Re: RCU nocb list not reclaiming causing OOM
From: Paul E. McKenney
Date: Fri Jul 27 2018 - 19:48:09 EST
On Fri, Jul 27, 2018 at 11:16:39PM +0000, David Chen wrote:
> Hi Paul,
>
> Thanks for the advice.
> The bug is kind of hard to hit, so I can't say for certain yet.
Well, you can always remove the "tail == &rdp->nocb_follower_head" as an
extra belt-and-suspenders safety net. I am not putting that in mainline,
but in the privacy of your own copy of the kernel, I don't see any really
serious problem with it. (As long as you aren't going for absolute maximum
performance, but even then there are other more important tuning actions
and code changes you could make.)
> Though after another look at the code, I found out the `smp_mb__after_atomic();`
> seems to be only a compiler barrier on x86.
Yes, and that is because the locked xchg instruction used on x86 to
implement xchg() already provides full ordering. ;-)
> tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
> *tail = rdp->nocb_gp_head;
> smp_mb__after_atomic(); /* Store *tail before wakeup. */
> if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
> swake_up(&rdp->nocb_wq);
>
> But that wouldn't be enough right? Because from 6b5fc3a13318, it stated that
> wakeup operation don't guarantee ordering. And when the follower wakes up, it checks
> for nocb_follower_head, which is assigned by `*tail = rdp->nocb_gp_head;` which doesn't
> have LOCK prefix. So it's possible for follower to wake up and see the list is empty and go
> back to sleep.
Again, xchg() is defined to provide full ordering against all operations
before and after it. Each architecture is required to do whatever is
necessary to implement that full ordering, and x86 need only provide
its "lock xchg" instruction.
The smp_mb__after_atomic() has effect only after atomic read-modify-write
operations that do not return a value, for example, atomic_inc().
If you use it after a value-returning atomic read-modify-write operation
like xchg(), all you do is needlessly slow things down on platforms
that provide non-empty smp_mb__after_atomic() definitions. So again,
smp_mb__after_atomic() after xchg() is pointless.
Please take a look at Documentation/core-api/atomic_ops.rst in the
Linux-kernel source tree for more information. Or get a v4.17 kernel
source tree and check this using the memory model (tools/memory-model
in that version).
Thanx, Paul
> Thanks,
> David
>
> From: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> Sent: Friday, July 27, 2018 3:31 PM
> To: David Chen
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: RCU nocb list not reclaiming causing OOM
>
>
> On Fri, Jul 27, 2018 at 07:07:46PM +0000, David Chen wrote:
> > Hi Paul,
> >
> > I'd like to opinion again on this subject.
> >
> > So we are going to backport this patch:
> > 6b5fc3a13318 ("rcu: Add memory barriers for NOCB leader wakeup")
>
> Does this one solve the problem, or are you still seeing hangs?
> If you are no longer seeing hangs, my advice is "hands off keyboard",
> though some would no doubt point out that I should follow that advice
> more myself. ;-)
>
> > But the other one:
> > 8be6e1b15c54 ("rcu: Use timer as backstop for NOCB deferred wakeups")
> > It doesn't apply cleanly, and I'm not too comfortable porting it myself.
>
> Yeah, that one is a bit on the non-trivial side, no two ways about it.
>
> > So I'm wondering if I use the following change to always wake up follower thread
> > regardless if the list was empty or not, just to be on the safe side. Do you think
> > this change is reasonable? Do you see any problem it might cause?
> >
> > Thanks,
> > David
> >
> > diff -ru linux-4.9.37.orig/kernel/rcu/tree_plugin.h linux-4.9.37/kernel/rcu/tree_plugin.h
> > --- linux-4.9.37.orig/kernel/rcu/tree_plugin.h 2017-07-12 06:42:41.000000000 -0700
> > +++ linux-4.9.37/kernel/rcu/tree_plugin.h 2018-07-27 11:57:03.582134519 -0700
> > @@ -2077,7 +2077,7 @@
> > tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
> > *tail = rdp->nocb_gp_head;
> > smp_mb__after_atomic(); /* Store *tail before wakeup. */
> > - if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
> > + if (rdp != my_rdp) {
>
> This will burn a bit of extra CPU time, but it should be fine other than
> that.
>
> Thanx, Paul
>
> > /*
> > * List was empty, wake up the follower.
> > * Memory barriers supplied by atomic_long_add().
> >
> >
> > From: David Chen
> > Sent: Friday, July 20, 2018 5:12 PM
> > To: paulmck@xxxxxxxxxxxxxxxxxx
> > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > Subject: Re: RCU nocb list not reclaiming causing OOM
> >
> >
> > Hi Paul,
> >
> > Ok, I'll try those patches.
> >
> > Thanks,
> > David
> >
> > From: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> > Sent: Friday, July 20, 2018 4:32:12 PM
> > To: David Chen
> > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > Subject: Re: RCU nocb list not reclaiming causing OOM
> >
> >
> > On Fri, Jul 20, 2018 at 11:05:52PM +0000, David Chen wrote:
> > > Hi Paul,
> > >
> > > We hit an RCU issue on 4.9.37 kernel. One of the nocb_follower list grows too
> > > large, and not getting reclaimed, causing the system to OOM.
> > >
> > > Printing the culprit rcu_sched_data:
> > >
> > > nocb_q_count = {
> > > counter = 32369635
> > > },
> > > nocb_follower_head = 0xffff88ae901c0a00,
> > > nocb_follower_tail = 0xffff88af1538b8d8,
> > > nocb_kthread = 0xffff88b06d290000,
> > >
> > > As you can see here, the nocb_follower_head is not empty, so in theory, the
> > > nocb_kthread shouldn't go to sleep. However, if dump the stack of the kthread:
> > >
> > > crash> bt 0xffff88b06d290000
> > > PID: 21 TASK: ffff88b06d290000 CPU: 3 COMMAND: "rcuos/1"
> > > #0 [ffffafc9020b7dc0] __schedule at ffffffff8d8789dc
> > > #1 [ffffafc9020b7e38] schedule at ffffffff8d878e76
> > > #2 [ffffafc9020b7e50] rcu_nocb_kthread at ffffffff8d112337
> > > #3 [ffffafc9020b7ec8] kthread at ffffffff8d0c6ce7
> > > #4 [ffffafc9020b7f50] ret_from_fork at ffffffff8d87d755
> > >
> > > And if we dis the address at ffffffff8d112337:
> > >
> > > /usr/src/debug/kernel-4.9.37/linux-4.9.37-29.nutanix.07142017.el7.centos.x86_64/kernel/rcu/tree_plugin.h: 2106
> > > 0xffffffff8d11232d <rcu_nocb_kthread+381>: test %rax,%rax
> > > 0xffffffff8d112330 <rcu_nocb_kthread+384>: jne 0xffffffff8d112355 <rcu_nocb_kthread+421>
> > > 0xffffffff8d112332 <rcu_nocb_kthread+386>: callq 0xffffffff8d878e40 <schedule>
> > > 0xffffffff8d112337 <rcu_nocb_kthread+391>: lea -0x40(%rbp),%rsi
> > >
> > > So the kthread is blocked at swait_event_interruptible in the nocb_follower_wait.
> > > This contradict with the fact that nocb_follower_head was not empty. So I
> > > wonder if this is caused by the lack of memory barrier in the place shown below.
> > > If the head is set to NULL after doing xchg, it will overwrite the head set
> > > by leader. This caused the kthread to sleep the next iteration, and the leader
> > > won't wake him up as the tail doesn't point to head.
> > >
> > > Please tell me what do you think.
> > >
> > > Thanks,
> > > David
> > >
> > > diff -ru linux-4.9.37.orig/kernel/rcu/tree_plugin.h linux-4.9.37/kernel/rcu/tree_plugin.h
> > > --- linux-4.9.37.orig/kernel/rcu/tree_plugin.h 2017-07-12 06:42:41.000000000 -0700
> > > +++ linux-4.9.37/kernel/rcu/tree_plugin.h 2018-07-20 15:25:57.311206343 -0700
> > > @@ -2149,6 +2149,7 @@
> > > BUG_ON(!list);
> > > trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
> > > WRITE_ONCE(rdp->nocb_follower_head, NULL);
> > > + smp_mb();
> > > tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
> >
> > The xchg() operation implies full memory barriers both before and after,
> > so adding the smp_mb() before would have no effect.
> >
> > But let me take a look at post-4.9 changes to this code...
> >
> > I suggest trying out the following commit:
> >
> > 6b5fc3a13318 ("rcu: Add memory barriers for NOCB leader wakeup")
> >
> > If that one doesn't help, the following might be worth trying, but probably
> > a lot harder to backport:
> >
> > 8be6e1b15c54 ("rcu: Use timer as backstop for NOCB deferred wakeups")
> >
> > Please let me know how it goes!
> >
> > Thanx, Paul
> >
> > ------------------------------------------------------------------------
> >
> > commit 6b5fc3a1331810db407c9e0e673dc1837afdc9d0
> > Author: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> > Date: Fri Apr 28 20:11:09 2017 -0700
> >
> > rcu: Add memory barriers for NOCB leader wakeup
> >
> > Wait/wakeup operations do not guarantee ordering on their own. Instead,
> > either locking or memory barriers are required. This commit therefore
> > adds memory barriers to wake_nocb_leader() and nocb_leader_wait().
> >
> > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> > Tested-by: Krister Johansen <kjlx@xxxxxxxxxxxxxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx> # 4.6.x
> >
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index 0b1042545116..573fbe9640a0 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -1810,6 +1810,7 @@ static void wake_nocb_leader(struct rcu_data *rdp, bool force)
> > if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
> > /* Prior smp_mb__after_atomic() orders against prior enqueue. */
> > WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
> > + smp_mb(); /* ->nocb_leader_sleep before swake_up(). */
> > swake_up(&rdp_leader->nocb_wq);
> > }
> > }
> > @@ -2064,6 +2065,7 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)
> > * nocb_gp_head, where they await a grace period.
> > */
> > gotcbs = false;
> > + smp_mb(); /* wakeup before ->nocb_head reads. */
> > for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
> > rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
> > if (!rdp->nocb_gp_head)
> >
> >
>
>