580.94.16

kernel-open/nvidia-uvm/nvidia-uvm.Kbuild

@@ -62,6 +62,8 @@ NV_CONFTEST_TYPE_COMPILE_TESTS += migrate_vma_added_flags
 NV_CONFTEST_TYPE_COMPILE_TESTS += migrate_device_range
 NV_CONFTEST_TYPE_COMPILE_TESTS += handle_mm_fault_has_pt_regs_arg
 NV_CONFTEST_TYPE_COMPILE_TESTS += get_dev_pagemap_has_pgmap_arg
+NV_CONFTEST_TYPE_COMPILE_TESTS += dev_pagemap_ops_has_folio_free
+NV_CONFTEST_TYPE_COMPILE_TESTS += zone_device_page_init_has_order_arg
 NV_CONFTEST_TYPE_COMPILE_TESTS += mempolicy_has_unified_nodes
 NV_CONFTEST_TYPE_COMPILE_TESTS += mempolicy_has_home_node
 NV_CONFTEST_TYPE_COMPILE_TESTS += mpol_preferred_many_present

kernel-open/nvidia-uvm/uvm_hmm.c

@@ -74,6 +74,16 @@ module_param(uvm_disable_hmm, bool, 0444);
 #include "uvm_va_policy.h"
 #include "uvm_tools.h"
 
+//
+// Pass 0 as the order, when actual large order support is added this
+// function will need to be revisited
+//
+#if defined(NV_ZONE_DEVICE_PAGE_INIT_HAS_ORDER_ARG)
+#define ZONE_DEVICE_PAGE_INIT(page)   zone_device_page_init(page, 0)
+#else
+#define ZONE_DEVICE_PAGE_INIT(page)   zone_device_page_init(page)
+#endif
+
 // The function nv_PageSwapCache() wraps the check for page swap cache flag in
 // order to support a wide variety of kernel versions.
 // The function PageSwapCache() is removed after 32f51ead3d77 ("mm: remove
@@ -2146,7 +2156,7 @@ static void fill_dst_pfn(uvm_va_block_t *va_block,
 
         UVM_ASSERT(!page_count(dpage));
         UVM_ASSERT(!dpage->zone_device_data);
-        zone_device_page_init(dpage);
+        ZONE_DEVICE_PAGE_INIT(dpage);
         dpage->zone_device_data = gpu_chunk;
         atomic64_inc(&va_block->hmm.va_space->hmm.allocated_page_count);
     }

kernel-open/nvidia-uvm/uvm_pmm_gpu.c

@@ -3159,6 +3159,11 @@ static void devmem_page_free(struct page *page)
                                  &gpu->pmm.root_chunks.va_block_lazy_free_q_item);
 }
 
+static void devmem_folio_free(struct folio *folio)
+{
+    devmem_page_free(&folio->page);
+}
+
 // This is called by HMM when the CPU faults on a ZONE_DEVICE private entry.
 static vm_fault_t devmem_fault(struct vm_fault *vmf)
 {
@@ -3177,7 +3182,11 @@ static vm_fault_t devmem_fault_entry(struct vm_fault *vmf)
 
 static const struct dev_pagemap_ops uvm_pmm_devmem_ops =
 {
+#if defined(NV_PAGEMAP_OPS_HAS_FOLIO_FREE)
+    .folio_free = devmem_folio_free,
+#else
     .page_free = devmem_page_free,
+#endif
     .migrate_to_ram = devmem_fault_entry,
 };
 
@@ -3272,6 +3281,11 @@ static void device_p2p_page_free(struct page *page)
     page->zone_device_data = NULL;
     nv_kref_put(&p2p_mem->refcount, device_p2p_page_free_wake);
 }
+
+static void device_p2p_folio_free(struct folio *folio)
+{
+    device_p2p_page_free(&folio->page);
+}
 #endif
 
 #if UVM_CDMM_PAGES_SUPPORTED()
@@ -3280,9 +3294,18 @@ static void device_coherent_page_free(struct page *page)
     device_p2p_page_free(page);
 }
 
+static void device_coherent_folio_free(struct folio *folio)
+{
+    device_p2p_page_free(&folio->page);
+}
+
 static const struct dev_pagemap_ops uvm_device_coherent_pgmap_ops =
 {
+#if defined(NV_PAGEMAP_OPS_HAS_FOLIO_FREE)
+    .folio_free = device_coherent_folio_free,
+#else
     .page_free = device_coherent_page_free,
+#endif
 };
 
 static NV_STATUS uvm_pmm_cdmm_init(uvm_parent_gpu_t *parent_gpu)
@@ -3419,7 +3442,11 @@ static bool uvm_pmm_gpu_check_orphan_pages(uvm_pmm_gpu_t *pmm)
 
 static const struct dev_pagemap_ops uvm_device_p2p_pgmap_ops =
 {
+#if defined(NV_PAGEMAP_OPS_HAS_FOLIO_FREE)
+    .folio_free = device_p2p_folio_free,
+#else
     .page_free = device_p2p_page_free,
+#endif
 };
 
 void uvm_pmm_gpu_device_p2p_init(uvm_parent_gpu_t *parent_gpu)
@@ -3477,12 +3504,10 @@ void uvm_pmm_gpu_device_p2p_init(uvm_parent_gpu_t *parent_gpu)
 
 void uvm_pmm_gpu_device_p2p_deinit(uvm_parent_gpu_t *parent_gpu)
 {
-    unsigned long pci_start_pfn = pci_resource_start(parent_gpu->pci_dev,
-                                                     uvm_device_p2p_static_bar(parent_gpu)) >> PAGE_SHIFT;
-    struct page *p2p_page;
-
     if (parent_gpu->device_p2p_initialised && !uvm_parent_gpu_is_coherent(parent_gpu)) {
-        p2p_page = pfn_to_page(pci_start_pfn);
+        struct page *p2p_page = pfn_to_page(pci_resource_start(parent_gpu->pci_dev,
+                                            uvm_device_p2p_static_bar(parent_gpu)) >> PAGE_SHIFT);
+
         devm_memunmap_pages(&parent_gpu->pci_dev->dev, page_pgmap(p2p_page));
     }
 

kernel-open/nvidia-uvm/uvm_populate_pageable.c

@@ -21,6 +21,8 @@
 
 *******************************************************************************/
 
+#include <linux/sched.h>
+
 #include "uvm_common.h"
 #include "uvm_ioctl.h"
 #include "uvm_linux.h"
@@ -63,6 +65,12 @@ static NV_STATUS handle_fault(struct vm_area_struct *vma, unsigned long start, u
         ret = UVM_HANDLE_MM_FAULT(vma, start + (i * PAGE_SIZE), fault_flags);
         if (ret & VM_FAULT_ERROR)
             return errno_to_nv_status(vm_fault_to_errno(ret, fault_flags));
+
+       // Depending on the kernel version and the active preemption model,
+       // calls to handle_mm_fault might not have had a chance to check for
+       // for scheduling points. Insert an explicit yield point here to prevent
+       // large buffers from triggering soft lockup timeout.
+       cond_resched();
     }
 
     return NV_OK;