diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_svm.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 346 |
1 files changed, 235 insertions, 111 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 31e500859ab0..97c2270f278f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -309,12 +309,13 @@ static void svm_range_free(struct svm_range *prange, bool do_unmap) } static void -svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc, - uint8_t *granularity, uint32_t *flags) +svm_range_set_default_attributes(struct svm_range_list *svms, int32_t *location, + int32_t *prefetch_loc, uint8_t *granularity, + uint32_t *flags) { *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; - *granularity = 9; + *granularity = svms->default_granularity; *flags = KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT; } @@ -358,7 +359,7 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, bitmap_copy(prange->bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE); - svm_range_set_default_attributes(&prange->preferred_loc, + svm_range_set_default_attributes(svms, &prange->preferred_loc, &prange->prefetch_loc, &prange->granularity, &prange->flags); @@ -404,6 +405,27 @@ static void svm_range_bo_release(struct kref *kref) spin_lock(&svm_bo->list_lock); } spin_unlock(&svm_bo->list_lock); + + if (mmget_not_zero(svm_bo->eviction_fence->mm)) { + struct kfd_process_device *pdd; + struct kfd_process *p; + struct mm_struct *mm; + + mm = svm_bo->eviction_fence->mm; + /* + * The forked child process takes svm_bo device pages ref, svm_bo could be + * released after parent process is gone. + */ + p = kfd_lookup_process_by_mm(mm); + if (p) { + pdd = kfd_get_process_device_data(svm_bo->node, p); + if (pdd) + atomic64_sub(amdgpu_bo_size(svm_bo->bo), &pdd->vram_usage); + kfd_unref_process(p); + } + mmput(mm); + } + if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) /* We're not in the eviction worker. Signal the fence. */ dma_fence_signal(&svm_bo->eviction_fence->base); @@ -531,6 +553,7 @@ int svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, bool clear) { + struct kfd_process_device *pdd; struct amdgpu_bo_param bp; struct svm_range_bo *svm_bo; struct amdgpu_bo_user *ubo; @@ -540,7 +563,8 @@ svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, int r; p = container_of(prange->svms, struct kfd_process, svms); - pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms, + pr_debug("process pid: %d svms 0x%p [0x%lx 0x%lx]\n", + p->lead_thread->pid, prange->svms, prange->start, prange->last); if (svm_range_validate_svm_bo(node, prange)) @@ -622,6 +646,10 @@ svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, list_add(&prange->svm_bo_list, &svm_bo->range_list); spin_unlock(&svm_bo->list_lock); + pdd = svm_range_get_pdd_by_node(prange, node); + if (pdd) + atomic64_add(amdgpu_bo_size(bo), &pdd->vram_usage); + return 0; reserve_bo_failed: @@ -1051,6 +1079,7 @@ svm_range_split_adjust(struct svm_range *new, struct svm_range *old, new->mapped_to_gpu = old->mapped_to_gpu; bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); + atomic_set(&new->queue_refcount, atomic_read(&old->queue_refcount)); return 0; } @@ -1142,13 +1171,12 @@ svm_range_split_head(struct svm_range *prange, uint64_t new_start, } static void -svm_range_add_child(struct svm_range *prange, struct mm_struct *mm, - struct svm_range *pchild, enum svm_work_list_ops op) +svm_range_add_child(struct svm_range *prange, struct svm_range *pchild, enum svm_work_list_ops op) { pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n", pchild, pchild->start, pchild->last, prange, op); - pchild->work_item.mm = mm; + pchild->work_item.mm = NULL; pchild->work_item.op = op; list_add_tail(&pchild->child_list, &prange->child_list); } @@ -1161,23 +1189,23 @@ svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b) } static uint64_t -svm_range_get_pte_flags(struct kfd_node *node, +svm_range_get_pte_flags(struct kfd_node *node, struct amdgpu_vm *vm, struct svm_range *prange, int domain) { struct kfd_node *bo_node; uint32_t flags = prange->flags; uint32_t mapping_flags = 0; + uint32_t gc_ip_version = KFD_GC_VERSION(node); uint64_t pte_flags; bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); bool coherent = flags & (KFD_IOCTL_SVM_FLAG_COHERENT | KFD_IOCTL_SVM_FLAG_EXT_COHERENT); bool ext_coherent = flags & KFD_IOCTL_SVM_FLAG_EXT_COHERENT; - bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/ unsigned int mtype_local; if (domain == SVM_RANGE_VRAM_DOMAIN) bo_node = prange->svm_bo->node; - switch (amdgpu_ip_version(node->adev, GC_HWIP, 0)) { + switch (gc_ip_version) { case IP_VERSION(9, 4, 1): if (domain == SVM_RANGE_VRAM_DOMAIN) { if (bo_node == node) { @@ -1213,15 +1241,15 @@ svm_range_get_pte_flags(struct kfd_node *node, } break; case IP_VERSION(9, 4, 3): + case IP_VERSION(9, 4, 4): + case IP_VERSION(9, 5, 0): if (ext_coherent) - mtype_local = node->adev->rev_id ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_UC; + mtype_local = AMDGPU_VM_MTYPE_CC; else mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC : amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; snoop = true; - if (uncached) { - mapping_flags |= AMDGPU_VM_MTYPE_UC; - } else if (domain == SVM_RANGE_VRAM_DOMAIN) { + if (domain == SVM_RANGE_VRAM_DOMAIN) { /* local HBM region close to partition */ if (bo_node->adev == node->adev && (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id)) @@ -1231,9 +1259,13 @@ svm_range_get_pte_flags(struct kfd_node *node, */ else if (svm_nodes_in_same_hive(bo_node, node) && !ext_coherent) mapping_flags |= AMDGPU_VM_MTYPE_NC; - /* PCIe P2P or extended system scope coherence */ - else + /* PCIe P2P on GPUs pre-9.5.0 */ + else if (gc_ip_version < IP_VERSION(9, 5, 0) && + !svm_nodes_in_same_hive(bo_node, node)) mapping_flags |= AMDGPU_VM_MTYPE_UC; + /* Other remote memory */ + else + mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; /* system memory accessed by the APU */ } else if (node->adev->flags & AMD_IS_APU) { /* On NUMA systems, locality is determined per-page @@ -1245,26 +1277,34 @@ svm_range_get_pte_flags(struct kfd_node *node, mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; /* system memory accessed by the dGPU */ } else { - mapping_flags |= AMDGPU_VM_MTYPE_UC; + if (gc_ip_version < IP_VERSION(9, 5, 0) || ext_coherent) + mapping_flags |= AMDGPU_VM_MTYPE_UC; + else + mapping_flags |= AMDGPU_VM_MTYPE_NC; } break; + case IP_VERSION(12, 0, 0): + case IP_VERSION(12, 0, 1): + mapping_flags |= AMDGPU_VM_MTYPE_NC; + break; default: mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; } - mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE; - - if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO) - mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE; if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC) mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; pte_flags = AMDGPU_PTE_VALID; pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM; pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; + if (gc_ip_version >= IP_VERSION(12, 0, 0)) + pte_flags |= AMDGPU_PTE_IS_PTE; - pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags); + amdgpu_gmc_get_vm_pte(node->adev, vm, NULL, mapping_flags, &pte_flags); + pte_flags |= AMDGPU_PTE_READABLE; + if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO)) + pte_flags |= AMDGPU_PTE_WRITEABLE; return pte_flags; } @@ -1371,7 +1411,7 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n", last_start, prange->start + i, last_domain ? "GPU" : "CPU"); - pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain); + pte_flags = svm_range_get_pte_flags(pdd->dev, vm, prange, last_domain); if (readonly) pte_flags &= ~AMDGPU_PTE_WRITEABLE; @@ -1658,7 +1698,7 @@ static int svm_range_validate_and_map(struct mm_struct *mm, start = map_start << PAGE_SHIFT; end = (map_last + 1) << PAGE_SHIFT; for (addr = start; !r && addr < end; ) { - struct hmm_range *hmm_range; + struct amdgpu_hmm_range *range = NULL; unsigned long map_start_vma; unsigned long map_last_vma; struct vm_area_struct *vma; @@ -1673,16 +1713,39 @@ static int svm_range_validate_and_map(struct mm_struct *mm, next = min(vma->vm_end, end); npages = (next - addr) >> PAGE_SHIFT; + /* HMM requires at least READ permissions. If provided with PROT_NONE, + * unmap the memory. If it's not already mapped, this is a no-op + * If PROT_WRITE is provided without READ, warn first then unmap + */ + if (!(vma->vm_flags & VM_READ)) { + unsigned long e, s; + + svm_range_lock(prange); + if (vma->vm_flags & VM_WRITE) + pr_debug("VM_WRITE without VM_READ is not supported"); + s = max(start, prange->start); + e = min(end, prange->last); + if (e >= s) + r = svm_range_unmap_from_gpus(prange, s, e, + KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU); + svm_range_unlock(prange); + /* If unmap returns non-zero, we'll bail on the next for loop + * iteration, so just leave r and continue + */ + addr = next; + continue; + } + WRITE_ONCE(p->svms.faulting_task, current); - r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages, - readonly, owner, NULL, - &hmm_range); + range = amdgpu_hmm_range_alloc(NULL); + if (likely(range)) + r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages, + readonly, owner, range); + else + r = -ENOMEM; WRITE_ONCE(p->svms.faulting_task, NULL); - if (r) { + if (r) pr_debug("failed %d to get svm range pages\n", r); - if (r == -EBUSY) - r = -EAGAIN; - } } else { r = -EFAULT; } @@ -1690,17 +1753,25 @@ static int svm_range_validate_and_map(struct mm_struct *mm, if (!r) { offset = (addr >> PAGE_SHIFT) - prange->start; r = svm_range_dma_map(prange, ctx->bitmap, offset, npages, - hmm_range->hmm_pfns); + range->hmm_range.hmm_pfns); if (r) pr_debug("failed %d to dma map range\n", r); } svm_range_lock(prange); - if (!r && amdgpu_hmm_range_get_pages_done(hmm_range)) { + + /* Free backing memory of hmm_range if it was initialized + * Override return value to TRY AGAIN only if prior returns + * were successful + */ + if (range && !amdgpu_hmm_range_valid(range) && !r) { pr_debug("hmm update the range, need validate again\n"); r = -EAGAIN; } + /* Free the hmm range */ + amdgpu_hmm_range_free(range); + if (!r && !list_empty(&prange->child_list)) { pr_debug("range split by unmap in parallel, validate again\n"); r = -EAGAIN; @@ -1980,6 +2051,7 @@ static struct svm_range *svm_range_clone(struct svm_range *old) new->vram_pages = old->vram_pages; bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); + atomic_set(&new->queue_refcount, atomic_read(&old->queue_refcount)); return new; } @@ -2248,16 +2320,10 @@ static void svm_range_drain_retry_fault(struct svm_range_list *svms) { struct kfd_process_device *pdd; struct kfd_process *p; - int drain; uint32_t i; p = container_of(svms, struct kfd_process, svms); -restart: - drain = atomic_read(&svms->drain_pagefaults); - if (!drain) - return; - for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) { pdd = p->pdds[i]; if (!pdd) @@ -2277,8 +2343,6 @@ restart: pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms); } - if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain) - goto restart; } static void svm_range_deferred_list_work(struct work_struct *work) @@ -2300,17 +2364,8 @@ static void svm_range_deferred_list_work(struct work_struct *work) prange->start, prange->last, prange->work_item.op); mm = prange->work_item.mm; -retry: - mmap_write_lock(mm); - /* Checking for the need to drain retry faults must be inside - * mmap write lock to serialize with munmap notifiers. - */ - if (unlikely(atomic_read(&svms->drain_pagefaults))) { - mmap_write_unlock(mm); - svm_range_drain_retry_fault(svms); - goto retry; - } + mmap_write_lock(mm); /* Remove from deferred_list must be inside mmap write lock, for * two race cases: @@ -2366,15 +2421,17 @@ svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange, prange->work_item.op != SVM_OP_UNMAP_RANGE) prange->work_item.op = op; } else { - prange->work_item.op = op; - - /* Pairs with mmput in deferred_list_work */ - mmget(mm); - prange->work_item.mm = mm; - list_add_tail(&prange->deferred_list, - &prange->svms->deferred_range_list); - pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n", - prange, prange->start, prange->last, op); + /* Pairs with mmput in deferred_list_work. + * If process is exiting and mm is gone, don't update mmu notifier. + */ + if (mmget_not_zero(mm)) { + prange->work_item.mm = mm; + prange->work_item.op = op; + list_add_tail(&prange->deferred_list, + &prange->svms->deferred_range_list); + pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n", + prange, prange->start, prange->last, op); + } } spin_unlock(&svms->deferred_list_lock); } @@ -2388,8 +2445,7 @@ void schedule_deferred_list_work(struct svm_range_list *svms) } static void -svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent, - struct svm_range *prange, unsigned long start, +svm_range_unmap_split(struct svm_range *parent, struct svm_range *prange, unsigned long start, unsigned long last) { struct svm_range *head; @@ -2410,12 +2466,12 @@ svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent, svm_range_split(tail, last + 1, tail->last, &head); if (head != prange && tail != prange) { - svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); - svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); + svm_range_add_child(parent, head, SVM_OP_UNMAP_RANGE); + svm_range_add_child(parent, tail, SVM_OP_ADD_RANGE); } else if (tail != prange) { - svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE); + svm_range_add_child(parent, tail, SVM_OP_UNMAP_RANGE); } else if (head != prange) { - svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); + svm_range_add_child(parent, head, SVM_OP_UNMAP_RANGE); } else if (parent != prange) { prange->work_item.op = SVM_OP_UNMAP_RANGE; } @@ -2431,6 +2487,17 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, struct kfd_process *p; unsigned long s, l; bool unmap_parent; + uint32_t i; + + if (atomic_read(&prange->queue_refcount)) { + int r; + + pr_warn("Freeing queue vital buffer 0x%lx, queue evicted\n", + prange->start << PAGE_SHIFT); + r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM); + if (r) + pr_debug("failed %d to quiesce KFD queues\n", r); + } p = kfd_lookup_process_by_mm(mm); if (!p) @@ -2440,11 +2507,38 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms, prange, prange->start, prange->last, start, last); - /* Make sure pending page faults are drained in the deferred worker - * before the range is freed to avoid straggler interrupts on - * unmapped memory causing "phantom faults". + /* calculate time stamps that are used to decide which page faults need be + * dropped or handled before unmap pages from gpu vm */ - atomic_inc(&svms->drain_pagefaults); + for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) { + struct kfd_process_device *pdd; + struct amdgpu_device *adev; + struct amdgpu_ih_ring *ih; + uint32_t checkpoint_wptr; + + pdd = p->pdds[i]; + if (!pdd) + continue; + + adev = pdd->dev->adev; + + /* Check and drain ih1 ring if cam not available */ + if (adev->irq.ih1.ring_size) { + ih = &adev->irq.ih1; + checkpoint_wptr = amdgpu_ih_get_wptr(adev, ih); + if (ih->rptr != checkpoint_wptr) { + svms->checkpoint_ts[i] = + amdgpu_ih_decode_iv_ts(adev, ih, checkpoint_wptr, -1); + continue; + } + } + + /* check if dev->irq.ih_soft is not empty */ + ih = &adev->irq.ih_soft; + checkpoint_wptr = amdgpu_ih_get_wptr(adev, ih); + if (ih->rptr != checkpoint_wptr) + svms->checkpoint_ts[i] = amdgpu_ih_decode_iv_ts(adev, ih, checkpoint_wptr, -1); + } unmap_parent = start <= prange->start && last >= prange->last; @@ -2454,14 +2548,14 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, l = min(last, pchild->last); if (l >= s) svm_range_unmap_from_gpus(pchild, s, l, trigger); - svm_range_unmap_split(mm, prange, pchild, start, last); + svm_range_unmap_split(prange, pchild, start, last); mutex_unlock(&pchild->lock); } s = max(start, prange->start); l = min(last, prange->last); if (l >= s) svm_range_unmap_from_gpus(prange, s, l, trigger); - svm_range_unmap_split(mm, prange, prange, start, last); + svm_range_unmap_split(prange, prange, start, last); if (unmap_parent) svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE); @@ -2504,8 +2598,6 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, if (range->event == MMU_NOTIFY_RELEASE) return true; - if (!mmget_not_zero(mni->mm)) - return true; start = mni->interval_tree.start; last = mni->interval_tree.last; @@ -2532,7 +2624,6 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, } svm_range_unlock(prange); - mmput(mni->mm); return true; } @@ -2619,7 +2710,7 @@ svm_range_best_restore_location(struct svm_range *prange, return -1; } - if (node->adev->flags & AMD_IS_APU) + if (node->adev->apu_prefer_gtt) return 0; if (prange->preferred_loc == gpuid || @@ -2668,9 +2759,10 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr, *is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma); start_limit = max(vma->vm_start >> PAGE_SHIFT, - (unsigned long)ALIGN_DOWN(addr, 2UL << 8)); + (unsigned long)ALIGN_DOWN(addr, 1UL << p->svms.default_granularity)); end_limit = min(vma->vm_end >> PAGE_SHIFT, - (unsigned long)ALIGN(addr + 1, 2UL << 8)); + (unsigned long)ALIGN(addr + 1, 1UL << p->svms.default_granularity)); + /* First range that starts after the fault address */ node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX); if (node) { @@ -2885,7 +2977,7 @@ svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, uint32_t vmid, uint32_t node_id, - uint64_t addr, bool write_fault) + uint64_t addr, uint64_t ts, bool write_fault) { unsigned long start, last, size; struct mm_struct *mm = NULL; @@ -2895,7 +2987,7 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, ktime_t timestamp = ktime_get_boottime(); struct kfd_node *node; int32_t best_loc; - int32_t gpuidx = MAX_GPU_INSTANCE; + int32_t gpuid, gpuidx = MAX_GPU_INSTANCE; bool write_locked = false; struct vm_area_struct *vma; bool migration = false; @@ -2906,7 +2998,7 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, return -EFAULT; } - p = kfd_lookup_process_by_pasid(pasid); + p = kfd_lookup_process_by_pasid(pasid, NULL); if (!p) { pr_debug("kfd process not founded pasid 0x%x\n", pasid); return 0; @@ -2916,11 +3008,25 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr); if (atomic_read(&svms->drain_pagefaults)) { - pr_debug("draining retry fault, drop fault 0x%llx\n", addr); + pr_debug("page fault handling disabled, drop fault 0x%llx\n", addr); r = 0; goto out; } + node = kfd_node_by_irq_ids(adev, node_id, vmid); + if (!node) { + pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id, + vmid); + r = -EFAULT; + goto out; + } + + if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) { + pr_debug("failed to get gpuid/gpuidex for node_id: %d\n", node_id); + r = -EFAULT; + goto out; + } + if (!p->xnack_enabled) { pr_debug("XNACK not enabled for pasid 0x%x\n", pasid); r = -EFAULT; @@ -2937,16 +3043,26 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, goto out; } - node = kfd_node_by_irq_ids(adev, node_id, vmid); - if (!node) { - pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id, - vmid); - r = -EFAULT; - goto out; - } mmap_read_lock(mm); retry_write_locked: mutex_lock(&svms->lock); + + /* check if this page fault time stamp is before svms->checkpoint_ts */ + if (svms->checkpoint_ts[gpuidx] != 0) { + if (amdgpu_ih_ts_after_or_equal(ts, svms->checkpoint_ts[gpuidx])) { + pr_debug("draining retry fault, drop fault 0x%llx\n", addr); + if (write_locked) + mmap_write_downgrade(mm); + r = -EAGAIN; + goto out_unlock_svms; + } else { + /* ts is after svms->checkpoint_ts now, reset svms->checkpoint_ts + * to zero to avoid following ts wrap around give wrong comparing + */ + svms->checkpoint_ts[gpuidx] = 0; + } + } + prange = svm_range_from_addr(svms, addr, NULL); if (!prange) { pr_debug("failed to find prange svms 0x%p address [0x%llx]\n", @@ -3028,8 +3144,6 @@ retry_write_locked: start = max_t(unsigned long, ALIGN_DOWN(addr, size), prange->start); last = min_t(unsigned long, ALIGN(addr + 1, size) - 1, prange->last); if (prange->actual_loc != 0 || best_loc != 0) { - migration = true; - if (best_loc) { r = svm_migrate_to_vram(prange, best_loc, start, last, mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU); @@ -3052,7 +3166,9 @@ retry_write_locked: if (r) { pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n", r, svms, start, last); - goto out_unlock_range; + goto out_migrate_fail; + } else { + migration = true; } } @@ -3062,6 +3178,7 @@ retry_write_locked: pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", r, svms, start, last); +out_migrate_fail: kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr, migration); @@ -3071,7 +3188,8 @@ out_unlock_svms: mutex_unlock(&svms->lock); mmap_read_unlock(mm); - svm_range_count_fault(node, p, gpuidx); + if (r != -EAGAIN) + svm_range_count_fault(node, p, gpuidx); mmput(mm); out: @@ -3148,7 +3266,8 @@ void svm_range_list_fini(struct kfd_process *p) struct svm_range *prange; struct svm_range *next; - pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms); + pr_debug("process pid %d svms 0x%p\n", p->lead_thread->pid, + &p->svms); cancel_delayed_work_sync(&p->svms.restore_work); @@ -3158,8 +3277,9 @@ void svm_range_list_fini(struct kfd_process *p) /* * Ensure no retry fault comes in afterwards, as page fault handler will * not find kfd process and take mm lock to recover fault. + * stop kfd page fault handing, then wait pending page faults got drained */ - atomic_inc(&p->svms.drain_pagefaults); + atomic_set(&p->svms.drain_pagefaults, 1); svm_range_drain_retry_fault(&p->svms); list_for_each_entry_safe(prange, next, &p->svms.list, list) { @@ -3170,7 +3290,8 @@ void svm_range_list_fini(struct kfd_process *p) mutex_destroy(&p->svms.lock); - pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms); + pr_debug("process pid %d svms 0x%p done\n", + p->lead_thread->pid, &p->svms); } int svm_range_list_init(struct kfd_process *p) @@ -3193,6 +3314,12 @@ int svm_range_list_init(struct kfd_process *p) if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev)) bitmap_set(svms->bitmap_supported, i, 1); + /* Value of default granularity cannot exceed 0x1B, the + * number of pages supported by a 4-level paging table + */ + svms->default_granularity = min_t(u8, amdgpu_svm_default_granularity, 0x1B); + pr_debug("Default SVM Granularity to use: %d\n", svms->default_granularity); + return 0; } @@ -3337,7 +3464,7 @@ svm_range_best_prefetch_location(struct svm_range *prange) goto out; } - if (bo_node->adev->flags & AMD_IS_APU) { + if (bo_node->adev->apu_prefer_gtt) { best_loc = 0; goto out; } @@ -3527,8 +3654,8 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, bool flush_tlb; int r, ret = 0; - pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n", - p->pasid, &p->svms, start, start + size - 1, size); + pr_debug("process pid %d svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n", + p->lead_thread->pid, &p->svms, start, start + size - 1, size); r = svm_range_check_attr(p, nattr, attrs); if (r) @@ -3566,6 +3693,8 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping); /* TODO: unmap ranges from GPU that lost access */ } + update_mapping |= !p->xnack_enabled && !list_empty(&remap_list); + list_for_each_entry_safe(prange, next, &remove_list, update_list) { pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, prange->start, @@ -3636,8 +3765,8 @@ out_unlock_range: out: mutex_unlock(&process_info->lock); - pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid, - &p->svms, start, start + size - 1, r); + pr_debug("process pid %d svms 0x%p [0x%llx 0x%llx] done, r=%d\n", + p->lead_thread->pid, &p->svms, start, start + size - 1, r); return ret ? ret : r; } @@ -3720,7 +3849,7 @@ svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm, node = interval_tree_iter_first(&svms->objects, start, last); if (!node) { pr_debug("range attrs not found return default values\n"); - svm_range_set_default_attributes(&location, &prefetch_loc, + svm_range_set_default_attributes(svms, &location, &prefetch_loc, &granularity, &flags_and); flags_or = flags_and; if (p->xnack_enabled) @@ -3975,8 +4104,8 @@ exit: return ret; } -int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, - uint64_t *svm_priv_data_size) +void svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, + uint64_t *svm_priv_data_size) { uint64_t total_size, accessibility_size, common_attr_size; int nattr_common = 4, nattr_accessibility = 1; @@ -3988,8 +4117,6 @@ int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, *svm_priv_data_size = 0; svms = &p->svms; - if (!svms) - return -EINVAL; mutex_lock(&svms->lock); list_for_each_entry(prange, &svms->list, list) { @@ -4031,7 +4158,6 @@ int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges, *svm_priv_data_size); - return 0; } int kfd_criu_checkpoint_svm(struct kfd_process *p, @@ -4048,8 +4174,6 @@ int kfd_criu_checkpoint_svm(struct kfd_process *p, struct mm_struct *mm; svms = &p->svms; - if (!svms) - return -EINVAL; mm = get_task_mm(p->lead_thread); if (!mm) { @@ -4147,7 +4271,7 @@ svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, r = svm_range_get_attr(p, mm, start, size, nattrs, attrs); break; default: - r = EINVAL; + r = -EINVAL; break; } |