Re: [PATCH v2] mm: shmem: limit forced anonymous shmem huge orders to MAX_PAGECACHE_ORDER

From: David Hildenbrand (Arm)

Date: Wed Nov 18 2026 - 04:31:27 EST


On 11/18/26 03:45, Jingyi Sun wrote:
> When the top-level shmem_enabled knob is set to "force",
> shmem_allowable_huge_orders() returns huge_shmem_orders_inherit for
> anonymous shmem without masking it with THP_ORDERS_ALL_FILE_DEFAULT,
> unlike the other return path at the end of the same function.
>
> huge_shmem_orders_inherit defaults to BIT(HPAGE_PMD_ORDER). On arm64
> with 64K base pages HPAGE_PMD_ORDER is 13, while MAX_PAGECACHE_ORDER
> is 11 because of the xarray split limit. With shmem_enabled=force, an
> anonymous MAP_SHARED mapping can therefore get a 512M (order-13) folio
> in the page cache. Splitting it later, e.g. on swap-out from
> shmem_writeout(), triggers
>
> WARNING: lib/xarray.c:1051 at xas_split_alloc
>
> and the split fails, so the folio can never be swapped out.
>
> This is the same problem that commit 9fd154ba926b ("mm/shmem: disable
> PMD-sized page cache if needed") and commit b66b1b71d7ff ("mm: shmem:
> avoid allocating huge pages larger than MAX_PAGECACHE_ORDER for shmem")
> fixed for the other paths; the force branch was left out. Mask it with
> THP_ORDERS_ALL_FILE_DEFAULT as well.
>
> Reproduced on v7.3-rc2 in QEMU (arm64, 64K pages, -cpu cortex-a72,
> 4G RAM, zram swap): with shmem_enabled=force, a PMD-aligned
> MAP_SHARED|MAP_ANONYMOUS mapping of one PMD size got a single
> PMD-mapped folio (ShmemPmdMapped: 524288 kB) and MADV_PAGEOUT
> triggered the warning above in 3/3 runs. With this change: 0/3.
> shmem_enabled=always and 4K pages did not show the warning. The code is
> unchanged in v7.4-rc3 and in mm-unstable, but the test has not been
> re-run on those trees.
>
> The issue was found with a research static analyzer for
> cross-architecture dependency bugs and confirmed with the QEMU
> reproducer described above.
>
> Fixes: e7a2ab7b3bb5 ("mm: shmem: add mTHP support for anonymous shmem")
> Cc: stable@xxxxxxxxxxxxxxx
> Cc: Luiz Capitulino <luizcap@xxxxxxxxxx>
> Reviewed-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> Signed-off-by: Jingyi Sun <1045704958@xxxxxx>
> ---
> Changes in v2:
> - Dropped the Assisted-by tag (Baolin).
> - Collected Reviewed-by from Baolin.
> - Dropped the note about the overlap with Luiz's mTHP series; Luiz
> confirmed his series will land after this patch.
>
> v1: https://lore.kernel.org/all/20261117121626.1058-1-1045704958@xxxxxx/
>
> mm/shmem.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 10cef89b00b0..41bb3f2c6b63 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -2114,7 +2114,8 @@ unsigned long shmem_allowable_huge_orders(struct inode *inode,
> * means non-PMD sized THP can not override 'huge' mount option now.
> */
> if (shmem_huge == SHMEM_HUGE_FORCE)
> - return READ_ONCE(huge_shmem_orders_inherit);
> + return THP_ORDERS_ALL_FILE_DEFAULT &
> + READ_ONCE(huge_shmem_orders_inherit);

could we instead set

mask = READ_ONCE(huge_shmem_orders_inherit);

And sue a goto to the end of the function where we have

return THP_ORDERS_ALL_FILE_DEFAULT & mask;

Avoiding the goto means more churn I guess.

--
Cheers,

David