From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-2.2 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, MAILING_LIST_MULTI,SPF_PASS,URIBL_BLOCKED,USER_AGENT_MUTT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 6A5D6C6778F for ; Sat, 28 Jul 2018 08:29:54 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 2708C2088E for ; Sat, 28 Jul 2018 08:29:54 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 2708C2088E Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=linux.vnet.ibm.com Authentication-Results: mail.kernel.org; spf=none smtp.mailfrom=linux-kernel-owner@vger.kernel.org Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1726546AbeG1Jzc (ORCPT ); Sat, 28 Jul 2018 05:55:32 -0400 Received: from mx0b-001b2d01.pphosted.com ([148.163.158.5]:54384 "EHLO mx0a-001b2d01.pphosted.com" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1726061AbeG1Jzb (ORCPT ); Sat, 28 Jul 2018 05:55:31 -0400 Received: from pps.filterd (m0098417.ppops.net [127.0.0.1]) by mx0a-001b2d01.pphosted.com (8.16.0.22/8.16.0.22) with SMTP id w6S8Sl0l124435 for ; Sat, 28 Jul 2018 04:29:48 -0400 Received: from e17.ny.us.ibm.com (e17.ny.us.ibm.com [129.33.205.207]) by mx0a-001b2d01.pphosted.com with ESMTP id 2kgmca12tn-1 (version=TLSv1.2 cipher=AES256-GCM-SHA384 bits=256 verify=NOT) for ; Sat, 28 Jul 2018 04:29:47 -0400 Received: from localhost by e17.ny.us.ibm.com with IBM ESMTP SMTP Gateway: Authorized Use Only! Violators will be prosecuted for from ; Sat, 28 Jul 2018 04:29:47 -0400 Received: from b01cxnp23032.gho.pok.ibm.com (9.57.198.27) by e17.ny.us.ibm.com (146.89.104.204) with IBM ESMTP SMTP Gateway: Authorized Use Only! Violators will be prosecuted; (version=TLSv1/SSLv3 cipher=AES256-GCM-SHA384 bits=256/256) Sat, 28 Jul 2018 04:29:45 -0400 Received: from b01ledav003.gho.pok.ibm.com (b01ledav003.gho.pok.ibm.com [9.57.199.108]) by b01cxnp23032.gho.pok.ibm.com (8.14.9/8.14.9/NCO v10.0) with ESMTP id w6S8Tj1c17039838 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-GCM-SHA384 bits=256 verify=FAIL); Sat, 28 Jul 2018 08:29:45 GMT Received: from b01ledav003.gho.pok.ibm.com (unknown [127.0.0.1]) by IMSVA (Postfix) with ESMTP id 2E105B2066; Sat, 28 Jul 2018 04:29:23 -0400 (EDT) Received: from b01ledav003.gho.pok.ibm.com (unknown [127.0.0.1]) by IMSVA (Postfix) with ESMTP id E3CCFB2065; Sat, 28 Jul 2018 04:29:22 -0400 (EDT) Received: from paulmck-ThinkPad-W541 (unknown [9.85.169.156]) by b01ledav003.gho.pok.ibm.com (Postfix) with ESMTP; Sat, 28 Jul 2018 04:29:22 -0400 (EDT) Received: by paulmck-ThinkPad-W541 (Postfix, from userid 1000) id B04FA16CA3C8; Sat, 28 Jul 2018 01:29:44 -0700 (PDT) Date: Sat, 28 Jul 2018 01:29:44 -0700 From: "Paul E. McKenney" To: David Chen Cc: "linux-kernel@vger.kernel.org" Subject: Re: RCU nocb list not reclaiming causing OOM Reply-To: paulmck@linux.vnet.ibm.com References: <20180720233212.GC12945@linux.vnet.ibm.com> <20180727223125.GJ24813@linux.vnet.ibm.com> <20180727234723.GN24813@linux.vnet.ibm.com> <20180728014700.GO24813@linux.vnet.ibm.com> MIME-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Disposition: inline Content-Transfer-Encoding: 8bit In-Reply-To: <20180728014700.GO24813@linux.vnet.ibm.com> User-Agent: Mutt/1.5.21 (2010-09-15) X-TM-AS-GCONF: 00 x-cbid: 18072808-0040-0000-0000-00000454F1A3 X-IBM-SpamModules-Scores: X-IBM-SpamModules-Versions: BY=3.00009443; HX=3.00000241; KW=3.00000007; PH=3.00000004; SC=3.00000266; SDB=6.01067238; UDB=6.00548396; IPR=6.00845123; MB=3.00022365; MTD=3.00000008; XFM=3.00000015; UTC=2018-07-28 08:29:47 X-IBM-AV-DETECTION: SAVI=unused REMOTE=unused XFE=unused x-cbparentid: 18072808-0041-0000-0000-0000085B29D3 Message-Id: <20180728082944.GA23760@linux.vnet.ibm.com> X-Proofpoint-Virus-Version: vendor=fsecure engine=2.50.10434:,, definitions=2018-07-28_04:,, signatures=0 X-Proofpoint-Spam-Details: rule=outbound_notspam policy=outbound score=0 priorityscore=1501 malwarescore=0 suspectscore=0 phishscore=0 bulkscore=0 spamscore=0 clxscore=1015 lowpriorityscore=0 mlxscore=0 impostorscore=0 mlxlogscore=999 adultscore=0 classifier=spam adjust=0 reason=mlx scancount=1 engine=8.0.1-1806210000 definitions=main-1807280091 Sender: linux-kernel-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Fri, Jul 27, 2018 at 06:47:00PM -0700, Paul E. McKenney wrote: > On Sat, Jul 28, 2018 at 12:07:19AM +0000, David Chen wrote: > > Hi Paul, > > > > I wasn't talking about the xchg() though. > > > > The smp_mb__after_atomic() is not for xchg(), it's for `*tail = rdp->nocb_gp_head;` > > it's stated so in the comment. And I do think we need ordering between > > `*tail = rdp->nocb_gp_head;` and wake up, because the waiter is checking on head not tail. > > > > swait_event_interruptible(rdp->nocb_wq, > > READ_ONCE(rdp->nocb_follower_head)); > > > > So what I'm saying is that since we need to maintain ordering between `*tail = rdp->nocb_gp_head;` > > and wake up, we need to change the smp_mb__after_atomic() to smp_mb(). Because > > smp_mb__after_atomic() wouldn't guarantee > > > > So this is what I'm proposing. > > Good eyes! > > Hmmm... What do I do about this in mainline? Ah, I introduced a > ->nocb_lock in mainline to prevent this from happening. In that same > commit that you didn't want to use because it is hard to backport. ;-) > > So yes, but there might well be other misorderings fixed by the > hard-to-backport commit that your change below does not cover. Still I > do agree that you need full ordering at that point. Another approach would be to backport only the ->nocb_lock portions of that patch. This would still potentially leave failure-to-wake (as opposed to misordering-on-wake) issues, but it should cover all of the misordering-on-wake issues. Thanx, Paul > > 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 16:23:41.349044259 -0700 > > @@ -2076,7 +2076,7 @@ > > /* Append callbacks to follower's "done" list. */ > > tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); > > *tail = rdp->nocb_gp_head; > > - smp_mb__after_atomic(); /* Store *tail before wakeup. */ > > + smp_mb(); /* Store *tail before wakeup. */ > > if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { > > /* > > * List was empty, wake up the follower. > > > > Thanks, > > David > > > > From: Paul E. McKenney > > Sent: Friday, July 27, 2018 4:47 PM > > To: David Chen > > Cc: linux-kernel@vger.kernel.org > > Subject: Re: RCU nocb list not reclaiming causing OOM > >   > > > > 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 > > > Sent: Friday, July 27, 2018 3:31 PM > > > To: David Chen > > > Cc: linux-kernel@vger.kernel.org > > > 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@linux.vnet.ibm.com > > > > Cc: linux-kernel@vger.kernel.org > > > > Subject: Re: RCU nocb list not reclaiming causing OOM > > > >    > > > > > > > > Hi Paul, > > > > > > > > Ok, I'll try those patches. > > > > > > > > Thanks, > > > > David > > > >   > > > > From: Paul E. McKenney > > > > Sent: Friday, July 20, 2018 4:32:12 PM > > > > To: David Chen > > > > Cc: linux-kernel@vger.kernel.org > > > > 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 :      test   %rax,%rax > > > > > 0xffffffff8d112330 :      jne    0xffffffff8d112355 > > > > > 0xffffffff8d112332 :      callq  0xffffffff8d878e40 > > > > > 0xffffffff8d112337 :      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 > > > > 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 > > > >     Tested-by: Krister Johansen > > > >     Cc: # 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) > > > > > > > >      > > > > > >     > > > >