Re: [PATCH 1/3] mm: handle swap page faults under VMA lock if page is uncontended

["Huang, Ying" <ying.huang@intel.com>]

Yosry Ahmed <yosryahmed@google.com> writes:

> On Wed, May 3, 2023 at 12:57 PM Suren Baghdasaryan <surenb@google.com> wrote:
>>
>> On Wed, May 3, 2023 at 1:34 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>> >
>> > On Tue, May 2, 2023 at 4:05 PM Suren Baghdasaryan <surenb@google.com> wrote:
>> > >
>> > > On Tue, May 2, 2023 at 3:31 PM Matthew Wilcox <willy@infradead.org> wrote:
>> > > >
>> > > > On Tue, May 02, 2023 at 09:36:03AM -0700, Suren Baghdasaryan wrote:
>> > > > > On Tue, May 2, 2023 at 8:03 AM Matthew Wilcox <willy@infradead.org> wrote:
>> > > > > >
>> > > > > > On Mon, May 01, 2023 at 10:04:56PM -0700, Suren Baghdasaryan wrote:
>> > > > > > > On Mon, May 1, 2023 at 8:22 PM Matthew Wilcox <willy@infradead.org> wrote:
>> > > > > > > >
>> > > > > > > > On Mon, May 01, 2023 at 07:30:13PM -0700, Suren Baghdasaryan wrote:
>> > > > > > > > > On Mon, May 1, 2023 at 7:02 PM Matthew Wilcox <willy@infradead.org> wrote:
>> > > > > > > > > >
>> > > > > > > > > > On Mon, May 01, 2023 at 10:50:23AM -0700, Suren Baghdasaryan wrote:
>> > > > > > > > > > > +++ b/mm/memory.c
>> > > > > > > > > > > @@ -3711,11 +3711,6 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
>> > > > > > > > > > >       if (!pte_unmap_same(vmf))
>> > > > > > > > > > >               goto out;
>> > > > > > > > > > >
>> > > > > > > > > > > -     if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
>> > > > > > > > > > > -             ret = VM_FAULT_RETRY;
>> > > > > > > > > > > -             goto out;
>> > > > > > > > > > > -     }
>> > > > > > > > > > > -
>> > > > > > > > > > >       entry = pte_to_swp_entry(vmf->orig_pte);
>> > > > > > > > > > >       if (unlikely(non_swap_entry(entry))) {
>> > > > > > > > > > >               if (is_migration_entry(entry)) {
>> > > > > > > > > >
>> > > > > > > > > > You're missing the necessary fallback in the (!folio) case.
>> > > > > > > > > > swap_readpage() is synchronous and will sleep.
>> > > > > > > > >
>> > > > > > > > > True, but is it unsafe to do that under VMA lock and has to be done
>> > > > > > > > > under mmap_lock?
>> > > > > > > >
>> > > > > > > > ... you were the one arguing that we didn't want to wait for I/O with
>> > > > > > > > the VMA lock held?
>> > > > > > >
>> > > > > > > Well, that discussion was about waiting in folio_lock_or_retry() with
>> > > > > > > the lock being held. I argued against it because currently we drop
>> > > > > > > mmap_lock lock before waiting, so if we don't drop VMA lock we would
>> > > > > > > be changing the current behavior which might introduce new
>> > > > > > > regressions. In the case of swap_readpage and swapin_readahead we
>> > > > > > > already wait with mmap_lock held, so waiting with VMA lock held does
>> > > > > > > not introduce new problems (unless there is a need to hold mmap_lock).
>> > > > > > >
>> > > > > > > That said, you are absolutely correct that this situation can be
>> > > > > > > improved by dropping the lock in these cases too. I just didn't want
>> > > > > > > to attack everything at once. I believe after we agree on the approach
>> > > > > > > implemented in https://lore.kernel.org/all/20230501175025.36233-3-surenb@google.com
>> > > > > > > for dropping the VMA lock before waiting, these cases can be added
>> > > > > > > easier. Does that make sense?
>> > > > > >
>> > > > > > OK, I looked at this path some more, and I think we're fine.  This
>> > > > > > patch is only called for SWP_SYNCHRONOUS_IO which is only set for
>> > > > > > QUEUE_FLAG_SYNCHRONOUS devices, which are brd, zram and nvdimms
>> > > > > > (both btt and pmem).  So the answer is that we don't sleep in this
>> > > > > > path, and there's no need to drop the lock.
>> > > > >
>> > > > > Yes but swapin_readahead does sleep, so I'll have to handle that case
>> > > > > too after this.
>> > > >
>> > > > Sleeping is OK, we do that in pXd_alloc()!  Do we block on I/O anywhere
>> > > > in swapin_readahead()?  It all looks like async I/O to me.
>> > >
>> > > Hmm. I thought that we have synchronous I/O in the following paths:
>> > >     swapin_readahead()->swap_cluster_readahead()->swap_readpage()
>> > >     swapin_readahead()->swap_vma_readahead()->swap_readpage()
>> > > but just noticed that in both cases swap_readpage() is called with the
>> > > synchronous parameter being false. So you are probably right here...
>> >
>> > In both swap_cluster_readahead() and swap_vma_readahead() it looks
>> > like if the readahead window is 1 (aka we are not reading ahead), then
>> > we jump to directly calling read_swap_cache_async() passing do_poll =
>> > true, which means we may end up calling swap_readpage() passing
>> > synchronous = true.
>> >
>> > I am not familiar with readahead heuristics, so I am not sure how
>> > common this is, but it's something to think about.
>>
>> Uh, you are correct. If this branch is common, we could use the same
>> "drop the lock and retry" pattern inside read_swap_cache_async(). That
>> would be quite easy to implement.
>> Thanks for checking on it!
>
>
> I am honestly not sure how common this is.
>
> +Ying who might have a better idea.

Checked the code and related history.  It seems that we can just pass
"synchronous = false" to swap_readpage() in read_swap_cache_async().
"synchronous = true" was introduced in commit 23955622ff8d ("swap: add
block io poll in swapin path") to reduce swap read latency for block
devices that can be polled.  But in commit 9650b453a3d4 ("block: ignore
RWF_HIPRI hint for sync dio"), the polling is deleted.  So, we don't
need to pass "synchronous = true" to swap_readpage() during
swapin_readahead(), because we will wait the IO to complete in
folio_lock_or_retry().

Best Regards,
Huang, Ying

>>
>>
>> >
>> > > Does that mean swapin_readahead() might return a page which does not
>> > > have its content swapped-in yet?
>> > >