Re: [BUG] mm/gup: WARNING: mm/page-writeback.c:2678 at __folio_mark_dirty on v7.4-rc3
From: David Hildenbrand (Arm)
Date: Sat Nov 21 2026 - 11:44:47 EST
On 11/21/26 15:49, Yi Wen wrote:
> Hi,
>
> I am reporting a kernel WARNING triggered by a syzkaller reproducer on
> Linux v7.4-rc3.
>
> The reproducer triggers a WARNING in ext4_dirty_folio() when
> process_vm_writev() writes to a file-backed mapping.
>
> The call trace shows that the warning occurs during the GUP page
> unpinning and dirtying path:
>
> process_vm_writev()
> -> process_vm_rw_core()
> -> unpin_user_pages_dirty_lock()
> -> folio_mark_dirty()
> -> ext4_dirty_folio()
>
> A second WARNING is subsequently triggered in __folio_mark_dirty() in
> mm/page-writeback.c.
>
> This appears to be related to the interaction between GUP-based page
> dirtying and filesystem-managed folio metadata. In particular, the GUP
> dirtying path may mark a file-backed folio dirty without ensuring that
> the filesystem-specific metadata required by its dirty_folio()
> implementation remains available.
>
> This may be related to the known issue where GUP-based page dirtying
> does not properly notify the filesystem through page_mkwrite(), as
> well as the interaction between GUP dirtying and
> filemap_release_folio().
>
> Before these warnings, all four ext4 loop devices also report delayed
> block allocation errors and abnormal free block accounting. These
> errors may indicate an additional filesystem inconsistency and are
> included in the console log for further investigation.
>
Hi,
thanks for the report.
Did you only find this through syzcaller or by the help of an LLM?
I'm asking because the timing is interesting; this should be one of the problems
discussed by:
https://lore.kernel.org/r/20261119-gup-dirty-what-v1-0-236d69f25c24@xxxxxxx
CCing Pedro, John und Jan.
--
Cheers,
David