Re: [PATCH v5 1/6] drm/i915/gem: Count mapped pages in a folio

[Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>]

Hi Krzysztof,

While I'm finally more or less OK with your commit description :-), I
still have some comments to code changes you propose.

On Mon, 2026-08-17 at 09:56 +0000, Krzysztof Karas wrote:
> Before addition of commit 029ae067431a
> ("drm/i915: Fix potential overflow of shmem scatterlist length")
> and after folios were introduced complete folios were always
> allocated, possibly overloading the scatterlist capacity which
> was never truly limited to PAGE_SIZE when requested via
> max_segment. The above commit addressed scatterlist overloading,
> but unintentionally disabled PAGE_SIZE as a valid max_segment
> value and failed to take care of remaining pages from folios
> above max_segment boundary.
> 
> This created a state, where multitude of scatterlists were used
> for the same folio, but never counting enough of its pages to
> jump to the next folio.
> 
> Track how many pages have already been counted in a folio and
> use that number as an offset on consecutive allocations from the
> same folio to ensure it is fully covered before reading next
> folio.
> 
> Fixes: 029ae067431a ("drm/i915: Fix potential overflow of shmem scatterlist length")
> Closes: https://gitlab.freedesktop.org/drm/i915/kernel/-/work_items/15816
> Signed-off-by: Krzysztof Karas <krzysztof.karas@intel.com>
> ---
> v5:
>  * Moved max_segment value validation before allocating shmem table
>  (Sebastian).
>  * Aligned comments (Sebastian).
> 
>  drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 121 ++++++++++++++--------
>  1 file changed, 76 insertions(+), 45 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> index 06543ae60706..f338dc39fad1 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> @@ -68,10 +68,13 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st,
>  			 unsigned int max_segment)
>  {
>  	unsigned int page_count; /* restricted by sg_alloc_table */
> -	unsigned long i;
> +	unsigned long next_pfn = 0; /* suppress gcc warning */
> +	unsigned long folio_start = 0;
> +	unsigned long folio_end = 0;
> +	struct folio *folio = NULL;
>  	struct scatterlist *sg;
> -	unsigned long next_pfn = 0;	/* suppress gcc warning */
>  	gfp_t noreclaim;
> +	unsigned long i;
>  	int ret;
>  
>  	if (overflows_type(size / PAGE_SIZE, page_count))
> @@ -85,6 +88,9 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st,
>  	if (size > resource_size(&mr->region))
>  		return -ENOMEM;
>  
> +	if (max_segment < PAGE_SIZE)
> +		return -EINVAL;
> +
>  	if (sg_alloc_table(st, page_count, GFP_KERNEL | __GFP_NOWARN))
>  		return -ENOMEM;
>  
> @@ -101,7 +107,7 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st,
>  	sg = st->sgl;
>  	st->nents = 0;
>  	for (i = 0; i < page_count; i++) {
> -		struct folio *folio;
> +		unsigned long folio_page_index = 0;
>  		unsigned long nr_pages;
>  		const unsigned int shrink[] = {
>  			I915_SHRINK_BOUND | I915_SHRINK_UNBOUND,
> @@ -109,71 +115,95 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st,
>  		}, *s = shrink;
>  		gfp_t gfp = noreclaim;
>  
> -		do {
> -			cond_resched();
> -			folio = shmem_read_folio_gfp(mapping, i, gfp);
> -			if (!IS_ERR(folio))
> -				break;
> +		/* Grab the next folio if we exhausted the current one. */

I think that even with folio_start < i < folio_end, 
shmem_read_folio_gfp(..., i, ...) should always return the right folio, so 
we shouldn't worry about getting the same folio again in a relatively rare 
case of the folio size exceeding a scatterlist capacity, we only need to 
take care of tracking an offset within that folio.  Then, ...

> +		if (!i || i > folio_end) {
> +			do {
> +				cond_resched();
> +				folio = shmem_read_folio_gfp(mapping, i, gfp);
> +				if (!IS_ERR(folio))
> +					break;
>  
> -			if (!*s) {
> -				ret = PTR_ERR(folio);
> -				goto err_sg;
> -			}
> +				if (!*s) {
> +					ret = PTR_ERR(folio);
> +					goto err_sg;
> +				}
>  
> -			i915_gem_shrink(NULL, i915, 2 * page_count, NULL, *s++);
> -
> -			/*
> -			 * We've tried hard to allocate the memory by reaping
> -			 * our own buffer, now let the real VM do its job and
> -			 * go down in flames if truly OOM.
> -			 *
> -			 * However, since graphics tend to be disposable,
> -			 * defer the oom here by reporting the ENOMEM back
> -			 * to userspace.
> -			 */
> -			if (!*s) {
> -				/* reclaim and warn, but no oom */
> -				gfp = mapping_gfp_mask(mapping);
> +				i915_gem_shrink(NULL, i915, 2 * page_count, NULL, *s++);
>  
>  				/*
> -				 * Our bo are always dirty and so we require
> -				 * kswapd to reclaim our pages (direct reclaim
> -				 * does not effectively begin pageout of our
> -				 * buffers on its own). However, direct reclaim
> -				 * only waits for kswapd when under allocation
> -				 * congestion. So as a result __GFP_RECLAIM is
> -				 * unreliable and fails to actually reclaim our
> -				 * dirty pages -- unless you try over and over
> -				 * again with !__GFP_NORETRY. However, we still
> -				 * want to fail this allocation rather than
> -				 * trigger the out-of-memory killer and for
> -				 * this we want __GFP_RETRY_MAYFAIL.
> +				 * We've tried hard to allocate the memory by reaping
> +				 * our own buffer, now let the real VM do its job and
> +				 * go down in flames if truly OOM.
> +				 *
> +				 * However, since graphics tend to be disposable,
> +				 * defer the oom here by reporting the ENOMEM back
> +				 * to userspace.
>  				 */
> -				gfp |= __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
> -			}
> -		} while (1);
> +				if (!*s) {
> +					/* reclaim and warn, but no oom */
> +					gfp = mapping_gfp_mask(mapping);
> +
> +					/*
> +					 * Our bo are always dirty and so we require
> +					 * kswapd to reclaim our pages (direct reclaim
> +					 * does not effectively begin pageout of our
> +					 * buffers on its own). However, direct reclaim
> +					 * only waits for kswapd when under allocation
> +					 * congestion. So as a result __GFP_RECLAIM is
> +					 * unreliable and fails to actually reclaim our
> +					 * dirty pages -- unless you try over and over
> +					 * again with !__GFP_NORETRY. However, we still
> +					 * want to fail this allocation rather than
> +					 * trigger the out-of-memory killer and for
> +					 * this we want __GFP_RETRY_MAYFAIL.
> +					 */
> +					gfp |= __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
> +				}
> +			} while (1);
> +
> +			folio_start = folio_pgoff(folio);
> +			folio_end = folio_start + folio_nr_pages(folio) - 1;
> +		}
> +
> +		folio_page_index = i - folio_start;
> +		if (WARN_ON_ONCE(folio_page_index >= folio_nr_pages(folio))) {
> +			ret = -EINVAL;
> +			folio_put(folio);
> +			goto err_sg;
> +		}

... the existing code above, including the existing 'do' loop, may be left 
untouched, I believe, an your folio_page_index can easily be calculated 
from an already maintained next_pfn as:

+		folio_page_index = next_pfn - folio_pfn(folio);
+		if (folio_page_index < 0 || folio_page_index >= folio_nr_pages(folio))
+			folio_page_index = 0;

IOW, we need to calculate and apply an offset within the folio only if 
next_pfn is still within the folio's PFN range, otherwise that must be a 
new folio and the offset we apply must be 0.

Then, unless I'm missing something, I believe the patch could be much more 
compact while still correct with my approach.  However, if other reviewers 
are more OK with your proposed changes rather than what I suggest then I 
won't oppose.

>  
>  		nr_pages = min_array(((unsigned long[]) {
> -					folio_nr_pages(folio),
> +					folio_nr_pages(folio) - folio_page_index,
>  					page_count - i,
>  					max_segment / PAGE_SIZE,
>  				      }), 3);
>  
>  		if (!i ||
>  		    sg->length >= max_segment ||
> -		    folio_pfn(folio) != next_pfn) {
> +		    folio_pfn(folio) + folio_page_index != next_pfn) {

I think the existing condition was correct, and still applicable even when 
tracking potential offsets within folios: when there is still some room in 
the current scatterlist, enter the else part that fills up the scatterlist 
with more data only if we've got a new folio that starts exactly where 
the previous one ended to have contiguity of pages preserved, otherwise 
switch to a new scatterlist.

>  			if (i)
>  				sg = sg_next(sg);
>  
>  			st->nents++;
> -			sg_set_folio(sg, folio, nr_pages * PAGE_SIZE, 0);
> +			sg_set_page(sg, folio_page(folio, folio_page_index),
> +				    nr_pages * PAGE_SIZE, 0);
>  		} else {
> +			/*
> +			 * If our prediction about folio placement is true and
> +			 * scatterlist still has space left for more pages,
> +			 * then we land here.
> +			 */
>  			nr_pages = min_t(unsigned long, nr_pages,
>  					 (max_segment - sg->length) / PAGE_SIZE);
>  
>  			sg->length += nr_pages * PAGE_SIZE;
>  		}
> -		next_pfn = folio_pfn(folio) + nr_pages;
> +
> +		/*
> +		 * We assume folios are placed one after the other in memory
> +		 * and predict where the next folio begins.
> +		 */
> +		next_pfn = folio_pfn(folio) + folio_page_index + nr_pages;


>  		i += nr_pages - 1;
>  
>  		/* Check that the i965g/gm workaround works. */
> @@ -186,6 +216,7 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st,
>  	i915_sg_trim(st);
>  
>  	return 0;
> +

Not related, I wouldn't mix it in, unless you have a good justification.

Thanks,
Janusz

>  err_sg:
>  	sg_mark_end(sg);
>  	if (sg != st->sgl) {