Re: [PATCH net] tcp: Add READ_ONCE() to read tcp_orphan_count

[Marco Elver]

On Thu, 12 May 2022 at 22:06, Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Thu, May 12, 2022 at 3:32 AM Liu Jian <liujian56@xxxxxxxxxx> wrote:
> >
> > The tcp_orphan_count per-CPU variable is read locklessly, so this commit
> > add the READ_ONCE() to a load in order to avoid below KCSAN warnning:
> >
> > BUG: KCSAN: data-race in tcp_orphan_count_sum net/ipv4/tcp.c:2476 [inline]
> > BUG: KCSAN: data-race in tcp_orphan_update+0x64/0x100 net/ipv4/tcp.c:2487
> >
> > race at unknown origin, with read to 0xffff9c63bbdac7a8 of 4 bytes by interrupt on cpu 2:
> >  tcp_orphan_count_sum net/ipv4/tcp.c:2476 [inline]
> >  tcp_orphan_update+0x64/0x100 net/ipv4/tcp.c:2487
> >  call_timer_fn+0x33/0x210 kernel/time/timer.c:1414
> >
> > Fixes: 19757cebf0c5 ("tcp: switch orphan_count to bare per-cpu counters")
> > Signed-off-by: Liu Jian <liujian56@xxxxxxxxxx>
> > ---
> >  net/ipv4/tcp.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> > index cf18fbcbf123..7245609f41e6 100644
> > --- a/net/ipv4/tcp.c
> > +++ b/net/ipv4/tcp.c
> > @@ -2718,7 +2718,7 @@ int tcp_orphan_count_sum(void)
> >         int i, total = 0;
> >
> >         for_each_possible_cpu(i)
> > -               total += per_cpu(tcp_orphan_count, i);
> > +               total += READ_ONCE(per_cpu(tcp_orphan_count, i));
>
> We might raise the discussion to lkml and/or KCSAN supporters.
>
> Presumably, all per_cpu() uses in the kernel will have the same issue ?
>
> By definition per-cpu data can be changed by other cpus.
>
> So maybe per_cpu() should contain the annotation, instead of having to
> annotate all users.

I guess the question is, is it the norm that per_cpu() retrieves data
that can legally be modified concurrently, or not. If not, and in most
cases it's a bug, the annotations should be here.

Paul, was there any guidance/documentation on this, but I fail to find
it right now? (access-marking.txt doesn't say much about per-CPU
data.)

Thanks,
-- Marco