Re: [PATCH] drm/i915/gem: Skip writeback of shmem folios with extra references
From: Andi Shyti
Date: Thu Nov 19 2026 - 12:43:22 EST
Hi Kiryl,
On Thu, Nov 19, 2026 at 01:42:41PM +0000, Kiryl Shutsemau (Meta) wrote:
> __shmem_writeback() writes out every dirty folio that is not mapped. A
> folio stops being mapped when munmap() removes its rmap, but the TLB
> flush comes later. Until then another thread can still store to it
> through a stale TLB entry:
>
> CPU0 (munmap) CPU1 (same mm) CPU2 (i915 shrinker)
> zap_huge_pmd()
> folio_remove_rmap_pmd()
> __shmem_writeback()
> clear dirty
> !folio_mapped()
> shmem_write_folio()
> store via stale TLB
> tlb_finish_mmu()
>
> The store leaves the folio clean in the swap cache. Once reclaim frees
> it, the object reads back from swap without that store.
>
> The legacy mmap ioctl maps the shmem file directly, and i915 mounts it
> with huge=within_size on some platforms, so the folio can be PMD-mapped.
> The core mm keeps the rmap until after the flush for PTEs, but not for
> shmem PMDs: shmem has no writeback, and pageout() refuses a folio with
> an extra reference.
>
> Do the same here. The unmapping side holds its reference until the TLB
> is flushed, so a folio with references beyond the page cache and
> writeback_iter() is left to reclaim, as a mapped one already is.
>
> Fixes: 2d6692e642e7 ("drm/i915: Start writeback from the shrinker")
> Signed-off-by: Kiryl Shutsemau (Meta) <kas@xxxxxxxxxx>
> ---
> Found while reviewing David's fix for the same race on file PMDs, which
> leaves shmem out because shmem has no writeback:
> https://patch.msgid.link/20261116-dirty_file_pmd-v2-2-571883053174@xxxxxxxxxx
>
> Build-tested only.
>
> drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 8 +++++++-
> 1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> index bf9e7480f0f5..748772c358ac 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> @@ -316,9 +316,15 @@ void __shmem_writeback(size_t size, struct address_space *mapping)
> * leaving only CPU mmapings around) and add those folios to the LRU
> * instead of invoking writeback so they are aged and paged out
> * as normal.
> + *
> + * The same goes for a folio with references beyond the page cache
> + * and writeback_iter(). A folio that was just unmapped can still be
> + * written through a stale TLB entry until the unmapping side flushes
> + * the TLB and drops its reference. pageout() checks the same.
> */
> while ((folio = writeback_iter(mapping, &wbc, folio, &error))) {
> - if (folio_mapped(folio))
> + if (folio_mapped(folio) ||
> + folio_ref_count(folio) != 1 + folio_nr_pages(folio))
I don't think we need this fix in the current tree. The i915
writeback iterator already skips the shmem folio in the case you
describe, so the new refcount check doesn't change the outcome.
Thanks,
Andi
> folio_redirty_for_writepage(&wbc, folio);
> else
> error = shmem_write_folio(folio);
>
> base-commit: 4405c18c95d2d0f99f6c581ef0ee8591a1aa6ccd
> --
> 2.54.0
>