Revert "drm/amdkfd: Use partial migrations in GPU page faults"

This reverts commit dc427a47.

The change prevents migrating the entire range to VRAM because retry
fault restore_pages map the remaining system memory range to GPUs. It
will work correctly to submit together with partial mapping to GPU
patch later.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdkfd/kfd_migrate.h

@@ -41,13 +41,9 @@ enum MIGRATION_COPY_DIR {
 };
 
 int svm_migrate_to_vram(struct svm_range *prange,  uint32_t best_loc,
-			unsigned long start, unsigned long last,
 			struct mm_struct *mm, uint32_t trigger);
-
 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
-			unsigned long start, unsigned long last,
-			uint32_t trigger, struct page *fault_page);
-
+			    uint32_t trigger, struct page *fault_page);
 unsigned long
 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr);
 

drivers/gpu/drm/amd/amdkfd/kfd_svm.c

@@ -158,13 +158,12 @@ svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr)
 static int
 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
 		      unsigned long offset, unsigned long npages,
-		      unsigned long *hmm_pfns, uint32_t gpuidx, uint64_t *vram_pages)
+		      unsigned long *hmm_pfns, uint32_t gpuidx)
 {
 	enum dma_data_direction dir = DMA_BIDIRECTIONAL;
 	dma_addr_t *addr = prange->dma_addr[gpuidx];
 	struct device *dev = adev->dev;
 	struct page *page;
-	uint64_t vram_pages_dev;
 	int i, r;
 
 	if (!addr) {
@@ -174,7 +173,6 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
 		prange->dma_addr[gpuidx] = addr;
 	}
 
-	vram_pages_dev = 0;
 	addr += offset;
 	for (i = 0; i < npages; i++) {
 		if (svm_is_valid_dma_mapping_addr(dev, addr[i]))
@@ -184,7 +182,6 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange,
 		if (is_zone_device_page(page)) {
 			struct amdgpu_device *bo_adev = prange->svm_bo->node->adev;
 
-			vram_pages_dev++;
 			addr[i] = (hmm_pfns[i] << PAGE_SHIFT) +
 				   bo_adev->vm_manager.vram_base_offset -
 				   bo_adev->kfd.pgmap.range.start;
@@ -201,14 +198,13 @@ 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));
 	}
-	*vram_pages = vram_pages_dev;
 	return 0;
 }
 
 static int
 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
 		  unsigned long offset, unsigned long npages,
-		  unsigned long *hmm_pfns, uint64_t *vram_pages)
+		  unsigned long *hmm_pfns)
 {
 	struct kfd_process *p;
 	uint32_t gpuidx;
@@ -227,7 +223,7 @@ svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
 		}
 
 		r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages,
-					  hmm_pfns, gpuidx, vram_pages);
+					  hmm_pfns, gpuidx);
 		if (r)
 			break;
 	}
@@ -353,7 +349,6 @@ 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);
 
@@ -400,8 +395,6 @@ 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_ON(prange->actual_loc);
 		mutex_unlock(&prange->lock);
 
 		spin_lock(&svm_bo->list_lock);
@@ -982,11 +975,6 @@ svm_range_split_nodes(struct svm_range *new, struct svm_range *old,
 	new->svm_bo = svm_range_bo_ref(old->svm_bo);
 	new->ttm_res = old->ttm_res;
 
-	/* set new's vram_pages as old range's now, the acurate vram_pages
-	 * will be updated during mapping
-	 */
-	new->vram_pages = min(old->vram_pages, new->npages);
-
 	spin_lock(&new->svm_bo->list_lock);
 	list_add(&new->svm_bo_list, &new->svm_bo->range_list);
 	spin_unlock(&new->svm_bo->list_lock);
@@ -1630,7 +1618,6 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
 	struct svm_validate_context *ctx;
 	unsigned long start, end, addr;
 	struct kfd_process *p;
-	uint64_t vram_pages;
 	void *owner;
 	int32_t idx;
 	int r = 0;
@@ -1699,13 +1686,11 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
 		}
 	}
 
-	vram_pages = 0;
 	start = prange->start << PAGE_SHIFT;
 	end = (prange->last + 1) << PAGE_SHIFT;
 	for (addr = start; !r && addr < end; ) {
 		struct hmm_range *hmm_range;
 		struct vm_area_struct *vma;
-		uint64_t vram_pages_vma;
 		unsigned long next = 0;
 		unsigned long offset;
 		unsigned long npages;
@@ -1734,11 +1719,9 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
 		if (!r) {
 			offset = (addr - start) >> PAGE_SHIFT;
 			r = svm_range_dma_map(prange, ctx->bitmap, offset, npages,
-					      hmm_range->hmm_pfns, &vram_pages_vma);
+					      hmm_range->hmm_pfns);
 			if (r)
 				pr_debug("failed %d to dma map range\n", r);
-			else
-				vram_pages += vram_pages_vma;
 		}
 
 		svm_range_lock(prange);
@@ -1764,19 +1747,6 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
 		addr = next;
 	}
 
-	if (addr == end) {
-		prange->vram_pages = vram_pages;
-
-		/* if prange does not include any vram page and it
-		 * has not released svm_bo drop its svm_bo reference
-		 * and set its actaul_loc to sys ram
-		 */
-		if (!vram_pages && prange->ttm_res) {
-			prange->actual_loc = 0;
-			svm_range_vram_node_free(prange);
-		}
-	}
-
 	svm_range_unreserve_bos(ctx);
 	if (!r)
 		prange->validate_timestamp = ktime_get_boottime();
@@ -2029,7 +1999,6 @@ 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);
 
@@ -2937,7 +2906,6 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 			uint32_t vmid, uint32_t node_id,
 			uint64_t addr, bool write_fault)
 {
-	unsigned long start, last, size;
 	struct mm_struct *mm = NULL;
 	struct svm_range_list *svms;
 	struct svm_range *prange;
@@ -3073,35 +3041,32 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 	kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
 				       write_fault, timestamp);
 
-	if (prange->actual_loc != 0 || best_loc != 0) {
+	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 (best_loc) {
-			r = svm_migrate_to_vram(prange, best_loc, start, last,
-					mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU);
+			r = svm_migrate_to_vram(prange, best_loc, 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 && prange->actual_loc != best_loc)
-					r = svm_migrate_vram_to_ram(prange, mm, start, last,
-						KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL);
+				if (prange->actual_loc)
+					r = svm_migrate_vram_to_ram(prange, mm,
+					   KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
+					   NULL);
 				else
 					r = 0;
 			}
 		} else {
-			r = svm_migrate_vram_to_ram(prange, mm, start, last,
-					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL);
+			r = svm_migrate_vram_to_ram(prange, mm,
+					KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU,
+					NULL);
 		}
 		if (r) {
 			pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
-				 r, svms, start, last);
+				 r, svms, prange->start, prange->last);
 			goto out_unlock_range;
 		}
 	}
@@ -3455,24 +3420,18 @@ svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,
 	*migrated = false;
 	best_loc = svm_range_best_prefetch_location(prange);
 
-	/* 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))
+	if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED ||
+	    best_loc == prange->actual_loc)
 		return 0;
 
 	if (!best_loc) {
-		r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last,
+		r = svm_migrate_vram_to_ram(prange, mm,
 					KFD_MIGRATE_TRIGGER_PREFETCH, NULL);
 		*migrated = !r;
 		return r;
 	}
 
-	r = svm_migrate_to_vram(prange, best_loc, prange->start, prange->last,
-				mm, KFD_MIGRATE_TRIGGER_PREFETCH);
+	r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH);
 	*migrated = !r;
 
 	return r;
@@ -3527,11 +3486,7 @@ 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);
 

drivers/gpu/drm/amd/amdkfd/kfd_svm.h

@@ -78,7 +78,6 @@ struct svm_work_list_item {
  * @update_list:link list node used to add to update_list
  * @mapping:    bo_va mapping structure to create and update GPU page table
  * @npages:     number of pages
- * @vram_pages: vram pages number in this svm_range
  * @dma_addr:   dma mapping address on each GPU for system memory physical page
  * @ttm_res:    vram ttm resource map
  * @offset:     range start offset within mm_nodes
@@ -89,9 +88,7 @@ struct svm_work_list_item {
  * @flags:      flags defined as KFD_IOCTL_SVM_FLAG_*
  * @perferred_loc: perferred location, 0 for CPU, or GPU id
  * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id
- * @actual_loc: this svm_range location. 0: all pages are from sys ram;
- *              GPU id: this svm_range may include vram pages from GPU with
- *              id actual_loc.
+ * @actual_loc: the actual location, 0 for CPU, or GPU id
  * @granularity:migration granularity, log2 num pages
  * @invalid:    not 0 means cpu page table is invalidated
  * @validate_timestamp: system timestamp when range is validated
@@ -115,7 +112,6 @@ struct svm_range {
 	struct list_head		list;
 	struct list_head		update_list;
 	uint64_t			npages;
-	uint64_t			vram_pages;
 	dma_addr_t			*dma_addr[MAX_GPU_INSTANCE];
 	struct ttm_resource		*ttm_res;
 	uint64_t			offset;