Linux-mm Archive on lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH] mm/shmem: fix swapin livelock on a mismatched large swap cache folio
@ 2026-10-27 20:33 Mani Tofigh via B4 Relay
  2026-10-30  8:44 ` Baolin Wang
  2026-11-18 23:06 ` David Hildenbrand (Arm)
  0 siblings, 2 replies; 3+ messages in thread
From: Mani Tofigh via B4 Relay @ 2026-10-27 20:33 UTC (permalink / raw)
  To: Hugh Dickins, Baolin Wang, Andrew Morton, Kairui Song
  Cc: linux-mm, linux-kernel, tz2294, Chris Li, Kemeng Shi, Nhat Pham,
	Baoquan He, Barry Song, Youngjun Park, David Hildenbrand, stable,
	Mani Tofigh

From: Mani Tofigh <mt4022@columbia.edu>

On a SWP_SYNCHRONOUS_IO swap device, a large shmem folio can be left in
the swap cache over swap slots that another file now uses.  Reading
that other file then loops forever in shmem_get_folio_gfp():
shmem_swapin_folio() finds the large folio, rounds the index and the
swap entry down to the folio's size, finds no matching entry in the
file and returns -EEXIST, and the retry finds the same folio again.
Since the task keeps the mapping locked, a munmap() of the same mapping
waits forever.

The folio is left behind when the slots are freed and reused while the
swapin is still allocating it:

  Task A (reads X)                      Task B
  ----------------                      ------
  shmem_swapin_folio()
    allocate a large folio
    for X's slots
                                        punch a hole in X
                                        /* X's slots freed */
                                        swap out file Y
                                        /* Y reuses those slots */
    add the folio to the swap
    cache over those slots
    shmem_confirm_swap() fails
    /* -EEXIST, the folio stays
       in the swap cache */

  Task C (reads Y)
  ----------------
  shmem_swapin_folio()
    finds the large folio
    rounds down to its start,
    where Y has no entry
    /* -EEXIST, retried forever */

Before commit c246d236b18b ("mm/shmem: never bypass the swap cache for
SWP_SYNCHRONOUS_IO"), a direct swapin that failed this way freed its
folio instead of leaving it in the swap cache.

do_swap_page() already drops a large swap cache folio that does not
fit for anonymous memory.  Do the same in shmem_swapin_folio() and on
a -EEXIST, remove a large folio that is not
anonymous/dirty/under-writeback from the swap cache so that the retry
swaps in a folio that fits.

The reproducer below fills a small brd swap device so that a 64K shmem
folio and the 4K pages of another file land in the same slots, and races
a read of the first file with a hole punch and a swapout of the second.
This bug was identified and confirmed with the help of a deterministic
hypervisor currently under development, but the non-deterministic C
reproducer can typically trigger it in the 2-4s time range.  With the
fix, nothing is triggered after 600 seconds.

Fixes: c246d236b18b ("mm/shmem: never bypass the swap cache for SWP_SYNCHRONOUS_IO")
Cc: stable@vger.kernel.org
Signed-off-by: Mani Tofigh <mt4022@columbia.edu>
---
Two patches in David Hildenbrand's v2 shmem fixes series are related
to this one.  "mm/shmem: fix truncate hang in large swap entry" [1]
makes truncation retry -EEXIST directly, so it can also loop on such a
folio.  "mm/shmem: preserve allocation errors when swapping in large
entries" [2] fixes a different -EEXIST loop, where a failed allocation
is reported as -EEXIST.

[1] https://lore.kernel.org/all/20261027-fixes_shmem-v2-10-86a00952acb9@kernel.org/
[2] https://lore.kernel.org/all/20261027-fixes_shmem-v2-11-86a00952acb9@kernel.org/

Reproducer: https://github.com/manitofigh/files/blob/main/linux-repros/shmemloop.c
---
 mm/shmem.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/mm/shmem.c b/mm/shmem.c
index 13e0786436cd..f1da8432d55c 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -2568,11 +2568,14 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
 	 * and swap cache folios are never partially freed.
 	 */
 	folio_lock(folio);
-	if (!folio_matches_swap_entry(folio, swap) ||
-	    shmem_confirm_swap(mapping, index, swap) < 0) {
+	if (!folio_matches_swap_entry(folio, swap)) {
 		error = -EEXIST;
 		goto unlock;
 	}
+	if (shmem_confirm_swap(mapping, index, swap) < 0) {
+		error = -EEXIST;
+		goto failed;
+	}
 	if (!folio_test_uptodate(folio)) {
 		error = -EIO;
 		goto failed;
@@ -2613,6 +2616,15 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
 		error = -EEXIST;
 	if (error == -EIO)
 		shmem_set_folio_swapin_error(inode, index, folio, swap);
+	/*
+	 * A large folio that does not fit the mapping, e.g. because its swap
+	 * entries were freed and reused, would be found again on every retry.
+	 * Drop it from the swap cache so the retry can swap in a fitting one.
+	 */
+	if (error == -EEXIST && folio && folio_test_large(folio) &&
+	    !folio_test_anon(folio) && !folio_test_dirty(folio) &&
+	    !folio_test_writeback(folio))
+		swap_cache_del_folio(folio);
 unlock:
 	if (folio)
 		folio_unlock(folio);

---
base-commit: bae9a042b03b27924c6fd5bdc00a20d6a44d6041
change-id: 20261026-shmem-swapin-mismatch-fd7aaae7152b

Best regards,
--  
Mani Tofigh <mt4022@columbia.edu>




^ permalink raw reply related	[flat|nested] 3+ messages in thread

* Re: [PATCH] mm/shmem: fix swapin livelock on a mismatched large swap cache folio
  2026-10-27 20:33 [PATCH] mm/shmem: fix swapin livelock on a mismatched large swap cache folio Mani Tofigh via B4 Relay
@ 2026-10-30  8:44 ` Baolin Wang
  2026-11-18 23:06 ` David Hildenbrand (Arm)
  1 sibling, 0 replies; 3+ messages in thread
From: Baolin Wang @ 2026-10-30  8:44 UTC (permalink / raw)
  To: mt4022, Hugh Dickins, Andrew Morton, Kairui Song
  Cc: linux-mm, linux-kernel, tz2294, Chris Li, Kemeng Shi, Nhat Pham,
	Baoquan He, Barry Song, Youngjun Park, David Hildenbrand, stable



On 10/28/26 4:33 AM, Mani Tofigh via B4 Relay wrote:
> From: Mani Tofigh <mt4022@columbia.edu>
> 
> On a SWP_SYNCHRONOUS_IO swap device, a large shmem folio can be left in
> the swap cache over swap slots that another file now uses.  Reading
> that other file then loops forever in shmem_get_folio_gfp():
> shmem_swapin_folio() finds the large folio, rounds the index and the
> swap entry down to the folio's size, finds no matching entry in the
> file and returns -EEXIST, and the retry finds the same folio again.
> Since the task keeps the mapping locked, a munmap() of the same mapping
> waits forever.
> 
> The folio is left behind when the slots are freed and reused while the
> swapin is still allocating it:
> 
>    Task A (reads X)                      Task B
>    ----------------                      ------
>    shmem_swapin_folio()
>      allocate a large folio
>      for X's slots
>                                          punch a hole in X
>                                          /* X's slots freed */
>                                          swap out file Y
>                                          /* Y reuses those slots */
>      add the folio to the swap
>      cache over those slots
>      shmem_confirm_swap() fails
>      /* -EEXIST, the folio stays
>         in the swap cache */
> 
>    Task C (reads Y)
>    ----------------
>    shmem_swapin_folio()
>      finds the large folio
>      rounds down to its start,
>      where Y has no entry
>      /* -EEXIST, retried forever */
> 
> Before commit c246d236b18b ("mm/shmem: never bypass the swap cache for
> SWP_SYNCHRONOUS_IO"), a direct swapin that failed this way freed its
> folio instead of leaving it in the swap cache.
> 
> do_swap_page() already drops a large swap cache folio that does not
> fit for anonymous memory.  Do the same in shmem_swapin_folio() and on
> a -EEXIST, remove a large folio that is not
> anonymous/dirty/under-writeback from the swap cache so that the retry
> swaps in a folio that fits.
> 
> The reproducer below fills a small brd swap device so that a 64K shmem
> folio and the 4K pages of another file land in the same slots, and races
> a read of the first file with a hole punch and a swapout of the second.
> This bug was identified and confirmed with the help of a deterministic
> hypervisor currently under development, but the non-deterministic C
> reproducer can typically trigger it in the 2-4s time range.  With the
> fix, nothing is triggered after 600 seconds.
> 
> Fixes: c246d236b18b ("mm/shmem: never bypass the swap cache for SWP_SYNCHRONOUS_IO")
> Cc: stable@vger.kernel.org
> Signed-off-by: Mani Tofigh <mt4022@columbia.edu>
> ---
> Two patches in David Hildenbrand's v2 shmem fixes series are related
> to this one.  "mm/shmem: fix truncate hang in large swap entry" [1]
> makes truncation retry -EEXIST directly, so it can also loop on such a
> folio.  "mm/shmem: preserve allocation errors when swapping in large
> entries" [2] fixes a different -EEXIST loop, where a failed allocation
> is reported as -EEXIST.
> 
> [1] https://lore.kernel.org/all/20261027-fixes_shmem-v2-10-86a00952acb9@kernel.org/
> [2] https://lore.kernel.org/all/20261027-fixes_shmem-v2-11-86a00952acb9@kernel.org/
> 
> Reproducer: https://github.com/manitofigh/files/blob/main/linux-repros/shmemloop.c
> ---

Thanks for the fix. It's a really complicated race and I spent some time 
understanding it:) . I tried several different ways to fix it, but your 
patch looks like the simplest one. So,

Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>


^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] mm/shmem: fix swapin livelock on a mismatched large swap cache folio
  2026-10-27 20:33 [PATCH] mm/shmem: fix swapin livelock on a mismatched large swap cache folio Mani Tofigh via B4 Relay
  2026-10-30  8:44 ` Baolin Wang
@ 2026-11-18 23:06 ` David Hildenbrand (Arm)
  1 sibling, 0 replies; 3+ messages in thread
From: David Hildenbrand (Arm) @ 2026-11-18 23:06 UTC (permalink / raw)
  To: mt4022, Hugh Dickins, Baolin Wang, Andrew Morton, Kairui Song
  Cc: linux-mm, linux-kernel, tz2294, Chris Li, Kemeng Shi, Nhat Pham,
	Baoquan He, Barry Song, Youngjun Park, stable

On 10/27/26 21:33, Mani Tofigh via B4 Relay wrote:
> From: Mani Tofigh <mt4022@columbia.edu>
> 
> On a SWP_SYNCHRONOUS_IO swap device, a large shmem folio can be left in
> the swap cache over swap slots that another file now uses.  Reading
> that other file then loops forever in shmem_get_folio_gfp():
> shmem_swapin_folio() finds the large folio, rounds the index and the
> swap entry down to the folio's size, finds no matching entry in the
> file and returns -EEXIST, and the retry finds the same folio again.
> Since the task keeps the mapping locked, a munmap() of the same mapping
> waits forever.
> 
> The folio is left behind when the slots are freed and reused while the
> swapin is still allocating it:
> 
>   Task A (reads X)                      Task B
>   ----------------                      ------
>   shmem_swapin_folio()
>     allocate a large folio
>     for X's slots
>                                         punch a hole in X
>                                         /* X's slots freed */
>                                         swap out file Y
>                                         /* Y reuses those slots */
>     add the folio to the swap
>     cache over those slots
>     shmem_confirm_swap() fails
>     /* -EEXIST, the folio stays
>        in the swap cache */
> 
>   Task C (reads Y)
>   ----------------
>   shmem_swapin_folio()
>     finds the large folio
>     rounds down to its start,
>     where Y has no entry
>     /* -EEXIST, retried forever */
> 
> Before commit c246d236b18b ("mm/shmem: never bypass the swap cache for
> SWP_SYNCHRONOUS_IO"), a direct swapin that failed this way freed its
> folio instead of leaving it in the swap cache.
> 
> do_swap_page() already drops a large swap cache folio that does not
> fit for anonymous memory.  Do the same in shmem_swapin_folio() and on
> a -EEXIST, remove a large folio that is not
> anonymous/dirty/under-writeback from the swap cache so that the retry
> swaps in a folio that fits.
> 
> The reproducer below fills a small brd swap device so that a 64K shmem
> folio and the 4K pages of another file land in the same slots, and races
> a read of the first file with a hole punch and a swapout of the second.
> This bug was identified and confirmed with the help of a deterministic
> hypervisor currently under development, but the non-deterministic C
> reproducer can typically trigger it in the 2-4s time range.  With the
> fix, nothing is triggered after 600 seconds.
> 
> Fixes: c246d236b18b ("mm/shmem: never bypass the swap cache for SWP_SYNCHRONOUS_IO")
> Cc: stable@vger.kernel.org
> Signed-off-by: Mani Tofigh <mt4022@columbia.edu>
> ---
> Two patches in David Hildenbrand's v2 shmem fixes series are related
> to this one.  "mm/shmem: fix truncate hang in large swap entry" [1]
> makes truncation retry -EEXIST directly, so it can also loop on such a
> folio.  "mm/shmem: preserve allocation errors when swapping in large
> entries" [2] fixes a different -EEXIST loop, where a failed allocation
> is reported as -EEXIST.
> 
> [1] https://lore.kernel.org/all/20261027-fixes_shmem-v2-10-86a00952acb9@kernel.org/
> [2] https://lore.kernel.org/all/20261027-fixes_shmem-v2-11-86a00952acb9@kernel.org/
> 
> Reproducer: https://github.com/manitofigh/files/blob/main/linux-repros/shmemloop.c
> ---
>  mm/shmem.c | 16 ++++++++++++++--
>  1 file changed, 14 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 13e0786436cd..f1da8432d55c 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -2568,11 +2568,14 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>  	 * and swap cache folios are never partially freed.
>  	 */
>  	folio_lock(folio);
> -	if (!folio_matches_swap_entry(folio, swap) ||
> -	    shmem_confirm_swap(mapping, index, swap) < 0) {
> +	if (!folio_matches_swap_entry(folio, swap)) {
>  		error = -EEXIST;
>  		goto unlock;
>  	}
> +	if (shmem_confirm_swap(mapping, index, swap) < 0) {
> +		error = -EEXIST;
> +		goto failed;
> +	}
>  	if (!folio_test_uptodate(folio)) {
>  		error = -EIO;
>  		goto failed;
> @@ -2613,6 +2616,15 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>  		error = -EEXIST;
>  	if (error == -EIO)
>  		shmem_set_folio_swapin_error(inode, index, folio, swap);
> +	/*
> +	 * A large folio that does not fit the mapping, e.g. because its swap
> +	 * entries were freed and reused, would be found again on every retry.
> +	 * Drop it from the swap cache so the retry can swap in a fitting one.
> +	 */
> +	if (error == -EEXIST && folio && folio_test_large(folio) &&
> +	    !folio_test_anon(folio) && !folio_test_dirty(folio) &&
> +	    !folio_test_writeback(folio))
> +		swap_cache_del_folio(folio);
>  unlock:
>  	if (folio)
>  		folio_unlock(folio);
> 
> ---

I have an AI fix here:

Author: Chris Mason <mason@kernel.org>
Date:   Wed Nov 11 18:44:55 2026 +0000

    mm/shmem: don't drop a folio whose swap write failed

    If the swap write of a large shmem folio fails while a swap-in of
    another file finds that folio in the swap cache, its data can be
    silently lost. A read of the file then returns whatever the swap device
    holds.

    "mm/shmem: fix swapin livelock on a mismatched large swap cache folio",
    earlier in this series, lets shmem_swapin_folio() drop a large folio
    from the swap cache, if the folio does not fit the file that is swapped
    in and is neither dirty nor under writeback. A folio that does not fit
    can belong to another file, with its swap write still in flight.

    shmem_swapin_folio() holds the folio lock, but the code that completes a
    write does not take that lock. And a failed write sets dirty before it
    ends writeback. So the folio can look clean and written, although its
    write has failed:

        shmem_swapin_folio()              swap_write_end(), failed write
        ====================              ==============================
        folio_test_dirty() == false
                                          folio_mark_dirty()
                                          folio_end_writeback()
        folio_test_writeback() == false
        swap_cache_del_folio()
          /* the only copy of the data */

    The fix used here is to test writeback first, and dirty second.

    Signed-off-by: Chris Mason <mason@kernel.org>

diff --git a/mm/shmem.c b/mm/shmem.c
index 915b312e14175..457a502353a9a 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -2809,10 +2809,13 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
         * A large folio that does not fit the mapping, e.g. because its swap
         * entries were freed and reused, would be found again on every retry.
         * Drop it from the swap cache so the retry can swap in a fitting one.
+        * A failed write sets dirty and then ends writeback, with no folio
+        * lock, so test writeback first. Both flags share folio->flags, so the
+        * order of the two loads holds.
         */
        if (error == -EEXIST && folio && folio_test_large(folio) &&
-           !folio_test_anon(folio) && !folio_test_dirty(folio) &&
-           !folio_test_writeback(folio))
+           !folio_test_anon(folio) && !folio_test_writeback(folio) &&
+           !folio_test_dirty(folio))
                swap_cache_del_folio(folio);
 unlock:
        if (folio)


@Mani, can you double check if that makes sense (can we actually trigger that), and
if so, send it out officially, with Fixes tag ?

Feel free to rework the patch as you please.



-- 
Cheers,

David


^ permalink raw reply related	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2026-11-18 23:06 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-10-27 20:33 [PATCH] mm/shmem: fix swapin livelock on a mismatched large swap cache folio Mani Tofigh via B4 Relay
2026-10-30  8:44 ` Baolin Wang
2026-11-18 23:06 ` David Hildenbrand (Arm)

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox