[PATCH v2 07/14] mm/shmem: preserve data in clean folios when splitting fails

["David Hildenbrand (Arm)" <david@kernel.org>]

From: Chris Mason <mason@kernel.org>

A shared writable mapping can have dirty PTEs while its shmem folio is
still clean. A read fault can install a writable PTE, and a later store
sets the PTE's dirty bit without marking the folio dirty.

After shmem_undo_range() has handled the boundary folios, a concurrent
read fault into a hole below the new EOF can allocate a large folio.
For example, on a tmpfs mount with huge=always, the fault's offset must
be within the file, but the allocated folio can extend beyond i_size.
The final scan then finds this folio and calls
truncate_inode_partial_folio() without marking it dirty.

Splitting normally collects the PTE dirty bits when unmapping the
folio. An extra reference, for example from vmsplice(), can make it
fail before that step. The fallback unmapping on split failure excludes
shmem, so the helper still sees a clean folio and drops it, losing data
outside the range.

CPU0: ftruncate(fd, 3M)                CPU1: MAP_SHARED, tmpfs huge=always
─────────────────────────────────────  ──────────────────────────────────
i_size = 3M
shmem_undo_range(3M, -1)
  first pass: hole, nothing to do
  boundary folio at 3M: hole, none
                                       read fault at 2M+x (below EOF)
                                         PMD folio [2M,4M), crosses EOF
                                         writable PTE, folio clean
                                       store to 2M+x
                                         PTE dirty, folio still clean
                                       vmsplice(2M+x)
                                         extra folio reference
  whole_folios: finds [2M,4M) folio
  truncate_inode_partial_folio()
    zero [3M,4M)
    folio_split()
      unmap_folio(): extra ref
        → -EAGAIN, PTE dirty bit
          never reaches the folio
    folio clean
      → truncate_inode_folio(whole)
        [2M,3M) dropped
                                       read 2M+x → refault → zeroes

Mark large folios dirty before partial truncation in the final scan,
as the boundary handling already does. Exclude failed-fallocate
rollback, which must discard the folios it added.

Fix generated by an LLM and sent by Chris Mason in private.

Fixes: b9a8a4195c7d ("truncate,shmem: Handle truncates that split large folios")
Cc: stable@vger.kernel.org
Signed-off-by: Chris Mason <mason@kernel.org>
Acked-by: Kiryl Shutsemau (Meta) <kas@kernel.org>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Tested-by: Gregory Price (Meta) <gourry@gourry.net>
[ david: simplify descriptions, adjust comments ]
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
---
 mm/shmem.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/mm/shmem.c b/mm/shmem.c
index 5d21cdf5fb0a5..b615301a311e0 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1464,6 +1464,15 @@ static void shmem_undo_range(struct inode *inode, loff_t lstart, uoff_t lend,
 				VM_BUG_ON_FOLIO(folio_test_writeback(folio),
 						folio);
 
+				/*
+				 * A shared mapping can have dirty PTEs while
+				 * the folio is still clean. Mark the folio
+				 * dirty so a failed split cannot discard data
+				 * outside the range.
+				 */
+				if (!unfalloc && folio_test_large(folio))
+					folio_mark_dirty(folio);
+
 				if (!folio_test_large(folio)) {
 					truncate_inode_folio(mapping, folio);
 				} else if (truncate_inode_partial_folio(folio,

-- 
2.43.0