Skip to contentSkip to conversation

nouveau/svm: Fix to migrate all requested pages

PATCH In review Alistair Popple 3 participants 3 messages Jul 20, 2022 – Jul 28, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nouveau/svm: Fix to migrate all requested pages

Users may request that pages from an OpenCL SVM allocation be migrated to the GPU with clEnqueueSVMMigrateMem(). In Nouveau this will call into nouveau_dmem_migrate_vma() to do the migration. If the total range to be migrated exceeds SG_MAX_SINGLE_ALLOC the pages will be migrated in chunks of size SG_MAX_SINGLE_ALLOC. However a typo in updating the starting address means that only the first chunk will get migrated.

Fix the calculation so that the entire range will get migrated if possible.

Fixes e3d8b0890469 ("drm/nouveau/svm: map pages after migration")
---
 drivers/gpu/drm/nouveau/nouveau_dmem.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/nouveau/nouveau_dmem.c +5 -1
index 7ba66ad68a8a..16356611b5b9 100644
@@ -680,7 +680,11 @@ nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
 		goto out_free_dma;
 
 	for (i = 0; i < npages; i += max) {
-		args.end = start + (max << PAGE_SHIFT);
+		if (args.start + (max << PAGE_SHIFT) > end)
+			args.end = end;
+		else
+			args.end = args.start + (max << PAGE_SHIFT);
+
 		ret = migrate_vma_setup(&args);
 		if (ret)
 			goto out_free_pfns;
Signature · 1 line
2.35.1
Reviewed-by on v1 1/1
#1 lore raw
Quoted 7 lines · Alistair Popple

Users may request that pages from an OpenCL SVM allocation be migrated to the GPU with clEnqueueSVMMigrateMem(). In Nouveau this will call into nouveau_dmem_migrate_vma() to do the migration. If the total range to be migrated exceeds SG_MAX_SINGLE_ALLOC the pages will be migrated in chunks of size SG_MAX_SINGLE_ALLOC. However a typo in updating the starting address means that only the first chunk will get migrated.

Quoted 5 lines

Fix the calculation so that the entire range will get migrated if possible.

Fixes e3d8b0890469 ("drm/nouveau/svm: map pages after migration")

Thanks for fixing this!

maintainer Reviewed-by on v1 1/1
#2 lore raw
In reply to Alistair Popple

hi, in the future if it's taking a while for a review to come back feel free to just send a poke on the thread for the patch you submitted, I do my best to keep up with all of the patches coming in. But there's a lot of email that I get so every now and then one slips through the cracks.

Anyway, this patch looks good to me:

I will push it to the appropriate branch in a little bit

Quoted 34 lines · Alistair Popple

Users may request that pages from an OpenCL SVM allocation be migrated to the GPU with clEnqueueSVMMigrateMem(). In Nouveau this will call into nouveau_dmem_migrate_vma() to do the migration. If the total range to be migrated exceeds SG_MAX_SINGLE_ALLOC the pages will be migrated in chunks of size SG_MAX_SINGLE_ALLOC. However a typo in updating the starting address means that only the first chunk will get migrated.

Fix the calculation so that the entire range will get migrated if possible.

Fixes e3d8b0890469 ("drm/nouveau/svm: map pages after migration")

---
 drivers/gpu/drm/nouveau/nouveau_dmem.c | 6 +++++-

 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/nouveau/nouveau_dmem.c +0 -0
index 7ba66ad68a8a..16356611b5b9 100644
@@ -680,7 +680,11 @@ nouveau_dmem_migrate_vma(struct nouveau_drm *drm,
                 goto out_free_dma;


        for (i = 0; i < npages; i += max) {
-               args.end = start + (max << PAGE_SHIFT);
+               if (args.start + (max << PAGE_SHIFT) > end) +                       args.end = end;
+               else
+                       args.end = args.start + (max << PAGE_SHIFT); +
                ret = migrate_vma_setup(&args);
                if (ret)
                        goto out_free_pfns;

Signature · 3 lines
Cheers,
 Lyude Paul (she/her)
 Software Engineer at Red Hat