Messages in this thread |  | | From | David CARLIER <> | | Date | Wed, 1 Apr 2026 08:10:08 +0100 | | Subject | Re: [PATCH v2] mm/page_io: fix PSWPIN undercount for large folios in sio_read_complete() |
| |
On Tue, 31 Mar 2026 at 23:33, Barry Song <21cnbao@gmail.com> wrote: > > On Mon, Mar 30, 2026 at 3:12 PM David Carlier <devnexen@gmail.com> wrote: > > > > sio_read_complete() uses sio->pages to account global PSWPIN vm events, > > but sio->pages tracks the number of bvec entries (folios), not base > > pages. For large folios this undercounts compared to the per-memcg path > > which correctly uses folio_nr_pages(), and compared to the bdev read > > paths which also use folio_nr_pages(). > > > > Use sio->len >> PAGE_SHIFT instead, which gives the correct base page > > count since sio->len is accumulated via folio_size(folio). > > > > Fixes: a1a0dfd56f97 ("mm: handle THP in swap_*page_fs()") > > Signed-off-by: David Carlier <devnexen@gmail.com> > > The patch seems theoretically correct, but I’m wondering > where we can swap in mTHP for filesystem-based swap? > > In both do_swap_page() and shmem_swapin_folio(), we check > data_race(si->flags & SWP_SYNCHRONOUS_IO) before allocating > large folios. Am I missing something?
▎ The patch seems theoretically correct, but I'm wondering ▎ where we can swap in mTHP for filesystem-based swap?
▎ In both do_swap_page() and shmem_swapin_folio(), we check ▎ data_race(si->flags & SWP_SYNCHRONOUS_IO) before allocating ▎ large folios. Am I missing something?
You're right, I missed that. SWP_FS_OPS is only set by NFS and SMB which have no bdev, so SWP_SYNCHRONOUS_IO can never be set alongside it. Large folios can't currently reach this path since both do_swap_page() and shmem_swapin_folio() gate mTHP allocation on SWP_SYNCHRONOUS_IO.
That said, sio_read_complete() already calls count_mthp_stat() and the per-memcg accounting uses folio_nr_pages(), so the code seems written with large folios in mind even if the path is currently unreachable. Using sio->pages (bvec entry count) for a base-page count is still semantically wrong, but I understand the practical impact is nil today.
Happy to either drop this or keep it as a correctness cleanup, whatever you and Andrew prefer.
Cheers ! > > > --- > > mm/page_io.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/mm/page_io.c b/mm/page_io.c > > index 63b262f4c5a9..1389cd57ca88 100644 > > --- a/mm/page_io.c > > +++ b/mm/page_io.c > > @@ -497,7 +497,7 @@ static void sio_read_complete(struct kiocb *iocb, long ret) > > folio_mark_uptodate(folio); > > folio_unlock(folio); > > } > > - count_vm_events(PSWPIN, sio->pages); > > + count_vm_events(PSWPIN, sio->len >> PAGE_SHIFT); > > } else { > > for (p = 0; p < sio->pages; p++) { > > struct folio *folio = page_folio(sio->bvec[p].bv_page); > > -- > > 2.53.0 > > > > Thanks > Barry
|  |