Skip to contentSkip to conversation

drm/i915/selftests: mark huge_gem_object as not shrinkable

PATCH In review Matthew Auld 4 participants 5 messages Mar 23, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915/selftests: mark huge_gem_object as not shrinkable

It looks like the callers expect a non-volatile object, but it looks the shrinker will discard the object pages anyway, thinking that the pages can be swapped out if the object is marked as WILLNEED. If that's true then it might be better to mark it as volatile and fix the callers instead, but on the other hand huge_gem_objects are fairly unique in that they duplicate pages for the backing store, so maybe shrinking is not that applicable.

---
 drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +1 -2
index fa16f2c3f3ac..2b46c6530da9 100644
@@ -88,8 +88,7 @@ static void huge_put_pages(struct drm_i915_gem_object *obj,
 }
 
 static const struct drm_i915_gem_object_ops huge_ops = {
-	.flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
-		 I915_GEM_OBJECT_IS_SHRINKABLE,
+	.flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
 	.get_pages = huge_get_pages,
 	.put_pages = huge_put_pages,
 };
Signature · 6 lines
2.20.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 1/1
#1 lore raw

Quoting Matthew Auld (2020-03-23 13:08:21)

Quoted 7 lines

It looks like the callers expect a non-volatile object, but it looks the shrinker will discard the object pages anyway, thinking that the pages can be swapped out if the object is marked as WILLNEED. If that's true then it might be better to mark it as volatile and fix the callers instead, but on the other hand huge_gem_objects are fairly unique in that they duplicate pages for the backing store, so maybe shrinking is not that applicable.

Duplication of backing store is irrelevant for the shrinker -- it just deals with trying to make room by releasing objects. If we release the entire object, all duplicate references are released and the pages become recoverable.

Now as to whether the callers were expecting the object to be volatile (for the backing pages to be discarded on swapping) is another question. The answer would be that originally it was used with perma-pinned pages, so it was never a problem. But looking at the users, they do *not* expect to lose data on swapping.

So we need to fix the huge object to not gleefully throw away data, which also means that we cannot shrink it (as there is no backing storage to copy the pages to).

So both making the pages as DONTNEED and IS_SHRINKABLE are technically incorrect.

Quoted 10 lines
---
 drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +0 -0
index fa16f2c3f3ac..2b46c6530da9 100644

drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c static void huge_put_pages(struct drm_i915_gem_object *obj, · line 91

Quoted 6 lines
drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +1 -2
@@ -88,8 +88,7 @@ static void huge_put_pages(struct drm_i915_gem_object *obj,
 }
 
 static const struct drm_i915_gem_object_ops huge_ops = {
-       .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
-                I915_GEM_OBJECT_IS_SHRINKABLE,
+       .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,

but I think the changelog can be clarified and we either include the DONTNEED fixes or follow up.
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#2 lore raw
Quoted 10 lines · Chris Wilson

Quoting Matthew Auld (2020-03-23 13:08:21)

Quoted 7 lines

It looks like the callers expect a non-volatile object, but it looks the shrinker will discard the object pages anyway, thinking that the pages can be swapped out if the object is marked as WILLNEED. If that's true then it might be better to mark it as volatile and fix the callers instead, but on the other hand huge_gem_objects are fairly unique in that they duplicate pages for the backing store, so maybe shrinking is not that applicable.

Quoted 4 lines

Duplication of backing store is irrelevant for the shrinker -- it just deals with trying to make room by releasing objects. If we release the entire object, all duplicate references are released and the pages become recoverable.

Ok. My rough thinking was that the shrinker page accounting(i.e what we return) is currently based on obj->base.size, which might be pure lies for huge_gem_object.

Quoted 7 lines

Now as to whether the callers were expecting the object to be volatile (for the backing pages to be discarded on swapping) is another question. The answer would be that originally it was used with perma-pinned pages, so it was never a problem. But looking at the users, they do *not* expect to lose data on swapping.

Quoted 6 lines

So we need to fix the huge object to not gleefully throw away data, which also means that we cannot shrink it (as there is no backing storage to copy the pages to).

So both making the pages as DONTNEED and IS_SHRINKABLE are technically incorrect.

Do you mean WILLNEED and IS_SHRINKABLE, if object doesn't also support swapping? DONTNEED and IS_SHRINKABLE is correct for volatile objects.

GEM_BUG_ON(is_shrinkable(obj) && !is_swappable(obj) && obj->mm.madv == WILLNEED);

Quoted 20 lines
Quoted 11 lines
---
 drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +0 -0
index fa16f2c3f3ac..2b46c6530da9 100644

drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c static void huge_put_pages(struct drm_i915_gem_object *obj, · line 91

Quoted 6 lines
drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +1 -2
@@ -88,8 +88,7 @@ static void huge_put_pages(struct drm_i915_gem_object *obj,
 }
 
 static const struct drm_i915_gem_object_ops huge_ops = {
-       .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
-                I915_GEM_OBJECT_IS_SHRINKABLE,
+       .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
Quoted 7 lines

but I think the changelog can be clarified and we either include the DONTNEED fixes or follow up.
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw

Quoting Matthew Auld (2020-03-23 13:53:13)

Quoted 31 lines · Chris Wilson
Quoted 10 lines

Quoting Matthew Auld (2020-03-23 13:08:21)

Quoted 7 lines

It looks like the callers expect a non-volatile object, but it looks the shrinker will discard the object pages anyway, thinking that the pages can be swapped out if the object is marked as WILLNEED. If that's true then it might be better to mark it as volatile and fix the callers instead, but on the other hand huge_gem_objects are fairly unique in that they duplicate pages for the backing store, so maybe shrinking is not that applicable.

Quoted 4 lines

Duplication of backing store is irrelevant for the shrinker -- it just deals with trying to make room by releasing objects. If we release the entire object, all duplicate references are released and the pages become recoverable.

Ok. My rough thinking was that the shrinker page accounting(i.e what we return) is currently based on obj->base.size, which might be pure lies for huge_gem_object.

Quoted 11 lines

Now as to whether the callers were expecting the object to be volatile (for the backing pages to be discarded on swapping) is another question. The answer would be that originally it was used with perma-pinned pages, so it was never a problem. But looking at the users, they do *not* expect to lose data on swapping.

So we need to fix the huge object to not gleefully throw away data, which also means that we cannot shrink it (as there is no backing storage to copy the pages to).

Quoted 8 lines
Quoted 2 lines

So both making the pages as DONTNEED and IS_SHRINKABLE are technically incorrect.

Do you mean WILLNEED and IS_SHRINKABLE, if object doesn't also support swapping? DONTNEED and IS_SHRINKABLE is correct for volatile objects.

GEM_BUG_ON(is_shrinkable(obj) && !is_swappable(obj) && obj->mm.madv == WILLNEED);

For the moment, we can just drop IS_SHRINKABLE here. But there is room for a huge object that can take part in shrinker ops by being volatile.

But for the moment there's no pressing need, so we can defer the API problem. (I'm thinking we'll need to expose i915_gem_object_madvise.) -Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#4 lore raw

[Intel-gfx] ✗ Fi.CI.BAT: failure for drm/i915/selftests: mark huge_gem_object as not shrinkable

In reply to Matthew Auld
== Series Details ==

== Series Details ==

Series: drm/i915/selftests: mark huge_gem_object as not shrinkable
URL : https://patchwork.freedesktop.org/series/74972/
State : failure

== Summary ==

Applying: drm/i915/selftests: mark huge_gem_object as not shrinkable Using index info to reconstruct a base tree...
M drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c
Falling back to patching base and 3-way merge...
No changes -- Patch already applied.

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx