[PATCH] drm/i915/userptr: Convert page APIs to folio equivalents in put_pages

[PATCH] drm/i915/userptr: Convert page APIs to folio equivalents in put_pages

[Hongfu Li]

From: Hongfu Li <lihongfu@kylinos.cn>

Cache the folio with page_folio() at the start of the sg_table loop,
then use folio helpers directly to eliminate repeated compound_head()
lookups inside page API wrappers.

Signed-off-by: Hongfu Li <lihongfu@kylinos.cn>
---
 drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
index 043095f93ac6..0ae68c52cc12 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
@@ -160,6 +160,7 @@ i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
 			   struct sg_table *pages)
 {
 	struct sgt_iter sgt_iter;
 	struct page *page;
+	struct folio *folio;
 
 	if (!pages)
@@ -177,7 +178,9 @@ i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
 		obj->mm.dirty = false;
 
 	for_each_sgt_page(page, sgt_iter, pages) {
-		if (obj->mm.dirty && trylock_page(page)) {
+		folio = page_folio(page);
+
+		if (obj->mm.dirty && folio_trylock(folio)) {
 			/*
 			 * As this may not be anonymous memory (e.g. shmem)
 			 * but exist on a real mapping, we have to lock
@@ -193,14 +196,14 @@ i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
 			 * on the folio. Such a try_to_unmap() will result
 			 * in us calling put_pages() and so recursively try
 			 * to lock the page. We avoid that deadlock with
-			 * a trylock_page() and in exchange we risk missing
+			 * a folio_trylock() and in exchange we risk missing
 			 * some page dirtying.
 			 */
-			set_page_dirty(page);
-			unlock_page(page);
+			folio_mark_dirty(folio);
+			folio_unlock(folio);
 		}
 
-		mark_page_accessed(page);
+		folio_mark_accessed(folio);
 	}
 	obj->mm.dirty = false;
 
-- 
2.54.0

Re: [PATCH] drm/i915/userptr: Convert page APIs to folio equivalents in put_pages

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Unconditional PAGE_SIZE stepping over a scatterlist performs non-idempotent folio operations multiple times for large folios.
--

--- Patch [1]: [PATCH] drm/i915/userptr: Convert page APIs to folio equivalents in put_pages ---
commit a2440a520ca40b971ab3143efb8ecac51e56d073
Author: Hongfu Li <lihongfu@kylinos.cn>

drm/i915/userptr: Convert page APIs to folio equivalents in put_pages

This commit replaces page-based APIs with folio equivalents in
i915_gem_userptr_put_pages(). It caches the folio at the start of the loop
to avoid repeated compound head lookups inside page API wrappers.

> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
> index 043095f93ac60..bd91782f4bbc7 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
[ ... ]
> @@ -177,7 +178,9 @@ i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
>  		obj->mm.dirty = false;
>  
>  	for_each_sgt_page(page, sgt_iter, pages) {
> -		if (obj->mm.dirty && trylock_page(page)) {
> +		folio = page_folio(page);
> +
> +		if (obj->mm.dirty && folio_trylock(folio)) {

[Severity: High]
This isn't a bug introduced by this patch, but does this loop process the same
folio multiple times for large folios?

In i915_gem_userptr_put_pages(), the for_each_sgt_page() macro unconditionally
steps by PAGE_SIZE. This causes page_folio() to repeatedly return the same folio
for the tail pages of a large folio.

For non-idempotent operations, this causes the folio to be processed redundantly.
If the scatterlist is backed by a 2MB THP, the same folio is processed 512
times. This means the code unnecessarily locks and unlocks the folio 512 times,
and folio_mark_accessed() causes the folio to repeatedly toggle between
referenced and active states.

Because large folios bypass per-CPU LRU batches, this hammers the global zone
LRU spinlock repeatedly per THP, causing severe system-wide lock contention.

Could we track the last processed folio to avoid duplicate processing of tail
pages in this loop?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260723064651.4100-1-hongfu.li@linux.dev?part=1