Re: INFO: task hung in perf_trace_event_unreg

From: Paul E. McKenney
Date: Thu Apr 12 2018 - 11:06:26 EST


On Thu, Apr 12, 2018 at 11:39:42AM +0200, Dmitry Vyukov wrote:
> On Wed, Apr 11, 2018 at 9:36 PM, Paul E. McKenney
> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> >> >> >> >> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> >> >> >> >> >> >> >> >>
> >> >> >> >> >> >> >> >> > Hello,
> >> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> >> > syzbot hit the following crash on upstream commit
> >> >> >> >> >> >> >> >> > 0adb32858b0bddf4ada5f364a84ed60b196dbcda (Sun Apr 1 21:20:27 2018 +0000)
> >> >> >> >> >> >> >> >> > Linux 4.16
> >> >> >> >> >> >> >> >> > syzbot dashboard link:
> >> >> >> >> >> >> >> >> > https://syzkaller.appspot.com/bug?extid=2dbc55da20fa246378fd
> >> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> >> > Unfortunately, I don't have any reproducer for this crash yet.
> >> >> >> >> >> >> >> >> > Raw console output:
> >> >> >> >> >> >> >> >> > https://syzkaller.appspot.com/x/log.txt?id=5487937873510400
> >> >> >> >> >> >> >> >> > Kernel config:
> >> >> >> >> >> >> >> >> > https://syzkaller.appspot.com/x/.config?id=-2374466361298166459
> >> >> >> >> >> >> >> >> > compiler: gcc (GCC) 7.1.1 20170620
> >> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> >> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> >> >> >> >> >> >> >> >> > Reported-by: syzbot+2dbc55da20fa246378fd@xxxxxxxxxxxxxxxxxxxxxxxxx
> >> >> >> >> >> >> >> >> > It will help syzbot understand when the bug is fixed. See footer for
> >> >> >> >> >> >> >> >> > details.
> >> >> >> >> >> >> >> >> > If you forward the report, please keep this part and the footer.
> >> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> >> > REISERFS warning (device loop4): super-6502 reiserfs_getopt: unknown mount
> >> >> >> >> >> >> >> >> > option "g ï;eïKï×>pquota"
> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> > Might not hurt to look into the above, though perhaps this is just syzkaller
> >> >> >> >> >> >> >> > playing around with mount options.
> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> >> > INFO: task syz-executor3:10803 blocked for more than 120 seconds.
> >> >> >> >> >> >> >> >> > Not tainted 4.16.0+ #10
> >> >> >> >> >> >> >> >> > "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >> >> >> >> >> >> >> >> > syz-executor3 D20944 10803 4492 0x80000002
> >> >> >> >> >> >> >> >> > Call Trace:
> >> >> >> >> >> >> >> >> > context_switch kernel/sched/core.c:2862 [inline]
> >> >> >> >> >> >> >> >> > __schedule+0x8fb/0x1ec0 kernel/sched/core.c:3440
> >> >> >> >> >> >> >> >> > schedule+0xf5/0x430 kernel/sched/core.c:3499
> >> >> >> >> >> >> >> >> > schedule_timeout+0x1a3/0x230 kernel/time/timer.c:1777
> >> >> >> >> >> >> >> >> > do_wait_for_common kernel/sched/completion.c:86 [inline]
> >> >> >> >> >> >> >> >> > __wait_for_common kernel/sched/completion.c:107 [inline]
> >> >> >> >> >> >> >> >> > wait_for_common kernel/sched/completion.c:118 [inline]
> >> >> >> >> >> >> >> >> > wait_for_completion+0x415/0x770 kernel/sched/completion.c:139
> >> >> >> >> >> >> >> >> > __wait_rcu_gp+0x221/0x340 kernel/rcu/update.c:414
> >> >> >> >> >> >> >> >> > synchronize_sched.part.64+0xac/0x100 kernel/rcu/tree.c:3212
> >> >> >> >> >> >> >> >> > synchronize_sched+0x76/0xf0 kernel/rcu/tree.c:3213
> >> >> >> >> >> >> >> >>
> >> >> >> >> >> >> >> >> I don't think this is a perf issue. Looks like something is preventing
> >> >> >> >> >> >> >> >> rcu_sched from completing. If there's a CPU that is running in kernel
> >> >> >> >> >> >> >> >> space and never scheduling, that can cause this issue. Or if RCU
> >> >> >> >> >> >> >> >> somehow missed a transition into idle or user space.
> >> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> > The RCU CPU stall warning below strongly supports this position ...
> >> >> >> >> >> >> >>
> >> >> >> >> >> >> >> I think this is this guy then:
> >> >> >> >> >> >> >>
> >> >> >> >> >> >> >> https://syzkaller.appspot.com/bug?id=17f23b094cd80df750e5b0f8982c521ee6bcbf40
> >> >> >> >> >> >> >>
> >> >> >> >> >> >> >> #syz dup: INFO: rcu detected stall in __process_echoes
> >> >> >> >> >> >> >
> >> >> >> >> >> >> > Seems likely to me!
> >> >> >> >> >> >> >
> >> >> >> >> >> >> >> Looking retrospectively at the various hang/stall bugs that we have, I
> >> >> >> >> >> >> >> think we need some kind of priority between them. I.e. we have rcu
> >> >> >> >> >> >> >> stalls, spinlock stalls, workqueue hangs, task hangs, silent machine
> >> >> >> >> >> >> >> hang and maybe something else. It would be useful if they fire
> >> >> >> >> >> >> >> deterministically according to priorities. If there is an rcu stall,
> >> >> >> >> >> >> >> that's always detected as CPU stall. Then if there is no RCU stall,
> >> >> >> >> >> >> >> but a workqueue stall, then that's always detected as workqueue stall,
> >> >> >> >> >> >> >> etc.
> >> >> >> >> >> >> >> Currently if we have an RCU stall (effectively CPU stall), that can be
> >> >> >> >> >> >> >> detected either RCU stall or a task hung, producing 2 different bug
> >> >> >> >> >> >> >> reports (which is bad).
> >> >> >> >> >> >> >> One can say that it's only a matter of tuning timeouts, but at least
> >> >> >> >> >> >> >> task hung detector has a problem that if you set timeout to X, it can
> >> >> >> >> >> >> >> detect hung anywhere between X and 2*X. And on one hand we need quite
> >> >> >> >> >> >> >> large timeout (a minute may not be enough), and on the other hand we
> >> >> >> >> >> >> >> can't wait for an hour just to make sure that the machine is indeed
> >> >> >> >> >> >> >> dead (these things happen every few minutes).
> >> >> >> >> >> >> >
> >> >> >> >> >> >> > I suppose that we could have a global variable that was set to the
> >> >> >> >> >> >> > priority of the complaint in question, which would suppress all
> >> >> >> >> >> >> > lower-priority complaints. Might need to be opt-in, though -- I would
> >> >> >> >> >> >> > guess that not everyone is going to be happy with one complaint suppressing
> >> >> >> >> >> >> > others, especially given the possibility that the two complaints might
> >> >> >> >> >> >> > be about different things.
> >> >> >> >> >> >> >
> >> >> >> >> >> >> > Or did you have something more deft in mind?
> >> >> >> >> >> >>
> >> >> >> >> >> >>
> >> >> >> >> >> >> syzkaller generally looks only at the first report. One does not know
> >> >> >> >> >> >> if/when there will be a second one, or the second one can be induced
> >> >> >> >> >> >> by the first one, and we generally want clean reports on a non-tainted
> >> >> >> >> >> >> kernel. So we don't just need to suppress lower priority ones, we need
> >> >> >> >> >> >> to produce the right report first.
> >> >> >> >> >> >> I am thinking maybe setting:
> >> >> >> >> >> >> - rcu stalls at 1.5 minutes
> >> >> >> >> >> >> - workqueue stalls at 2 minutes
> >> >> >> >> >> >> - task hungs at 2.5 minutes
> >> >> >> >> >> >> - and no output whatsoever at 3 minutes
> >> >> >> >> >> >> Do I miss anything? I think at least spinlocks. Should they go before
> >> >> >> >> >> >> or after rcu?
> >> >> >> >> >> >
> >> >> >> >> >> > That is what I know of, but the Linux kernel being what it is, there is
> >> >> >> >> >> > probably something more out there. If not now, in a few months. The
> >> >> >> >> >> > RCU CPU stall timeout can be set on the kernel-boot command line, but
> >> >> >> >> >> > you probably already knew that.
> >> >> >> >> >>
> >> >> >> >> >> Well, it's all based solely on a large number of patches and stopgaps.
> >> >> >> >> >> If we fix main problems for today, it's already good.
> >> >> >> >> >
> >> >> >> >> > Fair enough!
> >> >> >> >> >
> >> >> >> >> >> > Just for comparison, back in DYNIX/ptx days the RCU CPU stall timeout
> >> >> >> >> >> > was 1.5 -seconds-. ;-)
> >> >> >> >> >>
> >> >> >> >> >> Have you tried to instrument every basic block with a function call to
> >> >> >> >> >> collect coverage, check every damn memory access for validity, enable
> >> >> >> >> >> all thinkable and unthinkable debug configs and put the insanest load
> >> >> >> >> >> one can imagine from a swarm of parallel threads? It makes things a
> >> >> >> >> >> bit slower ;)
> >> >> >> >> >
> >> >> >> >> > Given that we wouldn't have had enough CPU or memory to accommodate
> >> >> >> >> > all of that back in DYNIX/ptx days, I am forced to answer "no". ;-)
> >> >> >> >> >
> >> >> >> >> >> >> This will require fixing task hung. Have not yet looked at workqueue detector.
> >> >> >> >> >> >> Does at least RCU respect the given timeout more or less precisely?
> >> >> >> >> >> >
> >> >> >> >> >> > Assuming that there is at least one CPU capable of taking scheduling-clock
> >> >> >> >> >> > interrupts, it should respect the timeout to within a few jiffies.
> >> >> >> >>
> >> >> >> >>
> >> >> >> >> Hi Paul,
> >> >> >> >>
> >> >> >> >> Speaking of stalls and rcu, we are seeing lots of crashes that go like this:
> >> >> >> >>
> >> >> >> >> INFO: rcu_sched self-detected stall on CPU[ 404.992530] INFO:
> >> >> >> >> rcu_sched detected stalls on CPUs/tasks:
> >> >> >> >> INFO: rcu_sched self-detected stall on CPU[ 454.347448] INFO:
> >> >> >> >> rcu_sched detected stalls on CPUs/tasks:
> >> >> >> >> INFO: rcu_sched self-detected stall on CPU[ 396.073634] INFO:
> >> >> >> >> rcu_sched detected stalls on CPUs/tasks:
> >> >> >> >>
> >> >> >> >> or like this:
> >> >> >> >>
> >> >> >> >> INFO: rcu_sched self-detected stall on CPU
> >> >> >> >> INFO: rcu_sched detected stalls on CPUs/tasks:
> >> >> >> >> 0-....: (125000 ticks this GP) idle=0ba/1/4611686018427387906
> >> >> >> >> softirq=57641/57641 fqs=31151
> >> >> >> >> 0-....: (125000 ticks this GP) idle=0ba/1/4611686018427387906
> >> >> >> >> softirq=57641/57641 fqs=31151
> >> >> >> >> (t=125002 jiffies g=31656 c=31655 q=910)
> >> >> >> >>
> >> >> >> >> INFO: rcu_sched self-detected stall on CPU
> >> >> >> >> INFO: rcu_sched detected stalls on CPUs/tasks:
> >> >> >> >> 0-....: (125000 ticks this GP) idle=49a/1/4611686018427387906
> >> >> >> >> softirq=65194/65194 fqs=31231
> >> >> >> >> 0-....: (125000 ticks this GP) idle=49a/1/4611686018427387906
> >> >> >> >> softirq=65194/65194 fqs=31231
> >> >> >> >> (t=125002 jiffies g=34421 c=34420 q=1119)
> >> >> >> >> (detected by 1, t=125002 jiffies, g=34421, c=34420, q=1119)
> >> >> >> >>
> >> >> >> >>
> >> >> >> >> and then there is an unintelligible mess of 2 reports. Such crashes go
> >> >> >> >> to trash bin, because we can't even say which function hanged. It
> >> >> >> >> seems that in all cases 2 different rcu stall detection facilities
> >> >> >> >> race with each other. Is it possible to make them not race?
> >> >> >> >
> >> >> >> > How about the following (untested, not for mainline) patch? It suppresses
> >> >> >> > all but the "main" RCU flavor, which is rcu_sched for !PREEMPT builds and
> >> >> >> > rcu_preempt otherwise. Either way, this is the RCU flavor corresponding
> >> >> >> > to synchronize_rcu(). This works well in the common case where there
> >> >> >> > is almost always an RCU grace period in flight.
> >> >> >> >
> >> >> >> > One reason that this patch is not for mainline is that I am working on
> >> >> >> > merging the RCU-bh, RCU-preempt, and RCU-sched flavors into one thing,
> >> >> >> > at which point there won't be any races. But that might be a couple
> >> >> >> > merge windows away from now.
> >> >> >> >
> >> >> >> > Thanx, Paul
> >> >> >> >
> >> >> >> > ------------------------------------------------------------------------
> >> >> >> >
> >> >> >> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> >> >> >> > index 381b47a68ac6..31f7818f2d63 100644
> >> >> >> > --- a/kernel/rcu/tree.c
> >> >> >> > +++ b/kernel/rcu/tree.c
> >> >> >> > @@ -1552,7 +1552,7 @@ static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
> >> >> >> > struct rcu_node *rnp;
> >> >> >> >
> >> >> >> > if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
> >> >> >> > - !rcu_gp_in_progress(rsp))
> >> >> >> > + !rcu_gp_in_progress(rsp) || rsp != rcu_state_p)
> >> >> >> > return;
> >> >> >> > rcu_stall_kick_kthreads(rsp);
> >> >> >> > j = jiffies;
> >> >> >>
> >> >> >> But doesn't they both relate to the same rcu flavor? They both say
> >> >> >> rcu_sched. I assumed that the difference is "self-detected" vs "on
> >> >> >> CPUs/tasks", i.e. on the current CPU vs on other CPUs.
> >> >> >
> >> >> > Right you are!
> >> >> >
> >> >> > One approach would be to increase the value of RCU_STALL_RAT_DELAY,
> >> >> > which is currently two jiffies to (say) 20 jiffies. This is in
> >> >> > kernel/rcu/tree.h. But this would fail on a sufficiently overloaded
> >> >> > system -- and the failure of the two-jiffy delay is a bit of a surprise,
> >> >> > given interrupts disabled and all that. Are you by any chance loaded
> >> >> > heavily enough to see vCPU preemption?
> >> >> >
> >> >> > I could avoid at least some of these timing issues instead using cmpxchg()
> >> >> > on ->jiffies_stall to allow only one CPU in, but leave the non-atomic
> >> >> > update to discourage overly long stall prints from running into the
> >> >> > next one. This is not perfect, either, and is roughly equivalent to
> >> >> > setting RCU_STALL_RAT_DELAY to many second's worth of jiffies, but
> >> >> > avoiding that minute's delay. But it should get rid of the duplication
> >> >> > in almost all cases, though it could allow a stall warning to overlap
> >> >> > with a later stall warning for that same grace period. Which can
> >> >> > already happen anyway. Also, a tens-of-seconds vCPU preemption can
> >> >> > still cause concurrent stall warnings, but if that is happening to you,
> >> >> > the concurrent stall warnings are probably the least of your problems.
> >> >> > Besides, we do need at least one CPU to actually report the stall, which
> >> >> > won't happen if that CPU's vCPU is indefinitely preempted. So there is
> >> >> > only so much I can do about that particular corner case.
> >> >> >
> >> >> > So how does the following (untested) patch work for you?
> >> >>
> >> >> Looks good to me.
> >> >>
> >> >> We run on VMs, so we can well have vCPU preemption.
> >> >
> >> > Very good! Please do get me a Tested-by when you get to that point.
> >>
> >> Unfortunately I don't have a good way to test it until it's submitted
> >> upstream. While we are seeing thousands of such instances, they happen
> >> episodically on a farm of test machines. But they are still harmful,
> >> especially when the system tries to reproduce a bug, because it's
> >> mid-way through and thinks it got a hook, but then suddenly boom! it
> >> gets some mess that it can't parse and now it does not know if it's
> >> still the same bug, or maybe a different bug triggered by the same
> >> program, so it does not know how to properly attribute the reproducer.
> >> You can see these cases as they happen here (under report/log links in
> >> the table):
> >> https://syzkaller.appspot.com/bug?id=d5bc3e0c66d200d72216ab343a67c4327e4a3452
> >> When the patch is submitted, the rate should go down.
> >
> > OK, I will bite... How do you test fixes to problems that syzkaller finds?
>
> I don't. I can't. No one can test that many fixes.
>
> Normally syzbot provides reproducers for bugs. Then you have 2
> choices: (1) test it yourself (if you debugged it, you probably
> already have everything setup for this), or (2) ask syzbot to test the
> patch on this particular reproducer.
> Some bugs don't have reproducers. Then you either localize the bug and
> write a test, or go with the old good "it must be correct, right?".
> Even for the second case, syzbot will notify if the bug happens again
> after the fix is landed, or it's silent, then presumably the fix
> indeed fixed the bug.
>
> Now, this is not a syzbot bug (syzbot reports bugs itself from own
> email address). This is more like you looked at somebody else dmsg and
> like "oh, this looks bad, let me copy-paste and report it".
> So can also go with the old good "it must be correct, right?" and
> assess how well it goes after few weeks when it reaches syzbot, or
> someone needs to write a test for rcu.
>
> This could have been handled with some kind of "cluster-wide" test,
> but I don't see how it is feasible. See this for details:
> https://groups.google.com/d/msg/syzkaller-bugs/7ucgCkAJKSk/skZjgavRAQAJ
> Especially the part that someone will need to go through and triage
> hundreds of crashes and assess that they are not related to the new
> patch, and do something with then afterwards.

Fair enough, and apologies for the hassle. I don't expect that the
patch will be controversial, so it should go into the next merge
window.

Thanx, Paul