On Fri, Apr 14, 2023 at 11:43 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Fri, Apr 14, 2023 at 11:00:43AM -0700, Suren Baghdasaryan wrote:
> > When page fault is handled under VMA lock protection, all swap page
> > faults are retried with mmap_lock because folio_lock_or_retry
> > implementation has to drop and reacquire mmap_lock if folio could
> > not be immediately locked.
> > Instead of retrying all swapped page faults, retry only when folio
> > locking fails.
>
> Reviewed-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
Thank you for the reviews!
>
> Let's just review what can now be handled under the VMA lock instead of
> the mmap_lock, in case somebody knows better than me that it's not safe.
>
> - We can call migration_entry_wait(). This will wait for PG_locked to
> become clear (in migration_entry_wait_on_locked()). As previously
> discussed offline, I think this is safe to do while holding the VMA
> locked.
> - We can call remove_device_exclusive_entry(). That calls
> folio_lock_or_retry(), which will fail if it can't get the VMA lock.
> - We can call pgmap->ops->migrate_to_ram(). Perhaps somebody familiar
> with Nouveau and amdkfd could comment on how safe this is?
> - I believe we can't call handle_pte_marker() because we exclude UFFD
> VMAs earlier.
> - We can call swap_readpage() if we allocate a new folio. I haven't
> traced through all this code to tell if it's OK.
>
> So ... I believe this is all OK, but we're definitely now willing to
> wait for I/O from the swap device while holding the VMA lock when we
> weren't before. And maybe we should make a bigger deal of it in the
> changelog.
>
> And maybe we shouldn't just be failing the folio_lock_or_retry(),
> maybe we should be waiting for the folio lock with the VMA locked.
Wouldn't that cause holding the VMA lock for the duration of swap I/O
(something you said we want to avoid in the previous paragraph) and
effectively undo d065bd810b6d ("mm: retry page fault when blocking on
disk transfer") for VMA locks?
>
