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 | 1109 |
1 files changed, 727 insertions, 382 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 888209eb8cec..0e9cfe99ae9e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -26,12 +26,12 @@ #include <linux/sched/task.h> #include "amdgpu_object.h" +#include "amdgpu_gem.h" #include "amdgpu_vm.h" #include "amdgpu_amdkfd.h" #include "amdgpu_dma_buf.h" - -/* BO flag to indicate a KFD userptr BO */ -#define AMDGPU_AMDKFD_USERPTR_BO (1ULL << 63) +#include <uapi/linux/kfd_ioctl.h> +#include "amdgpu_xgmi.h" /* Userptr restore delay, just long enough to allow consecutive VM * changes to accumulate @@ -47,12 +47,6 @@ static struct { spinlock_t mem_limit_lock; } kfd_mem_limit; -/* Struct used for amdgpu_amdkfd_bo_validate */ -struct amdgpu_vm_parser { - uint32_t domain; - bool wait; -}; - static const char * const domain_bit_to_string[] = { "CPU", "GTT", @@ -72,20 +66,20 @@ static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd) return (struct amdgpu_device *)kgd; } -static bool check_if_add_bo_to_vm(struct amdgpu_vm *avm, +static bool kfd_mem_is_attached(struct amdgpu_vm *avm, struct kgd_mem *mem) { - struct kfd_bo_va_list *entry; + struct kfd_mem_attachment *entry; - list_for_each_entry(entry, &mem->bo_va_list, bo_list) + list_for_each_entry(entry, &mem->attachments, list) if (entry->bo_va->base.vm == avm) - return false; + return true; - return true; + return false; } /* Set memory usage limits. Current, limits are - * System (TTM + userptr) memory - 3/4th System RAM + * System (TTM + userptr) memory - 15/16th System RAM * TTM memory - 3/8th System RAM */ void amdgpu_amdkfd_gpuvm_init_mem_limits(void) @@ -94,17 +88,22 @@ 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); - kfd_mem_limit.max_system_mem_limit = (mem >> 1) + (mem >> 2); + kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4); kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3); pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n", (kfd_mem_limit.max_system_mem_limit >> 20), (kfd_mem_limit.max_ttm_mem_limit >> 20)); } +void amdgpu_amdkfd_reserve_system_mem(uint64_t size) +{ + kfd_mem_limit.system_mem_used += size; +} + /* Estimate page table size needed to represent a given memory size * * With 4KB pages, we need one 8 byte PTE for each 4KB of memory @@ -117,6 +116,16 @@ 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) { @@ -125,8 +134,7 @@ 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) { @@ -147,8 +155,12 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, spin_lock(&kfd_mem_limit.mem_limit_lock); + if (kfd_mem_limit.system_mem_used + system_mem_needed > + kfd_mem_limit.max_system_mem_limit) + pr_debug("Set no_system_mem_limit=1 if using shared memory\n"); + if ((kfd_mem_limit.system_mem_used + system_mem_needed > - kfd_mem_limit.max_system_mem_limit) || + kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) || (kfd_mem_limit.ttm_mem_used + ttm_mem_needed > kfd_mem_limit.max_ttm_mem_limit) || (adev->kfd.vram_used + vram_needed > @@ -169,8 +181,7 @@ 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) { @@ -202,7 +213,7 @@ void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo) u32 domain = bo->preferred_domains; bool sg = (bo->preferred_domains == AMDGPU_GEM_DOMAIN_CPU); - if (bo->flags & AMDGPU_AMDKFD_USERPTR_BO) { + if (bo->flags & AMDGPU_AMDKFD_CREATE_USERPTR_BO) { domain = AMDGPU_GEM_DOMAIN_CPU; sg = false; } @@ -230,12 +241,11 @@ static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo, if (!ef) return -EINVAL; - old = dma_resv_get_list(resv); + old = dma_resv_shared_list(resv); if (!old) return 0; - new = kmalloc(offsetof(typeof(*new), shared[old->shared_max]), - GFP_KERNEL); + new = kmalloc(struct_size(new, shared, old->shared_max), GFP_KERNEL); if (!new) return -ENOMEM; @@ -257,14 +267,12 @@ static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo, new->shared_count = k; /* Install the new fence list, seqcount provides the barriers */ - preempt_disable(); write_seqcount_begin(&resv->seq); RCU_INIT_POINTER(resv->fence, new); write_seqcount_end(&resv->seq); - preempt_enable(); /* Drop the references to the removed fences or move them to ef_list */ - for (i = j, k = 0; i < old->shared_count; ++i) { + for (i = j; i < old->shared_count; ++i) { struct dma_fence *f; f = rcu_dereference_protected(new->shared[i], @@ -276,6 +284,42 @@ static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo, return 0; } +int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo) +{ + struct amdgpu_bo *root = bo; + struct amdgpu_vm_bo_base *vm_bo; + struct amdgpu_vm *vm; + struct amdkfd_process_info *info; + struct amdgpu_amdkfd_fence *ef; + int ret; + + /* we can always get vm_bo from root PD bo.*/ + while (root->parent) + root = root->parent; + + vm_bo = root->vm_bo; + if (!vm_bo) + return 0; + + vm = vm_bo->vm; + if (!vm) + return 0; + + info = vm->process_info; + if (!info || !info->eviction_fence) + return 0; + + ef = container_of(dma_fence_get(&info->eviction_fence->base), + struct amdgpu_amdkfd_fence, base); + + BUG_ON(!dma_resv_trylock(bo->tbo.base.resv)); + ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef); + dma_resv_unlock(bo->tbo.base.resv); + + dma_fence_put(&ef->base); + return ret; +} + static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain, bool wait) { @@ -298,11 +342,9 @@ validate_fail: return ret; } -static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo) +static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo) { - struct amdgpu_vm_parser *p = param; - - return amdgpu_amdkfd_bo_validate(bo, p->domain, p->wait); + return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false); } /* vm_validate_pt_pd_bos - Validate page table and directory BOs @@ -314,33 +356,28 @@ static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo) */ static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm) { - struct amdgpu_bo *pd = vm->root.base.bo; + struct amdgpu_bo *pd = vm->root.bo; struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); - struct amdgpu_vm_parser param; int ret; - param.domain = AMDGPU_GEM_DOMAIN_VRAM; - param.wait = false; - - ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate, - ¶m); + ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL); if (ret) { - pr_err("amdgpu: failed to validate PT BOs\n"); + pr_err("failed to validate PT BOs\n"); return ret; } - ret = amdgpu_amdkfd_validate(¶m, pd); + ret = amdgpu_amdkfd_validate_vm_bo(NULL, pd); if (ret) { - pr_err("amdgpu: failed to validate PD\n"); + pr_err("failed to validate PD\n"); return ret; } - vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.base.bo); + vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo); if (vm->use_cpu_for_update) { ret = amdgpu_bo_kmap(pd, NULL); if (ret) { - pr_err("amdgpu: failed to kmap PD, ret=%d\n", ret); + pr_err("failed to kmap PD, ret=%d\n", ret); return ret; } } @@ -350,7 +387,7 @@ static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm) static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync) { - struct amdgpu_bo *pd = vm->root.base.bo; + struct amdgpu_bo *pd = vm->root.bo; struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); int ret; @@ -358,30 +395,58 @@ static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync) if (ret) return ret; - return amdgpu_sync_fence(NULL, sync, vm->last_update, false); + return amdgpu_sync_fence(sync, vm->last_update); } 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 & ALLOC_MEM_FLAGS_COHERENT; + 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 & ALLOC_MEM_FLAGS_WRITABLE) + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE) mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE; - if (mem->alloc_flags & ALLOC_MEM_FLAGS_EXECUTABLE) + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE) mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; switch (adev->asic_type) { case CHIP_ARCTURUS: - if (mem->alloc_flags & ALLOC_MEM_FLAGS_VRAM) { + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { if (bo_adev == adev) mapping_flags |= coherent ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; else - mapping_flags |= AMDGPU_VM_MTYPE_UC; + mapping_flags |= coherent ? + AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; + } else { + mapping_flags |= coherent ? + 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 |= coherent ? + AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; + if (adev->gmc.xgmi.connected_to_cpu) + snoop = true; + } else { + mapping_flags |= coherent ? + AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; + if (amdgpu_xgmi_same_hive(adev, bo_adev)) + snoop = true; + } } else { + snoop = true; mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; } @@ -391,90 +456,327 @@ static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem) 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; +} + +static int +kfd_mem_dmamap_userptr(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + enum dma_data_direction direction = + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DMA_BIDIRECTIONAL : DMA_TO_DEVICE; + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + struct amdgpu_device *adev = attachment->adev; + struct ttm_tt *src_ttm = mem->bo->tbo.ttm; + struct ttm_tt *ttm = bo->tbo.ttm; + int ret; + + ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL); + if (unlikely(!ttm->sg)) + return -ENOMEM; + + if (WARN_ON(ttm->num_pages != src_ttm->num_pages)) + return -EINVAL; + + /* Same sequence as in amdgpu_ttm_tt_pin_userptr */ + ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages, + ttm->num_pages, 0, + (u64)ttm->num_pages << PAGE_SHIFT, + GFP_KERNEL); + if (unlikely(ret)) + goto free_sg; + + ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0); + if (unlikely(ret)) + goto release_sg; + + drm_prime_sg_to_dma_addr_array(ttm->sg, ttm->dma_address, + ttm->num_pages); + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + if (ret) + goto unmap_sg; + + return 0; + +unmap_sg: + dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); +release_sg: + pr_err("DMA map userptr failed: %d\n", ret); + sg_free_table(ttm->sg); +free_sg: + kfree(ttm->sg); + ttm->sg = NULL; + return ret; +} + +static int +kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment) +{ + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); +} + +static int +kfd_mem_dmamap_attachment(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + switch (attachment->type) { + case KFD_MEM_ATT_SHARED: + return 0; + case KFD_MEM_ATT_USERPTR: + return kfd_mem_dmamap_userptr(mem, attachment); + case KFD_MEM_ATT_DMABUF: + return kfd_mem_dmamap_dmabuf(attachment); + default: + WARN_ON_ONCE(1); + } + return -EINVAL; +} + +static void +kfd_mem_dmaunmap_userptr(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + enum dma_data_direction direction = + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DMA_BIDIRECTIONAL : DMA_TO_DEVICE; + struct ttm_operation_ctx ctx = {.interruptible = false}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + struct amdgpu_device *adev = attachment->adev; + struct ttm_tt *ttm = bo->tbo.ttm; + + if (unlikely(!ttm->sg)) + return; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); + ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + + dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); + sg_free_table(ttm->sg); + kfree(ttm->sg); + ttm->sg = NULL; +} + +static void +kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment) +{ + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); + ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); +} + +static void +kfd_mem_dmaunmap_attachment(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + switch (attachment->type) { + case KFD_MEM_ATT_SHARED: + break; + case KFD_MEM_ATT_USERPTR: + kfd_mem_dmaunmap_userptr(mem, attachment); + break; + case KFD_MEM_ATT_DMABUF: + kfd_mem_dmaunmap_dmabuf(attachment); + break; + default: + WARN_ON_ONCE(1); + } +} + +static int +kfd_mem_attach_userptr(struct amdgpu_device *adev, struct kgd_mem *mem, + struct amdgpu_bo **bo) +{ + unsigned long bo_size = mem->bo->tbo.base.size; + struct drm_gem_object *gobj; + int ret; + + ret = amdgpu_bo_reserve(mem->bo, false); + if (ret) + return ret; + + ret = amdgpu_gem_object_create(adev, bo_size, 1, + AMDGPU_GEM_DOMAIN_CPU, + AMDGPU_GEM_CREATE_PREEMPTIBLE, + ttm_bo_type_sg, mem->bo->tbo.base.resv, + &gobj); + amdgpu_bo_unreserve(mem->bo); + if (ret) + return ret; + + *bo = gem_to_amdgpu_bo(gobj); + (*bo)->parent = amdgpu_bo_ref(mem->bo); + + return 0; +} + +static int +kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem, + struct amdgpu_bo **bo) +{ + struct drm_gem_object *gobj; + int ret; + + if (!mem->dmabuf) { + mem->dmabuf = amdgpu_gem_prime_export(&mem->bo->tbo.base, + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DRM_RDWR : 0); + if (IS_ERR(mem->dmabuf)) { + ret = PTR_ERR(mem->dmabuf); + mem->dmabuf = NULL; + return ret; + } + } + + gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf); + if (IS_ERR(gobj)) + return PTR_ERR(gobj); + + /* Import takes an extra reference on the dmabuf. Drop it now to + * avoid leaking it. We only need the one reference in + * kgd_mem->dmabuf. + */ + dma_buf_put(mem->dmabuf); + + *bo = gem_to_amdgpu_bo(gobj); + (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE; + (*bo)->parent = amdgpu_bo_ref(mem->bo); + + return 0; } -/* add_bo_to_vm - Add a BO to a VM +/* kfd_mem_attach - Add a BO to a VM * * Everything that needs to bo done only once when a BO is first added * to a VM. It can later be mapped and unmapped many times without * repeating these steps. * + * 0. Create BO for DMA mapping, if needed * 1. Allocate and initialize BO VA entry data structure * 2. Add BO to the VM * 3. Determine ASIC-specific PTE flags * 4. Alloc page tables and directories if needed * 4a. Validate new page tables and directories */ -static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem, - struct amdgpu_vm *vm, bool is_aql, - struct kfd_bo_va_list **p_bo_va_entry) +static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, + struct amdgpu_vm *vm, bool is_aql) { - int ret; - struct kfd_bo_va_list *bo_va_entry; - struct amdgpu_bo *bo = mem->bo; + struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); + unsigned long bo_size = mem->bo->tbo.base.size; uint64_t va = mem->va; - struct list_head *list_bo_va = &mem->bo_va_list; - unsigned long bo_size = bo->tbo.mem.size; + struct kfd_mem_attachment *attachment[2] = {NULL, NULL}; + struct amdgpu_bo *bo[2] = {NULL, NULL}; + int i, ret; if (!va) { pr_err("Invalid VA when adding BO to VM\n"); return -EINVAL; } - if (is_aql) - va += bo_size; - - bo_va_entry = kzalloc(sizeof(*bo_va_entry), GFP_KERNEL); - if (!bo_va_entry) - return -ENOMEM; + for (i = 0; i <= is_aql; i++) { + attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL); + if (unlikely(!attachment[i])) { + ret = -ENOMEM; + goto unwind; + } - pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va, - va + bo_size, vm); + pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va, + va + bo_size, vm); - /* Add BO to VM internal data structures*/ - bo_va_entry->bo_va = amdgpu_vm_bo_add(adev, vm, bo); - if (!bo_va_entry->bo_va) { - ret = -EINVAL; - pr_err("Failed to add BO object to VM. ret == %d\n", - ret); - goto err_vmadd; - } + if (adev == bo_adev || (mem->domain == AMDGPU_GEM_DOMAIN_VRAM && + amdgpu_xgmi_same_hive(adev, bo_adev))) { + /* Mappings on the local GPU and VRAM mappings in the + * local hive share the original BO + */ + attachment[i]->type = KFD_MEM_ATT_SHARED; + bo[i] = mem->bo; + drm_gem_object_get(&bo[i]->tbo.base); + } else if (i > 0) { + /* Multiple mappings on the same GPU share the BO */ + attachment[i]->type = KFD_MEM_ATT_SHARED; + bo[i] = bo[0]; + drm_gem_object_get(&bo[i]->tbo.base); + } else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { + /* Create an SG BO to DMA-map userptrs on other GPUs */ + attachment[i]->type = KFD_MEM_ATT_USERPTR; + ret = kfd_mem_attach_userptr(adev, mem, &bo[i]); + if (ret) + goto unwind; + } else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT && + mem->bo->tbo.type != ttm_bo_type_sg) { + /* GTT BOs use DMA-mapping ability of dynamic-attach + * DMA bufs. TODO: The same should work for VRAM on + * large-BAR GPUs. + */ + attachment[i]->type = KFD_MEM_ATT_DMABUF; + ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]); + if (ret) + goto unwind; + } else { + /* FIXME: Need to DMA-map other BO types: + * large-BAR VRAM, doorbells, MMIO remap + */ + attachment[i]->type = KFD_MEM_ATT_SHARED; + bo[i] = mem->bo; + drm_gem_object_get(&bo[i]->tbo.base); + } - bo_va_entry->va = va; - bo_va_entry->pte_flags = get_pte_flags(adev, mem); - bo_va_entry->kgd_dev = (void *)adev; - list_add(&bo_va_entry->bo_list, list_bo_va); + /* Add BO to VM internal data structures */ + attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]); + if (unlikely(!attachment[i]->bo_va)) { + ret = -ENOMEM; + pr_err("Failed to add BO object to VM. ret == %d\n", + ret); + goto unwind; + } - if (p_bo_va_entry) - *p_bo_va_entry = bo_va_entry; + attachment[i]->va = va; + attachment[i]->pte_flags = get_pte_flags(adev, mem); + attachment[i]->adev = adev; + list_add(&attachment[i]->list, &mem->attachments); - /* Allocate validate page tables if needed */ - ret = vm_validate_pt_pd_bos(vm); - if (ret) { - pr_err("validate_pt_pd_bos() failed\n"); - goto err_alloc_pts; + va += bo_size; } return 0; -err_alloc_pts: - amdgpu_vm_bo_rmv(adev, bo_va_entry->bo_va); - list_del(&bo_va_entry->bo_list); -err_vmadd: - kfree(bo_va_entry); +unwind: + for (; i >= 0; i--) { + if (!attachment[i]) + continue; + if (attachment[i]->bo_va) { + amdgpu_vm_bo_rmv(adev, attachment[i]->bo_va); + list_del(&attachment[i]->list); + } + if (bo[i]) + drm_gem_object_put(&bo[i]->tbo.base); + kfree(attachment[i]); + } return ret; } -static void remove_bo_from_vm(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, unsigned long size) +static void kfd_mem_detach(struct kfd_mem_attachment *attachment) { - pr_debug("\t remove VA 0x%llx - 0x%llx in entry %p\n", - entry->va, - entry->va + size, entry); - amdgpu_vm_bo_rmv(adev, entry->bo_va); - list_del(&entry->bo_list); - kfree(entry); + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + + pr_debug("\t remove VA 0x%llx in entry %p\n", + attachment->va, attachment); + amdgpu_vm_bo_rmv(attachment->adev, attachment->bo_va); + drm_gem_object_put(&bo->tbo.base); + list_del(&attachment->list); + kfree(attachment); } static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem, @@ -527,7 +829,7 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr) mutex_lock(&process_info->lock); - ret = amdgpu_ttm_tt_set_userptr(bo->tbo.ttm, user_addr, 0); + ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0); if (ret) { pr_err("%s: Failed to set userptr: %d\n", __func__, ret); goto out; @@ -623,15 +925,15 @@ static int reserve_bo_and_vm(struct kgd_mem *mem, ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list, false, &ctx->duplicates); - if (!ret) - ctx->reserved = true; - else { - pr_err("Failed to reserve buffers in ttm\n"); + if (ret) { + pr_err("Failed to reserve buffers in ttm.\n"); kfree(ctx->vm_pd); ctx->vm_pd = NULL; + return ret; } - return ret; + ctx->reserved = true; + return 0; } /** @@ -649,7 +951,7 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem, struct bo_vm_reservation_context *ctx) { struct amdgpu_bo *bo = mem->bo; - struct kfd_bo_va_list *entry; + struct kfd_mem_attachment *entry; unsigned int i; int ret; @@ -661,7 +963,7 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem, INIT_LIST_HEAD(&ctx->list); INIT_LIST_HEAD(&ctx->duplicates); - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { + list_for_each_entry(entry, &mem->attachments, list) { if ((vm && vm != entry->bo_va->base.vm) || (entry->is_mapped != map_type && map_type != BO_VM_ALL)) @@ -683,7 +985,7 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem, list_add(&ctx->kfd_bo.tv.head, &ctx->list); i = 0; - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { + list_for_each_entry(entry, &mem->attachments, list) { if ((vm && vm != entry->bo_va->base.vm) || (entry->is_mapped != map_type && map_type != BO_VM_ALL)) @@ -696,17 +998,15 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem, ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list, false, &ctx->duplicates); - if (!ret) - ctx->reserved = true; - else - pr_err("Failed to reserve buffers in ttm.\n"); - if (ret) { + pr_err("Failed to reserve buffers in ttm.\n"); kfree(ctx->vm_pd); ctx->vm_pd = NULL; + return ret; } - return ret; + ctx->reserved = true; + return 0; } /** @@ -739,47 +1039,56 @@ static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx, return ret; } -static int unmap_bo_from_gpuvm(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, +static void unmap_bo_from_gpuvm(struct kgd_mem *mem, + struct kfd_mem_attachment *entry, struct amdgpu_sync *sync) { struct amdgpu_bo_va *bo_va = entry->bo_va; + struct amdgpu_device *adev = entry->adev; struct amdgpu_vm *vm = bo_va->base.vm; amdgpu_vm_bo_unmap(adev, bo_va, entry->va); amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); - amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false); + amdgpu_sync_fence(sync, bo_va->last_pt_update); - return 0; + kfd_mem_dmaunmap_attachment(mem, entry); } -static int update_gpuvm_pte(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, - struct amdgpu_sync *sync) +static int update_gpuvm_pte(struct kgd_mem *mem, + struct kfd_mem_attachment *entry, + struct amdgpu_sync *sync, + bool *table_freed) { - int ret; struct amdgpu_bo_va *bo_va = entry->bo_va; + struct amdgpu_device *adev = entry->adev; + int ret; + + ret = kfd_mem_dmamap_attachment(mem, entry); + if (ret) + return ret; /* Update the page tables */ - ret = amdgpu_vm_bo_update(adev, bo_va, false); + ret = amdgpu_vm_bo_update(adev, bo_va, false, table_freed); if (ret) { pr_err("amdgpu_vm_bo_update failed\n"); return ret; } - return amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false); + return amdgpu_sync_fence(sync, bo_va->last_pt_update); } -static int map_bo_to_gpuvm(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, struct amdgpu_sync *sync, - bool no_update_pte) +static int map_bo_to_gpuvm(struct kgd_mem *mem, + struct kfd_mem_attachment *entry, + struct amdgpu_sync *sync, + bool no_update_pte, + bool *table_freed) { int ret; /* Set virtual address for the allocation */ - ret = amdgpu_vm_bo_map(adev, entry->bo_va, entry->va, 0, + ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0, amdgpu_bo_size(entry->bo_va->base.bo), entry->pte_flags); if (ret) { @@ -791,7 +1100,7 @@ static int map_bo_to_gpuvm(struct amdgpu_device *adev, if (no_update_pte) return 0; - ret = update_gpuvm_pte(adev, entry, sync); + ret = update_gpuvm_pte(mem, entry, sync, table_freed); if (ret) { pr_err("update_gpuvm_pte() failed\n"); goto update_gpuvm_pte_failed; @@ -800,7 +1109,7 @@ static int map_bo_to_gpuvm(struct amdgpu_device *adev, return 0; update_gpuvm_pte_failed: - unmap_bo_from_gpuvm(adev, entry, sync); + unmap_bo_from_gpuvm(mem, entry, sync); return ret; } @@ -845,11 +1154,11 @@ static int process_sync_pds_resv(struct amdkfd_process_info *process_info, list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { - struct amdgpu_bo *pd = peer_vm->root.base.bo; + struct amdgpu_bo *pd = peer_vm->root.bo; - ret = amdgpu_sync_resv(NULL, - sync, pd->tbo.base.resv, - AMDGPU_FENCE_OWNER_KFD, false); + ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv, + AMDGPU_SYNC_NE_OWNER, + AMDGPU_FENCE_OWNER_KFD); if (ret) return ret; } @@ -892,7 +1201,8 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, info->eviction_fence = amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), - current->mm); + current->mm, + NULL); if (!info->eviction_fence) { pr_err("Failed to create eviction fence\n"); ret = -ENOMEM; @@ -911,7 +1221,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, vm->process_info = *process_info; /* Validate page directory and attach eviction fence */ - ret = amdgpu_bo_reserve(vm->root.base.bo, true); + ret = amdgpu_bo_reserve(vm->root.bo, true); if (ret) goto reserve_pd_fail; ret = vm_validate_pt_pd_bos(vm); @@ -919,16 +1229,16 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, pr_err("validate_pt_pd_bos() failed\n"); goto validate_pd_fail; } - ret = amdgpu_bo_sync_wait(vm->root.base.bo, + ret = amdgpu_bo_sync_wait(vm->root.bo, AMDGPU_FENCE_OWNER_KFD, false); if (ret) goto wait_pd_fail; - ret = dma_resv_reserve_shared(vm->root.base.bo->tbo.base.resv, 1); + ret = dma_resv_reserve_shared(vm->root.bo->tbo.base.resv, 1); if (ret) goto reserve_shared_fail; - amdgpu_bo_fence(vm->root.base.bo, + amdgpu_bo_fence(vm->root.bo, &vm->process_info->eviction_fence->base, true); - amdgpu_bo_unreserve(vm->root.base.bo); + amdgpu_bo_unreserve(vm->root.bo); /* Update process info */ mutex_lock(&vm->process_info->lock); @@ -942,7 +1252,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, reserve_shared_fail: wait_pd_fail: validate_pd_fail: - amdgpu_bo_unreserve(vm->root.base.bo); + amdgpu_bo_unreserve(vm->root.bo); reserve_pd_fail: vm->process_info = NULL; if (info) { @@ -959,67 +1269,47 @@ create_evict_fence_fail: return ret; } -int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, unsigned int pasid, - void **vm, void **process_info, - struct dma_fence **ef) -{ - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *new_vm; - int ret; - - new_vm = kzalloc(sizeof(*new_vm), GFP_KERNEL); - if (!new_vm) - return -ENOMEM; - - /* Initialize AMDGPU part of the VM */ - ret = amdgpu_vm_init(adev, new_vm, AMDGPU_VM_CONTEXT_COMPUTE, pasid); - if (ret) { - pr_err("Failed init vm ret %d\n", ret); - goto amdgpu_vm_init_fail; - } - - /* Initialize KFD part of the VM and process info */ - ret = init_kfd_vm(new_vm, process_info, ef); - if (ret) - goto init_kfd_vm_fail; - - *vm = (void *) new_vm; - - return 0; - -init_kfd_vm_fail: - amdgpu_vm_fini(adev, new_vm); -amdgpu_vm_init_fail: - kfree(new_vm); - return ret; -} - int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd, - struct file *filp, unsigned int pasid, - void **vm, void **process_info, + struct file *filp, u32 pasid, + void **process_info, struct dma_fence **ef) { struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct drm_file *drm_priv = filp->private_data; - struct amdgpu_fpriv *drv_priv = drm_priv->driver_priv; - struct amdgpu_vm *avm = &drv_priv->vm; + struct amdgpu_fpriv *drv_priv; + struct amdgpu_vm *avm; int ret; + ret = amdgpu_file_to_fpriv(filp, &drv_priv); + if (ret) + return ret; + avm = &drv_priv->vm; + /* Already a compute VM? */ if (avm->process_info) return -EINVAL; + /* Free the original amdgpu allocated pasid, + * will be replaced with kfd allocated pasid. + */ + if (avm->pasid) { + amdgpu_pasid_free(avm->pasid); + amdgpu_vm_set_pasid(adev, avm, 0); + } + /* Convert VM into a compute VM */ - ret = amdgpu_vm_make_compute(adev, avm, pasid); + ret = amdgpu_vm_make_compute(adev, avm); if (ret) return ret; + ret = amdgpu_vm_set_pasid(adev, avm, pasid); + if (ret) + return ret; /* Initialize KFD part of the VM and process info */ ret = init_kfd_vm(avm, process_info, ef); if (ret) return ret; - *vm = (void *)avm; + amdgpu_vm_set_task_info(avm); return 0; } @@ -1028,7 +1318,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, struct amdgpu_vm *vm) { struct amdkfd_process_info *process_info = vm->process_info; - struct amdgpu_bo *pd = vm->root.base.bo; + struct amdgpu_bo *pd = vm->root.bo; if (!process_info) return; @@ -1044,6 +1334,8 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, list_del(&vm->vm_list_node); mutex_unlock(&process_info->lock); + vm->process_info = NULL; + /* Release per-process resources when last compute VM is destroyed */ if (!process_info->n_vms) { WARN_ON(!list_empty(&process_info->kfd_bo_list)); @@ -1058,44 +1350,31 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, } } -void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm) +void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *drm_priv) { struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + struct amdgpu_vm *avm; - if (WARN_ON(!kgd || !vm)) + if (WARN_ON(!kgd || !drm_priv)) return; - pr_debug("Destroying process vm %p\n", vm); - - /* Release the VM context */ - amdgpu_vm_fini(adev, avm); - kfree(vm); -} - -void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *vm) -{ - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; - - if (WARN_ON(!kgd || !vm)) - return; + avm = drm_priv_to_vm(drm_priv); - pr_debug("Releasing process vm %p\n", vm); + pr_debug("Releasing process vm %p\n", avm); - /* The original pasid of amdgpu vm has already been - * released during making a amdgpu vm to a compute vm - * The current pasid is managed by kfd and will be - * released on kfd process destroy. Set amdgpu pasid - * to 0 to avoid duplicate release. - */ + /* The original pasid of amdgpu vm has already been + * released during making a amdgpu vm to a compute vm + * The current pasid is managed by kfd and will be + * released on kfd process destroy. Set amdgpu pasid + * to 0 to avoid duplicate release. + */ amdgpu_vm_release_compute(adev, avm); } -uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm) +uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv) { - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; - struct amdgpu_bo *pd = avm->root.base.bo; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + struct amdgpu_bo *pd = avm->root.bo; struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); if (adev->asic_type < CHIP_VEGA10) @@ -1105,16 +1384,16 @@ uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm) int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( struct kgd_dev *kgd, uint64_t va, uint64_t size, - void *vm, struct kgd_mem **mem, + void *drm_priv, struct kgd_mem **mem, uint64_t *offset, uint32_t flags) { struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); enum ttm_bo_type bo_type = ttm_bo_type_device; struct sg_table *sg = NULL; uint64_t user_addr = 0; struct amdgpu_bo *bo; - struct amdgpu_bo_param bp; + struct drm_gem_object *gobj; u32 domain, alloc_domain; u64 alloc_flags; int ret; @@ -1122,24 +1401,23 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( /* * Check on which domain to allocate BO */ - if (flags & ALLOC_MEM_FLAGS_VRAM) { + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM; alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; - alloc_flags |= (flags & ALLOC_MEM_FLAGS_PUBLIC) ? - AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : - AMDGPU_GEM_CREATE_NO_CPU_ACCESS; - } else if (flags & ALLOC_MEM_FLAGS_GTT) { + alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ? + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0; + } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT; alloc_flags = 0; - } else if (flags & ALLOC_MEM_FLAGS_USERPTR) { + } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { domain = AMDGPU_GEM_DOMAIN_GTT; alloc_domain = AMDGPU_GEM_DOMAIN_CPU; - alloc_flags = 0; + alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE; if (!offset || !*offset) return -EINVAL; user_addr = untagged_addr(*offset); - } else if (flags & (ALLOC_MEM_FLAGS_DOORBELL | - ALLOC_MEM_FLAGS_MMIO_REMAP)) { + } else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL | + KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) { domain = AMDGPU_GEM_DOMAIN_GTT; alloc_domain = AMDGPU_GEM_DOMAIN_CPU; bo_type = ttm_bo_type_sg; @@ -1158,9 +1436,9 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( ret = -ENOMEM; goto err; } - INIT_LIST_HEAD(&(*mem)->bo_va_list); + INIT_LIST_HEAD(&(*mem)->attachments); mutex_init(&(*mem)->lock); - (*mem)->aql_queue = !!(flags & ALLOC_MEM_FLAGS_AQL_QUEUE_MEM); + (*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM); /* Workaround for AQL queue wraparound bug. Map the same * memory twice. That means we only actually allocate half @@ -1175,26 +1453,26 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, alloc_domain, !!sg); if (ret) { - pr_debug("Insufficient system memory\n"); + pr_debug("Insufficient memory\n"); goto err_reserve_limit; } pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n", va, size, domain_string(alloc_domain)); - memset(&bp, 0, sizeof(bp)); - bp.size = size; - bp.byte_align = 1; - bp.domain = alloc_domain; - bp.flags = alloc_flags; - bp.type = bo_type; - bp.resv = NULL; - ret = amdgpu_bo_create(adev, &bp, &bo); + ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags, + bo_type, NULL, &gobj); if (ret) { pr_debug("Failed to create BO on domain %s. ret %d\n", - domain_string(alloc_domain), ret); + domain_string(alloc_domain), ret); goto err_bo_create; } + ret = drm_vma_node_allow(&gobj->vma_node, drm_priv); + if (ret) { + pr_debug("Failed to allow vma node access. ret %d\n", ret); + goto err_node_allow; + } + bo = gem_to_amdgpu_bo(gobj); if (bo_type == ttm_bo_type_sg) { bo->tbo.sg = sg; bo->tbo.ttm->sg = sg; @@ -1202,7 +1480,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( bo->kfd_bo = *mem; (*mem)->bo = bo; if (user_addr) - bo->flags |= AMDGPU_AMDKFD_USERPTR_BO; + bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO; (*mem)->va = va; (*mem)->domain = domain; @@ -1223,7 +1501,9 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( allocate_init_user_pages_failed: remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); - amdgpu_bo_unref(&bo); + drm_vma_node_revoke(&gobj->vma_node, drm_priv); +err_node_allow: + drm_gem_object_put(gobj); /* Don't unreserve system mem limit twice */ goto err_reserve_limit; err_bo_create: @@ -1240,31 +1520,31 @@ err: } int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( - struct kgd_dev *kgd, struct kgd_mem *mem) + struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv, + uint64_t *size) { struct amdkfd_process_info *process_info = mem->process_info; - unsigned long bo_size = mem->bo->tbo.mem.size; - struct kfd_bo_va_list *entry, *tmp; + unsigned long bo_size = mem->bo->tbo.base.size; + struct kfd_mem_attachment *entry, *tmp; struct bo_vm_reservation_context ctx; struct ttm_validate_buffer *bo_list_entry; + unsigned int mapped_to_gpu_memory; int ret; + bool is_imported = false; mutex_lock(&mem->lock); - - if (mem->mapped_to_gpu_memory > 0) { - pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n", - mem->va, bo_size); - mutex_unlock(&mem->lock); - return -EBUSY; - } - + mapped_to_gpu_memory = mem->mapped_to_gpu_memory; + is_imported = mem->is_imported; mutex_unlock(&mem->lock); /* lock is not needed after this, since mem is unused and will * be freed anyway */ - /* No more MMU notifiers */ - amdgpu_mn_unregister(mem->bo); + if (mapped_to_gpu_memory > 0) { + pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n", + mem->va, bo_size); + return -EBUSY; + } /* Make sure restore workers don't access the BO any more */ bo_list_entry = &mem->validate_list; @@ -1272,6 +1552,9 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( list_del(&bo_list_entry->head); mutex_unlock(&process_info->lock); + /* No more MMU notifiers */ + amdgpu_mn_unregister(mem->bo); + ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx); if (unlikely(ret)) return ret; @@ -1285,13 +1568,12 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue)); - /* Remove from VM internal data structures */ - list_for_each_entry_safe(entry, tmp, &mem->bo_va_list, bo_list) - remove_bo_from_vm((struct amdgpu_device *)entry->kgd_dev, - entry, bo_size); - ret = unreserve_bo_and_vms(&ctx, false, false); + /* Remove from VM internal data structures */ + list_for_each_entry_safe(entry, tmp, &mem->attachments, list) + kfd_mem_detach(entry); + /* Free the sync object */ amdgpu_sync_free(&mem->sync); @@ -1303,8 +1585,22 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( kfree(mem->bo->tbo.sg); } + /* Update the size of the BO being freed if it was allocated from + * VRAM and is not imported. + */ + if (size) { + if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) && + (!is_imported)) + *size = bo_size; + else + *size = 0; + } + /* Free the BO*/ - amdgpu_bo_unref(&mem->bo); + drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv); + if (mem->dmabuf) + dma_buf_put(mem->dmabuf); + drm_gem_object_put(&mem->bo->tbo.base); mutex_destroy(&mem->lock); kfree(mem); @@ -1312,17 +1608,16 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( } int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( - struct kgd_dev *kgd, struct kgd_mem *mem, void *vm) + struct kgd_dev *kgd, struct kgd_mem *mem, + void *drm_priv, bool *table_freed) { struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); int ret; struct amdgpu_bo *bo; uint32_t domain; - struct kfd_bo_va_list *entry; + struct kfd_mem_attachment *entry; struct bo_vm_reservation_context ctx; - struct kfd_bo_va_list *bo_va_entry = NULL; - struct kfd_bo_va_list *bo_va_entry_aql = NULL; unsigned long bo_size; bool is_invalid_userptr = false; @@ -1343,22 +1638,28 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( * concurrently and the queues are actually stopped */ if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { - down_write(¤t->mm->mmap_sem); + mmap_write_lock(current->mm); is_invalid_userptr = atomic_read(&mem->invalid); - up_write(¤t->mm->mmap_sem); + mmap_write_unlock(current->mm); } mutex_lock(&mem->lock); domain = mem->domain; - bo_size = bo->tbo.mem.size; + bo_size = bo->tbo.base.size; pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue), - vm, domain_string(domain)); + avm, domain_string(domain)); - ret = reserve_bo_and_vm(mem, vm, &ctx); + if (!kfd_mem_is_attached(avm, mem)) { + ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue); + if (ret) + goto out; + } + + ret = reserve_bo_and_vm(mem, avm, &ctx); if (unlikely(ret)) goto out; @@ -1368,25 +1669,12 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( * the next restore worker */ if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && - bo->tbo.mem.mem_type == TTM_PL_SYSTEM) + bo->tbo.resource->mem_type == TTM_PL_SYSTEM) is_invalid_userptr = true; - if (check_if_add_bo_to_vm(avm, mem)) { - ret = add_bo_to_vm(adev, mem, avm, false, - &bo_va_entry); - if (ret) - goto add_bo_to_vm_failed; - if (mem->aql_queue) { - ret = add_bo_to_vm(adev, mem, avm, - true, &bo_va_entry_aql); - if (ret) - goto add_bo_to_vm_failed_aql; - } - } else { - ret = vm_validate_pt_pd_bos(avm); - if (unlikely(ret)) - goto add_bo_to_vm_failed; - } + ret = vm_validate_pt_pd_bos(avm); + if (unlikely(ret)) + goto out_unreserve; if (mem->mapped_to_gpu_memory == 0 && !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { @@ -1397,51 +1685,51 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( ret = amdgpu_amdkfd_bo_validate(bo, domain, true); if (ret) { pr_debug("Validate failed\n"); - goto map_bo_to_gpuvm_failed; + goto out_unreserve; } } - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { - if (entry->bo_va->base.vm == vm && !entry->is_mapped) { - pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n", - entry->va, entry->va + bo_size, - entry); + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != avm || entry->is_mapped) + continue; - ret = map_bo_to_gpuvm(adev, entry, ctx.sync, - is_invalid_userptr); - if (ret) { - pr_err("Failed to map bo to gpuvm\n"); - goto map_bo_to_gpuvm_failed; - } + pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n", + entry->va, entry->va + bo_size, entry); - ret = vm_update_pds(vm, ctx.sync); - if (ret) { - pr_err("Failed to update page directories\n"); - goto map_bo_to_gpuvm_failed; - } + ret = map_bo_to_gpuvm(mem, entry, ctx.sync, + is_invalid_userptr, table_freed); + if (ret) { + pr_err("Failed to map bo to gpuvm\n"); + goto out_unreserve; + } - entry->is_mapped = true; - mem->mapped_to_gpu_memory++; - pr_debug("\t INC mapping count %d\n", - mem->mapped_to_gpu_memory); + ret = vm_update_pds(avm, ctx.sync); + if (ret) { + pr_err("Failed to update page directories\n"); + goto out_unreserve; } + + entry->is_mapped = true; + mem->mapped_to_gpu_memory++; + pr_debug("\t INC mapping count %d\n", + mem->mapped_to_gpu_memory); } - if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->pin_count) + if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count) amdgpu_bo_fence(bo, &avm->process_info->eviction_fence->base, true); ret = unreserve_bo_and_vms(&ctx, false, false); + /* Only apply no TLB flush on Aldebaran to + * workaround regressions on other Asics. + */ + if (table_freed && (adev->asic_type != CHIP_ALDEBARAN)) + *table_freed = true; + goto out; -map_bo_to_gpuvm_failed: - if (bo_va_entry_aql) - remove_bo_from_vm(adev, bo_va_entry_aql, bo_size); -add_bo_to_vm_failed_aql: - if (bo_va_entry) - remove_bo_from_vm(adev, bo_va_entry, bo_size); -add_bo_to_vm_failed: +out_unreserve: unreserve_bo_and_vms(&ctx, false, false); out: mutex_unlock(&mem->process_info->lock); @@ -1450,19 +1738,18 @@ out: } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( - struct kgd_dev *kgd, struct kgd_mem *mem, void *vm) + struct kgd_dev *kgd, struct kgd_mem *mem, void *drm_priv) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdkfd_process_info *process_info = - ((struct amdgpu_vm *)vm)->process_info; - unsigned long bo_size = mem->bo->tbo.mem.size; - struct kfd_bo_va_list *entry; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + struct amdkfd_process_info *process_info = avm->process_info; + unsigned long bo_size = mem->bo->tbo.base.size; + struct kfd_mem_attachment *entry; struct bo_vm_reservation_context ctx; int ret; mutex_lock(&mem->lock); - ret = reserve_bo_and_cond_vms(mem, vm, BO_VM_MAPPED, &ctx); + ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx); if (unlikely(ret)) goto out; /* If no VMs were reserved, it means the BO wasn't actually mapped */ @@ -1471,42 +1758,36 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( goto unreserve_out; } - ret = vm_validate_pt_pd_bos((struct amdgpu_vm *)vm); + ret = vm_validate_pt_pd_bos(avm); if (unlikely(ret)) goto unreserve_out; pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue), - vm); - - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { - if (entry->bo_va->base.vm == vm && entry->is_mapped) { - pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n", - entry->va, - entry->va + bo_size, - entry); - - ret = unmap_bo_from_gpuvm(adev, entry, ctx.sync); - if (ret == 0) { - entry->is_mapped = false; - } else { - pr_err("failed to unmap VA 0x%llx\n", - mem->va); - goto unreserve_out; - } + avm); - mem->mapped_to_gpu_memory--; - pr_debug("\t DEC mapping count %d\n", - mem->mapped_to_gpu_memory); - } + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != avm || !entry->is_mapped) + continue; + + pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n", + entry->va, entry->va + bo_size, entry); + + unmap_bo_from_gpuvm(mem, entry, ctx.sync); + entry->is_mapped = false; + + mem->mapped_to_gpu_memory--; + pr_debug("\t DEC mapping count %d\n", + mem->mapped_to_gpu_memory); } /* If BO is unmapped from all VMs, unfence it. It can be evicted if * required. */ if (mem->mapped_to_gpu_memory == 0 && - !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && !mem->bo->pin_count) + !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && + !mem->bo->tbo.pin_count) amdgpu_amdkfd_remove_eviction_fence(mem->bo, process_info->eviction_fence); @@ -1590,6 +1871,16 @@ bo_reserve_failed: return ret; } +void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_dev *kgd, struct kgd_mem *mem) +{ + struct amdgpu_bo *bo = mem->bo; + + amdgpu_bo_reserve(bo, true); + amdgpu_bo_kunmap(bo); + amdgpu_bo_unpin(bo); + amdgpu_bo_unreserve(bo); +} + int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd, struct kfd_vm_fault_info *mem) { @@ -1606,21 +1897,22 @@ int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd, int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd, struct dma_buf *dma_buf, - uint64_t va, void *vm, + uint64_t va, void *drm_priv, struct kgd_mem **mem, uint64_t *size, uint64_t *mmap_offset) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); struct drm_gem_object *obj; struct amdgpu_bo *bo; - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + int ret; if (dma_buf->ops != &amdgpu_dmabuf_ops) /* Can't handle non-graphics buffers */ return -EINVAL; obj = dma_buf->priv; - if (obj->dev->dev_private != adev) + if (drm_to_adev(obj->dev) != adev) /* Can't handle buffers from other devices */ return -EINVAL; @@ -1634,20 +1926,29 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd, if (!*mem) return -ENOMEM; + ret = drm_vma_node_allow(&obj->vma_node, drm_priv); + if (ret) { + kfree(mem); + return ret; + } + if (size) *size = amdgpu_bo_size(bo); if (mmap_offset) *mmap_offset = amdgpu_bo_mmap_offset(bo); - INIT_LIST_HEAD(&(*mem)->bo_va_list); + INIT_LIST_HEAD(&(*mem)->attachments); mutex_init(&(*mem)->lock); + (*mem)->alloc_flags = ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? - ALLOC_MEM_FLAGS_VRAM : ALLOC_MEM_FLAGS_GTT) | - ALLOC_MEM_FLAGS_WRITABLE | ALLOC_MEM_FLAGS_EXECUTABLE; + KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT) + | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE + | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE; - (*mem)->bo = amdgpu_bo_ref(bo); + drm_gem_object_get(&bo->tbo.base); + (*mem)->bo = bo; (*mem)->va = va; (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT; @@ -1655,6 +1956,7 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd, (*mem)->process_info = avm->process_info; add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false); amdgpu_sync_create(&(*mem)->sync); + (*mem)->is_imported = true; return 0; } @@ -1674,10 +1976,10 @@ int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm) { struct amdkfd_process_info *process_info = mem->process_info; - int invalid, evicted_bos; + int evicted_bos; int r = 0; - invalid = atomic_inc_return(&mem->invalid); + atomic_inc(&mem->invalid); evicted_bos = atomic_inc_return(&process_info->evicted_bos); if (evicted_bos == 1) { /* First eviction, stop the queues */ @@ -1749,19 +2051,26 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, /* Get updated user pages */ ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages); if (ret) { - pr_debug("%s: Failed to get user pages: %d\n", - __func__, ret); + pr_debug("Failed %d to get user pages\n", ret); + + /* Return -EFAULT bad address error as success. It will + * fail later with a VM fault if the GPU tries to access + * it. Better than hanging indefinitely with stalled + * user mode queues. + * + * Return other error -EBUSY or -ENOMEM to retry restore + */ + if (ret != -EFAULT) + return ret; + } else { - /* Return error -EBUSY or -ENOMEM, retry restore */ - return ret; + /* + * FIXME: Cannot ignore the return code, must hold + * notifier_lock + */ + amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm); } - /* - * FIXME: Cannot ignore the return code, must hold - * notifier_lock - */ - amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm); - /* Mark the BO as valid unless it was invalidated * again concurrently. */ @@ -1833,7 +2142,7 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_inval_list, validate_list.head) { - struct kfd_bo_va_list *bo_va_entry; + struct kfd_mem_attachment *attachment; bo = mem->bo; @@ -1856,13 +2165,12 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) * VM faults if the GPU tries to access the invalid * memory. */ - list_for_each_entry(bo_va_entry, &mem->bo_va_list, bo_list) { - if (!bo_va_entry->is_mapped) + list_for_each_entry(attachment, &mem->attachments, list) { + if (!attachment->is_mapped) continue; - ret = update_gpuvm_pte((struct amdgpu_device *) - bo_va_entry->kgd_dev, - bo_va_entry, &sync); + kfd_mem_dmaunmap_attachment(mem, attachment); + ret = update_gpuvm_pte(mem, attachment, &sync, NULL); if (ret) { pr_err("%s: update PTE failed\n", __func__); /* make sure this gets validated again */ @@ -1987,6 +2295,8 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) int ret = 0, i; struct list_head duplicate_save; struct amdgpu_sync sync_obj; + unsigned long failed_size = 0; + unsigned long total_size = 0; INIT_LIST_HEAD(&duplicate_save); INIT_LIST_HEAD(&ctx.list); @@ -2041,24 +2351,32 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) struct amdgpu_bo *bo = mem->bo; uint32_t domain = mem->domain; - struct kfd_bo_va_list *bo_va_entry; + struct kfd_mem_attachment *attachment; + + total_size += amdgpu_bo_size(bo); ret = amdgpu_amdkfd_bo_validate(bo, domain, false); if (ret) { - pr_debug("Memory eviction: Validate BOs failed. Try again\n"); - goto validate_map_fail; + pr_debug("Memory eviction: Validate BOs failed\n"); + failed_size += amdgpu_bo_size(bo); + ret = amdgpu_amdkfd_bo_validate(bo, + AMDGPU_GEM_DOMAIN_GTT, false); + if (ret) { + pr_debug("Memory eviction: Try again\n"); + goto validate_map_fail; + } } - ret = amdgpu_sync_fence(NULL, &sync_obj, bo->tbo.moving, false); + ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving); if (ret) { pr_debug("Memory eviction: Sync BO fence failed. Try again\n"); goto validate_map_fail; } - list_for_each_entry(bo_va_entry, &mem->bo_va_list, - bo_list) { - ret = update_gpuvm_pte((struct amdgpu_device *) - bo_va_entry->kgd_dev, - bo_va_entry, - &sync_obj); + list_for_each_entry(attachment, &mem->attachments, list) { + if (!attachment->is_mapped) + continue; + + kfd_mem_dmaunmap_attachment(mem, attachment); + ret = update_gpuvm_pte(mem, attachment, &sync_obj, NULL); if (ret) { pr_debug("Memory eviction: update PTE failed. Try again\n"); goto validate_map_fail; @@ -2066,6 +2384,9 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) } } + if (failed_size) + pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size); + /* Update page directories */ ret = process_update_pds(process_info, &sync_obj); if (ret) { @@ -2082,7 +2403,8 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) */ new_fence = amdgpu_amdkfd_fence_create( process_info->eviction_fence->base.context, - process_info->eviction_fence->mm); + process_info->eviction_fence->mm, + NULL); if (!new_fence) { pr_err("Failed to create eviction fence\n"); ret = -ENOMEM; @@ -2101,7 +2423,7 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) /* Attach eviction fence to PD / PT BOs */ list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { - struct amdgpu_bo *bo = peer_vm->root.base.bo; + struct amdgpu_bo *bo = peer_vm->root.bo; amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true); } @@ -2129,6 +2451,7 @@ int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem return -ENOMEM; mutex_init(&(*mem)->lock); + INIT_LIST_HEAD(&(*mem)->attachments); (*mem)->bo = amdgpu_bo_ref(gws_bo); (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS; (*mem)->process_info = process_info; @@ -2203,3 +2526,25 @@ int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem) kfree(mem); return 0; } + +/* Returns GPU-specific tiling mode information */ +int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd, + struct tile_config *config) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + + config->gb_addr_config = adev->gfx.config.gb_addr_config; + config->tile_config_ptr = adev->gfx.config.tile_mode_array; + config->num_tile_configs = + ARRAY_SIZE(adev->gfx.config.tile_mode_array); + config->macro_tile_config_ptr = + adev->gfx.config.macrotile_mode_array; + config->num_macro_tile_configs = + ARRAY_SIZE(adev->gfx.config.macrotile_mode_array); + + /* Those values are not set from GFX9 onwards */ + config->num_banks = adev->gfx.config.num_banks; + config->num_ranks = adev->gfx.config.num_ranks; + + return 0; +} |