Skip to contentSkip to conversation

drm/amdkfd: remove unused code

PATCH In review James Zhu 4 participants 11 messages May 28, 2025 – Jul 28, 2025

Subsystems (3)
lore
Share & tools
lore

James Zhu

author
#0 lore raw

PATCH 1/2 drm/amdkfd: remove unused code

upages is assigned under cpages = 0, so it isn't really used in this function.

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 1 -
 1 file changed, 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -1
index 79251f22b702..f0b690d4bb46 100644
@@ -736,7 +736,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 	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)
Signature · 1 line
2.34.1
author
#1 lore raw

PATCH 2/2 drm/amdkfd: add svm_migrate_successful_pages

to get migration pages. When migrating pages from system to vram, needn't check bit MIGRATE_PFN_VALID, since the system page could be allocated, but not be accessed.

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 25 ++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +13 -12
index f0b690d4bb46..131838bf8751 100644
@@ -260,17 +260,20 @@ static void svm_migrate_put_sys_page(unsigned long addr)
 	put_page(page);
 }
 
-static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
+static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
 {
-	unsigned long upages = 0;
+	unsigned long mpages = 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++;
+                if (migrate->flags == MIGRATE_VMA_SELECT_SYSTEM &&
+			migrate->src[i] & MIGRATE_PFN_MIGRATE)
+                        mpages++;
+                 else if (migrate->src[i] & MIGRATE_PFN_VALID &&
+				migrate->src[i] & MIGRATE_PFN_MIGRATE)
+                        mpages++;
 	}
-	return upages;
+	return mpages;
 }
 
 static int
@@ -447,7 +450,7 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
+	mpages = svm_migrate_successful_pages(&migrate);
 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
 			 mpages, cpages, migrate.npages);
 
@@ -688,7 +691,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 {
 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
 	unsigned long cpages = 0;
 	unsigned long mpages = 0;
 	struct amdgpu_device *adev = node->adev;
@@ -748,9 +750,9 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 				    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);
+	mpages = svm_migrate_successful_pages(&migrate);
+	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
 
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
@@ -764,7 +766,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 				    node->id, 0, trigger, r);
 out:
 	if (!r && cpages) {
-		mpages = cpages - upages;
 		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
Signature · 1 line
2.34.1
#2 lore raw

Re: [PATCH 2/2] drm/amdkfd: add svm_migrate_successful_pages

Quoted 3 lines · James Zhu

to get migration pages. When migrating pages from system to vram, needn't check bit MIGRATE_PFN_VALID, since the system page could be allocated, but not be accessed.

I think the corner case is vram_pages becomes negative value when migrating prange from vram to system memory, if the src page is on system, src MIGRATE_PFN_MIGRATE set, MIGRATE_PFN_VALID not set, this count as successful migrate page by svm_migrate_unsuccessful_page.

As discussed offline, we could check dst MIGRATE_PFN_MIGRATE flag for both system to vram or vram to system cases.

Regards,

Philip

Quoted 75 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 25 ++++++++++++------------
  1 file changed, 13 insertions(+), 12 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +13 -12
index f0b690d4bb46..131838bf8751 100644
@@ -260,17 +260,20 @@ static void svm_migrate_put_sys_page(unsigned long addr)
  	put_page(page);
  }
 
-static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
+static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
  {
-	unsigned long upages = 0;
+	unsigned long mpages = 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++;
+                if (migrate->flags == MIGRATE_VMA_SELECT_SYSTEM &&
+			migrate->src[i] & MIGRATE_PFN_MIGRATE)
+                        mpages++;
+                 else if (migrate->src[i] & MIGRATE_PFN_VALID &&
+				migrate->src[i] & MIGRATE_PFN_MIGRATE)
+                        mpages++;
  	}
-	return upages;
+	return mpages;
  }
 
  static int
@@ -447,7 +450,7 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
+	mpages = svm_migrate_successful_pages(&migrate);
  	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
  			 mpages, cpages, migrate.npages);
 
@@ -688,7 +691,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  {
  	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
  	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
  	unsigned long cpages = 0;
  	unsigned long mpages = 0;
  	struct amdgpu_device *adev = node->adev;
@@ -748,9 +750,9 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  				    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);
+	mpages = svm_migrate_successful_pages(&migrate);
+	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
 
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
@@ -764,7 +766,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  				    node->id, 0, trigger, r);
  out:
  	if (!r && cpages) {
-		mpages = cpages - upages;
  		pdd = svm_range_get_pdd_by_node(prange, node);
  		if (pdd)
  			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
author
#3 lore raw

PATCH v2 2/2 drm/amdkfd: add svm_migrate_successful_pages

to get migration pages. dst bit MIGRATE_PFN_VALID and src bit MIGRATE_PFN_MIGRATE should always be set when success.

-v2 use dst to check MIGRATE_PFN_VALID bit(suggested-by philip)

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 22 ++++++++++------------
 1 file changed, 10 insertions(+), 12 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +10 -12
index f0b690d4bb46..a41db1009f58 100644
@@ -260,17 +260,17 @@ static void svm_migrate_put_sys_page(unsigned long addr)
 	put_page(page);
 }
 
-static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
+static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
 {
-	unsigned long upages = 0;
+	unsigned long mpages = 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++;
+		if (migrate->src[i] & MIGRATE_PFN_MIGRATE &&
+			migrate->dst[i] & MIGRATE_PFN_VALID)
+                        mpages++;
 	}
-	return upages;
+	return mpages;
 }
 
 static int
@@ -447,7 +447,7 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
+	mpages = svm_migrate_successful_pages(&migrate);
 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
 			 mpages, cpages, migrate.npages);
 
@@ -688,7 +688,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 {
 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
 	unsigned long cpages = 0;
 	unsigned long mpages = 0;
 	struct amdgpu_device *adev = node->adev;
@@ -748,9 +747,9 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 				    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);
+	mpages = svm_migrate_successful_pages(&migrate);
+	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
 
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
@@ -764,7 +763,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 				    node->id, 0, trigger, r);
 out:
 	if (!r && cpages) {
-		mpages = cpages - upages;
 		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
Signature · 1 line
2.34.1
author
#4 lore raw

PATCH v3 2/2 drm/amdkfd: return migration pages from copy function

In reply to James Zhu

dst MIGRATE_PFN_VALID bit and src MIGRATE_PFN_MIGRATE bit should always be set when migration success. cpage includes src MIGRATE_PFN_MIGRATE bit set and MIGRATE_PFN_VALID bit unset pages for both ram and vram when memory is only allocated without access before migration, those pages should be count as migrate_unsuccessful_pages.

-v2 use dst to check MIGRATE_PFN_VALID bit(suggested-by philip) -v3 add warning when vram pages is less than migration pages
return migration pages directly from copy function

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 44 +++++++++---------------
 1 file changed, 16 insertions(+), 28 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +16 -28
index f0b690d4bb46..aad1346bde79 100644
@@ -260,29 +260,15 @@ static void svm_migrate_put_sys_page(unsigned long addr)
 	put_page(page);
 }
 
-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 kfd_node *node, struct svm_range *prange,
 			 struct migrate_vma *migrate, struct dma_fence **mfence,
-			 dma_addr_t *scratch, uint64_t ttm_res_offset)
+			 dma_addr_t *scratch, uint64_t ttm_res_offset, unsigned long *mpages)
 {
 	uint64_t npages = migrate->npages;
 	struct amdgpu_device *adev = node->adev;
 	struct device *dev = adev->dev;
 	struct amdgpu_res_cursor cursor;
-	uint64_t mpages = 0;
 	dma_addr_t *src;
 	uint64_t *dst;
 	uint64_t i, j;
@@ -296,7 +282,8 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 
 	amdgpu_res_first(prange->ttm_res, ttm_res_offset,
 			 npages << PAGE_SHIFT, &cursor);
-	for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) {
+	*mpages = 0;
+	for (i = j = 0; (i < npages) && (*mpages < migrate->cpages); i++) {
 		struct page *spage;
 
 		if (migrate->src[i] & MIGRATE_PFN_MIGRATE) {
@@ -304,7 +291,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, 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]);
-			mpages++;
+			(*mpages)++;
 		}
 		spage = migrate_pfn_to_page(migrate->src[i]);
 		if (spage && !is_zone_device_page(spage)) {
@@ -356,12 +343,12 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 out_free_vram_pages:
 	if (r) {
 		pr_debug("failed %d to copy memory to vram\n", r);
-		for (i = 0; i < npages && mpages; i++) {
+		for (i = 0; i < npages && *mpages; i++) {
 			if (!dst[i])
 				continue;
 			svm_migrate_put_vram_page(adev, dst[i]);
 			migrate->dst[i] = 0;
-			mpages--;
+			(*mpages)--;
 		}
 	}
 
@@ -441,13 +428,12 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 	else
 		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
+	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset, &mpages);
 	migrate_vma_pages(&migrate);
 
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
 	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
 			 mpages, cpages, migrate.npages);
 
@@ -580,7 +566,7 @@ 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, uint64_t npages)
+			dma_addr_t *scratch, uint64_t npages, unsigned long *mpages)
 {
 	struct device *dev = adev->dev;
 	uint64_t *src;
@@ -598,6 +584,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 	src = (uint64_t *)(scratch + npages);
 	dst = scratch;
 
+	*mpages = 0;
 	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
 		struct page *spage;
 
@@ -646,6 +633,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 
 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
+		(*mpages)++;
 		j++;
 	}
 
@@ -688,7 +676,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 {
 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
 	unsigned long cpages = 0;
 	unsigned long mpages = 0;
 	struct amdgpu_device *adev = node->adev;
@@ -745,12 +732,11 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 		pr_debug("0x%lx pages collected\n", cpages);
 
 	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
-				    scratch, npages);
+				    scratch, npages, &mpages);
 	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);
+	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
 
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
@@ -764,7 +750,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 				    node->id, 0, trigger, r);
 out:
 	if (!r && cpages) {
-		mpages = cpages - upages;
 		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
@@ -847,6 +832,9 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
 	}
 
 	if (r >= 0) {
+		WARN_ONCE(prange->vram_pages < mpages,
+			"Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).",
+			prange->vram_pages, mpages);
 		prange->vram_pages -= mpages;
 
 		/* prange does not have vram page set its actual_loc to system
Signature · 1 line
2.34.1
maintainer
#5 lore raw

Re: [PATCH v3 2/2] drm/amdkfd: return migration pages from copy function

Quoted 6 lines · James Zhu

dst MIGRATE_PFN_VALID bit and src MIGRATE_PFN_MIGRATE bit should always be set when migration success. cpage includes src MIGRATE_PFN_MIGRATE bit set and MIGRATE_PFN_VALID bit unset pages for both ram and vram when memory is only allocated without access before migration, those pages should be count as migrate_unsuccessful_pages.

I think the patch is correct but I'm not sure I agree with the explanation. Pages that were never accessed (and are not populated in system memory) should be counted as successfully migrated. It seems they were counted as unsuccessful by the old code.

Some more suggestions inline.

Quoted 26 lines
(quoted diff) +0 -2
-v2 use dst to check MIGRATE_PFN_VALID bit(suggested-by philip)
-v3 add warning when vram pages is less than migration pages
     return migration pages directly from copy function
---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 44 +++++++++---------------
  1 file changed, 16 insertions(+), 28 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -8
index f0b690d4bb46..aad1346bde79 100644
@@ -260,29 +260,15 @@ static void svm_migrate_put_sys_page(unsigned long addr)
  	put_page(page);
  }
 
-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))

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c static void svm_migrate_put_sys_page(unsigned long addr) · line 266

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -6
-			upages++;
-	}
-	return upages;
-}
-
  static int
  svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  			 struct migrate_vma *migrate, struct dma_fence **mfence,
-			 dma_addr_t *scratch, uint64_t ttm_res_offset)
+			 dma_addr_t *scratch, uint64_t ttm_res_offset, unsigned long *mpages)

You could return mpages as the return value. That would match the convention of svm_migrate_vma_to_vram.

Quoted 47 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +5 -5
  {
  	uint64_t npages = migrate->npages;
  	struct amdgpu_device *adev = node->adev;
  	struct device *dev = adev->dev;
  	struct amdgpu_res_cursor cursor;
-	uint64_t mpages = 0;
  	dma_addr_t *src;
  	uint64_t *dst;
  	uint64_t i, j;
@@ -296,7 +282,8 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 
  	amdgpu_res_first(prange->ttm_res, ttm_res_offset,
  			 npages << PAGE_SHIFT, &cursor);
-	for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) {
+	*mpages = 0;
+	for (i = j = 0; (i < npages) && (*mpages < migrate->cpages); i++) {
  		struct page *spage;
 
  		if (migrate->src[i] & MIGRATE_PFN_MIGRATE) {
@@ -304,7 +291,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, 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]);
-			mpages++;
+			(*mpages)++;
  		}
  		spage = migrate_pfn_to_page(migrate->src[i]);
  		if (spage && !is_zone_device_page(spage)) {
@@ -356,12 +343,12 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  out_free_vram_pages:
  	if (r) {
  		pr_debug("failed %d to copy memory to vram\n", r);
-		for (i = 0; i < npages && mpages; i++) {
+		for (i = 0; i < npages && *mpages; i++) {
  			if (!dst[i])
  				continue;
  			svm_migrate_put_vram_page(adev, dst[i]);
  			migrate->dst[i] = 0;
-			mpages--;
+			(*mpages)--;
  		}
  	}
 
@@ -441,13 +428,12 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  	else
  		pr_debug("0x%lx pages collected\n", cpages);
 

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, · line 438

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -2
-	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
+	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset, &mpages);
  	migrate_vma_pages(&migrate);
 
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
  	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
  			 mpages, cpages, migrate.npages);

Maybe change "successful/cpages/npages" to "migrated/collected/requested". I think that would be a better explanation of what these numbers actually mean.

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c static void svm_migrate_page_free(struct page *page) · line 569

Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -1
@@ -580,7 +566,7 @@ 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, uint64_t npages)
+			dma_addr_t *scratch, uint64_t npages, unsigned long *mpages)

You could return mpages as the return value. That would match the convention of svm_migrate_vma_to_ram.

Quoted 30 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +2 -1
  {
  	struct device *dev = adev->dev;
  	uint64_t *src;
@@ -598,6 +584,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  	src = (uint64_t *)(scratch + npages);
  	dst = scratch;
 
+	*mpages = 0;
  	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
  		struct page *spage;
 
@@ -646,6 +633,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 
  		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
+		(*mpages)++;
  		j++;
  	}
 
@@ -688,7 +676,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  {
  	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
  	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
  	unsigned long cpages = 0;
  	unsigned long mpages = 0;
  	struct amdgpu_device *adev = node->adev;
@@ -745,12 +732,11 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  		pr_debug("0x%lx pages collected\n", cpages);
 

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, · line 739

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +3 -4
  	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
-				    scratch, npages);
+				    scratch, npages, &mpages);
  	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);
+	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);

Maybe change "successful/cpages/npages" to "migrated/collected/requested". I think that would be a better explanation of what these numbers actually mean.

Regards,
  Felix

Quoted 21 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +3 -1
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
@@ -764,7 +750,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  				    node->id, 0, trigger, r);
  out:
  	if (!r && cpages) {
-		mpages = cpages - upages;
  		pdd = svm_range_get_pdd_by_node(prange, node);
  		if (pdd)
  			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
@@ -847,6 +832,9 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
  	}
 
  	if (r >= 0) {
+		WARN_ONCE(prange->vram_pages < mpages,
+			"Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).",
+			prange->vram_pages, mpages);
  		prange->vram_pages -= mpages;
 
  		/* prange does not have vram page set its actual_loc to system
#6 lore raw

Re: [PATCH v3 2/2] drm/amdkfd: return migration pages from copy function

Hi Felix

Thanks!

Best Regadrs!

James Zhu

Quoted 9 lines · Felix Kuehling
Quoted 6 lines · James Zhu

dst MIGRATE_PFN_VALID bit and src MIGRATE_PFN_MIGRATE bit should always be set when migration success. cpage includes src MIGRATE_PFN_MIGRATE bit set and MIGRATE_PFN_VALID bit unset pages for both ram and vram when memory is only allocated without access before migration, those pages should be count as migrate_unsuccessful_pages.

I think the patch is correct but I'm not sure I agree with the explanation. Pages that were never accessed (and are not populated in system memory) should be counted as successfully migrated. It seems they were counted as unsuccessful by the old code.

[JZ] Yes, yo are right. I give wrong expiation . Since vram_pages count the un-accessed pages in dst migration,

Quoted 35 lines

Some more suggestions inline.

Quoted 31 lines
(quoted diff) +0 -2
-v2 use dst to check MIGRATE_PFN_VALID bit(suggested-by philip)
-v3 add warning when vram pages is less than migration pages

     return migration pages directly from copy function

---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 44 +++++++++---------------

  1 file changed, 16 insertions(+), 28 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -12
index f0b690d4bb46..aad1346bde79 100644
@@ -260,29 +260,15 @@ static void svm_migrate_put_sys_page(unsigned 
 long addr)
       put_page(page);
   }
   -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;
-}
-
Quoted 8 lines
Quoted 6 lines

  static int
  svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange, struct migrate_vma *migrate, struct dma_fence **mfence, -             dma_addr_t *scratch, uint64_t ttm_res_offset)
+             dma_addr_t *scratch, uint64_t ttm_res_offset, unsigned long *mpages)

You could return mpages as the return value. That would match the convention of svm_migrate_vma_to_vram.

[JZ] Sure

Quoted 71 lines
Quoted 49 lines

  {
      uint64_t npages = migrate->npages;
      struct amdgpu_device *adev = node->adev;
      struct device *dev = adev->dev;
      struct amdgpu_res_cursor cursor;
-    uint64_t mpages = 0;
      dma_addr_t *src;
      uint64_t *dst;
      uint64_t i, j;

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +3 -2
@@ -296,7 +282,8 @@ svm_migrate_copy_to_vram(struct kfd_node *node, 
 struct svm_range *prange,
         amdgpu_res_first(prange->ttm_res, ttm_res_offset,
                npages << PAGE_SHIFT, &cursor);
-    for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) {
+    *mpages = 0;
+    for (i = j = 0; (i < npages) && (*mpages < migrate->cpages); i++) {
           struct page *spage;
             if (migrate->src[i] & MIGRATE_PFN_MIGRATE) {
@@ -304,7 +291,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, 
 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]);
-            mpages++;
+            (*mpages)++;
           }
           spage = migrate_pfn_to_page(migrate->src[i]);

          if (spage && !is_zone_device_page(spage)) {

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -0
@@ -356,12 +343,12 @@ svm_migrate_copy_to_vram(struct kfd_node *node, 
 struct svm_range *prange,

  out_free_vram_pages:
      if (r) {
          pr_debug("failed %d to copy memory to vram\n", r);
-        for (i = 0; i < npages && mpages; i++) {
+        for (i = 0; i < npages && *mpages; i++) {
              if (!dst[i])
                  continue;
              svm_migrate_put_vram_page(adev, dst[i]);
              migrate->dst[i] = 0;
-            mpages--;
+            (*mpages)--;
          }
      }
  @@ -441,13 +428,12 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, else
          pr_debug("0x%lx pages collected\n", cpages);
  -    r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);

Quoted 8 lines

+    r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset, &mpages); migrate_vma_pages(&migrate); svm_migrate_copy_done(adev, mfence); migrate_vma_finalize(&migrate);
  -    mpages = cpages - svm_migrate_unsuccessful_pages(&migrate); pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n", mpages, cpages, migrate.npages);

Maybe change "successful/cpages/npages" to "migrated/collected/requested". I think that would be a better explanation of what these numbers actually mean.

[JZ]  Sure

Quoted 7 lines

  @@ -580,7 +566,7 @@ 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, uint64_t npages)
+            dma_addr_t *scratch, uint64_t npages, unsigned long *mpages)

Quoted 1 line

You could return mpages as the return value. That would match the convention of svm_migrate_vma_to_ram.

[JZ] Sure

Quoted 32 lines
Quoted 32 lines

  {
      struct device *dev = adev->dev;
      uint64_t *src;

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -0
@@ -598,6 +584,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device 
 *adev, struct svm_range *prange,
       src = (uint64_t *)(scratch + npages);
       dst = scratch;
   +    *mpages = 0;
       for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
           struct page *spage;

  @@ -646,6 +633,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
            migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
+        (*mpages)++;
          j++;
      }
  @@ -688,7 +676,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, {
      struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); uint64_t npages = (end - start) >> PAGE_SHIFT;
-    unsigned long upages = npages;
      unsigned long cpages = 0;
      unsigned long mpages = 0;
      struct amdgpu_device *adev = node->adev;

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1 -1
@@ -745,12 +732,11 @@ svm_migrate_vma_to_ram(struct kfd_node *node, 
 struct svm_range *prange,
           pr_debug("0x%lx pages collected\n", cpages);
         r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
-                    scratch, npages);
+                    scratch, npages, &mpages);
Quoted 8 lines

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, · line 739

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +2 -2
       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);
+    pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+         mpages, cpages, migrate.npages);

Maybe change "successful/cpages/npages" to "migrated/collected/requested". I think that would be a better explanation of what these numbers actually mean.

[JZ] Sure

Quoted 26 lines

Regards,
  Felix

Quoted 21 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -0
         svm_migrate_copy_done(adev, mfence);
       migrate_vma_finalize(&migrate);
@@ -764,7 +750,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, 
 struct svm_range *prange,
                       node->id, 0, trigger, r);

  out:
      if (!r && cpages) {
-        mpages = cpages - upages;
          pdd = svm_range_get_pdd_by_node(prange, node); if (pdd)
              WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);

drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +3 -0
@@ -847,6 +832,9 @@ int svm_migrate_vram_to_ram(struct svm_range 
 *prange, struct mm_struct *mm,
       }
         if (r >= 0) {
+        WARN_ONCE(prange->vram_pages < mpages,
+            "Recorded vram pages(0x%llx) should not be less than 
 migration pages(0x%lx).",
+            prange->vram_pages, mpages);
           prange->vram_pages -= mpages;
             /* prange does not have vram page set its actual_loc to system
author
#7 lore raw

PATCH v4 2/2 drm/amdkfd: return migration pages from copy function

In reply to James Zhu

dst MIGRATE_PFN_VALID bit and src MIGRATE_PFN_MIGRATE bit should always be set when migration success. cpage includes src MIGRATE_PFN_MIGRATE bit set and MIGRATE_PFN_VALID bit unset pages for both ram and vram when memory is only allocated without being populated before migration, those ram pages should be counted as migrate pages and those vram pages should not be counted as migrate pages. Here migration pages refer to how many vram pages invloved.

-v2 use dst to check MIGRATE_PFN_VALID bit (suggested-by Philip) -v3 add warning when vram pages is less than migration pages
return migration pages directly from copy function
-v4 correct comments and copy fucntion return mpage (suggested-by Felix)

---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 72 ++++++++++++------------
 1 file changed, 36 insertions(+), 36 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +36 -36
index f0b690d4bb46..5d7eb0234002 100644
@@ -260,20 +260,7 @@ static void svm_migrate_put_sys_page(unsigned long addr)
 	put_page(page);
 }
 
-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
+static long
 svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 			 struct migrate_vma *migrate, struct dma_fence **mfence,
 			 dma_addr_t *scratch, uint64_t ttm_res_offset)
@@ -282,7 +269,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 	struct amdgpu_device *adev = node->adev;
 	struct device *dev = adev->dev;
 	struct amdgpu_res_cursor cursor;
-	uint64_t mpages = 0;
+	long mpages;
 	dma_addr_t *src;
 	uint64_t *dst;
 	uint64_t i, j;
@@ -296,6 +283,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 
 	amdgpu_res_first(prange->ttm_res, ttm_res_offset,
 			 npages << PAGE_SHIFT, &cursor);
+	mpages = 0;
 	for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) {
 		struct page *spage;
 
@@ -356,13 +344,14 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 out_free_vram_pages:
 	if (r) {
 		pr_debug("failed %d to copy memory to vram\n", r);
-		for (i = 0; i < npages && mpages; i++) {
+		while (i-- && mpages) {
 			if (!dst[i])
 				continue;
 			svm_migrate_put_vram_page(adev, dst[i]);
 			migrate->dst[i] = 0;
 			mpages--;
 		}
+		mpages = r;
 	}
 
 #ifdef DEBUG_FORCE_MIXED_DOMAINS
@@ -380,7 +369,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 	}
 #endif
 
-	return r;
+	return mpages;
 }
 
 static long
@@ -395,7 +384,7 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 	struct dma_fence *mfence = NULL;
 	struct migrate_vma migrate = { 0 };
 	unsigned long cpages = 0;
-	unsigned long mpages = 0;
+	long mpages = 0;
 	dma_addr_t *scratch;
 	void *buf;
 	int r = -ENOMEM;
@@ -441,15 +430,17 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 	else
 		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
+	mpages = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
 	migrate_vma_pages(&migrate);
 
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
-	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+	if (mpages >= 0)
+		pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n",
 			 mpages, cpages, migrate.npages);
+	else
+		r = mpages;
 
 	svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages);
 
@@ -459,14 +450,13 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
 				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
 				    0, node->id, trigger, r);
 out:
-	if (!r && mpages) {
+	if (!r && mpages > 0) {
 		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_in, pdd->page_in + mpages);
-
-		return mpages;
 	}
-	return r;
+
+	return r ? r : mpages;
 }
 
 /**
@@ -577,7 +567,7 @@ static void svm_migrate_page_free(struct page *page)
 	}
 }
 
-static int
+static long
 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, uint64_t npages)
@@ -586,6 +576,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 	uint64_t *src;
 	dma_addr_t *dst;
 	struct page *dpage;
+	long mpages;
 	uint64_t i = 0, j;
 	uint64_t addr;
 	int r = 0;
@@ -598,6 +589,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 	src = (uint64_t *)(scratch + npages);
 	dst = scratch;
 
+	mpages = 0;
 	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
 		struct page *spage;
 
@@ -646,6 +638,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 
 		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
+		mpages++;
 		j++;
 	}
 
@@ -655,13 +648,17 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 out_oom:
 	if (r) {
 		pr_debug("failed %d copy to ram\n", r);
-		while (i--) {
+		while (i-- && mpages) {
+			if (!migrate->dst[i])
+				continue;
 			svm_migrate_put_sys_page(dst[i]);
 			migrate->dst[i] = 0;
+			mpages--;
 		}
+		mpages = r;
 	}
 
-	return r;
+	return mpages;
 }
 
 /**
@@ -688,9 +685,8 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 {
 	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
 	unsigned long cpages = 0;
-	unsigned long mpages = 0;
+	long mpages = 0;
 	struct amdgpu_device *adev = node->adev;
 	struct kfd_process_device *pdd;
 	struct dma_fence *mfence = NULL;
@@ -744,13 +740,15 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 	else
 		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
+	mpages = 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);
+	if (mpages >= 0)
+		pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
+	else
+		r = mpages;
 
 	svm_migrate_copy_done(adev, mfence);
 	migrate_vma_finalize(&migrate);
@@ -763,8 +761,7 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
 				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
 				    node->id, 0, trigger, r);
 out:
-	if (!r && cpages) {
-		mpages = cpages - upages;
+	if (!r && mpages > 0) {
 		pdd = svm_range_get_pdd_by_node(prange, node);
 		if (pdd)
 			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
@@ -847,6 +844,9 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
 	}
 
 	if (r >= 0) {
+		WARN_ONCE(prange->vram_pages < mpages,
+			"Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).",
+			prange->vram_pages, mpages);
 		prange->vram_pages -= mpages;
 
 		/* prange does not have vram page set its actual_loc to system
Signature · 1 line
2.34.1
#8 lore raw

Re: [PATCH v4 2/2] drm/amdkfd: return migration pages from copy function

Ping ...

Quoted 239 lines · James Zhu

dst MIGRATE_PFN_VALID bit and src MIGRATE_PFN_MIGRATE bit should always be set when migration success. cpage includes src MIGRATE_PFN_MIGRATE bit set and MIGRATE_PFN_VALID bit unset pages for both ram and vram when memory is only allocated without being populated before migration, those ram pages should be counted as migrate pages and those vram pages should not be counted as migrate pages. Here migration pages refer to how many vram pages invloved.

(quoted diff) +0 -3
-v2 use dst to check MIGRATE_PFN_VALID bit (suggested-by Philip)
-v3 add warning when vram pages is less than migration pages
     return migration pages directly from copy function
-v4 correct comments and copy fucntion return mpage (suggested-by Felix)
---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 72 ++++++++++++------------
  1 file changed, 36 insertions(+), 36 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +36 -36
index f0b690d4bb46..5d7eb0234002 100644
@@ -260,20 +260,7 @@ static void svm_migrate_put_sys_page(unsigned long addr)
  	put_page(page);
  }
 
-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
+static long
  svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  			 struct migrate_vma *migrate, struct dma_fence **mfence,
  			 dma_addr_t *scratch, uint64_t ttm_res_offset)
@@ -282,7 +269,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  	struct amdgpu_device *adev = node->adev;
  	struct device *dev = adev->dev;
  	struct amdgpu_res_cursor cursor;
-	uint64_t mpages = 0;
+	long mpages;
  	dma_addr_t *src;
  	uint64_t *dst;
  	uint64_t i, j;
@@ -296,6 +283,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 
  	amdgpu_res_first(prange->ttm_res, ttm_res_offset,
  			 npages << PAGE_SHIFT, &cursor);
+	mpages = 0;
  	for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) {
  		struct page *spage;
 
@@ -356,13 +344,14 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  out_free_vram_pages:
  	if (r) {
  		pr_debug("failed %d to copy memory to vram\n", r);
-		for (i = 0; i < npages && mpages; i++) {
+		while (i-- && mpages) {
  			if (!dst[i])
  				continue;
  			svm_migrate_put_vram_page(adev, dst[i]);
  			migrate->dst[i] = 0;
  			mpages--;
  		}
+		mpages = r;
  	}
 
  #ifdef DEBUG_FORCE_MIXED_DOMAINS
@@ -380,7 +369,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  	}
  #endif
 
-	return r;
+	return mpages;
  }
 
  static long
@@ -395,7 +384,7 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  	struct dma_fence *mfence = NULL;
  	struct migrate_vma migrate = { 0 };
  	unsigned long cpages = 0;
-	unsigned long mpages = 0;
+	long mpages = 0;
  	dma_addr_t *scratch;
  	void *buf;
  	int r = -ENOMEM;
@@ -441,15 +430,17 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  	else
  		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
+	mpages = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
  	migrate_vma_pages(&migrate);
 
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
-	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+	if (mpages >= 0)
+		pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n",
  			 mpages, cpages, migrate.npages);
+	else
+		r = mpages;
 
  	svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages);
 
@@ -459,14 +450,13 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
  				    0, node->id, trigger, r);
  out:
-	if (!r && mpages) {
+	if (!r && mpages > 0) {
  		pdd = svm_range_get_pdd_by_node(prange, node);
  		if (pdd)
  			WRITE_ONCE(pdd->page_in, pdd->page_in + mpages);
-
-		return mpages;
  	}
-	return r;
+
+	return r ? r : mpages;
  }
 
  /**
@@ -577,7 +567,7 @@ static void svm_migrate_page_free(struct page *page)
  	}
  }
 
-static int
+static long
  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, uint64_t npages)
@@ -586,6 +576,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  	uint64_t *src;
  	dma_addr_t *dst;
  	struct page *dpage;
+	long mpages;
  	uint64_t i = 0, j;
  	uint64_t addr;
  	int r = 0;
@@ -598,6 +589,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  	src = (uint64_t *)(scratch + npages);
  	dst = scratch;
 
+	mpages = 0;
  	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
  		struct page *spage;
 
@@ -646,6 +638,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 
  		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
+		mpages++;
  		j++;
  	}
 
@@ -655,13 +648,17 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  out_oom:
  	if (r) {
  		pr_debug("failed %d copy to ram\n", r);
-		while (i--) {
+		while (i-- && mpages) {
+			if (!migrate->dst[i])
+				continue;
  			svm_migrate_put_sys_page(dst[i]);
  			migrate->dst[i] = 0;
+			mpages--;
  		}
+		mpages = r;
  	}
 
-	return r;
+	return mpages;
  }
 
  /**
@@ -688,9 +685,8 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  {
  	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
  	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
  	unsigned long cpages = 0;
-	unsigned long mpages = 0;
+	long mpages = 0;
  	struct amdgpu_device *adev = node->adev;
  	struct kfd_process_device *pdd;
  	struct dma_fence *mfence = NULL;
@@ -744,13 +740,15 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  	else
  		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
+	mpages = 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);
+	if (mpages >= 0)
+		pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
+	else
+		r = mpages;
 
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
@@ -763,8 +761,7 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
  				    node->id, 0, trigger, r);
  out:
-	if (!r && cpages) {
-		mpages = cpages - upages;
+	if (!r && mpages > 0) {
  		pdd = svm_range_get_pdd_by_node(prange, node);
  		if (pdd)
  			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
@@ -847,6 +844,9 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
  	}
 
  	if (r >= 0) {
+		WARN_ONCE(prange->vram_pages < mpages,
+			"Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).",
+			prange->vram_pages, mpages);
  		prange->vram_pages -= mpages;
 
  		/* prange does not have vram page set its actual_loc to system
maintainer Reviewed-by on v4 2/2
#9 lore raw

Re: [PATCH v4 2/2] drm/amdkfd: return migration pages from copy function

In reply to James Zhu
Quoted 14 lines · James Zhu

dst MIGRATE_PFN_VALID bit and src MIGRATE_PFN_MIGRATE bit should always be set when migration success. cpage includes src MIGRATE_PFN_MIGRATE bit set and MIGRATE_PFN_VALID bit unset pages for both ram and vram when memory is only allocated without being populated before migration, those ram pages should be counted as migrate pages and those vram pages should not be counted as migrate pages. Here migration pages refer to how many vram pages invloved.

(quoted diff) +0 -3
-v2 use dst to check MIGRATE_PFN_VALID bit (suggested-by Philip)
-v3 add warning when vram pages is less than migration pages
     return migration pages directly from copy function
-v4 correct comments and copy fucntion return mpage (suggested-by Felix)
Quoted 1 line
Quoted 224 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 72 ++++++++++++------------
  1 file changed, 36 insertions(+), 36 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +36 -36
index f0b690d4bb46..5d7eb0234002 100644
@@ -260,20 +260,7 @@ static void svm_migrate_put_sys_page(unsigned long addr)
  	put_page(page);
  }
 
-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
+static long
  svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  			 struct migrate_vma *migrate, struct dma_fence **mfence,
  			 dma_addr_t *scratch, uint64_t ttm_res_offset)
@@ -282,7 +269,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  	struct amdgpu_device *adev = node->adev;
  	struct device *dev = adev->dev;
  	struct amdgpu_res_cursor cursor;
-	uint64_t mpages = 0;
+	long mpages;
  	dma_addr_t *src;
  	uint64_t *dst;
  	uint64_t i, j;
@@ -296,6 +283,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
 
  	amdgpu_res_first(prange->ttm_res, ttm_res_offset,
  			 npages << PAGE_SHIFT, &cursor);
+	mpages = 0;
  	for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) {
  		struct page *spage;
 
@@ -356,13 +344,14 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  out_free_vram_pages:
  	if (r) {
  		pr_debug("failed %d to copy memory to vram\n", r);
-		for (i = 0; i < npages && mpages; i++) {
+		while (i-- && mpages) {
  			if (!dst[i])
  				continue;
  			svm_migrate_put_vram_page(adev, dst[i]);
  			migrate->dst[i] = 0;
  			mpages--;
  		}
+		mpages = r;
  	}
 
  #ifdef DEBUG_FORCE_MIXED_DOMAINS
@@ -380,7 +369,7 @@ svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange,
  	}
  #endif
 
-	return r;
+	return mpages;
  }
 
  static long
@@ -395,7 +384,7 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  	struct dma_fence *mfence = NULL;
  	struct migrate_vma migrate = { 0 };
  	unsigned long cpages = 0;
-	unsigned long mpages = 0;
+	long mpages = 0;
  	dma_addr_t *scratch;
  	void *buf;
  	int r = -ENOMEM;
@@ -441,15 +430,17 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  	else
  		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
+	mpages = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset);
  	migrate_vma_pages(&migrate);
 
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
 
-	mpages = cpages - svm_migrate_unsuccessful_pages(&migrate);
-	pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+	if (mpages >= 0)
+		pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n",
  			 mpages, cpages, migrate.npages);
+	else
+		r = mpages;
 
  	svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages);
 
@@ -459,14 +450,13 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange,
  				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
  				    0, node->id, trigger, r);
  out:
-	if (!r && mpages) {
+	if (!r && mpages > 0) {
  		pdd = svm_range_get_pdd_by_node(prange, node);
  		if (pdd)
  			WRITE_ONCE(pdd->page_in, pdd->page_in + mpages);
-
-		return mpages;
  	}
-	return r;
+
+	return r ? r : mpages;
  }
 
  /**
@@ -577,7 +567,7 @@ static void svm_migrate_page_free(struct page *page)
  	}
  }
 
-static int
+static long
  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, uint64_t npages)
@@ -586,6 +576,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  	uint64_t *src;
  	dma_addr_t *dst;
  	struct page *dpage;
+	long mpages;
  	uint64_t i = 0, j;
  	uint64_t addr;
  	int r = 0;
@@ -598,6 +589,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  	src = (uint64_t *)(scratch + npages);
  	dst = scratch;
 
+	mpages = 0;
  	for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
  		struct page *spage;
 
@@ -646,6 +638,7 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  				     dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 
  		migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
+		mpages++;
  		j++;
  	}
 
@@ -655,13 +648,17 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
  out_oom:
  	if (r) {
  		pr_debug("failed %d copy to ram\n", r);
-		while (i--) {
+		while (i-- && mpages) {
+			if (!migrate->dst[i])
+				continue;
  			svm_migrate_put_sys_page(dst[i]);
  			migrate->dst[i] = 0;
+			mpages--;
  		}
+		mpages = r;
  	}
 
-	return r;
+	return mpages;
  }
 
  /**
@@ -688,9 +685,8 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  {
  	struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
  	uint64_t npages = (end - start) >> PAGE_SHIFT;
-	unsigned long upages = npages;
  	unsigned long cpages = 0;
-	unsigned long mpages = 0;
+	long mpages = 0;
  	struct amdgpu_device *adev = node->adev;
  	struct kfd_process_device *pdd;
  	struct dma_fence *mfence = NULL;
@@ -744,13 +740,15 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  	else
  		pr_debug("0x%lx pages collected\n", cpages);
 
-	r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
+	mpages = 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);
+	if (mpages >= 0)
+		pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n",
+		 mpages, cpages, migrate.npages);
+	else
+		r = mpages;
 
  	svm_migrate_copy_done(adev, mfence);
  	migrate_vma_finalize(&migrate);
@@ -763,8 +761,7 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  				    start >> PAGE_SHIFT, end >> PAGE_SHIFT,
  				    node->id, 0, trigger, r);
  out:
-	if (!r && cpages) {
-		mpages = cpages - upages;
+	if (!r && mpages > 0) {
  		pdd = svm_range_get_pdd_by_node(prange, node);
  		if (pdd)
  			WRITE_ONCE(pdd->page_out, pdd->page_out + mpages);
@@ -847,6 +844,9 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
  	}
 
  	if (r >= 0) {
+		WARN_ONCE(prange->vram_pages < mpages,
+			"Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).",
+			prange->vram_pages, mpages);
  		prange->vram_pages -= mpages;
 
  		/* prange does not have vram page set its actual_loc to system
Reviewed-by Philip.Yang on v1 1/2
#10 lore raw
In reply to James Zhu
Quoted 3 lines · James Zhu

upages is assigned under cpages = 0, so it isn't really used in this function.

Quoted 16 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 1 -
  1 file changed, 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +0 -1
index 79251f22b702..f0b690d4bb46 100644
@@ -736,7 +736,6 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange,
  	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)