On Wed, Sep 28, 2016 at 07:11:00PM -0700, Paul E. McKenney wrote:
> On Thu, Sep 29, 2016 at 10:40:24AM +0900, Joonsoo Kim wrote:
> > On Wed, Sep 28, 2016 at 10:15:01AM +0300, Nikolay Borisov wrote:
> > >
> > >
> > > On 09/28/2016 08:31 AM, Joonsoo Kim wrote:
> > > > Hello,
> > > >
> > > > Ccing Paul, because it looks like RCU problem.
> > > >
> > > > On Mon, Sep 26, 2016 at 10:46:56AM +0300, Nikolay Borisov wrote:
> > > >> Hello,
> > > >>
> > > >> On 4.4.14 stable kernel I observed the following soft-lockup, however I
> > > >> also checked that the code is the same in 4.8-rc so the problem is
> > > >> present there as well:
> > > >>
> > > >> [434575.862377] NMI watchdog: BUG: soft lockup - CPU#13 stuck for 23s! [swapper/13:0]
> > > >> [434575.866352] CPU: 13 PID: 0 Comm: swapper/13 Tainted: P O 4.4.14-clouder5 #2
> > > >> [434575.866643] Hardware name: Supermicro X9DRD-iF/LF/X9DRD-iF, BIOS 3.0b 12/05/2013
> > > >> [434575.866932] task: ffff8803714aadc0 ti: ffff8803714c4000 task.ti: ffff8803714c4000
> > > >> [434575.867221] RIP: 0010:[<ffffffff81613f4c>] [<ffffffff81613f4c>] _raw_spin_unlock_irqrestore+0x1c/0x30
> > > >> [434575.867566] RSP: 0018:ffff880373ce3dc0 EFLAGS: 00000203
> > > >> [434575.867736] RAX: ffff88066e0c9a40 RBX: 0000000000000203 RCX: 0000000000000000
> > > >> [434575.868023] RDX: 0000000000000008 RSI: 0000000000000203 RDI: ffff88066e0c9a40
> > > >> [434575.868311] RBP: ffff880373ce3dc8 R08: ffff8803e5c1d118 R09: ffff8803e5c1d538
> > > >> [434575.868609] R10: 0000000000000000 R11: ffffea000f970600 R12: ffff88066e0c9a40
> > > >> [434575.868895] R13: ffffea000f970600 R14: 000000000046cf3b R15: ffff88036f8e3200
> > > >> [434575.869183] FS: 0000000000000000(0000) GS:ffff880373ce0000(0000) knlGS:0000000000000000
> > > >> [434575.869472] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > >> [434575.869643] CR2: ffffffffff600400 CR3: 0000000367201000 CR4: 00000000001406e0
> > > >> [434575.869931] Stack:
> > > >> [434575.870095] ffff88066e0c9a40 ffff880373ce3e78 ffffffff8117ea8a ffff880373ce3e08
> > > >> [434575.870567] 000000000046bd03 0000000100170017 ffff8803e5c1d118 ffff8803e5c1d118
> > > >> [434575.871037] 00ff000100000000 0000000000000203 0000000000000000 ffffffff8123d9ac
> > > >> [434575.874253] Call Trace:
> > > >> [434575.874418] <IRQ>
> > > >> [434575.874473] [<ffffffff8117ea8a>] __slab_free+0xca/0x290
> > > >> [434575.874806] [<ffffffff8123d9ac>] ? ext4_i_callback+0x1c/0x20
> > > >> [434575.874978] [<ffffffff8117ee3a>] kmem_cache_free+0x1ea/0x200
> > > >> [434575.875149] [<ffffffff8123d9ac>] ext4_i_callback+0x1c/0x20
> > > >> [434575.875325] [<ffffffff810ad09b>] rcu_process_callbacks+0x21b/0x620
> > > >> [434575.875506] [<ffffffff81057337>] __do_softirq+0x147/0x310
> > > >> [434575.875680] [<ffffffff8105764f>] irq_exit+0x5f/0x70
> > > >> [434575.875851] [<ffffffff81616a82>] smp_apic_timer_interrupt+0x42/0x50
> > > >> [434575.876025] [<ffffffff816151e9>] apic_timer_interrupt+0x89/0x90
> > > >> [434575.876197] <EOI>
> > > >> [434575.876250] [<ffffffff81510601>] ? cpuidle_enter_state+0x141/0x2c0
> > > >> [434575.876583] [<ffffffff815105f6>] ? cpuidle_enter_state+0x136/0x2c0
> > > >> [434575.876755] [<ffffffff815107b7>] cpuidle_enter+0x17/0x20
> > > >> [434575.876929] [<ffffffff810949fc>] cpu_startup_entry+0x2fc/0x360
> > > >> [434575.877105] [<ffffffff810330e3>] start_secondary+0xf3/0x100
> > > >>
> > > >> The ip in __slab_free points to this piece of code (in mm/slub.c):
> > > >>
> > > >> if (unlikely(n)) {
> > > >> spin_unlock_irqrestore(&n->list_lock, flags);
> > > >> n = NULL;
> > > >> }
> > > >>
> > > >> I think it's a pure chance that the spin_unlock_restore is being shown in this trace,
> > > >> do you think that a cond_resched is needed in this unlikely if clause? Apparently there
> > > >> are cases where this loop can take a considerable amount of time.
> > > >
> > > > I think that __slab_free() doesn't take too long time even if there is
> > > > lock contention. And, cond_resched() is valid on softirq context?
> > >
> > > Now that I think of it - it's not valid since it might sleep and softirq
> > > is atomic context. So my suggestion is actually invalid, too bad.
> > >
> > > >
> > > > I think that problem would be caused by too many rcu callback is
> > > > executed without scheduling. Paul?
> > >
> > > I don't think it's an RCU problem per-se since ext4_i_callback is being
> > > called from RCU due to the way inodes are being freed.
> >
> > That doesn't mean that RCU has no problem. IIUC, the fact is that RCU
> > has no scheduling point in rcu_process_callbacks() and it would be
> > problematic. It just depends on workload.
>
> You mean rcu_do_batch()? It does limit the callbacks invoked per call
> to rcu_do_batch() under normal conditions, see the "++count >= bl" check.
Sorry, in fact, I didn't check anything deeply. What I checked is that
rcu_process_callbacks() has for_each_rcu_flavor() which is a loop.
But, now I find that it is a small loop just for rcu type. Sorry about
my careless mention.
Now, I checked that rcu_do_batch() has a finish point as you mentioned.
Hmm... I still think that there is no problem on __slab_free().
There maybe someone to corrupt the memory and it would cause looping in
slab code.
Nikolay, could you share your .config?
>
> Now, if you dump a huge number of callbacks down call_rcu()'s throat,
> it will stop being Mr. Nice Guy and will start executing the callbacks
> as fast as it can for potentially quite some time. But a huge number
> will be in the millions. Per CPU. In which case I just might have a
> few questions about exactly what you are trying to do.
>
> Nevertheless, it is entirely possible that RCU's callback-invocation
> throttling strategy needs improvement. So why not turn on ftrace and
> have the machine tell you exactly where it is spending all of its time?
> That would allow us to work out what the right fix should be.
Enabling ftrace looks a good idea!
> > > On a slightly different note - on a different physical server, running
> > > zfsonlinux I experienced a very similar issue but without being in an
> > > RCU context, here what the stacktrace looked there:
> > >
> > > Call Trace:
> > > [<ffffffff8117ea8a>] __slab_free+0xca/0x290
> > > [<ffffffff8117ee3a>] ? kmem_cache_free+0x1ea/0x200
> > > [<ffffffff8117ee3a>] kmem_cache_free+0x1ea/0x200
> > > [<ffffffffa0245c97>] spl_kmem_cache_free+0x147/0x1d0 [spl]
> > > [<ffffffffa02cbecc>] dnode_destroy+0x1dc/0x230 [zfs]
> > > [<ffffffffa02cbf64>] dnode_buf_pageout+0x44/0xc0 [zfs]
> > > [<ffffffffa0248301>] taskq_thread+0x291/0x4e0 [spl]
> > > [<ffffffff8107ccd0>] ? wake_up_q+0x70/0x70
> > > [<ffffffffa0248070>] ? taskq_thread_spawn+0x50/0x50 [spl]
> > > [<ffffffff810715bf>] kthread+0xef/0x110
> > > [<ffffffff810714d0>] ? kthread_park+0x60/0x60
> > > [<ffffffff8161483f>] ret_from_fork+0x3f/0x70
> > > [<ffffffff810714d0>] ? kthread_park+0x60/0x60
> > >
> > > It's hard to believe this is a coincidence. I've inspected the callpaths
> > > in the taskq_thread/dnode_buf_pageout/dnode_destroy and there doesn't
> > > seem to be a loop apart from the one in __slab_free.
> >
> > I checked zfsonlinux and found that there are loops.
> > First, taskq_thread() doesn't call schedule() if there is pending
> > work item. Second, dnode_buf_pageout() is just simple loop.
>
> Then shouldn't these be fixed?
These are out of tree modules and I didn't look at it deeply. My
analysis would be incorrect.
Thanks.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>