Re: [PATCH] libceph: protect pending flags in ceph_con_keepalive()

[Ilya Dryomov]

On Tue, Jan 15, 2019 at 7:56 AM Myungho Jung <mhjungk@xxxxxxxxx> wrote:
>
> On Mon, Jan 14, 2019 at 09:37:25PM +0100, Ilya Dryomov wrote:
> > On Thu, Jan 3, 2019 at 4:50 AM Myungho Jung <mhjungk@xxxxxxxxx> wrote:
> > > I reproduced on vm using syzkaller utils and verified the fix by syzbot.
> >
> > Hi Myungho,
> >
> > I think this might be a better fix:
> >
> > diff --git a/net/ceph/messenger.c b/net/ceph/messenger.c
> > index d5718284db57..c5f5313e3537 100644
> > --- a/net/ceph/messenger.c
> > +++ b/net/ceph/messenger.c
> > @@ -3205,10 +3205,11 @@ void ceph_con_keepalive(struct ceph_connection *con)
> >  {
> >         dout("con_keepalive %p\n", con);
> >         mutex_lock(&con->mutex);
> > +       con_flag_set(con, CON_FLAG_KEEPALIVE_PENDING);
> >         clear_standby(con);
> >         mutex_unlock(&con->mutex);
> > -       if (con_flag_test_and_set(con, CON_FLAG_KEEPALIVE_PENDING) == 0 &&
> > -           con_flag_test_and_set(con, CON_FLAG_WRITE_PENDING) == 0)
> > +
> > +       if (con_flag_test_and_set(con, CON_FLAG_WRITE_PENDING) == 0)
> >                 queue_con(con);
> >  }
> >  EXPORT_SYMBOL(ceph_con_keepalive);
> >
> > WRITE_PENDING can be set without con->mutex held from socket callbacks.
> > This is the reason we use atomic bit ops here, so testing WRITE_PENDING
> > under the lock didn't make sense to me.
> >
> > At the same time, KEEPALIVE_PENDING could have been a non-atomic flag.
> > I spent some time trying to make sense of conditioning queue_con() call
> > on the previous value of KEEPALIVE_PENDING and couldn't see any, so I'm
> > setting it with con_flag_set(), making ceph_con_keepalive() symmetric
> > with ceph_con_send().
> >
> > Thanks,
> >
> >                 Ilya
>
> Hi Ilya,
>
> Yes, it looks clear and makes sense to have an atomic operation in if statement
> but it still triggers warning. KEEPALIVE_PENDING should be set after
> clear_standby() because con_fault() can be called right before acquiring the
> lock here which sets the flag in standby state. I tesed the change with syzbot
> and confirmed there was no warning.

Right, it still triggers one of the warnings.  I was too focused on
WRITE_PENDING and missed that in plain sight.  I'll update the patch.

Thanks for testing!

                Ilya