Re: [PATCH v3] mm: fix race between kmem_cache destroy, create and deactivate

[Paul E. McKenney]

On Sun, Jun 10, 2018 at 10:40:17AM -0700, Shakeel Butt wrote:
> On Sun, Jun 10, 2018 at 9:32 AM Paul E. McKenney
> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Sun, Jun 10, 2018 at 07:52:50AM -0700, Shakeel Butt wrote:
> > > On Sat, Jun 9, 2018 at 3:20 AM Vladimir Davydov <vdavydov.dev@xxxxxxxxx> wrote:
> > > >
> > > > On Tue, May 29, 2018 at 05:12:04PM -0700, Shakeel Butt wrote:
> > > > > The memcg kmem cache creation and deactivation (SLUB only) is
> > > > > asynchronous. If a root kmem cache is destroyed whose memcg cache is in
> > > > > the process of creation or deactivation, the kernel may crash.
> > > > >
> > > > > Example of one such crash:
> > > > >       general protection fault: 0000 [#1] SMP PTI
> > > > >       CPU: 1 PID: 1721 Comm: kworker/14:1 Not tainted 4.17.0-smp
> > > > >       ...
> > > > >       Workqueue: memcg_kmem_cache kmemcg_deactivate_workfn
> > > > >       RIP: 0010:has_cpu_slab
> > > > >       ...
> > > > >       Call Trace:
> > > > >       ? on_each_cpu_cond
> > > > >       __kmem_cache_shrink
> > > > >       kmemcg_cache_deact_after_rcu
> > > > >       kmemcg_deactivate_workfn
> > > > >       process_one_work
> > > > >       worker_thread
> > > > >       kthread
> > > > >       ret_from_fork+0x35/0x40
> > > > >
> > > > > To fix this race, on root kmem cache destruction, mark the cache as
> > > > > dying and flush the workqueue used for memcg kmem cache creation and
> > > > > deactivation.
> > > >
> > > > > @@ -845,6 +862,8 @@ void kmem_cache_destroy(struct kmem_cache *s)
> > > > >       if (unlikely(!s))
> > > > >               return;
> > > > >
> > > > > +     flush_memcg_workqueue(s);
> > > > > +
> > > >
> > > > This should definitely help against async memcg_kmem_cache_create(),
> > > > but I'm afraid it doesn't eliminate the race with async destruction,
> > > > unfortunately, because the latter uses call_rcu_sched():
> > > >
> > > >   memcg_deactivate_kmem_caches
> > > >    __kmem_cache_deactivate
> > > >     slab_deactivate_memcg_cache_rcu_sched
> > > >      call_rcu_sched
> > > >                                             kmem_cache_destroy
> > > >                                              shutdown_memcg_caches
> > > >                                               shutdown_cache
> > > >       memcg_deactivate_rcufn
> > > >        <dereference destroyed cache>
> > > >
> > > > Can we somehow flush those pending rcu requests?
> > >
> > > You are right and thanks for catching that. Now I am wondering if
> > > synchronize_sched() just before flush_workqueue() should be enough.
> > > Otherwise we might have to replace call_sched_rcu with
> > > synchronize_sched() in kmemcg_deactivate_workfn which I would not
> > > prefer as that would holdup the kmem_cache workqueue.
> > >
> > > +Paul
> > >
> > > Paul, we have a situation something similar to the following pseudo code.
> > >
> > > CPU0:
> > > lock(l)
> > > if (!flag)
> > >   call_rcu_sched(callback);
> > > unlock(l)
> > > ------
> > > CPU1:
> > > lock(l)
> > > flag = true
> > > unlock(l)
> > > synchronize_sched()
> > > ------
> > >
> > > If CPU0 has called already called call_rchu_sched(callback) then later
> > > if CPU1 calls synchronize_sched(). Is there any guarantee that on
> > > return from synchronize_sched(), the rcu callback scheduled by CPU0
> > > has already been executed?
> >
> > No.  There is no such guarantee.
> >
> > You instead want rcu_barrier_sched(), which waits for the callbacks from
> > all prior invocations of call_rcu_sched() to be invoked.
> >
> > Please note that synchronize_sched() is -not- sufficient.  It is only
> > guaranteed to wait for a grace period, not necessarily for all prior
> > callbacks.  This goes both directions because if there are no callbacks
> > in the system, then rcu_barrier_sched() is within its rights to return
> > immediately.
> >
> > So please make sure you use each of synchronize_sched() and
> > rcu_barrier_sched() to do the job that it was intended to do!  ;-)
> >
> > If your lock(l) is shorthand for spin_lock(&l), it looks to me like you
> > actually only need rcu_barrier_sched():
> >
> >         CPU0:
> >         spin_lock(&l);
> >         if (!flag)
> >           call_rcu_sched(callback);
> >         spin_unlock(&l);
> >
> >         CPU1:
> >         spin_lock(&l);
> >         flag = true;
> >         spin_unlock(&l);
> >         /* At this point, no more callbacks will be registered. */
> >         rcu_barrier_sched();
> >         /* At this point, all registered callbacks will have been invoked. */
> >
> > On the other hand, if your "lock(l)" was instead shorthand for
> > rcu_read_lock_sched(), then you need -both- synchronize_sched() -and-
> > rcu_barrier().  And even then, you will be broken in -rt kernels.
> > (Which might or might not be a concern, depending on whether your code
> > matters to -rt kernels.
> >
> > Make sense?
> 
> Thanks a lot, that was really helpful. The lock is actually
> mutex_lock. So, I think rcu_barrier_sched() should be sufficient.

Yes, with either spin_lock() or mutex_lock(), this should work.  Mutual
exclusion and all that.  ;-)

							Thanx, Paul