Re: [PATCH] mm: page_alloc: dump migrate-failed pages only at -EBUSY

[Minchan Kim <minchan@kernel.org>]

On Fri, May 21, 2021 at 10:08:15AM +0200, David Hildenbrand wrote:
> On 20.05.21 22:51, Minchan Kim wrote:
> > On Thu, May 20, 2021 at 09:28:09PM +0200, David Hildenbrand wrote:
> > > Minchan Kim <minchan@kernel.org> schrieb am Do. 20. Mai 2021 um 21:20:
> > > 
> > > > On Wed, May 19, 2021 at 02:33:41PM -0700, Minchan Kim wrote:
> > > > > alloc_contig_dump_pages aims for helping debugging page migration
> > > > > failure by page refcount mismatch or something else of page itself
> > > > > from migration handler function. However, in -ENOMEM case, there is
> > > > > nothing to get clue from page descriptor information so just
> > > > > dump pages only when -EBUSY happens.
> > > > > 
> > > > > Signed-off-by: Minchan Kim <minchan@kernel.org>
> > > > > ---
> > > > >   mm/page_alloc.c | 3 ++-
> > > > >   1 file changed, 2 insertions(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > > index 3100fcb08500..c0a2971dc755 100644
> > > > > --- a/mm/page_alloc.c
> > > > > +++ b/mm/page_alloc.c
> > > > > @@ -8760,7 +8760,8 @@ static int __alloc_contig_migrate_range(struct
> > > > compact_control *cc,
> > > > > 
> > > > >        lru_cache_enable();
> > > > >        if (ret < 0) {
> > > > > -             alloc_contig_dump_pages(&cc->migratepages);
> > > > > +             if (ret == -EBUSY)
> > > > > +                     alloc_contig_dump_pages(&cc->migratepages);
> > > > >                putback_movable_pages(&cc->migratepages);
> > > > >                return ret;
> > > > >        }
> > > > > --
> > > > > 2.31.1.751.gd2f1c929bd-goog
> > > > > 
> > > > 
> > > > Resend with a little modifying description.
> > > > 
> > > >  From c5a2fea291cf46079b87cc9ac9a25fc7f819d0fd Mon Sep 17 00:00:00 2001
> > > > From: Minchan Kim <minchan@kernel.org>
> > > > Date: Wed, 19 May 2021 14:22:18 -0700
> > > > Subject: [PATCH] mm: page_alloc: dump migrate-failed pages only at -EBUSY
> > > > 
> > > > alloc_contig_dump_pages aims for helping debugging page migration
> > > > failure by elevated page refcount compared to expected_count.
> > > > (for the detail, please look at migrate_page_move_mapping)
> > > > 
> > > > However, -ENOMEM is just the case that system is under memory
> > > > pressure state, not relevant with page refcount at all. Thus,
> > > > the dumping page list is not helpful for the debugging point of view.
> > > > 
> > > 
> > > what about -ENOMEM when migrating empty/free huge pages? I think there is
> > > value in having the pages dumped to identify something like that. And it
> > > doesn‘t require heavy memory pressure to fail allocating a huge page.
> > > 
> > 
> > -ENOMEM means there is no memory to alloate destination page.
> > How could it help dumping source pages in those case from dump_page
> > content point of view?
> 
> You would spot a huge page in the source list (usually at first position)
> without any obvious migration blockers I assume?

It was not a huge page case.

> 
> I'm wondering, did you actually run into this being suboptimal? If it's a
> real problem dumping too many stuff when running into -ENOMEM, fine with me.
> If it's a theoretical issue, I'd prefer to just keep it simple as is.

That's exactly what I encountered. With -ENOMEM, it dumped bunch of
pages on migratepages list. It was just useless with just consuming
logbuffer since there are nothing much to investigate with dumping
source pages.