Re: [PATCH v3] mm/huge_memory: prevent losing PG_waiters in __split_folio_to_order()

[Zi Yan <ziy@nvidia.com>]

On 9 Nov 2026, at 16:05, Zi Yan wrote:

> On 9 Nov 2026, at 15:21, David Hildenbrand (Arm) wrote:
>
>> On 11/9/26 20:32, Zi Yan wrote:
>>> __split_folio_to_order() calls prep_compound_page() on each new head page
>>> to create a compound page after clear_compound_head(). prep_compound_page()
>>> uses non atomic __SetPageHead(). It is mostly OK, since the folio is locked
>>> and frozen and no one else is supposed to modify the folio's flags.
>>>
>>> However, code calling migration_entry_wait() can reach the folio by PFN,
>>> then access its flags. It is because softleaf_entry_wait_on_locked() used
>>> by migration_entry_wait() sets PG_waiters concurrently without taking a
>>> reference and the frozen refcount during split cannot block it. Meanwhile,
>>> __SetPageHead() does a plain read and write, so PG_waiters atomically set
>>> by softleaf_entry_wait_on_locked() will be lost if the set happens after
>>> __SetPageHead()'s read and before its write.
>>>
>>> One example code interleaving:
>>>
>>> CPU0: __split_folio_to_order(new_order>0)  CPU1: do_swap_page()
>>>
>>> 1. clear_compound_head(new_head)
>>>    new_head->compound_info = 0
>>>    (page_folio(new_head) == new_head now)
>>>
>>>                                            2. migration_entry_wait()
>>>                                              softleaf_to_folio(entry)
>>>                                              page_folio() -> new_head
>>>                                              (not the old head any more)
>>>
>>> 3. prep_compound_page(new_head, new_order)
>>>      __SetPageHead(new_head)
>>>        tmp = new_head->flags.f  <-- load; PG_waiters == 0 in tmp
>>>
>>>                                            4. folio_set_waiters(new_head)
>>>                                              set_bit(PG_waiters, &flags.f)
>>>                                              enqueue on waitqueue, sleep
>>>
>>> 5.     new_head->flags.f = tmp | (1UL << PG_head)
>>>        <-- store; overwrites step 4, PG_waiters == 0 again
>>>
>>> Later, folio_unlock(new_folio) sees no PG_waiters, skips folio_wake_bit(),
>>> and CPU1 can remain blocked indefinitely.
>>>
>>> Fix it by setting PG_head along with other flags in new_folio->flags before
>>> smp_wmb() and clear_compound_head(). The remaining compound page metadata
>>> is set after clear_compound_head(). In that way, after
>>> clear_compound_head() makes new_folio visible, nothing overwrites the
>>> PG_waiters from softleaf_entry_wait_on_locked() by modifying
>>> new_folio->flags non-atomically.
>>>
>>> While at it, use data_race() to suppress the KCSAN issue reported by Ye
>>> Liu's patch (the second Closes) on reading ->flags during a folio split.
>>>
>>> Suggested-by: Kiryl Shutsemau <kas@kernel.org>
>>> Reported-by: David Hildenbrand <david@kernel.org>
>>> Closes: https://lore.kernel.org/all/DMHE8BEEER1H.2BIY52NXATPIA@nvidia.com/
>>> Reported-by: Ye Liu <ye.liu@linux.dev>
>>> Closes: https://lore.kernel.org/all/20261029-data_race_huge_mem-v1-1-112e331291c1@linux.dev/
>>> Fixes: c010d47f107f6 ("mm: thp: split huge page to any lower order pages")
>>> Cc: stable@vger.kernel.org
>>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>>> ---
>>> The issue was raised by David when reviewing Ye Liu's "mm/huge_memory: use
>>> READ_ONCE() when cloning folio flags during split" [1]. This patch also
>>> supersedes Ye Liu's patch.
>>>
>>> [1] https://lore.kernel.org/all/20261029-data_race_huge_mem-v1-1-112e331291c1@linux.dev/
>>> ---
>>> Changes in v3:
>>> - set PG_head before smp_wmb() like Kiryl suggested instead of
>>>   prep_compound_page_atomic().
>>> - used data_race() to suppress KCSAN reports while at it.
>>> - dropped Lorenzo's ack, since the patch is changed.
>>> - Link to v2: https://patch.msgid.link/20261106-fix-split_folio_to_order-race-v2-1-1b59c67b8af3@nvidia.com
>>>
>>> Changes in v2:
>>> - renamed __prep_compound_head_and_tail() to __prep_compound_page().
>>> - added comments to prep_compound_page() and __prep_compound_page().
>>> - Link to v1: https://patch.msgid.link/20261029-fix-split_folio_to_order-race-v1-1-a483c7451b0d@nvidia.com
>>> ---
>>>  mm/huge_memory.c | 32 ++++++++++++++++++++++++++++----
>>>  mm/internal.h    |  2 ++
>>>  mm/page_alloc.c  | 17 +++++++++++++++--
>>>  3 files changed, 45 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>> index bd44d58e86845..99b48d7c988db 100644
>>> --- a/mm/huge_memory.c
>>> +++ b/mm/huge_memory.c
>>> @@ -3714,6 +3714,7 @@ static void __split_folio_to_order(struct folio *folio, int old_order,
>>>  		 * Don't pass it around before clear_compound_head().
>>>  		 */
>>>  		struct folio *new_folio = (struct folio *)new_head;
>>> +		memdesc_flags_t new_flags;
>>>
>>>  		VM_BUG_ON_PAGE(atomic_read(&new_folio->_mapcount) != -1, new_head);
>>>
>>> @@ -3730,8 +3731,16 @@ static void __split_folio_to_order(struct folio *folio, int old_order,
>>>  		 * unreferenced sub-pages of an anonymous THP: we can simply drop
>>>  		 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
>>>  		 */
>>> -		new_folio->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
>>> -		new_folio->flags.f |= (folio->flags.f &
>>> +		/*
>>> +		 * ->flags are stable since the folio is locked and frozen.
>>> +		 * data_race() is used because reference-less writers,
>>> +		 * 1. PG_waiters from softleaf_entry_wait_on_locked() and
>>
>> Hm? I thought that case would be handled by your patch? How can they still write
>> to the page?
>
> This is V3 of the patch. Instead of using prep_compound_page_atomic(),
> PG_head is written before clear_compound_page(), after which new_head
> is visible to the reference-less writers.
>
>>
>>> +		 * 2. PG_hwpoison from memory_failure(),
>>> +		 * can still write to it.
>>
>> We lose PG_hwpoison left and right when racing with non-atomic updates. It needs
>> a fundamental fix. Is this really something we should take care of in this series?
>
> IMHO, memory_failure()’s reference-less access violates the frozen
> page contract and should be fixed. But I do not think it needs to go
> along with this patch. We discussed HWPoison flag race with
> Michael S. Tsirkin and Miaohe before [1]:
>
> 1. Miaohe proposed to reset HWPoison in memory_failure() as a best effort,
> 2. mst used RCU for all non-atomic flag accesses,
> 3. you also proposed to move HWPoison out of ->flags.

We probably can add a per-section bitmap like pageblock_flags for
HWPoison and only use atomic ops for that.



Best Regards,
Yan, Zi