Re: [PATCH -next] mm/page_counter: mark intentional data races

[Marco Elver <elver@google.com>]

On Wed, 29 Jan 2020 at 09:51, Michal Hocko <mhocko@kernel.org> wrote:
>
> On Tue 28-01-20 23:20:19, Qian Cai wrote:
> > The commit 3e32cb2e0a12 ("mm: memcontrol: lockless page counters")
> > had memcg->memsw->failcnt and ->watermark could be accessed concurrently
> > as reported by KCSAN,
> >
> >  Reported by Kernel Concurrency Sanitizer on:
> >  BUG: KCSAN: data-race in page_counter_try_charge / page_counter_try_charge
> >
> >  read to 0xffff8fb18c4cd190 of 8 bytes by task 1081 on cpu 59:
> >   page_counter_try_charge+0x4d/0x150 mm/page_counter.c:138
> >   try_charge+0x131/0xd50

Why are the line numbers for the remaining symbols missing?  Doesn't
scripts/decode_stacktrace.sh give you all line numbers?

[ As an aside: if you want to use what syzbot uses to put line numbers
on symbols, which is a bit faster:
https://github.com/google/syzkaller/tree/master/tools/syz-symbolize
https://github.com/google/syzkaller/blob/master/docs/linux/setup.md
then 'go build tools/syz-symbolize'. ]

> >   __memcg_kmem_charge_memcg+0x58/0x140
> >   __memcg_kmem_charge+0xcc/0x280
> >   __alloc_pages_nodemask+0x1e1/0x450
> >   alloc_pages_current+0xa6/0x120
> >   pte_alloc_one+0x17/0xd0
> >   __pte_alloc+0x3a/0x1f0
> >   copy_p4d_range+0xc36/0x1990
> >   copy_page_range+0x21d/0x360
> >   dup_mmap+0x5f5/0x7a0
> >   dup_mm+0xa2/0x240
> >   copy_process+0x1b3f/0x3460
> >   _do_fork+0xaa/0xa20
> >   __x64_sys_clone+0x13b/0x170
> >   do_syscall_64+0x91/0xb47
> >   entry_SYSCALL_64_after_hwframe+0x49/0xbe
> >
> >  write to 0xffff8fb18c4cd190 of 8 bytes by task 1153 on cpu 120:
> >   page_counter_try_charge+0x5b/0x150 mm/page_counter.c:139
> >   try_charge+0x131/0xd50
> >   mem_cgroup_try_charge+0x159/0x460
> >   mem_cgroup_try_charge_delay+0x3d/0xa0
> >   wp_page_copy+0x14d/0x930
> >   do_wp_page+0x107/0x7b0
> >   __handle_mm_fault+0xce6/0xd40
> >   handle_mm_fault+0xfc/0x2f0
> >   do_page_fault+0x263/0x6f9
> >   page_fault+0x34/0x40
> >
> > Since the failcnt and watermark are tolerant of some inaccuracy, a data
> > race will not be harmful, thus mark them as intentional data races with
> > the data_race() macro.
>
> I am not familiar with KCSAN and git grep for data_race on the current
> linux-next doesn't really show any users of this macro. Is there a
> general consensus that data_race is going to be used to silence all
> KCSAN false positives?

It was discussed here:
https://lore.kernel.org/linux-fsdevel/CAHk-=wg5CkOEF8DTez1Qu0XTEFw_oHhxN98bDnFqbY7HL5AB2g@mail.gmail.com/

If they're intentional data races that should remain, data_race() is
one option. There are 4 options (other than address the data race) to
deal with 'false positives':
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/Documentation/dev-tools/kcsan.rst#n101

That being said, every use of data_race() needs to be justified, and
not just applied without understanding the issue. See below.

> > Signed-off-by: Qian Cai <cai@lca.pw>
> > ---
> >  mm/page_counter.c | 10 +++++-----
> >  1 file changed, 5 insertions(+), 5 deletions(-)
> >
> > diff --git a/mm/page_counter.c b/mm/page_counter.c
> > index de31470655f6..13934636eafd 100644
> > --- a/mm/page_counter.c
> > +++ b/mm/page_counter.c
> > @@ -82,8 +82,8 @@ void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
> >                * This is indeed racy, but we can live with some
> >                * inaccuracy in the watermark.
> >                */
> > -             if (new > c->watermark)
> > -                     c->watermark = new;
> > +             if (data_race(new > c->watermark))
> > +                     data_race(c->watermark = new);

These should be using 'READ_ONCE' and 'WRITE_ONCE' for c->watermark.
Store or load tearing would change the logic here, since the
comparison might see garbage.

> >       }
> >  }
> >
> > @@ -126,7 +126,7 @@ bool page_counter_try_charge(struct page_counter *counter,
> >                        * This is racy, but we can live with some
> >                        * inaccuracy in the failcnt.
> >                        */
> > -                     c->failcnt++;
> > +                     data_race(c->failcnt++);

This is probably fine.

> >                       *fail = c;
> >                       goto failed;
> >               }
> > @@ -135,8 +135,8 @@ bool page_counter_try_charge(struct page_counter *counter,
> >                * Just like with failcnt, we can live with some
> >                * inaccuracy in the watermark.
> >                */
> > -             if (new > c->watermark)
> > -                     c->watermark = new;
> > +             if (data_race(new > c->watermark))
> > +                     data_race(c->watermark = new);

This should be READ_ONCE / WRITE_ONCE.

> >       }
> >       return true;
> >
> > --
> > 2.21.0 (Apple Git-122.2)
>
> --
> Michal Hocko
> SUSE Labs