Re: [PATCH mm-unstable v1] mm/truncate: batch-clear shadow entries

[Yu Zhao]

On Mon, Jul 8, 2024 at 4:16 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Mon,  8 Jul 2024 15:27:53 -0600 Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
>
> > Make clear_shadow_entry() clear shadow entries in `struct folio_batch`
> > so that it can reduce contention on i_lock and i_pages locks, e.g.,
> >
> >   watchdog: BUG: soft lockup - CPU#29 stuck for 11s! [fio:2701649]
> >     clear_shadow_entry+0x3d/0x100
> >     mapping_try_invalidate+0x117/0x1d0
> >     invalidate_mapping_pages+0x10/0x20
> >     invalidate_bdev+0x3c/0x50
> >     blkdev_common_ioctl+0x5f7/0xa90
> >     blkdev_ioctl+0x109/0x270
>
> This will clearly reduce lock traffic a lot, but does it truly fix the
> issue?  Is it the case that sufficiently extreme loads will still run
> into problems?

I think Bharata was running extreme loads. So I'd say it's good enough
for now, considering truncation doesn't happen that often.

> > --- a/mm/truncate.c
> > +++ b/mm/truncate.c
> > @@ -39,12 +39,24 @@ static inline void __clear_shadow_entry(struct address_space *mapping,
> >       xas_store(&xas, NULL);
> >  }
> >
> > -static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
> > -                            void *entry)
> > +static void clear_shadow_entry(struct address_space *mapping,
> > +                            struct folio_batch *fbatch, pgoff_t *indices)
> >  {
> > +     int i;
> > +
> > +     if (shmem_mapping(mapping) || dax_mapping(mapping))
> > +             return;
>
> We lost the comment which was in invalidate_exceptional_entry() and
> elsewhere.  It wasn't a terribly good one but I do think a few words
> which explain why we're testing for these things would be helpful.

I'll put the original comment back. It seems to me it was stating the
obvious, and I don't really know how to improve it since it's obvious
;)

> I expect we should backport this.  But identifying a Fixes: target
> looks to be challenging.

I wouldn't worry about backporting, nobody else has run into this
scalability issue (not even a day-to-day performance problem).