lkml.org 
[lkml]   [2020]   [Nov]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] kfence: Avoid stalling work queue task without allocations
On Thu, Nov 12, 2020 at 12:00PM -0800, Paul E. McKenney wrote:
> On Thu, Nov 12, 2020 at 07:12:54PM +0100, Marco Elver wrote:
> > On Thu, Nov 12, 2020 at 09:54AM -0800, Paul E. McKenney wrote:
> > > On Thu, Nov 12, 2020 at 05:14:39PM +0100, Marco Elver wrote:
> > > > On Thu, Nov 12, 2020 at 01:49PM +0100, Marco Elver wrote:
> > > > > On Thu, 12 Nov 2020 at 01:11, Paul E. McKenney <paulmck@kernel.org> wrote:
> > > > [...]
> > > > > > > This assert didn't fire yet, I just get more of the below. I'll keep
> > > > > > > rerunning, but am not too hopeful...
> > > > > >
> > > > > > Is bisection a possibility?
> > > > >
> > > > > I've been running a bisection for past ~12h, and am making slow
> > > > > progress. It might be another 12h, but I think it'll get there.
> > > >
> > > > Bisection gave me this:
> > > >
> > > > | git bisect start
> > > > | # bad: [c07b306d7fa5680777e2132662d2e6c19fb53579] kfence: Avoid stalling work queue task without allocations
> > > > | git bisect bad c07b306d7fa5680777e2132662d2e6c19fb53579
> > > > | # good: [3cea11cd5e3b00d91caf0b4730194039b45c5891] Linux 5.10-rc2
> > > > | git bisect good 27598e7e73260ed0b2917eb02d4a515ebb578313
> > > > | # good: [3e5acbea719e66ef3be64fe74c99cc905ca697dc] Merge remote-tracking branch 'wireless-drivers-next/master' into master
> > > > | git bisect good 3e5acbea719e66ef3be64fe74c99cc905ca697dc
> > > > | # good: [491a5a9a2fea28353d99621b8abb83b6928b4e36] Merge remote-tracking branch 'sound-asoc/for-next' into master
> > > > | git bisect good 491a5a9a2fea28353d99621b8abb83b6928b4e36
> > > > | # bad: [502f8643d6e21c7e370a0b75131130cc51609055] Merge remote-tracking branch 'phy-next/next' into master
> > > > | git bisect bad 502f8643d6e21c7e370a0b75131130cc51609055
> > > > | # good: [6693cb1fa5ea7b91ec00f9404776a095713face5] Merge remote-tracking branch 'tip/auto-latest' into master
> > > > | git bisect good 6693cb1fa5ea7b91ec00f9404776a095713face5
> > > > | # bad: [b790e3afead9357195b6d1e1b6cd9b3521503ad2] Merge branch 'tglx-pc.2020.10.30a' into HEAD
> > > > | git bisect bad b790e3afead9357195b6d1e1b6cd9b3521503ad2
> > > > | # bad: [765b512bb3d639bfad7dd43c288ee085236c7267] Merge branches 'cpuinfo.2020.11.06a', 'doc.2020.11.06a', 'fixes.2020.11.02a', 'lockdep.2020.11.02a', 'tasks.2020.11.06a' and 'torture.2020.11.06a' into HEAD
> > > > | git bisect bad 765b512bb3d639bfad7dd43c288ee085236c7267
> > > > | # good: [01f9e708d9eae6335ae9ff25ab09893c20727a55] tools/rcutorture: Fix BUG parsing of console.log
> > >
> > > So torture.2020.11.06a is OK.
> > >
> > > > | git bisect good 01f9e708d9eae6335ae9ff25ab09893c20727a55
> > > > | # good: [1be6ab91e2db157faedb7f16ab0636a80745a073] srcu: Take early exit on memory-allocation failure
> > >
> > > As is fixes.2020.11.02a.
> > >
> > > > | git bisect good 1be6ab91e2db157faedb7f16ab0636a80745a073
> > > > | # good: [65e9eb1ccfe56b41a0d8bfec651ea014968413cb] rcu: Prevent RCU_LOCKDEP_WARN() from swallowing the condition
> > >
> > > And lockdep.2020.11.02a.
> > >
> > > > | git bisect good 65e9eb1ccfe56b41a0d8bfec651ea014968413cb
> > > > | # good: [c386e29d43728778ddd642fa73cc582bee684171] docs/rcu: Update the call_rcu() API
> > >
> > > And doc.2020.11.06a.
> > >
> > > > | git bisect good c386e29d43728778ddd642fa73cc582bee684171
> > > > | # good: [27c0f1448389baf7f309b69e62d4b531c9395e88] rcutorture: Make grace-period kthread report match RCU flavor being tested
> > >
> > > And the first three commits of tasks.2020.11.06a.
> > >
> > > > | git bisect good 27c0f1448389baf7f309b69e62d4b531c9395e88
> > > > | # good: [3fcd6a230fa7d03bffcb831a81b40435c146c12b] x86/cpu: Avoid cpuinfo-induced IPIing of idle CPUs
> > >
> > > And cpuinfo.2020.11.06a.
> > >
> > > > | git bisect good 3fcd6a230fa7d03bffcb831a81b40435c146c12b
> > > > | # good: [75dc2da5ecd65bdcbfc4d59b9d9b7342c61fe374] rcu-tasks: Make the units of ->init_fract be jiffies
> > >
> > > And the remaining commit of tasks.2020.11.06a.
> > >
> > > > | git bisect good 75dc2da5ecd65bdcbfc4d59b9d9b7342c61fe374
> > > > | # first bad commit: [765b512bb3d639bfad7dd43c288ee085236c7267] Merge branches 'cpuinfo.2020.11.06a', 'doc.2020.11.06a', 'fixes.2020.11.02a', 'lockdep.2020.11.02a', 'tasks.2020.11.06a' and 'torture.2020.11.06a' into HEAD
> > > >
> > > > This doesn't look very satisfying, given it's the merge commit. :-/
> > >
> > > So each individual branch is just fine, but the merge of them is not. Fun.
> > >
> > > These have been passing quite a bit of rcutorture over here, including
> > > preemptible kernels running !SMP, but admittedly on x86 rather than ARMv8.
> >
> > Note that this is ARMv8 on QEMU on an x86 host i.e. emulated. And it's
> > really slow as a result. Together with a bunch of debug tools including
> > lockdep.
>
> Then I don't envy you the bisection process! ;-)
>
> > > One approach would be to binary-search the combinations of merges.
> > > Except that there are six of them, so there are 64 combinations, of
> > > which you have tested only 8 thus far (none, one each, and all).
> > >
> > > But are you sure that the bisection points labeled "good" really are good?
> > > For example, what is the distribution of first failure times in the
> > > points labeled "bad" vs. the runtime used to make a "good" determination?
> > > Alternatively, just try a longer run on each of the commits feeding into
> > > the merge point.
> >
> > Yeah, I'm having doubts, and this might be even more non-deterministic
> > that I thought and some 'good' could maybe be 'bad' if I had re-run
> > them? I don't know. One thing I can try is to make sure I run it more
> > than once, but I'm definitely not doing that manually, so let me try and
> > script something so I don't have to hand-hold the bisection overnight.
> > :-)
>
> I know that feeling. A similar experience motivated me to upgrade my
> tooling, with more upgrades in the queue.

[.....]

> > > > | [ 841.143527] arch_local_irq_restore+0x4/0x8
> > >
> > > So we are just now restoring interrupts, hence our getting the
> > > interrupt at this point..
> > >
> > > > | [ 841.147612] trace_preempt_on+0xf4/0x190
> > >
> > > From within the trace code, which is apparently recording the fact
> > > that preemption is being enabled.
> > >
> > > > | [ 841.151656] preempt_schedule_common+0x12c/0x1b0
> > > > | [ 841.155869] preempt_schedule.part.88+0x20/0x28
> > > > | [ 841.160036] preempt_schedule+0x20/0x28
> > >
> > > I was not aware that releasing a raw spinlock could result in a direct
> > > call to preempt_schedule().
> > >
> > > > | [ 841.164051] _raw_spin_unlock_irq+0x80/0x90
> > > > | [ 841.168139] rcu_gp_kthread+0xe5c/0x19a8
> > >
> > > So the RCU grace-period kthread has spent many seconds attempting to
> > > release a lock? Am I reading this correctly? Mark Rutland, am I missing
> > > something here?
>
> And yes, this is the RCU grace-period kthread releasing a lock.
>
> I have no idea why that would take so long. It is acting like a
> self-deadlock or similar hang, except that in that case, lockdep should
> have complained before the RCU CPU stall warning rather than after.
>
> The only thing I can suggest is sprinkling lockdep_assert_irqs_disabled()
> calls hither and yon. All of the code that lockdep is complaining about
> runs in the context of the scheduling-clock interrupt, so interrupts
> had jolly well be disabled! ;-)
>
> Rerunning some of the allegedly good bisects might be more productive.

Oof, so I reran bisection, and this time confirming 3x each good run.
This is what I get:

| git bisect start
| # bad: [c07b306d7fa5680777e2132662d2e6c19fb53579] kfence: Avoid stalling work queue task without allocations
| git bisect bad c07b306d7fa5680777e2132662d2e6c19fb53579
| # good: [3cea11cd5e3b00d91caf0b4730194039b45c5891] Linux 5.10-rc2
| git bisect good 27598e7e73260ed0b2917eb02d4a515ebb578313
| # good: [3e5acbea719e66ef3be64fe74c99cc905ca697dc] Merge remote-tracking branch 'wireless-drivers-next/master' into master
| git bisect good 3e5acbea719e66ef3be64fe74c99cc905ca697dc
| # good: [491a5a9a2fea28353d99621b8abb83b6928b4e36] Merge remote-tracking branch 'sound-asoc/for-next' into master
| git bisect good 491a5a9a2fea28353d99621b8abb83b6928b4e36
| # bad: [502f8643d6e21c7e370a0b75131130cc51609055] Merge remote-tracking branch 'phy-next/next' into master
| git bisect bad 502f8643d6e21c7e370a0b75131130cc51609055
| # good: [6693cb1fa5ea7b91ec00f9404776a095713face5] Merge remote-tracking branch 'tip/auto-latest' into master
| git bisect good 6693cb1fa5ea7b91ec00f9404776a095713face5
| # good: [b790e3afead9357195b6d1e1b6cd9b3521503ad2] Merge branch 'tglx-pc.2020.10.30a' into HEAD
| git bisect good b790e3afead9357195b6d1e1b6cd9b3521503ad2
| # bad: [7bd5bb161657717d576798f62b0e8d5b44653139] Merge remote-tracking branch 'drivers-x86/for-next' into master
| git bisect bad 7bd5bb161657717d576798f62b0e8d5b44653139
| # bad: [e71eb4c4d42bcf36a3a7ede30fd320d47b3c8cb8] Merge remote-tracking branch 'xen-tip/linux-next' into master
| git bisect bad e71eb4c4d42bcf36a3a7ede30fd320d47b3c8cb8
| # bad: [c0a41bf9dbc751692c8cb1a44bfd48e70e8bef7f] docs: Remove redundant "``" from Requirements.rst
| git bisect bad c0a41bf9dbc751692c8cb1a44bfd48e70e8bef7f
| # bad: [c293fb8f7de6c2fce11cb01a0218d668df326bcd] torture: Make --kcsan specify lockdep
| git bisect bad c293fb8f7de6c2fce11cb01a0218d668df326bcd
| # good: [5068ab7dcb6a526a401054ebe0d416f979efb3e1] rcutorture: Add testing for RCU's global memory ordering
| git bisect good 5068ab7dcb6a526a401054ebe0d416f979efb3e1
| # good: [bea68a13bbbdc575a2c868dabd7b454c2eddc618] rcu/segcblist: Add additional comments to explain smp_mb()
| git bisect good bea68a13bbbdc575a2c868dabd7b454c2eddc618
| # first bad commit: [c293fb8f7de6c2fce11cb01a0218d668df326bcd] torture: Make --kcsan specify lockdep

Which clearly is ridiculous! So my guess is this probably had existed
before, but something in -next is making it more visible.

Short of giving up, I can try your suggestion of sprinkling
lockdep_assert_irqs_disabled() everywhere, or if you have a patch to
apply that would give some other debug output you wanted I can run that
too.

Thanks,
-- Marco

\
 
 \ /
  Last update: 2020-11-13 12:09    [W:0.181 / U:1.012 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site