Re: [LSM/MM/BPF TOPIC] Do not hold mmap_lock following folio_lock failure in page faults

[Suren Baghdasaryan <surenb@google.com>]

On Fri, Feb 20, 2026 at 2:13 PM Barry Song <21cnbao@gmail.com> wrote:
>
> Currently, page faults use per-VMA locks whenever possible.
> However, when a page fault encounters an I/O read (e.g., in
> filemap_fault() or do_swap_page()), the handler releases the
> per-VMA lock and requests a retry after failing to acquire
> folio_lock. On the retry, the fault handler always acquires
> the mmap_lock in read mode unconditionally.
>
> This can occur frequently and may sometimes cause UI jank due
> to mmap_lock contention. A proposal suggests using the
> per-VMA lock in the page fault retry path instead of taking
> mmap_lock [1]. Oven reported that this can significantly
> improve performance by reducing mmap_lock wait time [2].
>
> Matthew appears to suggest removing the page fault retry
> entirely and instead waiting for I/O while holding the
> per-VMA lock, noting that folios under I/O read may be
> reclaimed by the LRU under severe memory pressure and
> with thousands of threads. As a result, folios might be
> reallocated, causing the page fault to repeatedly re-enter [3].
> However, holding the per-VMA lock during I/O wait raises
> some concerns: writers could experience long delays.
>
> On the other hand, retrying with the per-VMA lock may reduce
> mmap_lock contention, potentially shortening page fault retry
> time and giving the LRU fewer opportunities to reclaim folios
> waiting for I/O. By the time of the LSF/MM/BPF discussion, I
> may have already created similar cases and collected
> additional data.
>
> Another potential optimization is to remove the page fault retry
> without risking long I/O waits blocking writers. This occurs when
> folio_lock fails but the folio is already up-to-date. In this case,
> a parallel page fault may simply wait for the simultaneous PF to
> complete the PTE mapping.
>
> I would appreciate feedback from LSF/MM/BPF on approaches
> to eliminate mmap_lock acquisition in page faults triggered
> by folio_lock failure, for both I/O wait and PTE-mapping
> wait cases, for mainline inclusion.

Would love to participate in this discussion. Hope Barry can attend in
person this time :)

>
> [1] https://lore.kernel.org/linux-mm/20251127011438.6918-1-21cnbao@gmail.com/
> [2] https://lore.kernel.org/linux-mm/cccf352a-1a68-430d-83fa-a14bb5e37464@oppo.com/
> [3] https://lore.kernel.org/linux-mm/aSip2mWX13sqPW_l@casper.infradead.org/
>
> Thanks
> Barry