Re: RCU stalls on heavy socket traffic

[Paul E. McKenney]

On Mon, Mar 09, 2015 at 05:39:15PM +0100, Sebastian Andrzej Siewior wrote:
> * Paul E. McKenney | 2015-03-09 08:59:47 [-0700]:
> 
> >On Mon, Mar 09, 2015 at 04:26:51PM +0100, Sebastian Andrzej Siewior wrote:
> >> So I run "hackbench -g 600 -l 350 -s 250" which takes approx 77 seconds
> >> to complete. Then this popped up:
> >> 
> >> |INFO: rcu_preempt detected stalls on CPUs/tasks:
> >> | Tasks blocked on level-0 rcu_node (CPUs 0-3): P24858 P28514 P25185 P25184 P28713 P19549 P3139 P25275 P28474 P29062 P6703 P10 106 P14309 P27910 P4514 P14834 P28385 P21073 P27701 P642 P10340 P16939 P19147 P16949 P16945 P16952 P16941 P16937 P16938 P16946 P16954 P28664  P18701 P17782 P4875 P8873
> >> | Tasks blocked on level-0 rcu_node (CPUs 0-3): P24858 P28514 P25185 P25184 P28713 P19549 P3139 P25275 P28474 P29062 P6703 P10 106 P14309 P27910 P4514 P14834 P28385 P21073 P27701 P642 P10340 P16939 P19147 P16949 P16945 P16952 P16941 P16937 P16938 P16946 P16954 P28664  P18701 P17782 P4875 P8873
> >
> >Definitely lots of tasks recently preempted within RCU read-side
> >critical sections!
> >
> >For example, the P24858 says that the task with PID 24858 was preempted
> >(or blocked) while within an RCU read-side critical section that blocks
> >the current grace period, and that this task has not yet exited that
> >critical section.  This might be because that task has been preempted
> >throughout by a higher-priority task, that there is an infinite loop in
> >that critical section, or that the task is blocked on a spinlock that
> >is being held indefinitely by some other task (possibly because that
> >other task is being indefinitely preempted).
> 
> From what it looks in the backtrace of each task is that it got
> preempted within the RCU read section (once hackbach is done, everything
> gets back to normal). Nothing loops indefinitely. Once hackbench is done
> with sending messages, everything gets back to normal. I'm just not sure
> if everything in terms of RCU works okay on -RT or of something is
> missing. But as you suggested that it is rcurc/$CPU job to do this, I
> will check if this process runsâ

This does indeed sound like what RCU_BOOST is supposed to do.

Hmmm...  There was a race condition that Lai Jianshan provided a
patch for, which is in 3.19:

abaf3f9d275b (rcu: Revert "Allow post-unlock reference for rt_mutex" to avoid priority-inversion)

This patch fixed a priority-inversion problem, so might well fix your
RCU CPU stall-warning problem.

> >> | (detected by 0, t=5252 jiffies, g=2385, c=2384, q=41180)
> >> |hackbench       R  running task        0 24858   1995 0x00000000
> >> | ffff880058d9bb18 0000000000000082 0000000000012e40 ffff880058d9bfd8
> >> | 0000000000012e40 ffff8804226627c0 ffff880058d9bb38 ffff880058d9bfd8
> >> | 0000000000000001 ffff880058d5ccf8 0000000000000292 0000000000000002
> >> |Call Trace:
> >> | [<ffffffff81568e9f>] preempt_schedule+0x3f/0x60
> >> | [<ffffffff812d84d7>] ___preempt_schedule+0x35/0x67
> >> | [<ffffffff8156b725>] ? _raw_spin_unlock_irqrestore+0x25/0x30
> >> | [<ffffffff81088503>] try_to_wake_up+0x63/0x2f0
> >> | [<ffffffff8108881d>] default_wake_function+0xd/0x10
> >> | [<ffffffff8109b011>] autoremove_wake_function+0x11/0x40
> >> | [<ffffffff8109aa05>] __wake_up_common+0x55/0x90
> >> | [<ffffffff8109afd3>] __wake_up_sync_key+0x43/0x60
> >> | [<ffffffff8145c95e>] sock_def_readable+0x3e/0x70
> >> | [<ffffffff8150f9d1>] unix_stream_sendmsg+0x211/0x470
> >> | [<ffffffff81458b48>] sock_aio_write+0xf8/0x120
> >> | [<ffffffff8109ec19>] ? rt_up_read+0x19/0x20
> >> | [<ffffffff8116e855>] do_sync_write+0x55/0x90
> >> | [<ffffffff8116f435>] vfs_write+0x175/0x1f0
> >> | [<ffffffff8116fde4>] SyS_write+0x44/0xb0
> >> | [<ffffffff8156c1ed>] system_call_fastpath+0x16/0x1b
> >> 
> >> The other processes look more or less the same. I have the full splat
> >> here [0]. My understanding is that sock_def_readable() does a
> >> rcu_read_lock() which forbids a grace period. Since there are many
> >> processes (preempted) in this section, the grace period never starts
> >> since it never blocks new readers from getting into a read critical
> >> section.
> >> Is my understanding correct so far? Is it likely that -RT does not do
> >> something correctly to forbid such a situation or is it more or less
> >> "expected" ?
> >> 
> >> [0] https://breakpoint.cc/rt-rcu-stall.txt
> >
> >If you don't have CONFIG_RCU_BOOST=y with priority set above that of
> >any CPU-bound components of your workload, then yes, this is expected
> >behavior.  The priority boosting will be inherited, so that this covers
> >the case of indefinite preemption.  Of course, this won't help with
> >infinite loops.
> 
> I have
>  CONFIG_RCU_BOOST=y
>  CONFIG_RCU_BOOST_PRIO=6
> 
> and hackbach is a normal user process. How does this suppose to work?

Even the default CONFIG_RCU_BOOST_PRIO=1 should help in that case.

>                                                                       I
> have five tasks preempted in RCU read section. Then rcuc gets scheduled
> somehow. It boosts the five tasks one after the other until they get out
> of the RCU section. And since nothing has a higher priority than rcuc
> nothing can get on the CPU and into the RCU section again.
> What happens if the task got preempted because it was waiting for
> something? Say for a lock which was hold by another task on another CPU?
> It would then boost the task on the other CPU while this CPU would
> become "idle" and allow a task with a lower priority to get into RCU
> read section again.
> rcuc gets on the CPU again after the 5 tasks left the critical section
> but a new make it into the section. Is it possible and if so what would
> it do then?

If there is locking involved, the normal lock-based priority boosting
should fix things up.  This works because RCU's priority-boosting
procedure uses an rt_mutex.  Of course, if there is something like
a trylock()-loop between the boosted RCU read-side critical section,
then that will defeat priority boosting, just as it would with only
locks and no RCU involvement.

The rcuc task boosts only those tasks that are blocking the current
grace period.  If new tasks are preempted, they won't be boosted
unless their RCU read-side critical section started before the current
grace period started.  So tasks that entered their RCU read-side
critical section after the start of the current grace period won't
hold things up, at least not until the next grace period.

> >That said, someone was having trouble with the rcuoc kthreads that are
> >generated by CONFIG_RCU_BOOST=y when running NO_HZ_FULL.  So there might
> >need to be different kernel configs for different workloads.  Or we might
> >need some way to enable and disable boosting at boot time, I suppose.
> 
> I think I saw this with CONFIG_NO_HZ_IDLE, too.

Could you please try Lai Jiangshan's commit (abaf3f9d275b)?

							Thanx, Paul

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/