Re: [PATCH v5 6/8] mm/shmem, swap: simplify swapin path and result handling

[Hugh Dickins <hughd@google.com>]

[-- Attachment #1: Type: text/plain, Size: 2683 bytes --]

On Fri, 11 Jul 2025, Kairui Song wrote:
> On Fri, Jul 11, 2025 at 2:23 PM Baolin Wang
> <baolin.wang@linux.alibaba.com> wrote:
> >
> >
> >
> > On 2025/7/10 11:37, Kairui Song wrote:
> > > From: Kairui Song <kasong@tencent.com>
> > >
> > > Slightly tidy up the different handling of swap in and error handling
> > > for SWP_SYNCHRONOUS_IO and non-SWP_SYNCHRONOUS_IO devices. Now swapin
> > > will always use either shmem_swap_alloc_folio or shmem_swapin_cluster,
> > > then check the result.
> > >
> > > Simplify the control flow and avoid a redundant goto label.
> > >
> > > Signed-off-by: Kairui Song <kasong@tencent.com>
> >
> > LGTM, with a nit as follows.
> > Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
> >
> > > ---
> > >   mm/shmem.c | 45 +++++++++++++++++++--------------------------
> > >   1 file changed, 19 insertions(+), 26 deletions(-)
> > >
> > > diff --git a/mm/shmem.c b/mm/shmem.c
> > > index 847e6f128485..80f5b8c73eb8 100644
> > > --- a/mm/shmem.c
> > > +++ b/mm/shmem.c
> > > @@ -2320,40 +2320,33 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
> > >                       count_memcg_event_mm(fault_mm, PGMAJFAULT);
> > >               }
> > >
> > > -             /* Skip swapcache for synchronous device. */
> > >               if (data_race(si->flags & SWP_SYNCHRONOUS_IO)) {
> > > +                     /* Direct mTHP swapin skipping swap cache & readhaed */
> > >                       folio = shmem_swap_alloc_folio(inode, vma, index, swap, order, gfp);
> >
> > Nit: the 'mTHP' word can be confusing, since we will skip swapcache for
> > order 0 too. Please drop it.
> >
> 
> Yes, thanks for the review.

And a few words after that 'mTHP ', I keep wincing at 'readhaed':
Andrew, you already did a fix to remove the 'mTHP ', I hope we can
also persuade you to change 'readhaed' to 'readahead' there - thanks!

Kairui, I'm a little uneasy about the way this series does arithmetic
on swap.val, in the knowledge that swp_offset(entry) involves no shift.

Perhaps I haven't noticed, but I think this is the first place to
make that assumption; and a few years ago it was not true at all -
swp_type() was down the bottom.  Usually we would do it all with
swp_entry(swp_type(x), arithmetic_on(swp_offset(x))).

But I guess, let's just agree that it's easier to read and get right
the way you have it, and make no change: if I try to "correct" you,
or demand that you change it, we shall probably just bring in bugs.

I'm particularly glad that you now avoid SWP_SYNCHRONOUS_IO readahead:
that stupidity had very much annoyed me, once I realized it.

Thanks,
Hugh