1 files changed, 128 insertions, 60 deletions

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index f53e17a94ad8..9b9c2b9bf2ef 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
@@ -20,7 +20,6 @@
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  */
-
 #include <linux/types.h>
 #include <linux/hmm.h>
 #include <linux/dma-direction.h>
@@ -34,6 +33,11 @@
 #include "kfd_svm.h"
 #include "kfd_migrate.h"
 
+#ifdef dev_fmt
+#undef dev_fmt
+#endif
+#define dev_fmt(fmt) "kfd_migrate: %s: " fmt, __func__
+
 static uint64_t
 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, uint64_t addr)
 {
@@ -151,14 +155,14 @@ svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys,
 			gart_d = svm_migrate_direct_mapping_addr(adev, *vram);
 		}
 		if (r) {
-			pr_debug("failed %d to create gart mapping\n", r);
+			dev_err(adev->dev, "fail %d create gart mapping\n", r);
 			goto out_unlock;
 		}
 
 		r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE,
 				       NULL, &next, false, true, false);
 		if (r) {
-			pr_debug("failed %d to copy memory\n", r);
+			dev_err(adev->dev, "fail %d to copy memory\n", r);
 			goto out_unlock;
 		}
 
@@ -264,6 +268,32 @@ static void svm_migrate_put_sys_page(unsigned long addr)
 	put_page(page);
 }
 
+static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
+{
+	unsigned long cpages = 0;
+	unsigned long i;
+
+	for (i = 0; i < migrate->npages; i++) {
+		if (migrate->src[i] & MIGRATE_PFN_VALID &&
+		    migrate->src[i] & MIGRATE_PFN_MIGRATE)
+			cpages++;
+	}
+	return cpages;
+}
+
+static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
+{
+	unsigned long upages = 0;
+	unsigned long i;
+
+	for (i = 0; i < migrate->npages; i++) {
+		if (migrate->src[i] & MIGRATE_PFN_VALID &&
+		    !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
+			upages++;
+	}
+	return upages;
+}
+
 static int
 svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			 struct migrate_vma *migrate, struct dma_fence **mfence,
@@ -285,7 +315,7 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 
 	r = svm_range_vram_node_new(adev, prange, true);
 	if (r) {
-		pr_debug("failed %d get 0x%llx pages from vram\n", r, npages);
+		dev_err(adev->dev, "fail %d to alloc vram\n", r);
 		goto out;
 	}
 
@@ -300,12 +330,11 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]);
 			svm_migrate_get_vram_page(prange, migrate->dst[i]);
 			migrate->dst[i] = migrate_pfn(migrate->dst[i]);
-			migrate->dst[i] |= MIGRATE_PFN_LOCKED;
 			src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE,
 					      DMA_TO_DEVICE);
 			r = dma_mapping_error(dev, src[i]);
 			if (r) {
-				pr_debug("failed %d dma_map_page\n", r);
+				dev_err(adev->dev, "fail %d dma_map_page\n", r);
 				goto out_free_vram_pages;
 			}
 		} else {
@@ -325,8 +354,8 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			continue;
 		}
 
-		pr_debug("dma mapping src to 0x%llx, page_to_pfn 0x%lx\n",
-			 src[i] >> PAGE_SHIFT, page_to_pfn(spage));
+		pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n",
+				     src[i] >> PAGE_SHIFT, page_to_pfn(spage));
 
 		if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) {
 			r = svm_migrate_copy_memory_gart(adev, src + i - j,
@@ -372,7 +401,7 @@ out:
 	return r;
 }
 
-static int
+static long
 svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 			struct vm_area_struct *vma, uint64_t start,
 			uint64_t end)
@@ -381,6 +410,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
+	unsigned long cpages = 0;
 	dma_addr_t *scratch;
 	size_t size;
 	void *buf;
@@ -405,23 +435,31 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 
 	r = migrate_vma_setup(&migrate);
 	if (r) {
-		pr_debug("failed %d prepare migrate svms 0x%p [0x%lx 0x%lx]\n",
-			 r, prange->svms, prange->start, prange->last);
+		dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r,
+			prange->start, prange->last);
 		goto out_free;
 	}
-	if (migrate.cpages != npages) {
-		pr_debug("Partial migration. 0x%lx/0x%llx pages can be migrated\n",
-			 migrate.cpages,
-			 npages);
-	}
 
-	if (migrate.cpages) {
-		r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence,
-					     scratch);
-		migrate_vma_pages(&migrate);
-		svm_migrate_copy_done(adev, mfence);
-		migrate_vma_finalize(&migrate);
+	cpages = migrate.cpages;
+	if (!cpages) {
+		pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n",
+			 prange->start, prange->last);
+		goto out_free;
 	}
+	if (cpages != npages)
+		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
+			 cpages, npages);
+	else
+		pr_debug("0x%lx pages migrated\n", cpages);
+
+	r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch);
+	migrate_vma_pages(&migrate);
+
+	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		svm_migrate_successful_pages(&migrate), cpages, migrate.npages);
+
+	svm_migrate_copy_done(adev, mfence);
+	migrate_vma_finalize(&migrate);
 
 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
 	svm_range_free_dma_mappings(prange);
@@ -429,12 +467,13 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 out_free:
 	kvfree(buf);
 out:
-	if (!r) {
+	if (!r && cpages) {
 		pdd = svm_range_get_pdd_by_adev(prange, adev);
 		if (pdd)
-			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
-	}
+			WRITE_ONCE(pdd->page_in, pdd->page_in + cpages);
 
+		return cpages;
+	}
 	return r;
 }
 
@@ -456,7 +495,8 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
 	unsigned long addr, start, end;
 	struct vm_area_struct *vma;
 	struct amdgpu_device *adev;
-	int r = 0;
+	unsigned long cpages = 0;
+	long r = 0;
 
 	if (prange->actual_loc == best_loc) {
 		pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n",
@@ -488,17 +528,19 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,
 
 		next = min(vma->vm_end, end);
 		r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next);
-		if (r) {
-			pr_debug("failed to migrate\n");
+		if (r < 0) {
+			pr_debug("failed %ld to migrate\n", r);
 			break;
+		} else {
+			cpages += r;
 		}
 		addr = next;
 	}
 
-	if (!r)
+	if (cpages)
 		prange->actual_loc = best_loc;
 
-	return r;
+	return r < 0 ? r : 0;
 }
 
 static void svm_migrate_page_free(struct page *page)
@@ -506,7 +548,7 @@ static void svm_migrate_page_free(struct page *page)
 	struct svm_range_bo *svm_bo = page->zone_device_data;
 
 	if (svm_bo) {
-		pr_debug("svm_bo ref left: %d\n", kref_read(&svm_bo->kref));
+		pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref));
 		svm_range_bo_unref(svm_bo);
 	}
 }
@@ -572,15 +614,14 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 		dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 		r = dma_mapping_error(dev, dst[i]);
 		if (r) {
-			pr_debug("failed %d dma_map_page\n", r);
+			dev_err(adev->dev, "fail %d dma_map_page\n", r);
 			goto out_oom;
 		}
 
-		pr_debug("dma mapping dst to 0x%llx, page_to_pfn 0x%lx\n",
-			      dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
+		pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n",
+				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 
 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
-		migrate->dst[i] |= MIGRATE_PFN_LOCKED;
 		j++;
 	}
 
@@ -599,11 +640,13 @@ out_oom:
 	return r;
 }
 
-static int
+static long
 svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
 {
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
+	unsigned long upages = npages;
+	unsigned long cpages = 0;
 	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate;
@@ -631,36 +674,47 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 
 	r = migrate_vma_setup(&migrate);
 	if (r) {
-		pr_debug("failed %d prepare migrate svms 0x%p [0x%lx 0x%lx]\n",
-			 r, prange->svms, prange->start, prange->last);
+		dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r,
+			prange->start, prange->last);
 		goto out_free;
 	}
 
-	pr_debug("cpages %ld\n", migrate.cpages);
-
-	if (migrate.cpages) {
-		r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
-					    scratch, npages);
-		migrate_vma_pages(&migrate);
-		svm_migrate_copy_done(adev, mfence);
-		migrate_vma_finalize(&migrate);
-	} else {
+	cpages = migrate.cpages;
+	if (!cpages) {
 		pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
 			 prange->start, prange->last);
+		upages = svm_migrate_unsuccessful_pages(&migrate);
+		goto out_free;
 	}
+	if (cpages != npages)
+		pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
+			 cpages, npages);
+	else
+		pr_debug("0x%lx pages migrated\n", cpages);
+
+	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
+				    scratch, npages);
+	migrate_vma_pages(&migrate);
+
+	upages = svm_migrate_unsuccessful_pages(&migrate);
+	pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		 upages, cpages, migrate.npages);
 
+	svm_migrate_copy_done(adev, mfence);
+	migrate_vma_finalize(&migrate);
 	svm_range_dma_unmap(adev->dev, scratch, 0, npages);
 
 out_free:
 	kvfree(buf);
 out:
-	if (!r) {
+	if (!r && cpages) {
 		pdd = svm_range_get_pdd_by_adev(prange, adev);
 		if (pdd)
-			WRITE_ONCE(pdd->page_out,
-				   pdd->page_out + migrate.cpages);
+			WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
+
+		return upages;
 	}
-	return r;
+	return r ? r : upages;
 }
 
 /**
@@ -680,7 +734,8 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
 	unsigned long addr;
 	unsigned long start;
 	unsigned long end;
-	int r = 0;
+	unsigned long upages = 0;
+	long r = 0;
 
 	if (!prange->actual_loc) {
 		pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
@@ -711,18 +766,21 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
 
 		next = min(vma->vm_end, end);
 		r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next);
-		if (r) {
-			pr_debug("failed %d to migrate\n", r);
+		if (r < 0) {
+			pr_debug("failed %ld to migrate\n", r);
 			break;
+		} else {
+			upages += r;
 		}
 		addr = next;
 	}
 
-	if (!r) {
+	if (!upages) {
 		svm_range_vram_node_free(prange);
 		prange->actual_loc = 0;
 	}
-	return r;
+
+	return r < 0 ? r : 0;
 }
 
 /**
@@ -740,7 +798,7 @@ static int
 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
 			 struct mm_struct *mm)
 {
-	int r;
+	int r, retries = 3;
 
 	/*
 	 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip
@@ -749,9 +807,14 @@ svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
 
 	pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
 
-	r = svm_migrate_vram_to_ram(prange, mm);
-	if (r)
-		return r;
+	do {
+		r = svm_migrate_vram_to_ram(prange, mm);
+		if (r)
+			return r;
+	} while (prange->actual_loc && --retries);
+
+	if (prange->actual_loc)
+		return -EDEADLK;
 
 	return svm_migrate_ram_to_vram(prange, best_loc, mm);
 }
@@ -796,6 +859,11 @@ static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
 		pr_debug("failed find process at fault address 0x%lx\n", addr);
 		return VM_FAULT_SIGBUS;
 	}
+	if (READ_ONCE(p->svms.faulting_task) == current) {
+		pr_debug("skipping ram migration\n");
+		kfd_unref_process(p);
+		return 0;
+	}
 	addr >>= PAGE_SHIFT;
 	pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr);