Re: Question: RCU stall detected in memcg (Re: mmotm 2009-09-25-14-35 uploaded

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Sep 28, 2009 at 03:42:13PM +0900, KAMEZAWA Hiroyuki wrote:
> On Fri, 25 Sep 2009 14:35:46 -0700
> akpm@linux-foundation.org wrote:
> 
> > The mm-of-the-moment snapshot 2009-09-25-14-35 has been uploaded to
> > 
> >    http://userweb.kernel.org/~akpm/mmotm/
> > 
> > and will soon be available at
> > 
> >    git://git.zen-sources.org/zen/mmotm.git
> > 
> > It contains the following patches against 2.6.31:
> > 
> 
> At testing my (small) patch, with high memory pressure to
> memcg+hierarchy+softlimit, following is shown.
> ==
> INFO: RCU detected CPU 0 stall (t=10000 jiffies)
> sending NMI to all CPUs:
> NMI backtrace for cpu 0
> CPU 0:
> Modules linked in: sco bridge stp bnep l2cap crc16 bluetooth rfkill iptabl
> e_filter ip_tables ip6table_filter ip6_tables x_tables ipv6 cpufreq_ondemand acpi_cpufreq dm_mirror dm_region_hash dm_log d
> m_multipath dm_mod uinput ppdev i2c_i801 pcspkr i2c_core bnx2 sg e1000e parport_pc parport button shpchp megaraid_sas sd_mo
> d scsi_mod ext3 jbd uhci_hcd ohci_hcd ehci_hcd [last unloaded: microcode]
>  Pid: 2886, comm: ruby Not tainted 2.6.31-mm1 #2 PRIMERGY
>  RIP: 0010:[<ffffffff810878fe>]  [<ffffffff810878fe>] trace_hardirqs_off_ca
> ller+0x3e/0xb RSP: 0018:ffff88004fa03d98  EFLAGS: 00000006
>  RAX: 0000000000000046 RBX: 0000000000000c00 RCX: 000000000000e501
>  RDX: ffff8806133564f0 RSI: 0000000000000002 RDI: ffffffff8102a940
>  RBP: ffff88004fa03d98 R08: 0000000000000001 R09: 0000000000000000
>  R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000002
>  R13: 0000000000000046 R14: 00000000000000ff R15: ffff88004fa03f48
>  FS:  00007fdeca0856f0(0000) GS:ffff88004fa00000(0000) knlGS:0000000000000000
>  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>  CR2: 00007fdeca09e000 CR3: 0000000619fc6000 CR4: 00000000000006f0
>  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
>  DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
>  Call Trace:
>   <#DB[1]>  <<EOE>> Pid: 2886, comm: ruby Not tainted 2.6.31-mm1 #2
>  Call Trace:
>   <NMI>  [<ffffffff8100af79>] ? show_regs+0x49/0x50
>   [<ffffffff81429385>] nmi_watchdog_tick+0x1e5/0x210
>   [<ffffffff81428891>] do_nmi+0x1b1/0x2e0
>   [<ffffffff8142808a>] nmi+0x1a/0x2c
>   [<ffffffff8102a940>] ? flat_send_IPI_mask+0x90/0xb0
>   [<ffffffff810878fe>] ? trace_hardirqs_off_caller+0x3e/0xb0
>   <<EOE>>  <IRQ>  [<ffffffff810884bd>] trace_hardirqs_off+0xd/0x10
>   [<ffffffff8102a940>] flat_send_IPI_mask+0x90/0xb0
>   [<ffffffff8102a9c9>] flat_send_IPI_all+0x69/0x70
>   [<ffffffff81027372>] arch_trigger_all_cpu_backtrace+0x62/0xa0
>   [<ffffffff810bff8e>] __rcu_pending+0x7e/0x370
>   [<ffffffff810c02c7>] rcu_check_callbacks+0x47/0x130
>   [<ffffffff81063a26>] update_process_times+0x46/0x70
>   [<ffffffff81085930>] tick_sched_timer+0x60/0x160
>   [<ffffffff810858d0>] ? tick_sched_timer+0x0/0x160
>   [<ffffffff8107a03a>] __run_hrtimer+0xba/0x150
>   [<ffffffff8107a325>] hrtimer_interrupt+0xd5/0x1b0
>   [<ffffffff81426dfe>] ? trace_hardirqs_off_thunk+0x3a/0x3c
>   [<ffffffff8142cacd>] smp_apic_timer_interrupt+0x6d/0x9b
>   [<ffffffff8100cb33>] apic_timer_interrupt+0x13/0x20
>   <EOI>  [<ffffffff811317b6>] ? mem_cgroup_walk_tree+0x156/0x180
>   [<ffffffff811316d3>] ? mem_cgroup_walk_tree+0x73/0x180
>   [<ffffffff81131692>] ? mem_cgroup_walk_tree+0x32/0x180

mem_cgroup_walk_tree() does have a loop.  Not sure why it is showing
up on the stack twice.  But it might well be the culprit.  Also,
it repeatedly calls css_get_next(), which acquires ss->id_lock.
If some other CPU is holding this lock too long, or failed to release
it, then we could see a stall spinning on the lock.  (Though these
sorts of problems are often caught by other diagnostics.)

>   [<ffffffff81131a00>] ? mem_cgroup_get_local_stat+0x0/0x110

mem_cgroup_get_local_stat() is straight-line code, so should not be
the problem.

>   [<ffffffff81131d5b>] ? mem_control_stat_show+0x14b/0x330

mem_control_stat_show() has a pair of fixed-iteration loops, so
should not be the problem.

>   [<ffffffff810a57fd>] ? cgroup_seqfile_show+0x3d/0x60

Straight-line code, should not be a problem.

>   [<ffffffff810a5b90>] ? cgroup_map_add+0x0/0x30

Straight-line code, should not be a problem.

>   [<ffffffff8115de03>] ? seq_read+0xf3/0x420

I suppose that seq_read() could be repeatedly getting -EAGAIN out of
traverse(), but unless I am missing something, we would be seeing a
different stack trace in that case.

>   [<ffffffff811d9926>] ? security_file_permission+0x16/0x20

... and so on ...

>   [<ffffffff8113b7ec>] ? vfs_read+0xcc/0x190
>   [<ffffffff8113b9b5>] ? sys_read+0x55/0x90
>   [<ffffffff8100bf9b>] ? system_call_fastpath+0x16/0x1b
> .....
> ==
> mem_cgroup_walk_tree() is not very young function and I've been doing the
> same kind of tests but this is 1st messeage for me. So, I'm a bit confused.
> and not sure how I start debug from...
> 
> Does this mean mem_cgroup_walk_tree() blocks RCU's progress
> over 10000 jiffies ? What should I doubt ?

It means that this CPU has indeed been blocking RCU's progress for more
than 10 seconds (10,000 jiffies if HZ==1000).  One thing would be to
wait for another 30 seconds and see if you get another stall warning.
If not, then you could try decreasing RCU_SECONDS_TILL_STALL_RECHECK to
(say) 5*HZ.

Either way, if you get two stack traces during the same CPU-stall event,
it can give you a much better idea where in the callstack the stall is
actually occurring.

							Thanx, Paul