Re: [PATCH] mm/swap: fix swap_cache_get_folio() returning a folio that left the swap cache

[Jiayuan Chen]

On 10/15/26 2:05 PM, Jiayuan Chen wrote:
> From: Jiayuan Chen <jiayuan.chen@xxxxxxxxxx>
>
> swap_cache_get_folio() reads the folio from the swap table and takes a
> reference with folio_try_get(), without any lock. If the folio is
> removed from the swap cache and freed in between, and its memory gets
> reused, folio_try_get() still succeeds and we return a page that has
> nothing to do with the swap entry:
>
>    CPU 0                                  CPU 1
>    do_swap_page()
>      swap_cache_get_folio()
>        swp_tb = swap_table_get()
>        folio = swp_tb_to_folio(swp_tb)
>                                           madvise(MADV_DONTNEED)
>                                             zap_nonpresent_ptes()
>                                               swap_put_entries_direct()
>                                                 swap_put_entries_cluster()
>                                                   __try_to_reclaim_swap()
>                                                     swap_cache_del_folio()
>                                                     folio_put()   <- freed
>                                           page reused, e.g. as an anon
>                                           folio of another task
>        folio_try_get(folio)               <- succeeds
>        return folio
>      swap_update_readahead()              <- may clear PG_readahead on it
>      folio_lock_or_retry()                <- locks it
>      folio_matches_swap_entry() fails, unlock and put
>
> filemap_get_entry() does the same lockless lookup for the page cache,
> and after folio_try_get() it rechecks the xarray, because the folio may
> have been freed and reused in between. swap_cache_get_folio() lost
> that recheck when the swap cache moved to the swap table.
>
> So CPU 0 writes to a folio it doesn't own: it may clear PG_readahead and
> takes the folio lock of someone else's page, before
> folio_matches_swap_entry() finds out the folio is not ours.
>
> Recheck the slot after folio_try_get() and retry if it no longer holds
> this folio.
>
> Fixes: 8578e0c00dcf ("mm, swap: use the swap table for the swap cache and switch API")
> Cc: Jiayuan Chen <jiayuan.chen@xxxxxxxxx>
> Signed-off-by: Jiayuan Chen <jiayuan.chen@xxxxxxxxxx>
> ---
> https://sashiko.dev/#/patchset/20261014090821.227366-1-jiayuan.chen@xxxxxxxxx
> ---
>   mm/swap_state.c | 16 +++++++++++++---
>   1 file changed, 13 insertions(+), 3 deletions(-)
>
> diff --git a/mm/swap_state.c b/mm/swap_state.c
> index ebef568cd62e..d9e36c505398 100644
> --- a/mm/swap_state.c
> +++ b/mm/swap_state.c
> @@ -92,17 +92,27 @@ void show_swap_cache_info(void)
>    */
>   struct folio *swap_cache_get_folio(swp_entry_t entry)
>   {
> +	struct swap_cluster_info *ci = __swap_entry_to_cluster(entry);
> +	unsigned int ci_off = swp_cluster_offset(entry);
>   	unsigned long swp_tb;
>   	struct folio *folio;
>   
>   	for (;;) {
> -		swp_tb = swap_table_get(__swap_entry_to_cluster(entry),
> -					swp_cluster_offset(entry));
> +		swp_tb = swap_table_get(ci, ci_off);
>   		if (!swp_tb_is_folio(swp_tb))
>   			return NULL;
>   		folio = swp_tb_to_folio(swp_tb);
> -		if (likely(folio_try_get(folio)))
> +		if (!folio_try_get(folio))
> +			continue;
> +		/*
> +		 * The folio may have been freed and its memory reused before
> +		 * we got the reference. Make sure it is still in the slot.
> +		 */
> +		swp_tb = swap_table_get(ci, ci_off);
> +		if (likely(swp_tb_is_folio(swp_tb) &&
> +			   swp_tb_to_folio(swp_tb) == folio))
>   			return folio;
> +		folio_put(folio);
>   	}
>   
>   	return NULL;




I spent some more time on this and I'm not happy with any of the
options I have. Advice welcome.

  1. Trigger the worker where the zap gives up (this patch). On the
     swap entry side that works, __try_to_reclaim_swap() has the si.
     On the present PTE side it doesn't: free_swap_cache() lost the
     trylock, so it can't read folio->swap to find the si, as Kairui
     pointed out. And that trylock-and-give-up is not in one place,
     free_swap_cache(), do_wp_page() and __uprobe_write() all have
     their own copy. Chasing every one of them doesn't feel right.

  2. Clean up at memcg offline. Doesn't cover the root cgroup or a
     cgroup that stays alive, and on mm-new the folios are reparented
     at offline anyway, so there is nothing left to find there.

  3. A swapon flag, say --clean-cache, that enables the cleanup.
     Pro: covers everything, and the cleanup can live in its own
     function without touching the existing paths. Con: we know to
     turn it on because we hit this, others won't. Needs util-linux
     too.

  4. A per-device worker that walks the device on its own, say 1024
     clusters per second, backing off when a pass finds nothing.
     Pro: same as 3, and nothing to turn on. Con: a worker that wakes
     up on its own for this, I expect objections.

  5. A sysctl (or a drop_caches bit) to kick the cleanup by hand.
     Pro: same as 4, all in the worker. Con: same as 3, someone has
     to know to poke it.

4 is the one I'd pick if nobody objects. Thoughts?