NVIDIA/open-gpu-kernel-modules
1
0
Fork 0
mirror of https://github.com/NVIDIA/open-gpu-kernel-modules.git synced 2026-02-04 15:19:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

550.163.01

Browse Source
This commit is contained in:
Bernhard Stoeckner
2025-04-17 17:48:53 +02:00
parent ca09591fbd
commit 23e9e76214
62 changed files with 780 additions and 292 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
kernel-open/nvidia-uvm/uvm_conf_computing.c
View File

@@ -730,7 +730,12 @@ void uvm_conf_computing_disable_key_rotation(uvm_gpu_t *gpu)
bool uvm_conf_computing_is_key_rotation_enabled(uvm_gpu_t *gpu)
{
return gpu->channel_manager->conf_computing.key_rotation_enabled;
UVM_ASSERT(gpu);
// If the channel_manager is not set, we're in channel manager destroy
// path after the pointer was NULL-ed. Chances are that other key rotation
// infrastructure is not available either. Disallow the key rotation.
return gpu->channel_manager && gpu->channel_manager->conf_computing.key_rotation_enabled;
}
bool uvm_conf_computing_is_key_rotation_enabled_in_pool(uvm_channel_pool_t *pool)

2
kernel-open/nvidia-uvm/uvm_linux.h
View File

@@ -226,7 +226,7 @@ static inline const struct cpumask *uvm_cpumask_of_node(int node)
#define __GFP_NORETRY 0
#endif
#define NV_UVM_GFP_FLAGS (GFP_KERNEL)
#define NV_UVM_GFP_FLAGS (GFP_KERNEL | __GFP_NOMEMALLOC)
#ifndef NV_ALIGN_DOWN
#define NV_ALIGN_DOWN(v,g) ((v) & ~((g) - 1))

44
kernel-open/nvidia-uvm/uvm_migrate.c
View File

@@ -130,27 +130,12 @@ static NV_STATUS block_migrate_map_unmapped_pages(uvm_va_block_t *va_block,
uvm_tracker_t local_tracker = UVM_TRACKER_INIT();
NV_STATUS status = NV_OK;
NV_STATUS tracker_status;
uvm_prot_t prot = UVM_PROT_READ_WRITE_ATOMIC;
// Get the mask of unmapped pages because it will change after the
// first map operation
uvm_va_block_unmapped_pages_get(va_block, region, &va_block_context->caller_page_mask);
if (uvm_va_block_is_hmm(va_block) && !UVM_ID_IS_CPU(dest_id)) {
// Do not map pages that are already resident on the CPU. This is in
// order to avoid breaking system-wide atomic operations on HMM. HMM's
// implementation of system-side atomic operations involves restricting
// mappings to one processor (CPU or a GPU) at a time. If we were to
// grant a GPU a mapping to system memory, this gets into trouble
// because, on the CPU side, Linux can silently upgrade PTE permissions
// (move from read-only, to read-write, without any MMU notifiers
// firing), thus breaking the model by allowing simultaneous read-write
// access from two separate processors. To avoid that, just don't map
// such pages at all, when migrating.
uvm_page_mask_andnot(&va_block_context->caller_page_mask,
&va_block_context->caller_page_mask,
uvm_va_block_resident_mask_get(va_block, UVM_ID_CPU, NUMA_NO_NODE));
}
// Only map those pages that are not mapped anywhere else (likely due
// to a first touch or a migration). We pass
// UvmEventMapRemoteCauseInvalid since the destination processor of a
@@ -166,6 +151,31 @@ static NV_STATUS block_migrate_map_unmapped_pages(uvm_va_block_t *va_block,
if (status != NV_OK)
goto out;
if (uvm_va_block_is_hmm(va_block) && UVM_ID_IS_CPU(dest_id)) {
uvm_processor_id_t id;
// Do not atomically map pages that are resident on the CPU. This is in
// order to avoid breaking system-wide atomic operations on HMM. HMM's
// implementation of system-side atomic operations involves restricting
// mappings to one processor (CPU or a GPU) at a time. If we were to
// grant a GPU a mapping to system memory, this gets into trouble
// because, on the CPU side, Linux can silently upgrade PTE permissions
// (move from read-only, to read-write, without any MMU notifiers
// firing), thus breaking the model by allowing simultaneous read-write
// access from two separate processors. To avoid that, don't remote map
// such pages atomically, after migrating.
// Also note that HMM sets CPU mapping for resident pages so the mask
// of pages to be mapped needs to be recomputed without including the
// CPU mapping.
prot = UVM_PROT_READ_WRITE;
uvm_page_mask_region_fill(&va_block_context->caller_page_mask, region);
for_each_gpu_id_in_mask(id, &va_block->mapped) {
uvm_page_mask_andnot(&va_block_context->caller_page_mask,
&va_block_context->caller_page_mask,
uvm_va_block_map_mask_get(va_block, id));
}
}
// Add mappings for AccessedBy processors
//
// No mappings within this call will operate on dest_id, so we don't
@@ -176,7 +186,7 @@ static NV_STATUS block_migrate_map_unmapped_pages(uvm_va_block_t *va_block,
dest_id,
region,
&va_block_context->caller_page_mask,
UVM_PROT_READ_WRITE_ATOMIC,
prot,
NULL);
out:

2
kernel-open/nvidia-uvm/uvm_va_block.c
View File

@@ -1768,7 +1768,7 @@ static NV_STATUS block_alloc_cpu_chunk(uvm_va_block_t *block,
if (status == NV_OK)
break;
if (flags & UVM_CPU_CHUNK_ALLOC_FLAGS_STRICT) {
if ((flags & UVM_CPU_CHUNK_ALLOC_FLAGS_STRICT) && (num_possible_nodes() > 1)) {
flags &= ~UVM_CPU_CHUNK_ALLOC_FLAGS_STRICT;
numa_fallback = true;
status = block_alloc_cpu_chunk_inject_error(block, alloc_size, flags, NUMA_NO_NODE, chunk);