Re: [RFC PATCH 3/4] mm/shmem: optimize file read with folio batching

[Chi Zhiling]

On 5/18/26 14:57, Baolin Wang wrote:
> On 5/15/26 5:47 PM, Chi Zhiling wrote:
> > From: Chi Zhiling <chizhiling@xxxxxxxxxx>
> >
> > Optimize shmem file read by implementing folio batching in the read
> > iteration path. Only uptodate folios are added to the batch, ensuring
> > all folios in the batch are valid and ready for use without additional
> > checking.
> >
> > Signed-off-by: Chi Zhiling <chizhiling@xxxxxxxxxx>
> > ---
> >   mm/shmem.c | 102 ++++++++++++++++++++++++++++++++++++++++++++++++-----
> >   1 file changed, 93 insertions(+), 9 deletions(-)
> >
> > diff --git a/mm/shmem.c b/mm/shmem.c
> > index 767610f78d0d..4bc4e463ca97 100644
> > --- a/mm/shmem.c
> > +++ b/mm/shmem.c
> > @@ -3348,16 +3348,100 @@ shmem_write_end(const struct kiocb *iocb, 
> > struct address_space *mapping,
> >       return copied;
> >   }
> > +static pgoff_t shmem_get_read_batch(struct address_space *mapping,
> > +        pgoff_t index, pgoff_t max, struct folio_batch *fbatch)
> > +{
> > +    XA_STATE(xas, &mapping->i_pages, index);
> > +    struct folio *folio;
> > +    pgoff_t end = max;
> > +
> > +    rcu_read_lock();
> > +    xas_for_each(&xas, folio, max) {
> > +        if (xas_retry(&xas, folio))
> > +            continue;
> > +        if (xa_is_value(folio)) {
> > +            end = xas.xa_index;
> > +            break;
> > +        }
> > +        if (!folio_try_get(folio))
> > +            goto retry;
> > +
> > +        if (unlikely(folio != xas_reload(&xas)))
> > +            goto put_folio;
> > +
> > +        end = folio_next_index(folio);
> > +
> > +        if (!folio_test_uptodate(folio)) {
> > +            xas_advance(&xas, end - 1);
> > +            folio_put(folio);
> > +            continue;
> > +        }
> > +        if (!folio_batch_add(fbatch, folio))
> > +            break;
> > +        xas_advance(&xas, end - 1);
> > +        continue;
> > +put_folio:
> > +        folio_put(folio);
> > +retry:
> > +        xas_reset(&xas);
> > +    }
> > +    rcu_read_unlock();
> > +
> > +    return end;
> > +}
>
> I haven't looked at this patch in detail yet, but I feel this part 
> should belong in mm/filemap.c, or perhaps mm/filemap.c already has a 
> similar function implemented? Let's wait for Matthew's comments on this.

You are right. The existing filemap_get_folios_contig can handle this 
with clearer code, nice!


Thanks,