diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_svm.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 1134 |
1 files changed, 706 insertions, 428 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 64fdf63093a0..97c2270f278f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -23,10 +23,14 @@ #include <linux/types.h> #include <linux/sched/task.h> +#include <linux/dynamic_debug.h> +#include <drm/ttm/ttm_tt.h> +#include <drm/drm_exec.h> + #include "amdgpu_sync.h" #include "amdgpu_object.h" #include "amdgpu_vm.h" -#include "amdgpu_mn.h" +#include "amdgpu_hmm.h" #include "amdgpu.h" #include "amdgpu_xgmi.h" #include "kfd_priv.h" @@ -45,6 +49,13 @@ * page table is updated. */ #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC) +#if IS_ENABLED(CONFIG_DYNAMIC_DEBUG) +#define dynamic_svm_range_dump(svms) \ + _dynamic_func_call_no_desc("svm_range_dump", svm_range_debug_dump, svms) +#else +#define dynamic_svm_range_dump(svms) \ + do { if (0) svm_range_debug_dump(svms); } while (0) +#endif /* Giant svm range split into smaller ranges based on this, it is decided using * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to @@ -169,12 +180,11 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, page = hmm_pfn_to_page(hmm_pfns[i]); if (is_zone_device_page(page)) { - struct amdgpu_device *bo_adev = - amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); + struct amdgpu_device *bo_adev = prange->svm_bo->node->adev; addr[i] = (hmm_pfns[i] << PAGE_SHIFT) + bo_adev->vm_manager.vram_base_offset - - bo_adev->kfd.dev->pgmap.range.start; + bo_adev->kfd.pgmap.range.start; addr[i] |= SVM_RANGE_VRAM_DOMAIN; pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]); continue; @@ -188,6 +198,7 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n", addr[i] >> PAGE_SHIFT, page_to_pfn(page)); } + return 0; } @@ -221,7 +232,7 @@ svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap, return r; } -void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, +void svm_range_dma_unmap_dev(struct device *dev, dma_addr_t *dma_addr, unsigned long offset, unsigned long npages) { enum dma_data_direction dir = DMA_BIDIRECTIONAL; @@ -239,7 +250,7 @@ void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, } } -void svm_range_free_dma_mappings(struct svm_range *prange) +void svm_range_dma_unmap(struct svm_range *prange) { struct kfd_process_device *pdd; dma_addr_t *dma_addr; @@ -259,41 +270,52 @@ void svm_range_free_dma_mappings(struct svm_range *prange) pr_debug("failed to find device idx %d\n", gpuidx); continue; } - dev = &pdd->dev->pdev->dev; - svm_range_dma_unmap(dev, dma_addr, 0, prange->npages); - kvfree(dma_addr); - prange->dma_addr[gpuidx] = NULL; + dev = &pdd->dev->adev->pdev->dev; + + svm_range_dma_unmap_dev(dev, dma_addr, 0, prange->npages); } } -static void svm_range_free(struct svm_range *prange, bool update_mem_usage) +static void svm_range_free(struct svm_range *prange, bool do_unmap) { uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT; struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); + uint32_t gpuidx; pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, prange->start, prange->last); svm_range_vram_node_free(prange); - svm_range_free_dma_mappings(prange); + if (do_unmap) + svm_range_dma_unmap(prange); - if (update_mem_usage && !p->xnack_enabled) { + if (do_unmap && !p->xnack_enabled) { pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size); amdgpu_amdkfd_unreserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); + } + + /* free dma_addr array for each gpu */ + for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) { + if (prange->dma_addr[gpuidx]) { + kvfree(prange->dma_addr[gpuidx]); + prange->dma_addr[gpuidx] = NULL; + } } + mutex_destroy(&prange->lock); mutex_destroy(&prange->migrate_mutex); kfree(prange); } 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; } @@ -313,7 +335,7 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, p = container_of(svms, struct kfd_process, svms); if (!p->xnack_enabled && update_mem_usage && amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)) { + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) { pr_info("SVM mapping failed, exceeds resident system memory limit\n"); kfree(prange); return NULL; @@ -329,6 +351,7 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, INIT_LIST_HEAD(&prange->child_list); atomic_set(&prange->invalid, 0); prange->validate_timestamp = 0; + prange->vram_pages = 0; mutex_init(&prange->migrate_mutex); mutex_init(&prange->lock); @@ -336,7 +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); @@ -375,19 +398,37 @@ static void svm_range_bo_release(struct kref *kref) prange->start, prange->last); mutex_lock(&prange->lock); prange->svm_bo = NULL; + /* prange should not hold vram page now */ + WARN_ONCE(prange->actual_loc, "prange should not hold vram page"); mutex_unlock(&prange->lock); spin_lock(&svm_bo->list_lock); } spin_unlock(&svm_bo->list_lock); - if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) { - /* We're not in the eviction worker. - * Signal the fence and synchronize with any - * pending eviction work. + + 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. */ - dma_fence_signal(&svm_bo->eviction_fence->base); - cancel_work_sync(&svm_bo->eviction_work); + 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); dma_fence_put(&svm_bo->eviction_fence->base); amdgpu_bo_unref(&svm_bo->bo); kfree(svm_bo); @@ -423,10 +464,8 @@ static void svm_range_bo_unref(struct svm_range_bo *svm_bo) } static bool -svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange) +svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange) { - struct amdgpu_device *bo_adev; - mutex_lock(&prange->lock); if (!prange->svm_bo) { mutex_unlock(&prange->lock); @@ -439,12 +478,11 @@ svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange) } if (svm_bo_ref_unless_zero(prange->svm_bo)) { /* - * Migrate from GPU to GPU, remove range from source bo_adev - * svm_bo range list, and return false to allocate svm_bo from - * destination adev. + * Migrate from GPU to GPU, remove range from source svm_bo->node + * range list, and return false to allocate svm_bo from destination + * node. */ - bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); - if (bo_adev != adev) { + if (prange->svm_bo->node != node) { mutex_unlock(&prange->lock); spin_lock(&prange->svm_bo->list_lock); @@ -487,11 +525,11 @@ svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange) /* We need a new svm_bo. Spin-loop to wait for concurrent * svm_range_bo_release to finish removing this range from - * its range list. After this, it is safe to reuse the - * svm_bo pointer and svm_bo_list head. + * its range list and set prange->svm_bo to null. After this, + * it is safe to reuse the svm_bo pointer and svm_bo_list head. */ - while (!list_empty_careful(&prange->svm_bo_list)) - ; + while (!list_empty_careful(&prange->svm_bo_list) || prange->svm_bo) + cond_resched(); return false; } @@ -512,9 +550,10 @@ static struct svm_range_bo *svm_range_bo_new(void) } int -svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, +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; @@ -524,10 +563,11 @@ svm_range_vram_node_new(struct amdgpu_device *adev, 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(adev, prange)) + if (svm_range_validate_svm_bo(node, prange)) return 0; svm_bo = svm_range_bo_new(); @@ -541,6 +581,7 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, kfree(svm_bo); return -ESRCH; } + svm_bo->node = node; svm_bo->eviction_fence = amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), mm, @@ -557,19 +598,35 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE; bp.type = ttm_bo_type_device; bp.resv = NULL; + if (node->xcp) + bp.xcp_id_plus1 = node->xcp->id + 1; - r = amdgpu_bo_create_user(adev, &bp, &ubo); + r = amdgpu_bo_create_user(node->adev, &bp, &ubo); if (r) { pr_debug("failed %d to create bo\n", r); goto create_bo_failed; } bo = &ubo->bo; + + pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n", + bo->tbo.resource->start << PAGE_SHIFT, bp.size, + bp.xcp_id_plus1 - 1); + r = amdgpu_bo_reserve(bo, true); if (r) { pr_debug("failed %d to reserve bo\n", r); goto reserve_bo_failed; } + if (clear) { + r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false); + if (r) { + pr_debug("failed %d to sync bo\n", r); + amdgpu_bo_unreserve(bo); + goto reserve_bo_failed; + } + } + r = dma_resv_reserve_fences(bo->tbo.base.resv, 1); if (r) { pr_debug("failed %d to reserve bo\n", r); @@ -589,6 +646,10 @@ svm_range_vram_node_new(struct amdgpu_device *adev, 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: @@ -603,49 +664,41 @@ create_bo_failed: void svm_range_vram_node_free(struct svm_range *prange) { - svm_range_bo_unref(prange->svm_bo); - prange->ttm_res = NULL; + /* serialize prange->svm_bo unref */ + mutex_lock(&prange->lock); + /* prange->svm_bo has not been unref */ + if (prange->ttm_res) { + prange->ttm_res = NULL; + mutex_unlock(&prange->lock); + svm_range_bo_unref(prange->svm_bo); + } else + mutex_unlock(&prange->lock); } -struct amdgpu_device * -svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id) +struct kfd_node * +svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id) { - struct kfd_process_device *pdd; struct kfd_process *p; - int32_t gpu_idx; + struct kfd_process_device *pdd; p = container_of(prange->svms, struct kfd_process, svms); - - gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id); - if (gpu_idx < 0) { - pr_debug("failed to get device by id 0x%x\n", gpu_id); - return NULL; - } - pdd = kfd_process_device_from_gpuidx(p, gpu_idx); + pdd = kfd_process_device_data_by_id(p, gpu_id); if (!pdd) { - pr_debug("failed to get device by idx 0x%x\n", gpu_idx); + pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id); return NULL; } - return pdd->dev->adev; + return pdd->dev; } struct kfd_process_device * -svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev) +svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node) { struct kfd_process *p; - int32_t gpu_idx, gpuid; - int r; p = container_of(prange->svms, struct kfd_process, svms); - r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpu_idx); - if (r) { - pr_debug("failed to get device id by adev %p\n", adev); - return NULL; - } - - return kfd_process_device_from_gpuidx(p, gpu_idx); + return kfd_get_process_device_data(node, p); } static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo) @@ -725,7 +778,9 @@ svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, case KFD_IOCTL_SVM_ATTR_ACCESS: case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: case KFD_IOCTL_SVM_ATTR_NO_ACCESS: - *update_mapping = true; + if (!p->xnack_enabled) + *update_mapping = true; + gpuidx = kfd_process_gpuidx_from_gpuid(p, attrs[i].value); if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { @@ -748,7 +803,7 @@ svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, prange->flags &= ~attrs[i].value; break; case KFD_IOCTL_SVM_ATTR_GRANULARITY: - prange->granularity = attrs[i].value; + prange->granularity = min_t(uint32_t, attrs[i].value, 0x3F); break; default: WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); @@ -848,10 +903,56 @@ static void svm_range_debug_dump(struct svm_range_list *svms) } } +static void * +svm_range_copy_array(void *psrc, size_t size, uint64_t num_elements, + uint64_t offset, uint64_t *vram_pages) +{ + unsigned char *src = (unsigned char *)psrc + offset; + unsigned char *dst; + uint64_t i; + + dst = kvmalloc_array(num_elements, size, GFP_KERNEL); + if (!dst) + return NULL; + + if (!vram_pages) { + memcpy(dst, src, num_elements * size); + return (void *)dst; + } + + *vram_pages = 0; + for (i = 0; i < num_elements; i++) { + dma_addr_t *temp; + temp = (dma_addr_t *)dst + i; + *temp = *((dma_addr_t *)src + i); + if (*temp&SVM_RANGE_VRAM_DOMAIN) + (*vram_pages)++; + } + + return (void *)dst; +} + +static int +svm_range_copy_dma_addrs(struct svm_range *dst, struct svm_range *src) +{ + int i; + + for (i = 0; i < MAX_GPU_INSTANCE; i++) { + if (!src->dma_addr[i]) + continue; + dst->dma_addr[i] = svm_range_copy_array(src->dma_addr[i], + sizeof(*src->dma_addr[i]), src->npages, 0, NULL); + if (!dst->dma_addr[i]) + return -ENOMEM; + } + + return 0; +} + static int svm_range_split_array(void *ppnew, void *ppold, size_t size, uint64_t old_start, uint64_t old_n, - uint64_t new_start, uint64_t new_n) + uint64_t new_start, uint64_t new_n, uint64_t *new_vram_pages) { unsigned char *new, *old, *pold; uint64_t d; @@ -862,22 +963,17 @@ svm_range_split_array(void *ppnew, void *ppold, size_t size, if (!pold) return 0; - new = kvmalloc_array(new_n, size, GFP_KERNEL); + d = (new_start - old_start) * size; + /* get dma addr array for new range and calculte its vram page number */ + new = svm_range_copy_array(pold, size, new_n, d, new_vram_pages); if (!new) return -ENOMEM; - - d = (new_start - old_start) * size; - memcpy(new, pold + d, new_n * size); - - old = kvmalloc_array(old_n, size, GFP_KERNEL); + d = (new_start == old_start) ? new_n * size : 0; + old = svm_range_copy_array(pold, size, old_n, d, NULL); if (!old) { kvfree(new); return -ENOMEM; } - - d = (new_start == old_start) ? new_n * size : 0; - memcpy(old, pold + d, old_n * size); - kvfree(pold); *(void **)ppold = old; *(void **)ppnew = new; @@ -895,10 +991,13 @@ svm_range_split_pages(struct svm_range *new, struct svm_range *old, for (i = 0; i < MAX_GPU_INSTANCE; i++) { r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i], sizeof(*old->dma_addr[i]), old->start, - npages, new->start, new->npages); + npages, new->start, new->npages, + old->actual_loc ? &new->vram_pages : NULL); if (r) return r; } + if (old->actual_loc) + old->vram_pages -= new->vram_pages; return 0; } @@ -980,6 +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; } @@ -1041,126 +1141,80 @@ svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last, } static int -svm_range_split_tail(struct svm_range *prange, - uint64_t new_last, struct list_head *insert_list) +svm_range_split_tail(struct svm_range *prange, uint64_t new_last, + struct list_head *insert_list, struct list_head *remap_list) { - struct svm_range *tail; + struct svm_range *tail = NULL; int r = svm_range_split(prange, prange->start, new_last, &tail); - if (!r) + if (!r) { list_add(&tail->list, insert_list); + if (!IS_ALIGNED(new_last + 1, 1UL << prange->granularity)) + list_add(&tail->update_list, remap_list); + } return r; } static int -svm_range_split_head(struct svm_range *prange, - uint64_t new_start, struct list_head *insert_list) +svm_range_split_head(struct svm_range *prange, uint64_t new_start, + struct list_head *insert_list, struct list_head *remap_list) { - struct svm_range *head; + struct svm_range *head = NULL; int r = svm_range_split(prange, new_start, prange->last, &head); - if (!r) + if (!r) { list_add(&head->list, insert_list); + if (!IS_ALIGNED(new_start, 1UL << prange->granularity)) + list_add(&head->update_list, remap_list); + } return r; } 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); } -/** - * svm_range_split_by_granularity - collect ranges within granularity boundary - * - * @p: the process with svms list - * @mm: mm structure - * @addr: the vm fault address in pages, to split the prange - * @parent: parent range if prange is from child list - * @prange: prange to split - * - * Trims @prange to be a single aligned block of prange->granularity if - * possible. The head and tail are added to the child_list in @parent. - * - * Context: caller must hold mmap_read_lock and prange->lock - * - * Return: - * 0 - OK, otherwise error code - */ -int -svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, - unsigned long addr, struct svm_range *parent, - struct svm_range *prange) +static bool +svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b) { - struct svm_range *head, *tail; - unsigned long start, last, size; - int r; - - /* Align splited range start and size to granularity size, then a single - * PTE will be used for whole range, this reduces the number of PTE - * updated and the L1 TLB space used for translation. - */ - size = 1UL << prange->granularity; - start = ALIGN_DOWN(addr, size); - last = ALIGN(addr + 1, size) - 1; - - pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n", - prange->svms, prange->start, prange->last, start, last, size); - - if (start > prange->start) { - r = svm_range_split(prange, start, prange->last, &head); - if (r) - return r; - svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE); - } - - if (last < prange->last) { - r = svm_range_split(prange, prange->start, last, &tail); - if (r) - return r; - svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); - } - - /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */ - if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) { - prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP; - pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n", - prange, prange->start, prange->last, - SVM_OP_ADD_RANGE_AND_MAP); - } - return 0; + return (node_a->adev == node_b->adev || + amdgpu_xgmi_same_hive(node_a->adev, node_b->adev)); } static uint64_t -svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, - int domain) +svm_range_get_pte_flags(struct kfd_node *node, struct amdgpu_vm *vm, + struct svm_range *prange, int domain) { - struct amdgpu_device *bo_adev; + 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; + bool coherent = flags & (KFD_IOCTL_SVM_FLAG_COHERENT | KFD_IOCTL_SVM_FLAG_EXT_COHERENT); + bool ext_coherent = flags & KFD_IOCTL_SVM_FLAG_EXT_COHERENT; + unsigned int mtype_local; if (domain == SVM_RANGE_VRAM_DOMAIN) - bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); + bo_node = prange->svm_bo->node; - switch (KFD_GC_VERSION(adev->kfd.dev)) { + switch (gc_ip_version) { case IP_VERSION(9, 4, 1): if (domain == SVM_RANGE_VRAM_DOMAIN) { - if (bo_adev == adev) { + if (bo_node == node) { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; } else { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; - if (amdgpu_xgmi_same_hive(adev, bo_adev)) + if (svm_nodes_in_same_hive(node, bo_node)) snoop = true; } } else { @@ -1170,15 +1224,15 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, break; case IP_VERSION(9, 4, 2): if (domain == SVM_RANGE_VRAM_DOMAIN) { - if (bo_adev == adev) { + if (bo_node == node) { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; - if (adev->gmc.xgmi.connected_to_cpu) + if (node->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)) + if (svm_nodes_in_same_hive(node, bo_node)) snoop = true; } } else { @@ -1186,23 +1240,71 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; } break; + case IP_VERSION(9, 4, 3): + case IP_VERSION(9, 4, 4): + case IP_VERSION(9, 5, 0): + if (ext_coherent) + 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 (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)) + mapping_flags |= mtype_local; + /* local HBM region far from partition or remote XGMI GPU + * with regular system scope coherence + */ + else if (svm_nodes_in_same_hive(bo_node, node) && !ext_coherent) + mapping_flags |= AMDGPU_VM_MTYPE_NC; + /* 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 + * in amdgpu_gmc_override_vm_pte_flags + */ + if (num_possible_nodes() <= 1) + mapping_flags |= mtype_local; + else + mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; + /* system memory accessed by the dGPU */ + } else { + 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(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; } @@ -1215,7 +1317,7 @@ svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm, pr_debug("[0x%llx 0x%llx]\n", start, last); - return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start, + return amdgpu_vm_update_range(adev, vm, false, true, true, false, NULL, start, last, init_pte_value, 0, 0, NULL, NULL, fence); } @@ -1309,7 +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(adev, prange, last_domain); + pte_flags = svm_range_get_pte_flags(pdd->dev, vm, prange, last_domain); if (readonly) pte_flags &= ~AMDGPU_PTE_WRITEABLE; @@ -1318,8 +1420,12 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, pte_flags); - r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL, - last_start, prange->start + i, + /* For dGPU mode, we use same vm_manager to allocate VRAM for + * different memory partition based on fpfn/lpfn, we should use + * same vm_manager.vram_base_offset regardless memory partition. + */ + r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, true, + NULL, last_start, prange->start + i, pte_flags, (last_start - prange->start) << PAGE_SHIFT, bo_adev ? bo_adev->vm_manager.vram_base_offset : 0, @@ -1355,16 +1461,14 @@ svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, unsigned long *bitmap, bool wait, bool flush_tlb) { struct kfd_process_device *pdd; - struct amdgpu_device *bo_adev; + struct amdgpu_device *bo_adev = NULL; struct kfd_process *p; struct dma_fence *fence = NULL; uint32_t gpuidx; int r = 0; if (prange->svm_bo && prange->ttm_res) - bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); - else - bo_adev = NULL; + bo_adev = prange->svm_bo->node->adev; p = container_of(prange->svms, struct kfd_process, svms); for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { @@ -1413,37 +1517,34 @@ struct svm_validate_context { struct svm_range *prange; bool intr; DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); - struct ttm_validate_buffer tv[MAX_GPU_INSTANCE]; - struct list_head validate_list; - struct ww_acquire_ctx ticket; + struct drm_exec exec; }; -static int svm_range_reserve_bos(struct svm_validate_context *ctx) +static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr) { struct kfd_process_device *pdd; struct amdgpu_vm *vm; uint32_t gpuidx; int r; - INIT_LIST_HEAD(&ctx->validate_list); - for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { - pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); - if (!pdd) { - pr_debug("failed to find device idx %d\n", gpuidx); - return -EINVAL; - } - vm = drm_priv_to_vm(pdd->drm_priv); - - ctx->tv[gpuidx].bo = &vm->root.bo->tbo; - ctx->tv[gpuidx].num_shared = 4; - list_add(&ctx->tv[gpuidx].head, &ctx->validate_list); - } + drm_exec_init(&ctx->exec, intr ? DRM_EXEC_INTERRUPTIBLE_WAIT: 0, 0); + drm_exec_until_all_locked(&ctx->exec) { + for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { + pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); + if (!pdd) { + pr_debug("failed to find device idx %d\n", gpuidx); + r = -EINVAL; + goto unreserve_out; + } + vm = drm_priv_to_vm(pdd->drm_priv); - r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list, - ctx->intr, NULL); - if (r) { - pr_debug("failed %d to reserve bo\n", r); - return r; + r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(r)) { + pr_debug("failed %d to reserve bo\n", r); + goto unreserve_out; + } + } } for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { @@ -1454,9 +1555,9 @@ static int svm_range_reserve_bos(struct svm_validate_context *ctx) goto unreserve_out; } - r = amdgpu_vm_validate_pt_bos(pdd->dev->adev, - drm_priv_to_vm(pdd->drm_priv), - svm_range_bo_validate, NULL); + r = amdgpu_vm_validate(pdd->dev->adev, + drm_priv_to_vm(pdd->drm_priv), NULL, + svm_range_bo_validate, NULL); if (r) { pr_debug("failed %d validate pt bos\n", r); goto unreserve_out; @@ -1466,13 +1567,13 @@ static int svm_range_reserve_bos(struct svm_validate_context *ctx) return 0; unreserve_out: - ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list); + drm_exec_fini(&ctx->exec); return r; } static void svm_range_unreserve_bos(struct svm_validate_context *ctx) { - ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list); + drm_exec_fini(&ctx->exec); } static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) @@ -1480,6 +1581,8 @@ static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) struct kfd_process_device *pdd; pdd = kfd_process_device_from_gpuidx(p, gpuidx); + if (!pdd) + return NULL; return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev); } @@ -1509,51 +1612,64 @@ static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) * 5. Release page table (and SVM BO) reservation */ static int svm_range_validate_and_map(struct mm_struct *mm, + unsigned long map_start, unsigned long map_last, struct svm_range *prange, int32_t gpuidx, bool intr, bool wait, bool flush_tlb) { - struct svm_validate_context ctx; + struct svm_validate_context *ctx; unsigned long start, end, addr; struct kfd_process *p; void *owner; int32_t idx; int r = 0; - ctx.process = container_of(prange->svms, struct kfd_process, svms); - ctx.prange = prange; - ctx.intr = intr; + ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL); + if (!ctx) + return -ENOMEM; + ctx->process = container_of(prange->svms, struct kfd_process, svms); + ctx->prange = prange; + ctx->intr = intr; if (gpuidx < MAX_GPU_INSTANCE) { - bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE); - bitmap_set(ctx.bitmap, gpuidx, 1); - } else if (ctx.process->xnack_enabled) { - bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); + bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE); + bitmap_set(ctx->bitmap, gpuidx, 1); + } else if (ctx->process->xnack_enabled) { + bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); /* If prefetch range to GPU, or GPU retry fault migrate range to * GPU, which has ACCESS attribute to the range, create mapping * on that GPU. */ if (prange->actual_loc) { - gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process, + gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process, prange->actual_loc); if (gpuidx < 0) { WARN_ONCE(1, "failed get device by id 0x%x\n", prange->actual_loc); - return -EINVAL; + r = -EINVAL; + goto free_ctx; } if (test_bit(gpuidx, prange->bitmap_access)) - bitmap_set(ctx.bitmap, gpuidx, 1); + bitmap_set(ctx->bitmap, gpuidx, 1); + } + + /* + * If prange is already mapped or with always mapped flag, + * update mapping on GPUs with ACCESS attribute + */ + if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { + if (prange->mapped_to_gpu || + prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED) + bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); } } else { - bitmap_or(ctx.bitmap, prange->bitmap_access, + bitmap_or(ctx->bitmap, prange->bitmap_access, prange->bitmap_aip, MAX_GPU_INSTANCE); } - if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) { - if (!prange->mapped_to_gpu) - return 0; - - bitmap_copy(ctx.bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); + if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { + r = 0; + goto free_ctx; } if (prange->actual_loc && !prange->ttm_res) { @@ -1561,90 +1677,132 @@ static int svm_range_validate_and_map(struct mm_struct *mm, * svm_migrate_ram_to_vram after allocating a BO. */ WARN_ONCE(1, "VRAM BO missing during validation\n"); - return -EINVAL; + r = -EINVAL; + goto free_ctx; } - svm_range_reserve_bos(&ctx); + r = svm_range_reserve_bos(ctx, intr); + if (r) + goto free_ctx; p = container_of(prange->svms, struct kfd_process, svms); - owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap, + owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap, MAX_GPU_INSTANCE)); - for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) { + for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) { if (kfd_svm_page_owner(p, idx) != owner) { owner = NULL; break; } } - start = prange->start << PAGE_SHIFT; - end = (prange->last + 1) << PAGE_SHIFT; - for (addr = start; addr < end && !r; ) { - struct hmm_range *hmm_range; + start = map_start << PAGE_SHIFT; + end = (map_last + 1) << PAGE_SHIFT; + for (addr = start; !r && addr < end; ) { + struct amdgpu_hmm_range *range = NULL; + unsigned long map_start_vma; + unsigned long map_last_vma; struct vm_area_struct *vma; - unsigned long next; + unsigned long next = 0; unsigned long offset; unsigned long npages; bool readonly; - vma = find_vma(mm, addr); - if (!vma || addr < vma->vm_start) { - r = -EFAULT; - goto unreserve_out; - } - readonly = !(vma->vm_flags & VM_WRITE); + vma = vma_lookup(mm, addr); + if (vma) { + readonly = !(vma->vm_flags & VM_WRITE); - next = min(vma->vm_end, end); - npages = (next - addr) >> PAGE_SHIFT; - WRITE_ONCE(p->svms.faulting_task, current); - r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL, - addr, npages, &hmm_range, - readonly, true, owner); - WRITE_ONCE(p->svms.faulting_task, NULL); - if (r) { - pr_debug("failed %d to get svm range pages\n", r); - goto unreserve_out; + 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); + 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) + pr_debug("failed %d to get svm range pages\n", r); + } else { + r = -EFAULT; } - offset = (addr - start) >> PAGE_SHIFT; - r = svm_range_dma_map(prange, ctx.bitmap, offset, npages, - hmm_range->hmm_pfns); - if (r) { - pr_debug("failed %d to dma map range\n", r); - goto unreserve_out; + if (!r) { + offset = (addr >> PAGE_SHIFT) - prange->start; + r = svm_range_dma_map(prange, ctx->bitmap, offset, npages, + range->hmm_range.hmm_pfns); + if (r) + pr_debug("failed %d to dma map range\n", r); } svm_range_lock(prange); - if (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; - goto unlock_out; } - if (!list_empty(&prange->child_list)) { + + /* 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; - goto unlock_out; } - r = svm_range_map_to_gpus(prange, offset, npages, readonly, - ctx.bitmap, wait, flush_tlb); + if (!r) { + map_start_vma = max(map_start, prange->start + offset); + map_last_vma = min(map_last, prange->start + offset + npages - 1); + if (map_start_vma <= map_last_vma) { + offset = map_start_vma - prange->start; + npages = map_last_vma - map_start_vma + 1; + r = svm_range_map_to_gpus(prange, offset, npages, readonly, + ctx->bitmap, wait, flush_tlb); + } + } + + if (!r && next == end) + prange->mapped_to_gpu = true; -unlock_out: svm_range_unlock(prange); addr = next; } - if (addr == end) { - prange->validated_once = true; - prange->mapped_to_gpu = true; - } - -unreserve_out: - svm_range_unreserve_bos(&ctx); - + svm_range_unreserve_bos(ctx); if (!r) prange->validate_timestamp = ktime_get_boottime(); +free_ctx: + kfree(ctx); + return r; } @@ -1721,8 +1879,8 @@ static void svm_range_restore_work(struct work_struct *work) */ mutex_lock(&prange->migrate_mutex); - r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, - false, true, false); + r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, + MAX_GPU_INSTANCE, false, true, false); if (r) pr_debug("failed %d to map 0x%lx to gpus\n", r, prange->start); @@ -1759,7 +1917,7 @@ out_reschedule: /* If validation failed, reschedule another attempt */ if (evicted_ranges) { pr_debug("reschedule to restore svm range\n"); - schedule_delayed_work(&svms->restore_work, + queue_delayed_work(system_freezable_wq, &svms->restore_work, msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); kfd_smi_event_queue_restore_rescheduled(mm); @@ -1773,6 +1931,7 @@ out_reschedule: * @mm: current process mm_struct * @start: starting process queue number * @last: last process queue number + * @event: mmu notifier event when range is evicted or migrated * * Stop all queues of the process to ensure GPU doesn't access the memory, then * return to let CPU evict the buffer and proceed CPU pagetable update. @@ -1834,7 +1993,7 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm, pr_debug("failed to quiesce KFD\n"); pr_debug("schedule to restore svm %p ranges\n", svms); - schedule_delayed_work(&svms->restore_work, + queue_delayed_work(system_freezable_wq, &svms->restore_work, msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); } else { unsigned long s, l; @@ -1871,7 +2030,10 @@ static struct svm_range *svm_range_clone(struct svm_range *old) new = svm_range_new(old->svms, old->start, old->last, false); if (!new) return NULL; - + if (svm_range_copy_dma_addrs(new, old)) { + svm_range_free(new, false); + return NULL; + } if (old->svm_bo) { new->ttm_res = old->ttm_res; new->offset = old->offset; @@ -1886,8 +2048,10 @@ static struct svm_range *svm_range_clone(struct svm_range *old) new->actual_loc = old->actual_loc; new->granularity = old->granularity; new->mapped_to_gpu = old->mapped_to_gpu; + 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; } @@ -1896,14 +2060,23 @@ void svm_range_set_max_pages(struct amdgpu_device *adev) { uint64_t max_pages; uint64_t pages, _pages; + uint64_t min_pages = 0; + int i, id; + + for (i = 0; i < adev->kfd.dev->num_nodes; i++) { + if (adev->kfd.dev->nodes[i]->xcp) + id = adev->kfd.dev->nodes[i]->xcp->id; + else + id = -1; + pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17; + pages = clamp(pages, 1ULL << 9, 1ULL << 18); + pages = rounddown_pow_of_two(pages); + min_pages = min_not_zero(min_pages, pages); + } - /* 1/32 VRAM size in pages */ - pages = adev->gmc.real_vram_size >> 17; - pages = clamp(pages, 1ULL << 9, 1ULL << 18); - pages = rounddown_pow_of_two(pages); do { max_pages = READ_ONCE(max_svm_range_pages); - _pages = min_not_zero(max_pages, pages); + _pages = min_not_zero(max_pages, min_pages); } while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages); } @@ -1942,6 +2115,7 @@ svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last, * @update_list: output, the ranges need validate and update GPU mapping * @insert_list: output, the ranges need insert to svms * @remove_list: output, the ranges are replaced and need remove from svms + * @remap_list: output, remap unaligned svm ranges * * Check if the virtual address range has overlap with any existing ranges, * split partly overlapping ranges and add new ranges in the gaps. All changes @@ -1965,7 +2139,7 @@ static int svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, struct list_head *update_list, struct list_head *insert_list, - struct list_head *remove_list) + struct list_head *remove_list, struct list_head *remap_list) { unsigned long last = start + size - 1UL; struct svm_range_list *svms = &p->svms; @@ -1981,6 +2155,7 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, INIT_LIST_HEAD(insert_list); INIT_LIST_HEAD(remove_list); INIT_LIST_HEAD(&new_list); + INIT_LIST_HEAD(remap_list); node = interval_tree_iter_first(&svms->objects, start, last); while (node) { @@ -1994,7 +2169,8 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, next = interval_tree_iter_next(node, start, last); next_start = min(node->last, last) + 1; - if (svm_range_is_same_attrs(p, prange, nattr, attrs)) { + if (svm_range_is_same_attrs(p, prange, nattr, attrs) && + prange->mapped_to_gpu) { /* nothing to do */ } else if (node->start < start || node->last > last) { /* node intersects the update range and its attributes @@ -2016,14 +2192,14 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, if (node->start < start) { pr_debug("change old range start\n"); r = svm_range_split_head(prange, start, - insert_list); + insert_list, remap_list); if (r) goto out; } if (node->last > last) { pr_debug("change old range last\n"); r = svm_range_split_tail(prange, last, - insert_list); + insert_list, remap_list); if (r) goto out; } @@ -2144,16 +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) @@ -2162,11 +2332,17 @@ restart: pr_debug("drain retry fault gpu %d svms %p\n", i, svms); amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, - &pdd->dev->adev->irq.ih1); + pdd->dev->adev->irq.retry_cam_enabled ? + &pdd->dev->adev->irq.ih : + &pdd->dev->adev->irq.ih1); + + if (pdd->dev->adev->irq.retry_cam_enabled) + amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, + &pdd->dev->adev->irq.ih_soft); + + 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) @@ -2188,17 +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: @@ -2230,8 +2397,10 @@ retry: mutex_unlock(&svms->lock); mmap_write_unlock(mm); - /* Pairs with mmget in svm_range_add_list_work */ - mmput(mm); + /* Pairs with mmget in svm_range_add_list_work. If dropping the + * last mm refcount, schedule release work to avoid circular locking + */ + mmput_async(mm); spin_lock(&svms->deferred_list_lock); } @@ -2252,15 +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); } @@ -2274,8 +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; @@ -2296,12 +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; } @@ -2317,6 +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) @@ -2326,11 +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; @@ -2340,14 +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); @@ -2390,8 +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; @@ -2418,7 +2624,6 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, } svm_range_unlock(prange); - mmput(mni->mm); return true; } @@ -2489,29 +2694,31 @@ svm_range_from_addr(struct svm_range_list *svms, unsigned long addr, */ static int32_t svm_range_best_restore_location(struct svm_range *prange, - struct amdgpu_device *adev, + struct kfd_node *node, int32_t *gpuidx) { - struct amdgpu_device *bo_adev, *preferred_adev; + struct kfd_node *bo_node, *preferred_node; struct kfd_process *p; uint32_t gpuid; int r; p = container_of(prange->svms, struct kfd_process, svms); - r = kfd_process_gpuid_from_adev(p, adev, &gpuid, gpuidx); + r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx); if (r < 0) { pr_debug("failed to get gpuid from kgd\n"); return -1; } + if (node->adev->apu_prefer_gtt) + return 0; + if (prange->preferred_loc == gpuid || prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) { return prange->preferred_loc; } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) { - preferred_adev = svm_range_get_adev_by_id(prange, - prange->preferred_loc); - if (amdgpu_xgmi_same_hive(adev, preferred_adev)) + preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc); + if (preferred_node && svm_nodes_in_same_hive(node, preferred_node)) return prange->preferred_loc; /* fall through */ } @@ -2523,8 +2730,8 @@ svm_range_best_restore_location(struct svm_range *prange, if (!prange->actual_loc) return 0; - bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc); - if (amdgpu_xgmi_same_hive(adev, bo_adev)) + bo_node = svm_range_get_node_by_id(prange, prange->actual_loc); + if (bo_node && svm_nodes_in_same_hive(node, bo_node)) return prange->actual_loc; else return 0; @@ -2540,38 +2747,36 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr, { struct vm_area_struct *vma; struct interval_tree_node *node; + struct rb_node *rb_node; unsigned long start_limit, end_limit; - vma = find_vma(p->mm, addr << PAGE_SHIFT); - if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) { + vma = vma_lookup(p->mm, addr << PAGE_SHIFT); + if (!vma) { pr_debug("VMA does not exist in address [0x%llx]\n", addr); return -EFAULT; } - *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk && - vma->vm_end >= vma->vm_mm->start_brk) || - (vma->vm_start <= vma->vm_mm->start_stack && - vma->vm_end >= vma->vm_mm->start_stack); + *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) { end_limit = min(end_limit, node->start); /* Last range that ends before the fault address */ - node = container_of(rb_prev(&node->rb), - struct interval_tree_node, rb); + rb_node = rb_prev(&node->rb); } else { /* Last range must end before addr because * there was no range after addr */ - node = container_of(rb_last(&p->svms.objects.rb_root), - struct interval_tree_node, rb); + rb_node = rb_last(&p->svms.objects.rb_root); } - if (node) { + if (rb_node) { + node = container_of(rb_node, struct interval_tree_node, rb); if (node->last >= addr) { WARN(1, "Overlap with prev node and page fault addr\n"); return -EFAULT; @@ -2641,7 +2846,7 @@ svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last, } static struct -svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev, +svm_range *svm_range_create_unregistered_range(struct kfd_node *node, struct kfd_process *p, struct mm_struct *mm, int64_t addr) @@ -2676,7 +2881,7 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev, pr_debug("Failed to create prange in address [0x%llx]\n", addr); return NULL; } - if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) { + if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) { pr_debug("failed to get gpuid from kgd\n"); svm_range_free(prange, true); return NULL; @@ -2730,7 +2935,7 @@ static bool svm_range_skip_recover(struct svm_range *prange) } static void -svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p, +svm_range_count_fault(struct kfd_node *node, struct kfd_process *p, int32_t gpuidx) { struct kfd_process_device *pdd; @@ -2743,7 +2948,7 @@ svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p, uint32_t gpuid; int r; - r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx); + r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx); if (r < 0) return; } @@ -2771,26 +2976,29 @@ svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, - uint64_t addr, bool write_fault) + uint32_t vmid, uint32_t node_id, + uint64_t addr, uint64_t ts, bool write_fault) { + unsigned long start, last, size; struct mm_struct *mm = NULL; struct svm_range_list *svms; struct svm_range *prange; struct kfd_process *p; 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; int r = 0; - if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) { + if (!KFD_IS_SVM_API_SUPPORTED(adev)) { pr_debug("device does not support SVM\n"); 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; @@ -2800,11 +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; @@ -2824,6 +3046,23 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, 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", @@ -2839,7 +3078,7 @@ retry_write_locked: write_locked = true; goto retry_write_locked; } - prange = svm_range_create_unregistered_range(adev, p, mm, addr); + prange = svm_range_create_unregistered_range(node, p, mm, addr); if (!prange) { pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n", svms, addr); @@ -2854,7 +3093,7 @@ retry_write_locked: mutex_lock(&prange->migrate_mutex); if (svm_range_skip_recover(prange)) { - amdgpu_gmc_filter_faults_remove(adev, addr, pasid); + amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); r = 0; goto out_unlock_range; } @@ -2871,8 +3110,8 @@ retry_write_locked: /* __do_munmap removed VMA, return success as we are handling stale * retry fault. */ - vma = find_vma(mm, addr << PAGE_SHIFT); - if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) { + vma = vma_lookup(mm, addr << PAGE_SHIFT); + if (!vma) { pr_debug("address 0x%llx VMA is removed\n", addr); r = 0; goto out_unlock_range; @@ -2885,7 +3124,7 @@ retry_write_locked: goto out_unlock_range; } - best_loc = svm_range_best_restore_location(prange, adev, &gpuidx); + best_loc = svm_range_best_restore_location(prange, node, &gpuidx); if (best_loc == -1) { pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n", svms, prange->start, prange->last); @@ -2897,45 +3136,50 @@ retry_write_locked: svms, prange->start, prange->last, best_loc, prange->actual_loc); - kfd_smi_event_page_fault_start(adev->kfd.dev, p->lead_thread->pid, addr, + kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr, write_fault, timestamp); - if (prange->actual_loc != best_loc) { - migration = true; + /* Align migration range start and size to granularity size */ + size = 1UL << prange->granularity; + 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) { if (best_loc) { - r = svm_migrate_to_vram(prange, best_loc, mm, - KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU); + r = svm_migrate_to_vram(prange, best_loc, start, last, + mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU); if (r) { pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n", r, addr); /* Fallback to system memory if migration to * VRAM failed */ - if (prange->actual_loc) - r = svm_migrate_vram_to_ram(prange, mm, - KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, - NULL); + if (prange->actual_loc && prange->actual_loc != best_loc) + r = svm_migrate_vram_to_ram(prange, mm, start, last, + KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL); else r = 0; } } else { - r = svm_migrate_vram_to_ram(prange, mm, - KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, - NULL); + r = svm_migrate_vram_to_ram(prange, mm, start, last, + KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL); } if (r) { pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n", - r, svms, prange->start, prange->last); - goto out_unlock_range; + r, svms, start, last); + goto out_migrate_fail; + } else { + migration = true; } } - r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false); + r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false, + false, false); if (r) pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", - r, svms, prange->start, prange->last); + r, svms, start, last); - kfd_smi_event_page_fault_end(adev->kfd.dev, p->lead_thread->pid, addr, +out_migrate_fail: + kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr, migration); out_unlock_range: @@ -2944,7 +3188,8 @@ out_unlock_svms: mutex_unlock(&svms->lock); mmap_read_unlock(mm); - svm_range_count_fault(adev, p, gpuidx); + if (r != -EAGAIN) + svm_range_count_fault(node, p, gpuidx); mmput(mm); out: @@ -2952,7 +3197,7 @@ out: if (r == -EAGAIN) { pr_debug("recover vm fault later\n"); - amdgpu_gmc_filter_faults_remove(adev, addr, pasid); + amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); r = 0; } return r; @@ -2976,10 +3221,10 @@ svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled) size = (pchild->last - pchild->start + 1) << PAGE_SHIFT; if (xnack_enabled) { amdgpu_amdkfd_unreserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); } else { r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); if (r) goto out_unlock; reserved_size += size; @@ -2989,10 +3234,10 @@ svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled) size = (prange->last - prange->start + 1) << PAGE_SHIFT; if (xnack_enabled) { amdgpu_amdkfd_unreserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); } else { r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); if (r) goto out_unlock; reserved_size += size; @@ -3005,7 +3250,7 @@ out_unlock: if (r) amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); else /* Change xnack mode must be inside svms lock, to avoid race with * svm_range_deferred_list_work unreserve memory in parallel. @@ -3021,7 +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); @@ -3031,8 +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) { @@ -3043,7 +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) @@ -3063,9 +3311,15 @@ int svm_range_list_init(struct kfd_process *p) spin_lock_init(&svms->deferred_list_lock); for (i = 0; i < p->n_pdds; i++) - if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev)) + 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; } @@ -3152,9 +3406,8 @@ svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size) start <<= PAGE_SHIFT; end = start + (size << PAGE_SHIFT); do { - vma = find_vma(p->mm, start); - if (!vma || start < vma->vm_start || - (vma->vm_flags & device_vma)) + vma = vma_lookup(p->mm, start); + if (!vma || (vma->vm_flags & device_vma)) return -EFAULT; start = min(end, vma->vm_end); } while (start < end); @@ -3195,7 +3448,7 @@ svm_range_best_prefetch_location(struct svm_range *prange) DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); uint32_t best_loc = prange->prefetch_loc; struct kfd_process_device *pdd; - struct amdgpu_device *bo_adev; + struct kfd_node *bo_node; struct kfd_process *p; uint32_t gpuidx; @@ -3204,9 +3457,14 @@ svm_range_best_prefetch_location(struct svm_range *prange) if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) goto out; - bo_adev = svm_range_get_adev_by_id(prange, best_loc); - if (!bo_adev) { - WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc); + bo_node = svm_range_get_node_by_id(prange, best_loc); + if (!bo_node) { + WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc); + best_loc = 0; + goto out; + } + + if (bo_node->adev->apu_prefer_gtt) { best_loc = 0; goto out; } @@ -3224,10 +3482,10 @@ svm_range_best_prefetch_location(struct svm_range *prange) continue; } - if (pdd->dev->adev == bo_adev) + if (pdd->dev->adev == bo_node->adev) continue; - if (!amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) { + if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) { best_loc = 0; break; } @@ -3275,32 +3533,39 @@ svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange, *migrated = false; best_loc = svm_range_best_prefetch_location(prange); - if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED || - best_loc == prange->actual_loc) + /* when best_loc is a gpu node and same as prange->actual_loc + * we still need do migration as prange->actual_loc !=0 does + * not mean all pages in prange are vram. hmm migrate will pick + * up right pages during migration. + */ + if ((best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) || + (best_loc == 0 && prange->actual_loc == 0)) return 0; if (!best_loc) { - r = svm_migrate_vram_to_ram(prange, mm, + r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last, KFD_MIGRATE_TRIGGER_PREFETCH, NULL); *migrated = !r; return r; } - r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH); + r = svm_migrate_to_vram(prange, best_loc, prange->start, prange->last, + mm, KFD_MIGRATE_TRIGGER_PREFETCH); *migrated = !r; - return r; + return 0; } int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence) { - if (!fence) - return -EINVAL; - - if (dma_fence_is_signaled(&fence->base)) - return 0; - - if (fence->svm_bo) { + /* Dereferencing fence->svm_bo is safe here because the fence hasn't + * signaled yet and we're under the protection of the fence->lock. + * After the fence is signaled in svm_range_bo_release, we cannot get + * here any more. + * + * Reference is dropped in svm_range_evict_svm_bo_worker. + */ + if (svm_bo_ref_unless_zero(fence->svm_bo)) { WRITE_ONCE(fence->svm_bo->evicting, 1); schedule_work(&fence->svm_bo->eviction_work); } @@ -3315,8 +3580,6 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work) int r = 0; svm_bo = container_of(work, struct svm_range_bo, eviction_work); - if (!svm_bo_ref_unless_zero(svm_bo)) - return; /* svm_bo was freed while eviction was pending */ if (mmget_not_zero(svm_bo->eviction_fence->mm)) { mm = svm_bo->eviction_fence->mm; @@ -3341,7 +3604,11 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work) mutex_lock(&prange->migrate_mutex); do { + /* migrate all vram pages in this prange to sys ram + * after that prange->actual_loc should be zero + */ r = svm_migrate_vram_to_ram(prange, mm, + prange->start, prange->last, KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL); } while (!r && prange->actual_loc && --retries); @@ -3379,15 +3646,16 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, struct list_head update_list; struct list_head insert_list; struct list_head remove_list; + struct list_head remap_list; struct svm_range_list *svms; struct svm_range *prange; struct svm_range *next; bool update_mapping = false; bool flush_tlb; - int r = 0; + 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) @@ -3410,7 +3678,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, /* Add new range and split existing ranges as needed */ r = svm_range_add(p, start, size, nattr, attrs, &update_list, - &insert_list, &remove_list); + &insert_list, &remove_list, &remap_list); if (r) { mutex_unlock(&svms->lock); mmap_write_unlock(mm); @@ -3425,6 +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, @@ -3464,28 +3734,41 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu; - r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, - true, true, flush_tlb); + r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, + MAX_GPU_INSTANCE, true, true, flush_tlb); if (r) pr_debug("failed %d to map svm range\n", r); out_unlock_range: mutex_unlock(&prange->migrate_mutex); if (r) - break; + ret = r; } - svm_range_debug_dump(svms); + list_for_each_entry(prange, &remap_list, update_list) { + pr_debug("Remapping prange 0x%p [0x%lx 0x%lx]\n", + prange, prange->start, prange->last); + mutex_lock(&prange->migrate_mutex); + r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, + MAX_GPU_INSTANCE, true, true, prange->mapped_to_gpu); + if (r) + pr_debug("failed %d on remap svm range\n", r); + mutex_unlock(&prange->migrate_mutex); + if (r) + ret = r; + } + + dynamic_svm_range_dump(svms); mutex_unlock(&svms->lock); mmap_read_unlock(mm); 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 r; + return ret ? ret : r; } static int @@ -3566,7 +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) @@ -3821,8 +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; @@ -3834,8 +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) { @@ -3877,7 +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, @@ -3894,8 +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) { @@ -3993,7 +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; } |