[PATCH 06/10] drm/amdkfd: skip invalid pages during migrations

[Alex Sierra]

Invalid pages can be the result of pages that have been migrated
already due to copy-on-write procedure or pages that were never
migrated to VRAM in first place. This is not an issue anymore,
as pranges now support mixed memory domains (CPU/GPU).

Signed-off-by: Alex Sierra <alex.sierra at amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 38 +++++++++++-------------
 1 file changed, 18 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index b298aa8dea4d..6fd68528c425 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
@@ -419,7 +419,6 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 	size_t size;
 	void *buf;
 	int r = -ENOMEM;
-	int retry = 0;
 
 	memset(&migrate, 0, sizeof(migrate));
 	migrate.vma = vma;
@@ -438,7 +437,6 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 	migrate.dst = migrate.src + npages;
 	scratch = (dma_addr_t *)(migrate.dst + npages);
 
-retry:
 	r = migrate_vma_setup(&migrate);
 	if (r) {
 		pr_debug("failed %d prepare migrate svms 0x%p [0x%lx 0x%lx]\n",
@@ -446,17 +444,9 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 		goto out_free;
 	}
 	if (migrate.cpages != npages) {
-		pr_debug("collect 0x%lx/0x%llx pages, retry\n", migrate.cpages,
+		pr_debug("Partial migration. 0x%lx/0x%llx pages can be migrated\n",
+			 migrate.cpages,
 			 npages);
-		migrate_vma_finalize(&migrate);
-		if (retry++ >= 3) {
-			r = -ENOMEM;
-			pr_debug("failed %d migrate svms 0x%p [0x%lx 0x%lx]\n",
-				 r, prange->svms, prange->start, prange->last);
-			goto out_free;
-		}
-
-		goto retry;
 	}
 
 	if (migrate.cpages) {
@@ -547,9 +537,8 @@ static void svm_migrate_page_free(struct page *page)
 static int
 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 			struct migrate_vma *migrate, struct dma_fence **mfence,
-			dma_addr_t *scratch)
+			dma_addr_t *scratch, uint64_t npages)
 {
-	uint64_t npages = migrate->cpages;
 	struct device *dev = adev->dev;
 	uint64_t *src;
 	dma_addr_t *dst;
@@ -566,15 +555,23 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 	src = (uint64_t *)(scratch + npages);
 	dst = scratch;
 
-	for (i = 0, j = 0; i < npages; i++, j++, addr += PAGE_SIZE) {
+	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
 		struct page *spage;
 
 		spage = migrate_pfn_to_page(migrate->src[i]);
-		if (!spage) {
-			pr_debug("failed get spage svms 0x%p [0x%lx 0x%lx]\n",
+		if (!spage || !is_zone_device_page(spage)) {
+			pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n",
 				 prange->svms, prange->start, prange->last);
-			r = -ENOMEM;
-			goto out_oom;
+			if (j) {
+				r = svm_migrate_copy_memory_gart(adev, dst + i - j,
+								 src + i - j, j,
+								 FROM_VRAM_TO_RAM,
+								 mfence);
+				if (r)
+					goto out_oom;
+				j = 0;
+			}
+			continue;
 		}
 		src[i] = svm_migrate_addr(adev, spage);
 		if (i > 0 && src[i] != src[i - 1] + PAGE_SIZE) {
@@ -607,6 +604,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 
 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
 		migrate->dst[i] |= MIGRATE_PFN_LOCKED;
+		j++;
 	}
 
 	r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j,
@@ -664,7 +662,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 
 	if (migrate.cpages) {
 		r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
-					    scratch);
+					    scratch, npages);
 		migrate_vma_pages(&migrate);
 		svm_migrate_copy_done(adev, mfence);
 		migrate_vma_finalize(&migrate);
-- 
2.32.0