diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 53 |
1 files changed, 47 insertions, 6 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index ac0a432a9bf7..e93850f2f3b1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -31,6 +31,7 @@ #include "amdgpu_amdkfd.h" #include "amdgpu_dma_buf.h" #include <uapi/linux/kfd_ioctl.h> +#include "amdgpu_xgmi.h" /* BO flag to indicate a KFD userptr BO */ #define AMDGPU_AMDKFD_USERPTR_BO (1ULL << 63) @@ -96,7 +97,7 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void) uint64_t mem; si_meminfo(&si); - mem = si.totalram - si.totalhigh; + mem = si.freeram - si.freehigh; mem *= si.mem_unit; spin_lock_init(&kfd_mem_limit.mem_limit_lock); @@ -119,6 +120,16 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void) */ #define ESTIMATE_PT_SIZE(mem_size) ((mem_size) >> 14) +static size_t amdgpu_amdkfd_acc_size(uint64_t size) +{ + size >>= PAGE_SHIFT; + size *= sizeof(dma_addr_t) + sizeof(void *); + + return __roundup_pow_of_two(sizeof(struct amdgpu_bo)) + + __roundup_pow_of_two(sizeof(struct ttm_tt)) + + PAGE_ALIGN(size); +} + static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, uint64_t size, u32 domain, bool sg) { @@ -127,8 +138,7 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed; int ret = 0; - acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size, - sizeof(struct amdgpu_bo)); + acc_size = amdgpu_amdkfd_acc_size(size); vram_needed = 0; if (domain == AMDGPU_GEM_DOMAIN_GTT) { @@ -175,8 +185,7 @@ static void unreserve_mem_limit(struct amdgpu_device *adev, { size_t acc_size; - acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size, - sizeof(struct amdgpu_bo)); + acc_size = amdgpu_amdkfd_acc_size(size); spin_lock(&kfd_mem_limit.mem_limit_lock); if (domain == AMDGPU_GEM_DOMAIN_GTT) { @@ -404,7 +413,10 @@ static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem) { struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); bool coherent = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT; + bool uncached = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED; uint32_t mapping_flags; + uint64_t pte_flags; + bool snoop = false; mapping_flags = AMDGPU_VM_PAGE_READABLE; if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE) @@ -425,12 +437,41 @@ static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem) AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; } break; + case CHIP_ALDEBARAN: + if (coherent && uncached) { + if (adev->gmc.xgmi.connected_to_cpu || + !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)) + snoop = true; + mapping_flags |= AMDGPU_VM_MTYPE_UC; + } else if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { + if (bo_adev == adev) { + mapping_flags |= AMDGPU_VM_MTYPE_RW; + if (adev->gmc.xgmi.connected_to_cpu) + snoop = true; + } else { + mapping_flags |= AMDGPU_VM_MTYPE_NC; + if (amdgpu_xgmi_same_hive(adev, bo_adev)) + snoop = true; + } + } else { + snoop = true; + if (adev->gmc.xgmi.connected_to_cpu) + /* system memory uses NC on A+A */ + mapping_flags |= AMDGPU_VM_MTYPE_NC; + else + mapping_flags |= coherent ? + AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; + } + break; default: mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; } - return amdgpu_gem_va_map_flags(adev, mapping_flags); + pte_flags = amdgpu_gem_va_map_flags(adev, mapping_flags); + pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; + + return pte_flags; } /* add_bo_to_vm - Add a BO to a VM |