NVIDIA/open-gpu-kernel-modules
1
0
Fork 0
mirror of https://github.com/NVIDIA/open-gpu-kernel-modules.git synced 2026-02-01 05:59:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

580.65.06

Browse Source
This commit is contained in:
Maneet Singh
2025-08-04 11:15:02 -07:00
parent d890313300
commit 307159f262
1315 changed files with 477791 additions and 279973 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
kernel-open/nvidia-uvm/uvm_mem.c
View File

@@ -375,9 +375,10 @@ static void mem_free_sysmem_dma_chunks(uvm_mem_t *mem)
if (!mem->sysmem.va[i])
break;
uvm_parent_gpu_dma_free_page(mem->dma_owner->parent,
mem->sysmem.va[i],
mem->sysmem.dma_addrs[gpu_index][i]);
uvm_parent_gpu_dma_free(mem->chunk_size,
mem->dma_owner->parent,
mem->sysmem.va[i],
mem->sysmem.dma_addrs[gpu_index][i]);
}
end:
@@ -436,16 +437,16 @@ static NV_STATUS mem_alloc_dma_addrs(uvm_mem_t *mem, const uvm_gpu_t *gpu)
return NV_OK;
}
static gfp_t sysmem_allocation_gfp_flags(int order, bool zero)
static gfp_t sysmem_allocation_gfp_flags(uvm_mem_t *mem, bool zero)
{
gfp_t gfp_flags = NV_UVM_GFP_FLAGS;
if (zero)
gfp_flags |= __GFP_ZERO;
// High-order page allocations require the __GFP_COMP flag to work with
// High-order non-DMA page allocations require the __GFP_COMP flag to work with
// vm_insert_page.
if (order > 0)
if (get_order(mem->chunk_size) > 0 && !mem->dma_owner)
gfp_flags |= __GFP_COMP;
return gfp_flags;
@@ -468,10 +469,8 @@ static NV_STATUS mem_alloc_sysmem_dma_chunks(uvm_mem_t *mem, gfp_t gfp_flags)
NV_STATUS status;
NvU64 *dma_addrs;
UVM_ASSERT_MSG(mem->chunk_size == PAGE_SIZE,
"mem->chunk_size is 0x%llx. PAGE_SIZE is only supported.",
mem->chunk_size);
UVM_ASSERT(uvm_mem_is_sysmem_dma(mem));
UVM_ASSERT(!(gfp_flags & __GFP_COMP));
mem->sysmem.pages = uvm_kvmalloc_zero(sizeof(*mem->sysmem.pages) * mem->chunks_count);
mem->sysmem.va = uvm_kvmalloc_zero(sizeof(*mem->sysmem.va) * mem->chunks_count);
@@ -487,7 +486,7 @@ static NV_STATUS mem_alloc_sysmem_dma_chunks(uvm_mem_t *mem, gfp_t gfp_flags)
for (i = 0; i < mem->chunks_count; ++i) {
void *va;
status = uvm_gpu_dma_alloc_page(mem->dma_owner, gfp_flags, &va, &dma_addrs[i]);
status = uvm_gpu_dma_alloc(mem->chunk_size, mem->dma_owner, gfp_flags, &va, &dma_addrs[i]);
if (status != NV_OK)
goto error;
@@ -578,7 +577,7 @@ static NV_STATUS mem_alloc_chunks(uvm_mem_t *mem, struct mm_struct *mm, bool zer
NV_STATUS status;
UVM_ASSERT(PAGE_ALIGNED(mem->chunk_size));
gfp_flags = sysmem_allocation_gfp_flags(get_order(mem->chunk_size), zero);
gfp_flags = sysmem_allocation_gfp_flags(mem, zero);
if (UVM_CGROUP_ACCOUNTING_SUPPORTED() && mm)
gfp_flags |= NV_UVM_GFP_FLAGS_ACCOUNT;
@@ -961,6 +960,7 @@ static uvm_gpu_phys_address_t mem_gpu_physical_sysmem(uvm_mem_t *mem, uvm_gpu_t
UVM_ASSERT(uvm_mem_is_sysmem(mem));
UVM_ASSERT(sysmem_mapped_on_gpu_phys(mem, gpu));
// Sysmem should use coherent sys aperture
return uvm_gpu_phys_address(UVM_APERTURE_SYS, dma_addr + offset % mem->chunk_size);
}