc_ready(channel->pool->manager));
@@ -1373,10 +1402,28 @@ NV_STATUS uvm_channel_write_ctrl_gpfifo(uvm_channel_t *channel, NvU64 ctrl_fifo_
    if (status != NV_OK)
        return status;

+    if (g_uvm_global.conf_computing_enabled) {
+        // Rotating IV needs to idle the channel. However, there's no semaphore
+        // release after submitting a control entry. It is not possible to wait
+        // for in-flight entries after the GPFIFO submission.
+        // Instead, check for IV rotation early. Secure channels are locked for
+        // pushes after reservation so the IV space gained here can't be used
+        // up by concurrent pushes.
+        status = uvm_conf_computing_maybe_rotate_channel_ivs_retry_busy(channel);
+        if (status != NV_OK) {
+            uvm_channel_release(channel, 2);
+            return status;
+        }
+    }
+
    write_ctrl_gpfifo(channel, ctrl_fifo_entry_value);

    status = uvm_push_begin_on_reserved_channel(channel, &push, "write_ctrl_GPFIFO");
    if (status != NV_OK) {
+        uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
+
+        // One entry was consumed by GPFIFO entry
+        uvm_channel_release(channel, 1);
        UVM_ERR_PRINT("Failed to begin push on channel: %s, GPU %s\n", nvstatusToString(status), uvm_gpu_name(gpu));
        return status;
    }
@@ -1440,9 +1487,8 @@ NV_STATUS uvm_channel_reserve(uvm_channel_t *channel, NvU32 num_gpfifo_entries)
 {
    NV_STATUS status = NV_OK;
    uvm_spin_loop_t spin;
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);

-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return channel_reserve_and_lock(channel, num_gpfifo_entries);

    if (try_claim_channel(channel, num_gpfifo_entries))
@@ -1460,6 +1506,18 @@ NV_STATUS uvm_channel_reserve(uvm_channel_t *channel, NvU32 num_gpfifo_entries)
    return status;
 }

+void uvm_channel_release(uvm_channel_t *channel, NvU32 num_gpfifo_entries)
+{
+    channel_pool_lock(channel->pool);
+
+    UVM_ASSERT(uvm_channel_is_locked_for_push(channel));
+    unlock_channel_for_push(channel);
+
+    UVM_ASSERT(channel->current_gpfifo_count >= num_gpfifo_entries);
+    channel->current_gpfifo_count -= num_gpfifo_entries;
+    channel_pool_unlock(channel->pool);
+}
+
 // Get the first pending GPFIFO entry, if any.
 // This doesn't stop the entry from being reused.
 static uvm_gpfifo_entry_t *uvm_channel_get_first_pending_entry(uvm_channel_t *channel)
@@ -1483,14 +1541,14 @@ static uvm_gpfifo_entry_t *uvm_channel_get_first_pending_entry(uvm_channel_t *ch
 NV_STATUS uvm_channel_get_status(uvm_channel_t *channel)
 {
    uvm_gpu_t *gpu;
-    NvNotification *errorNotifier;
+    NvNotification *error_notifier;

    if (uvm_channel_is_proxy(channel))
-        errorNotifier = channel->proxy.channel_info.shadowErrorNotifier;
+        error_notifier = channel->proxy.channel_info.shadowErrorNotifier;
    else
-        errorNotifier = channel->channel_info.errorNotifier;
+        error_notifier = channel->channel_info.errorNotifier;

-    if (errorNotifier->status == 0)
+    if (error_notifier->status == 0)
        return NV_OK;

    // In case we hit a channel error, check the ECC error notifier as well so
@@ -1580,35 +1638,55 @@ NvU64 uvm_channel_update_completed_value(uvm_channel_t *channel)
    return uvm_gpu_tracking_semaphore_update_completed_value(&channel->tracking_sem);
 }

-static NV_STATUS csl_init(uvm_channel_t *channel)
+NV_STATUS uvm_channel_wait(uvm_channel_t *channel)
 {
-    NV_STATUS status;
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
+    NV_STATUS status = uvm_global_get_status();
+    uvm_spin_loop_t spin;

-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    if (uvm_channel_update_progress(channel) == 0 && status == NV_OK)
+        return uvm_channel_check_errors(channel);

-    uvm_mutex_init(&channel->csl.ctx_lock, UVM_LOCK_ORDER_LEAF);
+    uvm_spin_loop_init(&spin);
+    while (uvm_channel_update_progress(channel) > 0 && status == NV_OK) {
+        UVM_SPIN_LOOP(&spin);
+        status = uvm_global_get_status();

-    status = uvm_rm_locked_call(nvUvmInterfaceCslInitContext(&channel->csl.ctx, channel->handle));
-    if (status == NV_OK) {
-        channel->csl.is_ctx_initialized = true;
-    }
-    else {
-        UVM_DBG_PRINT("nvUvmInterfaceCslInitContext() failed: %s, GPU %s\n",
-                      nvstatusToString(status),
-                      uvm_gpu_name(gpu));
+        if (status == NV_OK)
+            status = uvm_channel_check_errors(channel);
    }

    return status;
 }

+static NV_STATUS csl_init(uvm_channel_t *channel)
+{
+    NV_STATUS status;
+
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
+
+    status = uvm_rm_locked_call(nvUvmInterfaceCslInitContext(&channel->csl.ctx, channel->handle));
+    if (status != NV_OK) {
+        uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
+
+        UVM_DBG_PRINT("nvUvmInterfaceCslInitContext() failed: %s, GPU %s\n",
+                      nvstatusToString(status),
+                      uvm_gpu_name(gpu));
+        return status;
+    }
+
+    uvm_mutex_init(&channel->csl.ctx_lock, UVM_LOCK_ORDER_CSL_CTX);
+    channel->csl.is_ctx_initialized = true;
+
+    return NV_OK;
+}
+
 static void csl_destroy(uvm_channel_t *channel)
 {
    if (!channel->csl.is_ctx_initialized)
        return;

    uvm_assert_mutex_unlocked(&channel->csl.ctx_lock);
-    UVM_ASSERT(!is_channel_locked_for_push(channel));
+    UVM_ASSERT(!uvm_channel_is_locked_for_push(channel));

    uvm_rm_locked_call_void(nvUvmInterfaceDeinitCslContext(&channel->csl.ctx));
    channel->csl.is_ctx_initialized = false;
@@ -1616,9 +1694,7 @@ static void csl_destroy(uvm_channel_t *channel)

 static void free_conf_computing_buffers(uvm_channel_t *channel)
 {
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
-
-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
    UVM_ASSERT(uvm_channel_is_ce(channel));

    uvm_rm_mem_free(channel->conf_computing.static_pb_protected_vidmem);
@@ -1650,7 +1726,7 @@ static NV_STATUS alloc_conf_computing_buffers_semaphore(uvm_channel_t *channel)
    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
    NV_STATUS status;

-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
    UVM_ASSERT(uvm_channel_is_ce(channel));

    status = uvm_rm_mem_alloc_and_map_cpu(gpu,
@@ -1770,9 +1846,8 @@ static NV_STATUS alloc_conf_computing_buffers_lcic(uvm_channel_t *channel)
 static NV_STATUS alloc_conf_computing_buffers(uvm_channel_t *channel)
 {
    NV_STATUS status;
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);

-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);
    UVM_ASSERT(uvm_channel_is_ce(channel));

    status = alloc_conf_computing_buffers_semaphore(channel);
@@ -1786,6 +1861,7 @@ static NV_STATUS alloc_conf_computing_buffers(uvm_channel_t *channel)
        status = alloc_conf_computing_buffers_lcic(channel);
    }
    else {
+        uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
        void *push_crypto_bundles = uvm_kvmalloc_zero(sizeof(*channel->conf_computing.push_crypto_bundles) *
                                                      channel->num_gpfifo_entries);

@@ -1806,8 +1882,6 @@ static NV_STATUS alloc_conf_computing_buffers(uvm_channel_t *channel)

 static void channel_destroy(uvm_channel_pool_t *pool, uvm_channel_t *channel)
 {
-    uvm_gpu_t *gpu = uvm_channel_get_gpu(channel);
-
    UVM_ASSERT(pool->num_channels > 0);

    if (channel->tracking_sem.queued_value > 0) {
@@ -1831,7 +1905,7 @@ static void channel_destroy(uvm_channel_pool_t *pool, uvm_channel_t *channel)

    uvm_kvfree(channel->gpfifo_entries);

-    if (uvm_conf_computing_mode_enabled(gpu)) {
+    if (g_uvm_global.conf_computing_enabled) {
        csl_destroy(channel);

        if (uvm_channel_is_ce(channel))
@@ -1889,7 +1963,7 @@ static uvmGpuTsgHandle channel_get_tsg(uvm_channel_t *channel)

    if (uvm_channel_pool_is_wlc(pool) || uvm_channel_pool_is_lcic(pool)) {
        if (uvm_channel_pool_is_lcic(pool)) {
-            channel = lcic_get_paired_wlc(channel);
+            channel = uvm_channel_lcic_get_paired_wlc(channel);
            pool = channel->pool;
        }

@@ -1906,7 +1980,6 @@ static NV_STATUS internal_channel_create(uvm_channel_t *channel)
    UvmGpuChannelAllocParams channel_alloc_params;
    UvmGpuChannelInfo *channel_info = &channel->channel_info;
    uvm_channel_manager_t *manager = channel->pool->manager;
-    uvm_gpu_t *gpu = manager->gpu;

    memset(&channel_alloc_params, 0, sizeof(channel_alloc_params));
    channel_alloc_params.numGpFifoEntries = channel_pool_type_num_gpfifo_entries(manager, channel->pool->pool_type);
@@ -1914,7 +1987,7 @@ static NV_STATUS internal_channel_create(uvm_channel_t *channel)
    channel_alloc_params.gpPutLoc = manager->conf.gpput_loc;

    if (uvm_channel_is_sec2(channel)) {
-        UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+        UVM_ASSERT(g_uvm_global.conf_computing_enabled);

        // SEC2 channels' GPPUT and GPFIFO must be allocated in sysmem.
        channel_alloc_params.gpFifoLoc = UVM_BUFFER_LOCATION_SYS;
@@ -1928,7 +2001,7 @@ static NV_STATUS internal_channel_create(uvm_channel_t *channel)
    if (status != NV_OK) {
        UVM_ERR_PRINT("nvUvmInterfaceChannelAllocate() failed: %s, GPU %s, type %s\n",
                      nvstatusToString(status),
-                      uvm_gpu_name(gpu),
+                      uvm_gpu_name(manager->gpu),
                      uvm_channel_pool_type_to_string(channel->pool->pool_type));
        return status;
    }
@@ -1994,7 +2067,7 @@ static NV_STATUS channel_create(uvm_channel_pool_t *pool, uvm_channel_t *channel
    channel->tools.pending_event_count = 0;
    INIT_LIST_HEAD(&channel->tools.channel_list_node);

-    if (uvm_conf_computing_mode_enabled(gpu) && uvm_channel_is_ce(channel))
+    if (g_uvm_global.conf_computing_enabled && uvm_channel_is_ce(channel))
        semaphore_pool = gpu->secure_semaphore_pool;

    status = uvm_gpu_tracking_semaphore_alloc(semaphore_pool, &channel->tracking_sem);
@@ -2020,7 +2093,7 @@ static NV_STATUS channel_create(uvm_channel_pool_t *pool, uvm_channel_t *channel
        goto error;
    }

-    if (uvm_conf_computing_mode_enabled(gpu)) {
+    if (g_uvm_global.conf_computing_enabled) {
        status = csl_init(channel);
        if (status != NV_OK)
            goto error;
@@ -2079,14 +2152,14 @@ static NV_STATUS channel_init(uvm_channel_t *channel)
    NV_STATUS status;
    NvU32 num_entries = 1;

-    if (uvm_gpu_has_pushbuffer_segments(gpu))
+    if (uvm_parent_gpu_needs_pushbuffer_segments(gpu->parent))
        num_entries++;

    status = uvm_channel_reserve(channel, num_entries);
    if (status != NV_OK)
        return status;

-    if (uvm_gpu_has_pushbuffer_segments(gpu)) {
+    if (uvm_parent_gpu_needs_pushbuffer_segments(gpu->parent)) {
        NvU64 gpfifo_entry;
        uvm_pushbuffer_t *pushbuffer = uvm_channel_get_pushbuffer(channel);
        NvU64 pb_base = uvm_pushbuffer_get_gpu_va_base(pushbuffer);
@@ -2102,6 +2175,10 @@ static NV_STATUS channel_init(uvm_channel_t *channel)

    status = uvm_push_begin_on_reserved_channel(channel, &push, "Init channel");
    if (status != NV_OK) {
+
+        // One entry was consumed by control GPFIFO entry above, release the
+        // second one.
+        uvm_channel_release(channel, 1);
        UVM_ERR_PRINT("Failed to begin push on channel: %s, GPU %s\n", nvstatusToString(status), uvm_gpu_name(gpu));
        return status;
    }
@@ -2126,7 +2203,7 @@ static NV_STATUS channel_init(uvm_channel_t *channel)

 static bool channel_manager_uses_proxy_pool(uvm_channel_manager_t *manager)
 {
-    return uvm_gpu_is_virt_mode_sriov_heavy(manager->gpu);
+    return uvm_parent_gpu_is_virt_mode_sriov_heavy(manager->gpu->parent);
 }

 // Number of channels to create in a pool of the given type.
@@ -2266,7 +2343,7 @@ static NV_STATUS channel_pool_add(uvm_channel_manager_t *channel_manager,
    num_channels = channel_pool_type_num_channels(pool_type);
    UVM_ASSERT(num_channels <= UVM_CHANNEL_MAX_NUM_CHANNELS_PER_POOL);

-    if (uvm_conf_computing_mode_enabled(channel_manager->gpu)) {
+    if (g_uvm_global.conf_computing_enabled) {
        // Use different order lock for SEC2 and WLC channels.
        // This allows reserving a SEC2 or WLC channel for indirect work
        // submission while holding a reservation for a channel.
@@ -2507,16 +2584,18 @@ out:

 // Return the pool corresponding to the given CE index
 //
-// This function cannot be used to access the proxy pool in SR-IOV heavy.
+// Used to retrieve pools of type UVM_CHANNEL_POOL_TYPE_CE only.
 static uvm_channel_pool_t *channel_manager_ce_pool(uvm_channel_manager_t *manager, NvU32 ce)
 {
-    uvm_channel_pool_t *pool;
+    uvm_channel_pool_t *pool = uvm_channel_pool_first(manager, UVM_CHANNEL_POOL_TYPE_CE);

+    UVM_ASSERT(pool != NULL);
    UVM_ASSERT(test_bit(ce, manager->ce_mask));

-    // The index of the pool associated with 'ce' is the number of usable CEs
-    // in [0, ce)
-    pool = manager->channel_pools + bitmap_weight(manager->ce_mask, ce);
+    // Pools of type UVM_CHANNEL_POOL_TYPE_CE are stored contiguously. The
+    // offset of the pool associated with 'ce' is the number of usable CEs in
+    // [0, ce).
+    pool += bitmap_weight(manager->ce_mask, ce);

    UVM_ASSERT(pool->pool_type == UVM_CHANNEL_POOL_TYPE_CE);
    UVM_ASSERT(pool->engine_index == ce);
@@ -2683,7 +2762,7 @@ static void init_channel_manager_conf(uvm_channel_manager_t *manager)
    // caches vidmem (and sysmem), we place GPFIFO and GPPUT on sysmem to avoid
    // cache thrash. The memory access latency is reduced, despite the required
    // access through the bus, because no cache coherence message is exchanged.
-    if (uvm_gpu_is_coherent(gpu->parent)) {
+    if (uvm_parent_gpu_is_coherent(gpu->parent)) {
        manager->conf.gpfifo_loc = UVM_BUFFER_LOCATION_SYS;

        // On GPUs with limited ESCHED addressing range, e.g., Volta on P9, RM
@@ -2721,11 +2800,11 @@ static unsigned channel_manager_get_max_pools(uvm_channel_manager_t *manager)
    num_channel_pools = bitmap_weight(manager->ce_mask, UVM_COPY_ENGINE_COUNT_MAX);

    // CE proxy channel pool.
-    if (uvm_gpu_uses_proxy_channel_pool(manager->gpu))
+    if (uvm_parent_gpu_needs_proxy_channel_pool(manager->gpu->parent))
        num_channel_pools++;

    // SEC2 pool, WLC pool, LCIC pool
-    if (uvm_conf_computing_mode_enabled(manager->gpu))
+    if (g_uvm_global.conf_computing_enabled)
        num_channel_pools += 3;

    return num_channel_pools;
@@ -2734,6 +2813,7 @@ static unsigned channel_manager_get_max_pools(uvm_channel_manager_t *manager)
 static NV_STATUS channel_manager_create_ce_pools(uvm_channel_manager_t *manager, unsigned *preferred_ce)
 {
    unsigned ce;
+    unsigned type;

    // A pool is created for each usable CE, even if it has not been selected as
    // the preferred CE for any type, because as more information is discovered
@@ -2741,18 +2821,20 @@ static NV_STATUS channel_manager_create_ce_pools(uvm_channel_manager_t *manager,
    // previously idle pools.
    for_each_set_bit(ce, manager->ce_mask, UVM_COPY_ENGINE_COUNT_MAX) {
        NV_STATUS status;
-        unsigned type;
        uvm_channel_pool_t *pool = NULL;

        status = channel_pool_add(manager, UVM_CHANNEL_POOL_TYPE_CE, ce, &pool);
        if (status != NV_OK)
            return status;
+    }

-        for (type = 0; type < UVM_CHANNEL_TYPE_CE_COUNT; type++) {
-            // Set pool type if it hasn't been set before.
-            if (preferred_ce[type] == ce && manager->pool_to_use.default_for_type[type] == NULL)
-                manager->pool_to_use.default_for_type[type] = pool;
-        }
+    for (type = 0; type < UVM_CHANNEL_TYPE_CE_COUNT; type++) {
+        // Avoid overwriting previously set defaults.
+        if (manager->pool_to_use.default_for_type[type] != NULL)
+            continue;
+
+        ce = preferred_ce[type];
+        manager->pool_to_use.default_for_type[type] = channel_manager_ce_pool(manager, ce);
    }

    return NV_OK;
@@ -3093,7 +3175,7 @@ static NV_STATUS channel_manager_create_conf_computing_pools(uvm_channel_manager
    uvm_channel_pool_t *wlc_pool = NULL;
    uvm_channel_pool_t *lcic_pool = NULL;

-    if (!uvm_conf_computing_mode_enabled(manager->gpu))
+    if (!g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = uvm_rm_mem_alloc(manager->gpu,
@@ -3173,7 +3255,7 @@ static NV_STATUS channel_manager_create_pools(uvm_channel_manager_t *manager)

    // In SR-IOV heavy, add an additional, single-channel, pool that is
    // dedicated to the MEMOPS type.
-    if (uvm_gpu_uses_proxy_channel_pool(manager->gpu)) {
+    if (uvm_parent_gpu_needs_proxy_channel_pool(manager->gpu->parent)) {
        uvm_channel_pool_t *proxy_pool = NULL;
        uvm_channel_type_t channel_type = uvm_channel_proxy_channel_type();

@@ -3295,7 +3377,7 @@ void uvm_channel_manager_destroy(uvm_channel_manager_t *channel_manager)

 bool uvm_channel_is_privileged(uvm_channel_t *channel)
 {
-    if (uvm_gpu_is_virt_mode_sriov_heavy(uvm_channel_get_gpu(channel)))
+    if (uvm_parent_gpu_is_virt_mode_sriov_heavy(uvm_channel_get_gpu(channel)->parent))
        return uvm_channel_is_proxy(channel);

    return true;
--- a/kernel-open/nvidia-uvm/uvm_channel.h
+++ b/kernel-open/nvidia-uvm/uvm_channel.h
@@ -497,6 +497,10 @@ static bool uvm_channel_is_lcic(uvm_channel_t *channel)
    return uvm_channel_pool_is_lcic(channel->pool);
 }

+uvm_channel_t *uvm_channel_lcic_get_paired_wlc(uvm_channel_t *lcic_channel);
+
+uvm_channel_t *uvm_channel_wlc_get_paired_lcic(uvm_channel_t *wlc_channel);
+
 static bool uvm_channel_pool_is_proxy(uvm_channel_pool_t *pool)
 {
    UVM_ASSERT(uvm_pool_type_is_valid(pool->pool_type));
@@ -603,6 +607,11 @@ bool uvm_channel_is_value_completed(uvm_channel_t *channel, NvU64 value);
 // Update and get the latest completed value by the channel
 NvU64 uvm_channel_update_completed_value(uvm_channel_t *channel);

+// Wait for the channel to idle
+// It waits for anything that is running, but doesn't prevent new work from
+// beginning.
+NV_STATUS uvm_channel_wait(uvm_channel_t *channel);
+
 // Select and reserve a channel with the specified type for a push
 NV_STATUS uvm_channel_reserve_type(uvm_channel_manager_t *manager,
                                   uvm_channel_type_t type,
@@ -617,6 +626,9 @@ NV_STATUS uvm_channel_reserve_gpu_to_gpu(uvm_channel_manager_t *channel_manager,
 // Reserve a specific channel for a push or for a control GPFIFO entry.
 NV_STATUS uvm_channel_reserve(uvm_channel_t *channel, NvU32 num_gpfifo_entries);

+// Release reservation on a specific channel
+void uvm_channel_release(uvm_channel_t *channel, NvU32 num_gpfifo_entries);
+
 // Set optimal CE for P2P transfers between manager->gpu and peer
 void uvm_channel_manager_set_p2p_ce(uvm_channel_manager_t *manager, uvm_gpu_t *peer, NvU32 optimal_ce);

@@ -648,6 +660,8 @@ NvU32 uvm_channel_get_available_gpfifo_entries(uvm_channel_t *channel);

 void uvm_channel_print_pending_pushes(uvm_channel_t *channel);

+bool uvm_channel_is_locked_for_push(uvm_channel_t *channel);
+
 static uvm_gpu_t *uvm_channel_get_gpu(uvm_channel_t *channel)
 {
    return channel->pool->manager->gpu;
--- a/kernel-open/nvidia-uvm/uvm_channel_test.c
+++ b/kernel-open/nvidia-uvm/uvm_channel_test.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2015-2022 NVIDIA Corporation
+    Copyright (c) 2015-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -24,6 +24,7 @@
 #include "uvm_global.h"
 #include "uvm_channel.h"
 #include "uvm_hal.h"
+#include "uvm_mem.h"
 #include "uvm_push.h"
 #include "uvm_test.h"
 #include "uvm_test_rng.h"
@@ -57,14 +58,14 @@ static NV_STATUS test_ordering(uvm_va_space_t *va_space)
    const NvU32 values_count = iters_per_channel_type_per_gpu;
    const size_t buffer_size = sizeof(NvU32) * values_count;

-    gpu = uvm_va_space_find_first_gpu(va_space);
-    TEST_CHECK_RET(gpu != NULL);
-
    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

+    gpu = uvm_va_space_find_first_gpu(va_space);
+    TEST_CHECK_RET(gpu != NULL);
+
    status = uvm_rm_mem_alloc_and_map_all(gpu, UVM_RM_MEM_TYPE_SYS, buffer_size, 0, &mem);
    TEST_CHECK_GOTO(status == NV_OK, done);

@@ -84,7 +85,7 @@ static NV_STATUS test_ordering(uvm_va_space_t *va_space)

    TEST_NV_CHECK_GOTO(uvm_tracker_add_push(&tracker, &push), done);

-    exclude_proxy_channel_type = uvm_gpu_uses_proxy_channel_pool(gpu);
+    exclude_proxy_channel_type = uvm_parent_gpu_needs_proxy_channel_pool(gpu->parent);

    for (i = 0; i < iters_per_channel_type_per_gpu; ++i) {
        for (j = 0; j < UVM_CHANNEL_TYPE_CE_COUNT; ++j) {
@@ -222,7 +223,7 @@ static NV_STATUS uvm_test_rc_for_gpu(uvm_gpu_t *gpu)
    // Check RC on a proxy channel (SR-IOV heavy) or internal channel (any other
    // mode). It is not allowed to use a virtual address in a memset pushed to
    // a proxy channel, so we use a physical address instead.
-    if (uvm_gpu_uses_proxy_channel_pool(gpu)) {
+    if (uvm_parent_gpu_needs_proxy_channel_pool(gpu->parent)) {
        uvm_gpu_address_t dst_address;

        // Save the line number the push that's supposed to fail was started on
@@ -314,6 +315,110 @@ static NV_STATUS test_rc(uvm_va_space_t *va_space)
    return NV_OK;
 }

+static NV_STATUS uvm_test_iommu_rc_for_gpu(uvm_gpu_t *gpu)
+{
+    NV_STATUS status = NV_OK;
+
+#if defined(NV_IOMMU_IS_DMA_DOMAIN_PRESENT) && defined(CONFIG_IOMMU_DEFAULT_DMA_STRICT)
+    // This test needs the DMA API to immediately invalidate IOMMU mappings on
+    // DMA unmap (as apposed to lazy invalidation). The policy can be changed
+    // on boot (e.g. iommu.strict=1), but there isn't a good way to check for
+    // the runtime setting. CONFIG_IOMMU_DEFAULT_DMA_STRICT checks for the
+    // default value.
+
+    uvm_push_t push;
+    uvm_mem_t *sysmem;
+    uvm_gpu_address_t sysmem_dma_addr;
+    char *cpu_ptr = NULL;
+    const size_t data_size = PAGE_SIZE;
+    size_t i;
+
+    struct device *dev = &gpu->parent->pci_dev->dev;
+    struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
+
+    // Check that the iommu domain is controlled by linux DMA API
+    if (!domain || !iommu_is_dma_domain(domain))
+        return NV_OK;
+
+    // Only run if ATS is enabled with 64kB base page.
+    // Otherwise the CE doesn't get response on writing to unmapped location.
+    if (!g_uvm_global.ats.enabled || PAGE_SIZE != UVM_PAGE_SIZE_64K)
+        return NV_OK;
+
+    status = uvm_mem_alloc_sysmem_and_map_cpu_kernel(data_size, NULL, &sysmem);
+    TEST_NV_CHECK_RET(status);
+
+    status = uvm_mem_map_gpu_phys(sysmem, gpu);
+    TEST_NV_CHECK_GOTO(status, done);
+
+    cpu_ptr = uvm_mem_get_cpu_addr_kernel(sysmem);
+    sysmem_dma_addr = uvm_mem_gpu_address_physical(sysmem, gpu, 0, data_size);
+
+    status = uvm_push_begin(gpu->channel_manager, UVM_CHANNEL_TYPE_GPU_TO_CPU, &push, "Test memset to IOMMU mapped sysmem");
+    TEST_NV_CHECK_GOTO(status, done);
+
+    gpu->parent->ce_hal->memset_8(&push, sysmem_dma_addr, 0, data_size);
+
+    status = uvm_push_end_and_wait(&push);
+    TEST_NV_CHECK_GOTO(status, done);
+
+    // Check that we have zeroed the memory
+    for (i = 0; i < data_size; ++i)
+        TEST_CHECK_GOTO(cpu_ptr[i] == 0, done);
+
+    // Unmap the buffer and try write again to the same address
+    uvm_mem_unmap_gpu_phys(sysmem, gpu);
+
+    status = uvm_push_begin(gpu->channel_manager, UVM_CHANNEL_TYPE_GPU_TO_CPU, &push, "Test memset after IOMMU unmap");
+    TEST_NV_CHECK_GOTO(status, done);
+
+    gpu->parent->ce_hal->memset_4(&push, sysmem_dma_addr, 0xffffffff, data_size);
+
+    status = uvm_push_end_and_wait(&push);
+
+    TEST_CHECK_GOTO(status == NV_ERR_RC_ERROR, done);
+    TEST_CHECK_GOTO(uvm_channel_get_status(push.channel) == NV_ERR_RC_ERROR, done);
+    TEST_CHECK_GOTO(uvm_global_reset_fatal_error() == NV_ERR_RC_ERROR, done);
+
+    // Check that writes after unmap did not succeed
+    for (i = 0; i < data_size; ++i)
+        TEST_CHECK_GOTO(cpu_ptr[i] == 0, done);
+
+    status = NV_OK;
+
+done:
+    uvm_mem_free(sysmem);
+#endif
+    return status;
+}
+
+static NV_STATUS test_iommu(uvm_va_space_t *va_space)
+{
+    uvm_gpu_t *gpu;
+
+    uvm_assert_mutex_locked(&g_uvm_global.global_lock);
+
+    for_each_va_space_gpu(gpu, va_space) {
+        NV_STATUS test_status, create_status;
+
+        // The GPU channel manager is destroyed and then re-created after
+        // testing ATS RC fault, so this test requires exclusive access to the GPU.
+        TEST_CHECK_RET(uvm_gpu_retained_count(gpu) == 1);
+
+        g_uvm_global.disable_fatal_error_assert = true;
+        test_status = uvm_test_iommu_rc_for_gpu(gpu);
+        g_uvm_global.disable_fatal_error_assert = false;
+
+        uvm_channel_manager_destroy(gpu->channel_manager);
+        create_status = uvm_channel_manager_create(gpu, &gpu->channel_manager);
+
+        TEST_NV_CHECK_RET(test_status);
+        TEST_NV_CHECK_RET(create_status);
+    }
+
+    return NV_OK;
+}
+
 typedef struct
 {
    uvm_push_t push;
@@ -403,7 +508,7 @@ static uvm_channel_type_t random_ce_channel_type_except(uvm_test_rng_t *rng, uvm

 static uvm_channel_type_t gpu_random_internal_ce_channel_type(uvm_gpu_t *gpu, uvm_test_rng_t *rng)
 {
-    if (uvm_gpu_uses_proxy_channel_pool(gpu))
+    if (uvm_parent_gpu_needs_proxy_channel_pool(gpu->parent))
        return random_ce_channel_type_except(rng, uvm_channel_proxy_channel_type());

    return random_ce_channel_type(rng);
@@ -586,12 +691,16 @@ static NV_STATUS stress_test_all_gpus_in_va(uvm_va_space_t *va_space,
            if (uvm_test_rng_range_32(&rng, 0, 1) == 0) {
                NvU32 random_stream_index = uvm_test_rng_range_32(&rng, 0, num_streams - 1);
                uvm_test_stream_t *random_stream = &streams[random_stream_index];
-                uvm_push_acquire_tracker(&stream->push, &random_stream->tracker);
-                snapshot_counter(&stream->push,
-                                 random_stream->counter_mem,
-                                 stream->other_stream_counter_snapshots_mem,
-                                 i,
-                                 random_stream->queued_counter_repeat);
+
+                if ((random_stream->push.gpu == gpu) || uvm_push_allow_dependencies_across_gpus()) {
+                    uvm_push_acquire_tracker(&stream->push, &random_stream->tracker);
+
+                    snapshot_counter(&stream->push,
+                                     random_stream->counter_mem,
+                                     stream->other_stream_counter_snapshots_mem,
+                                     i,
+                                     random_stream->queued_counter_repeat);
+                }
            }

            uvm_push_end(&stream->push);
@@ -693,9 +802,7 @@ NV_STATUS test_conf_computing_channel_selection(uvm_va_space_t *va_space)
    NvU32 i;
    NvU32 num_pushes;

-    gpu = uvm_va_space_find_first_gpu(va_space);
-
-    if (!uvm_conf_computing_mode_enabled(gpu))
+    if (!g_uvm_global.conf_computing_enabled)
        return NV_OK;

    uvm_thread_context_lock_disable_tracking();
@@ -746,6 +853,101 @@ error:
    return status;
 }

+NV_STATUS test_channel_iv_rotation(uvm_va_space_t *va_space)
+{
+    uvm_gpu_t *gpu;
+
+    if (!g_uvm_global.conf_computing_enabled)
+        return NV_OK;
+
+    for_each_va_space_gpu(gpu, va_space) {
+        uvm_channel_pool_t *pool;
+
+        uvm_for_each_pool(pool, gpu->channel_manager) {
+            NvU64 before_rotation_enc, before_rotation_dec, after_rotation_enc, after_rotation_dec;
+            NV_STATUS status = NV_OK;
+
+            // Check one (the first) channel per pool
+            uvm_channel_t *channel = pool->channels;
+
+            // Create a dummy encrypt/decrypt push to use few IVs.
+            // SEC2 used encrypt during initialization, no need to use a dummy
+            // push.
+            if (!uvm_channel_is_sec2(channel)) {
+                uvm_push_t push;
+                size_t data_size;
+                uvm_conf_computing_dma_buffer_t *cipher_text;
+                void *cipher_cpu_va, *plain_cpu_va, *tag_cpu_va;
+                uvm_gpu_address_t cipher_gpu_address, plain_gpu_address, tag_gpu_address;
+                uvm_channel_t *work_channel = uvm_channel_is_lcic(channel) ? uvm_channel_lcic_get_paired_wlc(channel) : channel;
+
+                plain_cpu_va = &status;
+                data_size = sizeof(status);
+
+                TEST_NV_CHECK_RET(uvm_conf_computing_dma_buffer_alloc(&gpu->conf_computing.dma_buffer_pool,
+                                                                      &cipher_text,
+                                                                      NULL));
+                cipher_cpu_va = uvm_mem_get_cpu_addr_kernel(cipher_text->alloc);
+                tag_cpu_va = uvm_mem_get_cpu_addr_kernel(cipher_text->auth_tag);
+
+                cipher_gpu_address = uvm_mem_gpu_address_virtual_kernel(cipher_text->alloc, gpu);
+                tag_gpu_address = uvm_mem_gpu_address_virtual_kernel(cipher_text->auth_tag, gpu);
+
+                TEST_NV_CHECK_GOTO(uvm_push_begin_on_channel(work_channel, &push, "Dummy push for IV rotation"), free);
+
+                (void)uvm_push_get_single_inline_buffer(&push,
+                                                        data_size,
+                                                        UVM_CONF_COMPUTING_BUF_ALIGNMENT,
+                                                        &plain_gpu_address);
+
+                uvm_conf_computing_cpu_encrypt(work_channel, cipher_cpu_va, plain_cpu_va, NULL, data_size, tag_cpu_va);
+                gpu->parent->ce_hal->decrypt(&push, plain_gpu_address, cipher_gpu_address, data_size, tag_gpu_address);
+
+                TEST_NV_CHECK_GOTO(uvm_push_end_and_wait(&push), free);
+
+free:
+                uvm_conf_computing_dma_buffer_free(&gpu->conf_computing.dma_buffer_pool, cipher_text, NULL);
+
+                if (status != NV_OK)
+                    return status;
+            }
+
+            // Reserve a channel to hold the push lock during rotation
+            if (!uvm_channel_is_lcic(channel))
+                TEST_NV_CHECK_RET(uvm_channel_reserve(channel, 1));
+
+            uvm_conf_computing_query_message_pools(channel, &before_rotation_enc, &before_rotation_dec);
+            TEST_NV_CHECK_GOTO(uvm_conf_computing_rotate_channel_ivs_below_limit(channel, -1, true), release);
+            uvm_conf_computing_query_message_pools(channel, &after_rotation_enc, &after_rotation_dec);
+
+release:
+            if (!uvm_channel_is_lcic(channel))
+                uvm_channel_release(channel, 1);
+
+            if (status != NV_OK)
+                return status;
+
+            // All channels except SEC2 used at least a single IV to release tracking.
+            // SEC2 doesn't support decrypt direction.
+            if (uvm_channel_is_sec2(channel))
+                TEST_CHECK_RET(before_rotation_dec == after_rotation_dec);
+            else
+                TEST_CHECK_RET(before_rotation_dec < after_rotation_dec);
+
+            // All channels used one CPU encrypt/GPU decrypt, either during
+            // initialization or in the push above, with the exception of LCIC.
+            // LCIC is used in tandem with WLC, but it never uses CPU encrypt/
+            // GPU decrypt ops.
+            if (uvm_channel_is_lcic(channel))
+                TEST_CHECK_RET(before_rotation_enc == after_rotation_enc);
+            else
+                TEST_CHECK_RET(before_rotation_enc < after_rotation_enc);
+        }
+    }
+
+    return NV_OK;
+}
+
 NV_STATUS test_write_ctrl_gpfifo_noop(uvm_va_space_t *va_space)
 {
    uvm_gpu_t *gpu;
@@ -845,11 +1047,9 @@ NV_STATUS test_write_ctrl_gpfifo_tight(uvm_va_space_t *va_space)
    NvU64 entry;
    uvm_push_t push;

-    gpu = uvm_va_space_find_first_gpu(va_space);
-
    // TODO: Bug 3839176: the test is waived on Confidential Computing because
    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(gpu))
+    if (g_uvm_global.conf_computing_enabled)
        return NV_OK;

    for_each_va_space_gpu(gpu, va_space) {
@@ -924,7 +1124,7 @@ static NV_STATUS test_channel_pushbuffer_extension_base(uvm_va_space_t *va_space
        uvm_channel_manager_t *manager;
        uvm_channel_pool_t *pool;

-        if (!uvm_gpu_has_pushbuffer_segments(gpu))
+        if (!uvm_parent_gpu_needs_pushbuffer_segments(gpu->parent))
            continue;

        // The GPU channel manager pushbuffer is destroyed and then re-created
@@ -999,6 +1199,10 @@ NV_STATUS uvm_test_channel_sanity(UVM_TEST_CHANNEL_SANITY_PARAMS *params, struct
    if (status != NV_OK)
        goto done;

+    status = test_channel_iv_rotation(va_space);
+    if (status != NV_OK)
+        goto done;
+
    // The following tests have side effects, they reset the GPU's
    // channel_manager.
    status = test_channel_pushbuffer_extension_base(va_space);
@@ -1019,6 +1223,10 @@ NV_STATUS uvm_test_channel_sanity(UVM_TEST_CHANNEL_SANITY_PARAMS *params, struct
            goto done;
    }

+    status = test_iommu(va_space);
+    if (status != NV_OK)
+        goto done;
+
 done:
    uvm_va_space_up_read_rm(va_space);
    uvm_mutex_unlock(&g_uvm_global.global_lock);
@@ -1034,23 +1242,22 @@ static NV_STATUS uvm_test_channel_stress_stream(uvm_va_space_t *va_space,
    if (params->iterations == 0 || params->num_streams == 0)
        return NV_ERR_INVALID_PARAMETER;

+    // TODO: Bug 3839176: the test is waived on Confidential Computing because
+    // it assumes that GPU can access system memory without using encryption.
+    if (g_uvm_global.conf_computing_enabled)
+        return NV_OK;
+
    // TODO: Bug 1764963: Rework the test to not rely on the global lock as that
    // serializes all the threads calling this at the same time.
    uvm_mutex_lock(&g_uvm_global.global_lock);
    uvm_va_space_down_read_rm(va_space);

-    // TODO: Bug 3839176: the test is waived on Confidential Computing because
-    // it assumes that GPU can access system memory without using encryption.
-    if (uvm_conf_computing_mode_enabled(uvm_va_space_find_first_gpu(va_space)))
-        goto done;
-
    status = stress_test_all_gpus_in_va(va_space,
                                        params->num_streams,
                                        params->iterations,
                                        params->seed,
                                        params->verbose);

-done:
    uvm_va_space_up_read_rm(va_space);
    uvm_mutex_unlock(&g_uvm_global.global_lock);

--- a/kernel-open/nvidia-uvm/uvm_common.c
+++ b/kernel-open/nvidia-uvm/uvm_common.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2013-2021 NVIDIA Corporation
+    Copyright (c) 2013-2023 NVIDIA Corporation

    This program is free software; you can redistribute it and/or
    modify it under the terms of the GNU General Public License
@@ -233,18 +233,6 @@ unsigned uvm_get_stale_thread_id(void)
    return (unsigned)task_pid_vnr(current);
 }

-//
-// A simple security rule for allowing access to UVM user space memory: if you
-// are the same user as the owner of the memory, or if you are root, then you
-// are granted access. The idea is to allow debuggers and profilers to work, but
-// without opening up any security holes.
-//
-NvBool uvm_user_id_security_check(uid_t euidTarget)
-{
-    return (NV_CURRENT_EUID() == euidTarget) ||
-           (UVM_ROOT_UID == euidTarget);
-}
-
 void on_uvm_test_fail(void)
 {
    (void)NULL;
@@ -330,10 +318,11 @@ int format_uuid_to_buffer(char *buffer, unsigned bufferLength, const NvProcessor
    unsigned i;
    unsigned dashMask = 1 << 4 | 1 << 6 | 1 << 8 | 1 << 10;

-    memcpy(buffer, "UVM-GPU-", 8);
    if (bufferLength < (8 /*prefix*/+ 16 * 2 /*digits*/ + 4 * 1 /*dashes*/ + 1 /*null*/))
        return *buffer = 0;

+    memcpy(buffer, "UVM-GPU-", 8);
+
    for (i = 0; i < 16; i++) {
        *str++ = uvm_digit_to_hex(pUuidStruct->uuid[i] >> 4);
        *str++ = uvm_digit_to_hex(pUuidStruct->uuid[i] & 0xF);
--- a/kernel-open/nvidia-uvm/uvm_common.h
+++ b/kernel-open/nvidia-uvm/uvm_common.h
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2013-2021 NVIDIA Corporation
+    Copyright (c) 2013-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -21,8 +21,8 @@

 *******************************************************************************/

-#ifndef _UVM_COMMON_H
-#define _UVM_COMMON_H
+#ifndef __UVM_COMMON_H__
+#define __UVM_COMMON_H__

 #ifdef DEBUG
    #define UVM_IS_DEBUG() 1
@@ -204,13 +204,6 @@ extern bool uvm_release_asserts_set_global_error_for_tests;
 #define UVM_ASSERT_MSG_RELEASE(expr, fmt, ...)  _UVM_ASSERT_MSG_RELEASE(expr, #expr, ": " fmt, ##__VA_ARGS__)
 #define UVM_ASSERT_RELEASE(expr)                _UVM_ASSERT_MSG_RELEASE(expr, #expr, "\n")

-// Provide a short form of UUID's, typically for use in debug printing:
-#define ABBREV_UUID(uuid) (unsigned)(uuid)
-
-static inline NvBool uvm_uuid_is_cpu(const NvProcessorUuid *uuid)
-{
-    return memcmp(uuid, &NV_PROCESSOR_UUID_CPU_DEFAULT, sizeof(*uuid)) == 0;
-}
 #define UVM_SIZE_1KB (1024ULL)
 #define UVM_SIZE_1MB (1024 * UVM_SIZE_1KB)
 #define UVM_SIZE_1GB (1024 * UVM_SIZE_1MB)
@@ -282,9 +275,6 @@ static inline void kmem_cache_destroy_safe(struct kmem_cache **ppCache)
    }
 }

-static const uid_t UVM_ROOT_UID = 0;
-
-
 typedef struct
 {
    NvU64 start_time_ns;
@@ -335,7 +325,6 @@ NV_STATUS errno_to_nv_status(int errnoCode);
 int nv_status_to_errno(NV_STATUS status);
 unsigned uvm_get_stale_process_id(void);
 unsigned uvm_get_stale_thread_id(void);
-NvBool uvm_user_id_security_check(uid_t euidTarget);

 extern int uvm_enable_builtin_tests;

@@ -413,4 +402,40 @@ static inline void uvm_touch_page(struct page *page)
 // Return true if the VMA is one used by UVM managed allocations.
 bool uvm_vma_is_managed(struct vm_area_struct *vma);

-#endif /* _UVM_COMMON_H */
+static bool uvm_platform_uses_canonical_form_address(void)
+{
+    if (NVCPU_IS_PPC64LE)
+        return false;
+
+    return true;
+}
+
+// Similar to the GPU MMU HAL num_va_bits(), it returns the CPU's num_va_bits().
+static NvU32 uvm_cpu_num_va_bits(void)
+{
+    return fls64(TASK_SIZE - 1) + 1;
+}
+
+// Return the unaddressable range in a num_va_bits-wide VA space, [first, outer)
+static void uvm_get_unaddressable_range(NvU32 num_va_bits, NvU64 *first, NvU64 *outer)
+{
+    UVM_ASSERT(num_va_bits < 64);
+    UVM_ASSERT(first);
+    UVM_ASSERT(outer);
+
+    if (uvm_platform_uses_canonical_form_address()) {
+        *first = 1ULL << (num_va_bits - 1);
+        *outer = (NvU64)((NvS64)(1ULL << 63) >> (64 - num_va_bits));
+    }
+    else {
+        *first = 1ULL << num_va_bits;
+        *outer = ~0Ull;
+    }
+}
+
+static void uvm_cpu_get_unaddressable_range(NvU64 *first, NvU64 *outer)
+{
+    return uvm_get_unaddressable_range(uvm_cpu_num_va_bits(), first, outer);
+}
+
+#endif /* __UVM_COMMON_H__ */
--- a/kernel-open/nvidia-uvm/uvm_conf_computing.c
+++ b/kernel-open/nvidia-uvm/uvm_conf_computing.c
@@ -33,22 +33,34 @@
 #include "nv_uvm_interface.h"
 #include "uvm_va_block.h"

+// The maximum number of secure operations per push is:
+// UVM_MAX_PUSH_SIZE / min(CE encryption size, CE decryption size)
+// + 1 (tracking semaphore) =  128 * 1024 / 56 + 1 = 2342
+#define UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MIN 2342lu
+
+// Channels use 32-bit counters so the value after rotation is 0xffffffff.
+// setting the limit to this value (or higher) will result in rotation
+// on every check. However, pre-emptive rotation when submitting control
+// GPFIFO entries relies on the fact that multiple successive checks after
+// rotation do not trigger more rotations if there was no IV used in between.
+#define UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MAX 0xfffffffelu
+
+// Attempt rotation when two billion IVs are left. IV rotation call can fail if
+// the necessary locks are not available, so multiple attempts may be need for
+// IV rotation to succeed.
+#define UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_DEFAULT (1lu << 31)
+
+// Start rotating after 500 encryption/decryptions when running tests.
+#define UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_TESTS ((1lu << 32) - 500lu)
+static ulong uvm_conf_computing_channel_iv_rotation_limit = UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_DEFAULT;
+
+module_param(uvm_conf_computing_channel_iv_rotation_limit, ulong, S_IRUGO);

 static UvmGpuConfComputeMode uvm_conf_computing_get_mode(const uvm_parent_gpu_t *parent)
 {
    return parent->rm_info.gpuConfComputeCaps.mode;
 }

-bool uvm_conf_computing_mode_enabled_parent(const uvm_parent_gpu_t *parent)
-{
-    return uvm_conf_computing_get_mode(parent) != UVM_GPU_CONF_COMPUTE_MODE_NONE;
-}
-
-bool uvm_conf_computing_mode_enabled(const uvm_gpu_t *gpu)
-{
-    return uvm_conf_computing_mode_enabled_parent(gpu->parent);
-}
-
 bool uvm_conf_computing_mode_is_hcc(const uvm_gpu_t *gpu)
 {
    return uvm_conf_computing_get_mode(gpu->parent) == UVM_GPU_CONF_COMPUTE_MODE_HCC;
@@ -56,24 +68,20 @@ bool uvm_conf_computing_mode_is_hcc(const uvm_gpu_t *gpu)

 void uvm_conf_computing_check_parent_gpu(const uvm_parent_gpu_t *parent)
 {
-    uvm_gpu_t *first_gpu;
+    uvm_parent_gpu_t *other_parent;
+    UvmGpuConfComputeMode parent_mode = uvm_conf_computing_get_mode(parent);

    uvm_assert_mutex_locked(&g_uvm_global.global_lock);

    // The Confidential Computing state of the GPU should match that of the
    // system.
-    UVM_ASSERT(uvm_conf_computing_mode_enabled_parent(parent) == g_uvm_global.conf_computing_enabled);
-
-    // TODO: Bug 2844714: since we have no routine to traverse parent GPUs,
-    // find first child GPU and get its parent.
-    first_gpu = uvm_global_processor_mask_find_first_gpu(&g_uvm_global.retained_gpus);
-    if (first_gpu == NULL)
-        return;
+    UVM_ASSERT((parent_mode != UVM_GPU_CONF_COMPUTE_MODE_NONE) == g_uvm_global.conf_computing_enabled);

    // All GPUs derive Confidential Computing status from their parent. By
    // current policy all parent GPUs have identical Confidential Computing
    // status.
-    UVM_ASSERT(uvm_conf_computing_get_mode(parent) == uvm_conf_computing_get_mode(first_gpu->parent));
+    for_each_parent_gpu(other_parent)
+        UVM_ASSERT(parent_mode == uvm_conf_computing_get_mode(other_parent));
 }

 static void dma_buffer_destroy_locked(uvm_conf_computing_dma_buffer_pool_t *dma_buffer_pool,
@@ -187,15 +195,11 @@ static void dma_buffer_pool_add(uvm_conf_computing_dma_buffer_pool_t *dma_buffer
 static NV_STATUS conf_computing_dma_buffer_pool_init(uvm_conf_computing_dma_buffer_pool_t *dma_buffer_pool)
 {
    size_t i;
-    uvm_gpu_t *gpu;
    size_t num_dma_buffers = 32;
    NV_STATUS status = NV_OK;

    UVM_ASSERT(dma_buffer_pool->num_dma_buffers == 0);
-
-    gpu = dma_buffer_pool_to_gpu(dma_buffer_pool);
-
-    UVM_ASSERT(uvm_conf_computing_mode_enabled(gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);

    INIT_LIST_HEAD(&dma_buffer_pool->free_dma_buffers);
    uvm_mutex_init(&dma_buffer_pool->lock, UVM_LOCK_ORDER_CONF_COMPUTING_DMA_BUFFER_POOL);
@@ -352,7 +356,7 @@ NV_STATUS uvm_conf_computing_gpu_init(uvm_gpu_t *gpu)
 {
    NV_STATUS status;

-    if (!uvm_conf_computing_mode_enabled(gpu))
+    if (!g_uvm_global.conf_computing_enabled)
        return NV_OK;

    status = conf_computing_dma_buffer_pool_init(&gpu->conf_computing.dma_buffer_pool);
@@ -363,6 +367,20 @@ NV_STATUS uvm_conf_computing_gpu_init(uvm_gpu_t *gpu)
    if (status != NV_OK)
        goto error;

+    if (uvm_enable_builtin_tests && uvm_conf_computing_channel_iv_rotation_limit == UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_DEFAULT)
+        uvm_conf_computing_channel_iv_rotation_limit = UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_TESTS;
+
+    if (uvm_conf_computing_channel_iv_rotation_limit < UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MIN ||
+        uvm_conf_computing_channel_iv_rotation_limit > UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MAX) {
+        UVM_ERR_PRINT("Value of uvm_conf_computing_channel_iv_rotation_limit: %lu is outside of the safe "
+                      "range: <%lu, %lu>. Using the default value instead (%lu)\n",
+                      uvm_conf_computing_channel_iv_rotation_limit,
+                      UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MIN,
+                      UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MAX,
+                      UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_DEFAULT);
+        uvm_conf_computing_channel_iv_rotation_limit = UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_DEFAULT;
+    }
+
    return NV_OK;

 error:
@@ -384,9 +402,8 @@ void uvm_conf_computing_log_gpu_encryption(uvm_channel_t *channel, UvmCslIv *iv)
    status = nvUvmInterfaceCslIncrementIv(&channel->csl.ctx, UVM_CSL_OPERATION_DECRYPT, 1, iv);
    uvm_mutex_unlock(&channel->csl.ctx_lock);

-    // TODO: Bug 4014720: If nvUvmInterfaceCslIncrementIv returns with
-    // NV_ERR_INSUFFICIENT_RESOURCES then the IV needs to be rotated via
-    // nvUvmInterfaceCslRotateIv.
+    // IV rotation is done preemptively as needed, so the above
+    // call cannot return failure.
    UVM_ASSERT(status == NV_OK);
 }

@@ -398,9 +415,8 @@ void uvm_conf_computing_acquire_encryption_iv(uvm_channel_t *channel, UvmCslIv *
    status = nvUvmInterfaceCslIncrementIv(&channel->csl.ctx, UVM_CSL_OPERATION_ENCRYPT, 1, iv);
    uvm_mutex_unlock(&channel->csl.ctx_lock);

-    // TODO: Bug 4014720: If nvUvmInterfaceCslIncrementIv returns with
-    // NV_ERR_INSUFFICIENT_RESOURCES then the IV needs to be rotated via
-    // nvUvmInterfaceCslRotateIv.
+    // IV rotation is done preemptively as needed, so the above
+    // call cannot return failure.
    UVM_ASSERT(status == NV_OK);
 }

@@ -424,8 +440,8 @@ void uvm_conf_computing_cpu_encrypt(uvm_channel_t *channel,
                                      (NvU8 *) auth_tag_buffer);
    uvm_mutex_unlock(&channel->csl.ctx_lock);

-    // nvUvmInterfaceCslEncrypt fails when a 64-bit encryption counter
-    // overflows. This is not supposed to happen on CC.
+    // IV rotation is done preemptively as needed, so the above
+    // call cannot return failure.
    UVM_ASSERT(status == NV_OK);
 }

@@ -438,6 +454,16 @@ NV_STATUS uvm_conf_computing_cpu_decrypt(uvm_channel_t *channel,
 {
    NV_STATUS status;

+    // The CSL context associated with a channel can be used by multiple
+    // threads. The IV sequence is thus guaranteed only while the channel is
+    // "locked for push". The channel/push lock is released in
+    // "uvm_channel_end_push", and at that time the GPU encryption operations
+    // have not executed, yet. Therefore the caller has to use
+    // "uvm_conf_computing_log_gpu_encryption" to explicitly store IVs needed
+    // to perform CPU decryption and pass those IVs to this function after the
+    // push that did the encryption completes.
+    UVM_ASSERT(src_iv);
+
    uvm_mutex_lock(&channel->csl.ctx_lock);
    status = nvUvmInterfaceCslDecrypt(&channel->csl.ctx,
                                      size,
@@ -466,7 +492,7 @@ NV_STATUS uvm_conf_computing_fault_decrypt(uvm_parent_gpu_t *parent_gpu,
    // decryption is invoked as part of fault servicing.
    UVM_ASSERT(uvm_sem_is_locked(&parent_gpu->isr.replayable_faults.service_lock));

-    UVM_ASSERT(!uvm_parent_gpu_replayable_fault_buffer_is_uvm_owned(parent_gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);

    status = nvUvmInterfaceCslDecrypt(&parent_gpu->fault_buffer_info.rm_info.replayable.cslCtx,
                                      parent_gpu->fault_buffer_hal->entry_size(parent_gpu),
@@ -478,7 +504,9 @@ NV_STATUS uvm_conf_computing_fault_decrypt(uvm_parent_gpu_t *parent_gpu,
                                      (const NvU8 *) auth_tag_buffer);

    if (status != NV_OK)
-        UVM_ERR_PRINT("nvUvmInterfaceCslDecrypt() failed: %s, GPU %s\n", nvstatusToString(status), parent_gpu->name);
+        UVM_ERR_PRINT("nvUvmInterfaceCslDecrypt() failed: %s, GPU %s\n",
+                      nvstatusToString(status),
+                      uvm_parent_gpu_name(parent_gpu));

    return status;
 }
@@ -490,7 +518,7 @@ void uvm_conf_computing_fault_increment_decrypt_iv(uvm_parent_gpu_t *parent_gpu,
    // See comment in uvm_conf_computing_fault_decrypt
    UVM_ASSERT(uvm_sem_is_locked(&parent_gpu->isr.replayable_faults.service_lock));

-    UVM_ASSERT(!uvm_parent_gpu_replayable_fault_buffer_is_uvm_owned(parent_gpu));
+    UVM_ASSERT(g_uvm_global.conf_computing_enabled);

    status = nvUvmInterfaceCslIncrementIv(&parent_gpu->fault_buffer_info.rm_info.replayable.cslCtx,
                                          UVM_CSL_OPERATION_DECRYPT,
@@ -499,3 +527,101 @@ void uvm_conf_computing_fault_increment_decrypt_iv(uvm_parent_gpu_t *parent_gpu,

    UVM_ASSERT(status == NV_OK);
 }
+
+void uvm_conf_computing_query_message_pools(uvm_channel_t *channel,
+                                            NvU64 *remaining_encryptions,
+                                            NvU64 *remaining_decryptions)
+{
+    NV_STATUS status;
+
+    UVM_ASSERT(channel);
+    UVM_ASSERT(remaining_encryptions);
+    UVM_ASSERT(remaining_decryptions);
+
+    uvm_mutex_lock(&channel->csl.ctx_lock);
+    status = nvUvmInterfaceCslQueryMessagePool(&channel->csl.ctx, UVM_CSL_OPERATION_ENCRYPT, remaining_encryptions);
+    UVM_ASSERT(status == NV_OK);
+    UVM_ASSERT(*remaining_encryptions <= NV_U32_MAX);
+
+    status = nvUvmInterfaceCslQueryMessagePool(&channel->csl.ctx, UVM_CSL_OPERATION_DECRYPT, remaining_decryptions);
+    UVM_ASSERT(status == NV_OK);
+    UVM_ASSERT(*remaining_decryptions <= NV_U32_MAX);
+
+    // LCIC channels never use CPU encrypt/GPU decrypt
+    if (uvm_channel_is_lcic(channel))
+        UVM_ASSERT(*remaining_encryptions == NV_U32_MAX);
+
+    uvm_mutex_unlock(&channel->csl.ctx_lock);
+}
+
+static NV_STATUS uvm_conf_computing_rotate_channel_ivs_below_limit_internal(uvm_channel_t *channel, NvU64 limit)
+{
+    NV_STATUS status = NV_OK;
+    NvU64 remaining_encryptions, remaining_decryptions;
+    bool rotate_encryption_iv, rotate_decryption_iv;
+
+    UVM_ASSERT(uvm_channel_is_locked_for_push(channel) ||
+               (uvm_channel_is_lcic(channel) && uvm_channel_manager_is_wlc_ready(channel->pool->manager)));
+
+    uvm_conf_computing_query_message_pools(channel, &remaining_encryptions, &remaining_decryptions);
+
+    // Ignore decryption limit for SEC2, only CE channels support
+    // GPU encrypt/CPU decrypt. However, RM reports _some_ decrementing
+    // value for SEC2 decryption counter.
+    rotate_decryption_iv = (remaining_decryptions <= limit) && uvm_channel_is_ce(channel);
+    rotate_encryption_iv = remaining_encryptions <= limit;
+
+    if (!rotate_encryption_iv && !rotate_decryption_iv)
+        return NV_OK;
+
+    // Wait for all in-flight pushes. The caller needs to guarantee that there
+    // are no concurrent pushes created, e.g. by only calling rotate after
+    // a channel is locked_for_push.
+    status = uvm_channel_wait(channel);
+    if (status != NV_OK)
+        return status;
+
+    uvm_mutex_lock(&channel->csl.ctx_lock);
+
+    if (rotate_encryption_iv)
+        status = nvUvmInterfaceCslRotateIv(&channel->csl.ctx, UVM_CSL_OPERATION_ENCRYPT);
+
+    if (status == NV_OK && rotate_decryption_iv)
+        status = nvUvmInterfaceCslRotateIv(&channel->csl.ctx, UVM_CSL_OPERATION_DECRYPT);
+
+    uvm_mutex_unlock(&channel->csl.ctx_lock);
+
+    // Change the error to out of resources if the available IVs are running
+    // too low
+    if (status == NV_ERR_STATE_IN_USE &&
+        (remaining_encryptions < UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MIN ||
+         remaining_decryptions < UVM_CONF_COMPUTING_IV_REMAINING_LIMIT_MIN))
+        return NV_ERR_INSUFFICIENT_RESOURCES;
+
+    return status;
+}
+
+NV_STATUS uvm_conf_computing_rotate_channel_ivs_below_limit(uvm_channel_t *channel, NvU64 limit, bool retry_if_busy)
+{
+    NV_STATUS status;
+
+    do {
+        status = uvm_conf_computing_rotate_channel_ivs_below_limit_internal(channel, limit);
+    } while (retry_if_busy && status == NV_ERR_STATE_IN_USE);
+
+    // Hide "busy" error. The rotation will be retried at the next opportunity.
+    if (!retry_if_busy && status == NV_ERR_STATE_IN_USE)
+        status = NV_OK;
+
+    return status;
+}
+
+NV_STATUS uvm_conf_computing_maybe_rotate_channel_ivs(uvm_channel_t *channel)
+{
+    return uvm_conf_computing_rotate_channel_ivs_below_limit(channel, uvm_conf_computing_channel_iv_rotation_limit, false);
+}
+
+NV_STATUS uvm_conf_computing_maybe_rotate_channel_ivs_retry_busy(uvm_channel_t *channel)
+{
+    return uvm_conf_computing_rotate_channel_ivs_below_limit(channel, uvm_conf_computing_channel_iv_rotation_limit, true);
+}
--- a/kernel-open/nvidia-uvm/uvm_conf_computing.h
+++ b/kernel-open/nvidia-uvm/uvm_conf_computing.h
@@ -62,8 +62,6 @@

 void uvm_conf_computing_check_parent_gpu(const uvm_parent_gpu_t *parent);

-bool uvm_conf_computing_mode_enabled_parent(const uvm_parent_gpu_t *parent);
-bool uvm_conf_computing_mode_enabled(const uvm_gpu_t *gpu);
 bool uvm_conf_computing_mode_is_hcc(const uvm_gpu_t *gpu);

 typedef struct
@@ -199,4 +197,21 @@ NV_STATUS uvm_conf_computing_fault_decrypt(uvm_parent_gpu_t *parent_gpu,
 //
 // Locking: this function must be invoked while holding the replayable ISR lock.
 void uvm_conf_computing_fault_increment_decrypt_iv(uvm_parent_gpu_t *parent_gpu, NvU64 increment);
+
+// Query the number of remaining messages before IV needs to be rotated.
+void uvm_conf_computing_query_message_pools(uvm_channel_t *channel,
+                                            NvU64 *remaining_encryptions,
+                                            NvU64 *remaining_decryptions);
+
+// Check if there are more than uvm_conf_computing_channel_iv_rotation_limit
+// messages available in the channel and try to rotate if not.
+NV_STATUS uvm_conf_computing_maybe_rotate_channel_ivs(uvm_channel_t *channel);
+
+// Check if there are more than uvm_conf_computing_channel_iv_rotation_limit
+// messages available in the channel and rotate if not.
+NV_STATUS uvm_conf_computing_maybe_rotate_channel_ivs_retry_busy(uvm_channel_t *channel);
+
+// Check if there are fewer than 'limit' messages available in either direction
+// and rotate if not.
+NV_STATUS uvm_conf_computing_rotate_channel_ivs_below_limit(uvm_channel_t *channel, NvU64 limit, bool retry_if_busy);
 #endif // __UVM_CONF_COMPUTING_H__
--- a/kernel-open/nvidia-uvm/uvm_fault_buffer_flush_test.c
+++ b/kernel-open/nvidia-uvm/uvm_fault_buffer_flush_test.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2016-2019 NVIDIA Corporation
+    Copyright (c) 2016-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -34,24 +34,27 @@ NV_STATUS uvm_test_fault_buffer_flush(UVM_TEST_FAULT_BUFFER_FLUSH_PARAMS *params
    NV_STATUS status = NV_OK;
    uvm_va_space_t *va_space = uvm_va_space_get(filp);
    uvm_gpu_t *gpu;
-    uvm_global_processor_mask_t retained_gpus;
+    uvm_processor_mask_t *retained_gpus;
    NvU64 i;

-    uvm_global_processor_mask_zero(&retained_gpus);
+    retained_gpus = uvm_processor_mask_cache_alloc();
+    if (!retained_gpus)
+        return NV_ERR_NO_MEMORY;
+
+    uvm_processor_mask_zero(retained_gpus);

    uvm_va_space_down_read(va_space);

-    for_each_va_space_gpu(gpu, va_space) {
-        if (gpu->parent->replayable_faults_supported)
-            uvm_global_processor_mask_set(&retained_gpus, gpu->global_id);
-    }
+    uvm_processor_mask_and(retained_gpus, &va_space->faultable_processors, &va_space->registered_gpus);

-    uvm_global_mask_retain(&retained_gpus);
+    uvm_global_gpu_retain(retained_gpus);

    uvm_va_space_up_read(va_space);

-    if (uvm_global_processor_mask_empty(&retained_gpus))
-        return NV_ERR_INVALID_DEVICE;
+    if (uvm_processor_mask_empty(retained_gpus)) {
+        status = NV_ERR_INVALID_DEVICE;
+        goto out;
+    }

    for (i = 0; i < params->iterations; i++) {
        if (fatal_signal_pending(current)) {
@@ -59,11 +62,12 @@ NV_STATUS uvm_test_fault_buffer_flush(UVM_TEST_FAULT_BUFFER_FLUSH_PARAMS *params
            break;
        }

-        for_each_global_gpu_in_mask(gpu, &retained_gpus)
+        for_each_gpu_in_mask(gpu, retained_gpus)
            TEST_CHECK_GOTO(uvm_gpu_fault_buffer_flush(gpu) == NV_OK, out);
    }

 out:
-    uvm_global_mask_release(&retained_gpus);
+    uvm_global_gpu_release(retained_gpus);
+    uvm_processor_mask_cache_free(retained_gpus);
    return status;
 }
--- a/kernel-open/nvidia-uvm/uvm_get_rm_ptes_test.c
+++ b/kernel-open/nvidia-uvm/uvm_get_rm_ptes_test.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2016-2021 NVidia Corporation
+    Copyright (c) 2016-2023 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -168,7 +168,8 @@ static NV_STATUS test_get_rm_ptes_single_gpu(uvm_va_space_t *va_space, UVM_TEST_
    client = params->hClient;
    memory = params->hMemory;

-    // Note: This check is safe as single GPU test does not run on SLI enabled devices.
+    // Note: This check is safe as single GPU test does not run on SLI enabled
+    // devices.
    memory_mapping_gpu = uvm_va_space_get_gpu_by_uuid_with_gpu_va_space(va_space, &params->gpu_uuid);
    if (!memory_mapping_gpu)
        return NV_ERR_INVALID_DEVICE;
@@ -180,7 +181,7 @@ static NV_STATUS test_get_rm_ptes_single_gpu(uvm_va_space_t *va_space, UVM_TEST_
    if (status != NV_OK)
        return status;

-    TEST_CHECK_GOTO(uvm_processor_uuid_eq(&memory_info.uuid, &params->gpu_uuid), done);
+    TEST_CHECK_GOTO(uvm_uuid_eq(&memory_info.uuid, &params->gpu_uuid), done);

    TEST_CHECK_GOTO((memory_info.size == params->size), done);

--- a/kernel-open/nvidia-uvm/uvm_global.c
+++ b/kernel-open/nvidia-uvm/uvm_global.c
@@ -1,5 +1,5 @@
 /*******************************************************************************
-    Copyright (c) 2015-2022 NVIDIA Corporation
+    Copyright (c) 2015-2024 NVIDIA Corporation

    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to
@@ -27,6 +27,7 @@
 #include "uvm_gpu_replayable_faults.h"
 #include "uvm_mem.h"
 #include "uvm_perf_events.h"
+#include "uvm_processors.h"
 #include "uvm_procfs.h"
 #include "uvm_thread_context.h"
 #include "uvm_va_range.h"
@@ -121,6 +122,12 @@ NV_STATUS uvm_global_init(void)
    g_uvm_global.num_simulated_devices = 0;
    g_uvm_global.conf_computing_enabled = platform_info.confComputingEnabled;

+    status = uvm_processor_mask_cache_init();
+    if (status != NV_OK) {
+        UVM_ERR_PRINT("uvm_processor_mask_cache_init() failed: %s\n", nvstatusToString(status));
+        goto error;
+    }
+
    status = uvm_gpu_init();
    if (status != NV_OK) {
        UVM_ERR_PRINT("uvm_gpu_init() failed: %s\n", nvstatusToString(status));
@@ -223,6 +230,7 @@ void uvm_global_exit(void)
    uvm_mem_global_exit();
    uvm_pmm_sysmem_exit();
    uvm_gpu_exit();
+    uvm_processor_mask_cache_exit();

    if (g_uvm_global.rm_session_handle != 0)
        uvm_rm_locked_call_void(nvUvmInterfaceSessionDestroy(g_uvm_global.rm_session_handle));
@@ -241,19 +249,19 @@ void uvm_global_exit(void)

 // Signal to the top-half ISR whether calls from the RM's top-half ISR are to
 // be completed without processing.
-static void uvm_gpu_set_isr_suspended(uvm_gpu_t *gpu, bool is_suspended)
+static void uvm_parent_gpu_set_isr_suspended(uvm_parent_gpu_t *parent_gpu, bool is_suspended)
 {
-    uvm_spin_lock_irqsave(&gpu->parent->isr.interrupts_lock);
+    uvm_spin_lock_irqsave(&parent_gpu->isr.interrupts_lock);

-    gpu->parent->isr.is_suspended = is_suspended;
+    parent_gpu->isr.is_suspended = is_suspended;

-    uvm_spin_unlock_irqrestore(&gpu->parent->isr.interrupts_lock);
+    uvm_spin_unlock_irqrestore(&parent_gpu->isr.interrupts_lock);
 }

 static NV_STATUS uvm_suspend(void)
 {
    uvm_va_space_t *va_space = NULL;
-    uvm_global_gpu_id_t gpu_id;
+    uvm_gpu_id_t gpu_id;
    uvm_gpu_t *gpu;

    // Upon entry into this function, the following is true:
@@ -287,7 +295,7 @@ static NV_STATUS uvm_suspend(void)
    // Though global_lock isn't held here, pm.lock indirectly prevents the
    // addition and removal of GPUs, since these operations can currently
    // only occur in response to ioctl() calls.
-    for_each_global_gpu_id_in_mask(gpu_id, &g_uvm_global.retained_gpus) {
+    for_each_gpu_id_in_mask(gpu_id, &g_uvm_global.retained_gpus) {
        gpu = uvm_gpu_get(gpu_id);

        // Since fault buffer state may be lost across sleep cycles, UVM must
@@ -306,9 +314,9 @@ static NV_STATUS uvm_suspend(void)
        // interrupts in the bottom half in the future, the bottom half flush
        // below will no longer be able to guarantee that all outstanding
        // notifications have been handled.
-        uvm_gpu_access_counters_set_ignore(gpu, true);
+        uvm_parent_gpu_access_counters_set_ignore(gpu->parent, true);

-        uvm_gpu_set_isr_suspended(gpu, true);
+        uvm_parent_gpu_set_isr_suspended(gpu->parent, true);

        nv_kthread_q_flush(&gpu->parent->isr.bottom_half_q);

@@ -341,7 +349,7 @@ NV_STATUS uvm_suspend_entry(void)
 static NV_STATUS uvm_resume(void)
 {
    uvm_va_space_t *va_space = NULL;
-    uvm_global_gpu_id_t gpu_id;
+    uvm_gpu_id_t gpu_id;
    uvm_gpu_t *gpu;

    g_uvm_global.pm.is_suspended = false;
@@ -360,18 +368,18 @@ static NV_STATUS uvm_resume(void)
    uvm_mutex_unlock(&g_uvm_global.va_spaces.lock);

    // pm.lock is held in lieu of global_lock to prevent GPU addition/removal
-    for_each_global_gpu_id_in_mask(gpu_id, &g_uvm_global.retained_gpus) {
+    for_each_gpu_id_in_mask(gpu_id, &g_uvm_global.retained_gpus) {
        gpu = uvm_gpu_get(gpu_id);

        // Bring the fault buffer software state back in sync with the
        // hardware state.
-        uvm_gpu_fault_buffer_resume(gpu->parent);
+        uvm_parent_gpu_fault_buffer_resume(gpu->parent);

-        uvm_gpu_set_isr_suspended(gpu, false);
+        uvm_parent_gpu_set_isr_suspended(gpu->parent, false);

        // Reenable access counter interrupt processing unless notifications
        // have been set to be suppressed.
-        uvm_gpu_access_counters_set_ignore(gpu, false);
+        uvm_parent_gpu_access_counters_set_ignore(gpu->parent, false);
    }

    uvm_up_write(&g_uvm_global.pm.lock);
@@ -425,35 +433,36 @@ NV_STATUS uvm_global_reset_fatal_error(void)
    return nv_atomic_xchg(&g_uvm_global.fatal_error, NV_OK);
 }

-void uvm_global_mask_retain(const uvm_global_processor_mask_t *mask)
+void uvm_global_gpu_retain(const uvm_processor_mask_t *mask)
 {
    uvm_gpu_t *gpu;
-    for_each_global_gpu_in_mask(gpu, mask)
+
+    for_each_gpu_in_mask(gpu, mask)
        uvm_gpu_retain(gpu);
 }

-void uvm_global_mask_release(const uvm_global_processor_mask_t *mask)
+void uvm_global_gpu_release(const uvm_processor_mask_t *mask)
 {
-    uvm_global_gpu_id_t gpu_id;
+    uvm_gpu_id_t gpu_id;

-    if (uvm_global_processor_mask_empty(mask))
+    if (uvm_processor_mask_empty(mask))
        return;

    uvm_mutex_lock(&g_uvm_global.global_lock);

-    // Do not use for_each_global_gpu_in_mask as it reads the GPU state and it
-    // might get destroyed
-    for_each_global_gpu_id_in_mask(gpu_id, mask)
+    // Do not use for_each_gpu_in_mask as it reads the GPU state and it
+    // might get destroyed.
+    for_each_gpu_id_in_mask(gpu_id, mask)
        uvm_gpu_release_locked(uvm_gpu_get(gpu_id));

    uvm_mutex_unlock(&g_uvm_global.global_lock);
 }

-NV_STATUS uvm_global_mask_check_ecc_error(uvm_global_processor_mask_t *gpus)
+NV_STATUS uvm_global_gpu_check_ecc_error(uvm_processor_mask_t *gpus)
 {
    uvm_gpu_t *gpu;

-    for_each_global_gpu_in_mask(gpu, gpus) {
+    for_each_gpu_in_mask(gpu, gpus) {
        NV_STATUS status = uvm_gpu_check_ecc_error(gpu);
        if (status != NV_OK)
            return status;
--- a/kernel-open/nvidia-uvm/uvm_global.h
+++ b/kernel-open/nvidia-uvm/uvm_global.h
@@ -40,13 +40,13 @@ struct uvm_global_struct
    // Note that GPUs are added to this mask as the last step of add_gpu() and
    // removed from it as the first step of remove_gpu() implying that a GPU
    // that's being initialized or deinitialized will not be in it.
-    uvm_global_processor_mask_t retained_gpus;
+    uvm_processor_mask_t retained_gpus;

    // Array of the parent GPUs registered with UVM. Note that GPUs will have
-    // ids offset by 1 to accomodate the UVM_GLOBAL_ID_CPU so e.g.
-    // parent_gpus[0] will have GPU id = 1. A GPU entry is unused iff it does
-    // not exist (is a NULL pointer) in this table.
-    uvm_parent_gpu_t *parent_gpus[UVM_MAX_GPUS];
+    // ids offset by 1 to accomodate the UVM_ID_CPU so e.g., parent_gpus[0]
+    // will have GPU id = 1. A GPU entry is unused iff it does not exist
+    // (is a NULL pointer) in this table.
+    uvm_parent_gpu_t *parent_gpus[UVM_PARENT_ID_MAX_GPUS];

    // A global RM session (RM client)
    // Created on module load and destroyed on module unload
@@ -172,7 +172,7 @@ NV_STATUS uvm_resume_entry(void);
 // LOCKING: requires that you hold the global lock and gpu_table_lock
 static void uvm_global_add_parent_gpu(uvm_parent_gpu_t *parent_gpu)
 {
-    NvU32 gpu_index = uvm_id_gpu_index(parent_gpu->id);
+    NvU32 gpu_index = uvm_parent_id_gpu_index(parent_gpu->id);

    uvm_assert_mutex_locked(&g_uvm_global.global_lock);
    uvm_assert_spinlock_locked(&g_uvm_global.gpu_table_lock);
@@ -186,7 +186,7 @@ static void uvm_global_add_parent_gpu(uvm_parent_gpu_t *parent_gpu)
 // LOCKING: requires that you hold the global lock and gpu_table_lock
 static void uvm_global_remove_parent_gpu(uvm_parent_gpu_t *parent_gpu)
 {
-    NvU32 gpu_index = uvm_id_gpu_index(parent_gpu->id);
+    NvU32 gpu_index = uvm_parent_id_gpu_index(parent_gpu->id);

    uvm_assert_mutex_locked(&g_uvm_global.global_lock);
    uvm_assert_spinlock_locked(&g_uvm_global.gpu_table_lock);
@@ -201,47 +201,29 @@ static void uvm_global_remove_parent_gpu(uvm_parent_gpu_t *parent_gpu)
 //
 // LOCKING: requires that you hold the gpu_table_lock, the global lock, or have
 // retained at least one of the child GPUs.
-static uvm_parent_gpu_t *uvm_parent_gpu_get(uvm_gpu_id_t id)
+static uvm_parent_gpu_t *uvm_parent_gpu_get(uvm_parent_gpu_id_t id)
 {
-    return g_uvm_global.parent_gpus[uvm_id_gpu_index(id)];
+    return g_uvm_global.parent_gpus[uvm_parent_id_gpu_index(id)];
 }

-// Get a gpu by its global id.
+// Get a gpu by its GPU id.
 // Returns a pointer to the GPU object, or NULL if not found.
 //
 // LOCKING: requires that you hold the gpu_table_lock, the global_lock, or have
 // retained the gpu.
-static uvm_gpu_t *uvm_gpu_get(uvm_global_gpu_id_t global_gpu_id)
+static uvm_gpu_t *uvm_gpu_get(uvm_gpu_id_t gpu_id)
 {
    uvm_parent_gpu_t *parent_gpu;

-    parent_gpu = g_uvm_global.parent_gpus[uvm_id_gpu_index_from_global_gpu_id(global_gpu_id)];
+    parent_gpu = g_uvm_global.parent_gpus[uvm_parent_id_gpu_index_from_gpu_id(gpu_id)];
    if (!parent_gpu)
        return NULL;

-    return parent_gpu->gpus[uvm_global_id_sub_processor_index(global_gpu_id)];
+    return parent_gpu->gpus[uvm_id_sub_processor_index(gpu_id)];
 }

-// Get a gpu by its processor id.
-// Returns a pointer to the GPU object, or NULL if not found.
-//
-// LOCKING: requires that you hold the gpu_table_lock, the global_lock, or have
-// retained the gpu.
-static uvm_gpu_t *uvm_gpu_get_by_processor_id(uvm_processor_id_t id)
+static uvmGpuSessionHandle uvm_global_session_handle(void)
 {
-    uvm_global_gpu_id_t global_id = uvm_global_gpu_id_from_gpu_id(id);
-    uvm_gpu_t *gpu = uvm_gpu_get(global_id);
-
-    if (gpu)
-        UVM_ASSERT(!gpu->parent->smc.enabled);
-
-    return gpu;
-}
-
-static uvmGpuSessionHandle uvm_gpu_session_handle(uvm_gpu_t *gpu)
-{
-    if (gpu->parent->smc.enabled)
-        return gpu->smc.rm_session_handle;
    return g_uvm_global.rm_session_handle;
 }

@@ -294,56 +276,57 @@ static NV_STATUS uvm_global_get_status(void)
 // reset call was made.
 NV_STATUS uvm_global_reset_fatal_error(void);

-static uvm_gpu_t *uvm_global_processor_mask_find_first_gpu(const uvm_global_processor_mask_t *global_gpus)
+static uvm_gpu_t *uvm_processor_mask_find_first_gpu(const uvm_processor_mask_t *gpus)
 {
    uvm_gpu_t *gpu;
-    uvm_global_gpu_id_t gpu_id = uvm_global_processor_mask_find_first_gpu_id(global_gpus);
+    uvm_gpu_id_t gpu_id = uvm_processor_mask_find_first_gpu_id(gpus);

-    if (UVM_GLOBAL_ID_IS_INVALID(gpu_id))
+    if (UVM_ID_IS_INVALID(gpu_id))
        return NULL;

    gpu = uvm_gpu_get(gpu_id);

    // If there is valid GPU id in the mask, assert that the corresponding
    // uvm_gpu_t is present. Otherwise it would stop a
-    // for_each_global_gpu_in_mask() loop pre-maturely. Today, this could only
+    // for_each_gpu_in_mask() loop pre-maturely. Today, this could only
    // happen in remove_gpu() because the GPU being removed is deleted from the
    // global table very early.
-    UVM_ASSERT_MSG(gpu, "gpu_id %u\n", uvm_global_id_value(gpu_id));
+    UVM_ASSERT_MSG(gpu, "gpu_id %u\n", uvm_id_value(gpu_id));

    return gpu;
 }

-static uvm_gpu_t *__uvm_global_processor_mask_find_next_gpu(const uvm_global_processor_mask_t *global_gpus, uvm_gpu_t *gpu)
+static uvm_gpu_t *__uvm_processor_mask_find_next_gpu(const uvm_processor_mask_t *gpus, uvm_gpu_t *gpu)
 {
-    uvm_global_gpu_id_t gpu_id;
+    uvm_gpu_id_t gpu_id;

    UVM_ASSERT(gpu);

-    gpu_id = uvm_global_processor_mask_find_next_id(global_gpus, uvm_global_gpu_id_next(gpu->global_id));
-    if (UVM_GLOBAL_ID_IS_INVALID(gpu_id))
+    gpu_id = uvm_processor_mask_find_next_id(gpus, uvm_gpu_id_next(gpu->id));
+    if (UVM_ID_IS_INVALID(gpu_id))
        return NULL;

    gpu = uvm_gpu_get(gpu_id);

-    // See comment in uvm_global_processor_mask_find_first_gpu().
-    UVM_ASSERT_MSG(gpu, "gpu_id %u\n", uvm_global_id_value(gpu_id));
+    // See comment in uvm_processor_mask_find_first_gpu().
+    UVM_ASSERT_MSG(gpu, "gpu_id %u\n", uvm_id_value(gpu_id));

    return gpu;
 }

 // Helper to iterate over all GPUs in the input mask
-#define for_each_global_gpu_in_mask(gpu, global_mask)                                         \
-    for (gpu = uvm_global_processor_mask_find_first_gpu(global_mask);                         \
-         gpu != NULL;                                                                         \
-         gpu = __uvm_global_processor_mask_find_next_gpu(global_mask, gpu))
+#define for_each_gpu_in_mask(gpu, mask)                         \
+    for (gpu = uvm_processor_mask_find_first_gpu(mask);         \
+         gpu != NULL;                                           \
+         gpu = __uvm_processor_mask_find_next_gpu(mask, gpu))

-// Helper to iterate over all GPUs retained by the UVM driver (across all va spaces)
-#define for_each_global_gpu(gpu)                                                              \
-    for (({uvm_assert_mutex_locked(&g_uvm_global.global_lock);                                \
-           gpu = uvm_global_processor_mask_find_first_gpu(&g_uvm_global.retained_gpus);});    \
-           gpu != NULL;                                                                       \
-           gpu = __uvm_global_processor_mask_find_next_gpu(&g_uvm_global.retained_gpus, gpu))
+// Helper to iterate over all GPUs retained by the UVM driver
+// (across all va spaces).
+#define for_each_gpu(gpu)                                                              \
+    for (({uvm_assert_mutex_locked(&g_uvm_global.global_lock);                         \
+           gpu = uvm_processor_mask_find_first_gpu(&g_uvm_global.retained_gpus);});    \
+           gpu != NULL;                                                                \
+           gpu = __uvm_processor_mask_find_next_gpu(&g_uvm_global.retained_gpus, gpu))

 // LOCKING: Must hold either the global_lock or the gpu_table_lock
 static uvm_parent_gpu_t *uvm_global_find_next_parent_gpu(uvm_parent_gpu_t *parent_gpu)
@@ -351,7 +334,7 @@ static uvm_parent_gpu_t *uvm_global_find_next_parent_gpu(uvm_parent_gpu_t *paren
    NvU32 i;

    if (parent_gpu) {
-        NvU32 gpu_index = uvm_id_gpu_index(parent_gpu->id);
+        NvU32 gpu_index = uvm_parent_id_gpu_index(parent_gpu->id);
        i = gpu_index + 1;
    }
    else {
@@ -360,7 +343,7 @@ static uvm_parent_gpu_t *uvm_global_find_next_parent_gpu(uvm_parent_gpu_t *paren

    parent_gpu = NULL;

-    while (i < UVM_MAX_GPUS) {
+    while (i < UVM_PARENT_ID_MAX_GPUS) {
        if (g_uvm_global.parent_gpus[i]) {
            parent_gpu = g_uvm_global.parent_gpus[i];
            break;
@@ -376,18 +359,18 @@ static uvm_parent_gpu_t *uvm_global_find_next_parent_gpu(uvm_parent_gpu_t *paren
 static uvm_gpu_t *uvm_gpu_find_next_valid_gpu_in_parent(uvm_parent_gpu_t *parent_gpu, uvm_gpu_t *cur_gpu)
 {
    uvm_gpu_t *gpu = NULL;
-    uvm_global_gpu_id_t global_gpu_id;
+    uvm_gpu_id_t gpu_id;
    NvU32 sub_processor_index;
    NvU32 cur_sub_processor_index;

    UVM_ASSERT(parent_gpu);

-    global_gpu_id = uvm_global_gpu_id_from_gpu_id(parent_gpu->id);
-    cur_sub_processor_index = cur_gpu ? uvm_global_id_sub_processor_index(cur_gpu->global_id) : -1;
+    gpu_id = uvm_gpu_id_from_parent_gpu_id(parent_gpu->id);
+    cur_sub_processor_index = cur_gpu ? uvm_id_sub_processor_index(cur_gpu->id) : -1;

-    sub_processor_index = find_next_bit(parent_gpu->valid_gpus, UVM_ID_MAX_SUB_PROCESSORS, cur_sub_processor_index + 1);
-    if (sub_processor_index < UVM_ID_MAX_SUB_PROCESSORS) {
-        gpu = uvm_gpu_get(uvm_global_id_from_value(uvm_global_id_value(global_gpu_id) + sub_processor_index));
+    sub_processor_index = find_next_bit(parent_gpu->valid_gpus, UVM_PARENT_ID_MAX_SUB_PROCESSORS, cur_sub_processor_index + 1);
+    if (sub_processor_index < UVM_PARENT_ID_MAX_SUB_PROCESSORS) {
+        gpu = uvm_gpu_get(uvm_id_from_value(uvm_id_value(gpu_id) + sub_processor_index));
        UVM_ASSERT(gpu != NULL);
    }

@@ -407,18 +390,18 @@ static uvm_gpu_t *uvm_gpu_find_next_valid_gpu_in_parent(uvm_parent_gpu_t *parent
         (gpu) != NULL;                                                                         \
         (gpu) = uvm_gpu_find_next_valid_gpu_in_parent((parent_gpu), (gpu)))

-// Helper which calls uvm_gpu_retain on each GPU in mask
-void uvm_global_mask_retain(const uvm_global_processor_mask_t *mask);
+// Helper which calls uvm_gpu_retain() on each GPU in mask.
+void uvm_global_gpu_retain(const uvm_processor_mask_t *mask);

 // Helper which calls uvm_gpu_release_locked on each GPU in mask.
 //
 // LOCKING: this function takes and releases the global lock if the input mask
 //          is not empty
-void uvm_global_mask_release(const uvm_global_processor_mask_t *mask);
+void uvm_global_gpu_release(const uvm_processor_mask_t *mask);

 // Check for ECC errors for all GPUs in a mask
 // Notably this check cannot be performed where it's not safe to call into RM.
-NV_STATUS uvm_global_mask_check_ecc_error(uvm_global_processor_mask_t *gpus);
+NV_STATUS uvm_global_gpu_check_ecc_error(uvm_processor_mask_t *gpus);

 // Pre-allocate fault service contexts.
 NV_STATUS uvm_service_block_context_init(void);
@@ -426,4 +409,10 @@ NV_STATUS uvm_service_block_context_init(void);
 // Release fault service contexts if any exist.
 void uvm_service_block_context_exit(void);

+// Allocate a service block context
+uvm_service_block_context_t *uvm_service_block_context_alloc(struct mm_struct *mm);
+
+// Free a servic block context
+void uvm_service_block_context_free(uvm_service_block_context_t *service_context);
+
 #endif // __UVM_GLOBAL_H__
--- a/kernel-open/nvidia-uvm/uvm_gpu.c
+++ b/kernel-open/nvidia-uvm/uvm_gpu.c
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Bernhard Stoeckner	1795a8bb20	555.58.02	2024-07-01 14:52:22 +02:00
Milos Tijanic	af77e083a2	555.58	2024-06-27 19:21:59 +02:00
Bernhard Stoeckner	78d807e001	555.52.04	2024-06-05 09:39:46 +02:00
Bernhard Stoeckner	5a1c474040	555.42.02	2024-05-21 15:12:40 +02:00
Bernhard Stoeckner	083cd9cf17	550.78	2024-04-25 16:24:58 +02:00
Bernhard Stoeckner	ea4c27fad6	550.76	2024-04-17 17:23:37 +02:00
Bernhard Stoeckner	3bf16b890c	550.67	2024-03-19 16:56:28 +01:00
Bernhard Stoeckner	12933b2d3c	550.54.15	2024-03-18 17:52:11 +01:00
Bernhard Stoeckner	476bd34534	550.54.14	2024-02-23 16:37:56 +01:00
Bernhard Stoeckner	91676d6628	550.40.07	2024-01-24 18:28:48 +01:00
Bernhard Stöckner	bb2dac1f20	Update 20_build_bug.yml	2024-01-23 15:30:14 +01:00
Maneet Singh	4c29105335	545.29.06	2023-11-21 13:38:23 -08:00
Andy Ritger	be3cd9abcb	545.29.02	2023-10-31 16:31:08 -07:00
Andy Ritger	a2f89d6b59	Revert "545.29.03" This reverts commit `f364378a65`. 545.29.03 and 545.29.02 are functionally the same for purposes of open-gpu-kernel-modules, but there was poor NVIDIA-internal communication about which driver would actually be released. Revert 545.29.03 so that a subsequent commit can provide 545.29.02 cleanly.	2023-10-31 16:28:17 -07:00
Maneet Singh	f364378a65	545.29.03	2023-10-31 09:44:03 -07:00
Andy Ritger	b5bf85a8e3	545.23.06	2023-10-17 09:25:29 -07:00