Messages in this thread Patch in this message | | | Date | Tue, 25 Jan 2022 10:53:38 -0800 | From | "Paul E. McKenney" <> | Subject | Re: Possible reproduction of CSD locking issue |
| |
On Tue, Jan 25, 2022 at 12:27:37PM -0600, Corey Minyard wrote: > We have a customer that had been seeing CSD lock issues on a Centos 7 > kernel (unfortunately). I couldn't find anything or any kernel changes > that might fix it, so I was consdering it was the CSD locking issue you > have been chasing for a while. > > So I backported the debug patches. And of course, they stopped seeing > the issue, at least as much, and they had trouble with the extra CPU > time the debug code took. But they just reproduced it. Here are the > logs: > > Jan 23 23:39:43 worker0 kernel: [285737.522743] csd: Detected non-responsive CSD lock (#1) on CPU#3, waiting 5000000042 ns for CPU#55 flush_tlb_func+0x0/0xb0(0xffff8e0b3e2afbe8). > Jan 23 23:39:43 worker0 kernel: [285737.522744] csd: CSD lock (#1) unresponsive. > Jan 23 23:39:43 worker0 kernel: [285737.522747] csd: cnt(0000000): 0000->0000 queue > Jan 23 23:39:43 worker0 kernel: [285737.522748] csd: cnt(0000001): ffff->0037 idle > Jan 23 23:39:43 worker0 kernel: [285737.522749] csd: cnt(63d8dd8): 0003->0037 ipi > Jan 23 23:39:43 worker0 kernel: [285737.522750] csd: cnt(63d8dd9): 0003->0037 ping > Jan 23 23:39:43 worker0 kernel: [285737.522750] csd: cnt(63d8dda): 0003->ffff pinged > Jan 23 23:39:43 worker0 kernel: [285737.522751] csd: cnt(63d8dea): 0035->0037 pinged > Jan 23 23:39:43 worker0 kernel: [285737.522752] csd: cnt(63d8deb): ffff->0037 gotipi > Jan 23 23:39:43 worker0 kernel: [285737.522752] csd: cnt(63d8dec): ffff->0037 handle > Jan 23 23:39:43 worker0 kernel: [285737.522753] csd: cnt(63d8ded): ffff->0037 dequeue (src CPU 0 == empty) > Jan 23 23:39:43 worker0 kernel: [285737.522754] csd: cnt(63d8dee): ffff->0037 hdlend (src CPU 0 == early) > Jan 23 23:39:43 worker0 kernel: [285737.522754] csd: cnt(63d8e1f): 0003->0037 queue > Jan 23 23:39:43 worker0 kernel: [285737.522755] csd: cnt(63d8e20): 0003->0037 ipi > Jan 23 23:39:43 worker0 kernel: [285737.522756] csd: cnt(63d8e21): 0003->0037 ping > Jan 23 23:39:43 worker0 kernel: [285737.522756] csd: cnt(63d8e22): 0003->0037 queue > Jan 23 23:39:43 worker0 kernel: [285737.522757] csd: cnt(63d8e23): 0003->0037 noipi > Jan 23 23:39:43 worker0 kernel: [285737.522757] csd: cnt now: 63fe4cd > Jan 23 23:39:43 worker0 kernel: [285737.522758] Task dump for CPU 55: > Jan 23 23:39:43 worker0 kernel: [285737.522761] kubelet R running task 0 277695 1 0x00080000 > Jan 23 23:39:43 worker0 kernel: [285737.522761] Call Trace: > Jan 23 23:39:43 worker0 kernel: [285737.522769] [<ffffffff84376b6a>] ? __schedule+0x46a/0x990 > Jan 23 23:39:43 worker0 kernel: [285737.522774] [<ffffffff83db6353>] ? context_tracking_user_enter+0x13/0x20 > Jan 23 23:39:43 worker0 kernel: [285737.522776] [<ffffffff843775b5>] ? schedule_user+0x45/0x50 > Jan 23 23:39:43 worker0 kernel: [285737.522779] [<ffffffff8437b518>] ? retint_careful+0x16/0x34
Long-running interrupt handler, maybe? Or am I misinterpreting this stack trace?
> Jan 23 23:39:43 worker0 kernel: [285737.522780] csd: Re-sending CSD lock (#1) IPI from CPU#03 to CPU#55 > Jan 23 23:39:43 worker0 kernel: [285737.522788] CPU: 3 PID: 54671 Comm: runc:[2:INIT] Kdump: loaded Tainted: G OE ------------ T 3.10.0-1062.12.1.rt56.1042.mvista.test.14.el7.x86_64 #1 > Jan 23 23:39:43 worker0 kernel: [285737.522789] Hardware name: Dell Inc. PowerEdge R740/0YWR7D, BIOS 2.9.4 11/06/2020 > Jan 23 23:39:43 worker0 kernel: [285737.522789] Call Trace: > Jan 23 23:39:43 worker0 kernel: [285737.522793] [<ffffffff843718ba>] dump_stack+0x19/0x1b > Jan 23 23:39:43 worker0 kernel: [285737.522798] [<ffffffff83d0bcd8>] __csd_lock_wait+0x1a8/0x2a0 > Jan 23 23:39:43 worker0 kernel: [285737.522800] [<ffffffff83c6d870>] ? leave_mm+0x120/0x120 > Jan 23 23:39:43 worker0 kernel: [285737.522802] [<ffffffff83d0bfa4>] smp_call_function_single+0xc4/0x1b0 > Jan 23 23:39:43 worker0 kernel: [285737.522804] [<ffffffff83c6d870>] ? leave_mm+0x120/0x120 > Jan 23 23:39:43 worker0 kernel: [285737.522809] [<ffffffff83e2684b>] ? page_counter_uncharge+0x3b/0x70 > Jan 23 23:39:43 worker0 kernel: [285737.522811] [<ffffffff83d0c614>] smp_call_function_many+0x344/0x380 > Jan 23 23:39:43 worker0 kernel: [285737.522813] [<ffffffff83c6d870>] ? leave_mm+0x120/0x120 > Jan 23 23:39:43 worker0 kernel: [285737.522816] [<ffffffff83c6da38>] native_flush_tlb_others+0xb8/0xc0 > Jan 23 23:39:43 worker0 kernel: [285737.522818] [<ffffffff83c6dc25>] flush_tlb_page+0x65/0xf0 > Jan 23 23:39:43 worker0 kernel: [285737.522821] [<ffffffff83dfdf98>] ptep_clear_flush+0x68/0xa0 > Jan 23 23:39:43 worker0 kernel: [285737.522825] [<ffffffff83de6806>] wp_page_copy.isra.83+0x3d6/0x650 > Jan 23 23:39:43 worker0 kernel: [285737.522828] [<ffffffff83de8cb4>] do_wp_page+0xb4/0x710 > Jan 23 23:39:43 worker0 kernel: [285737.522832] [<ffffffff83decbb4>] handle_mm_fault+0x884/0x1340 > Jan 23 23:39:43 worker0 kernel: [285737.522835] [<ffffffff83cd7799>] ? update_cfs_shares+0xa9/0xf0 > Jan 23 23:39:43 worker0 kernel: [285737.522839] [<ffffffff8437efc3>] __do_page_fault+0x213/0x5a0 > Jan 23 23:39:43 worker0 kernel: [285737.522841] [<ffffffff8437f385>] do_page_fault+0x35/0x90 > Jan 23 23:39:43 worker0 kernel: [285737.522842] [<ffffffff8437b728>] page_fault+0x28/0x30 > Jan 23 23:39:43 worker0 kernel: [285737.522845] csd: CSD lock (#1) got unstuck on CPU#03, CPU#55 released the lock. > > Hopefully this is the issue you are chasing and not something else. > I've been studying them to see what they mean, but I thought you might > be interested to get them asap.
Well, there have been several bugs causing these CSD lock issues, so what is one more? ;-)
More seriously, have you tried Frederic's patch? This fixes the issue described here: https://paulmck.livejournal.com/62071.html If your stack above was due to an interrupt storm rather than a long-running interrupt, this might well be the fix.
Oh, and Jürgen Groß reportedly found an issue about a year ago that could potentially be related, but I see that he is already on CC.
And, unfortunately, even more seriously, this CSD-lock diagnostic code will very likely continue to find problems, just as the infamous RCU CPU stall warnings and hard/soft lockup warnings do.
Thanx, Paul
------------------------------------------------------------------------
commit 53e87e3cdc155f20c3417b689df8d2ac88d79576 Author: Frederic Weisbecker <frederic@kernel.org> Date: Tue Oct 26 16:10:54 2021 +0200
timers/nohz: Last resort update jiffies on nohz_full IRQ entry When at least one CPU runs in nohz_full mode, a dedicated timekeeper CPU is guaranteed to stay online and to never stop its tick. Meanwhile on some rare case, the dedicated timekeeper may be running with interrupts disabled for a while, such as in stop_machine. If jiffies stop being updated, a nohz_full CPU may end up endlessly programming the next tick in the past, taking the last jiffies update monotonic timestamp as a stale base, resulting in an tick storm. Here is a scenario where it matters: 0) CPU 0 is the timekeeper and CPU 1 a nohz_full CPU. 1) A stop machine callback is queued to execute somewhere. 2) CPU 0 reaches MULTI_STOP_DISABLE_IRQ while CPU 1 is still in MULTI_STOP_PREPARE. Hence CPU 0 can't do its timekeeping duty. CPU 1 can still take IRQs. 3) CPU 1 receives an IRQ which queues a timer callback one jiffy forward. 4) On IRQ exit, CPU 1 schedules the tick one jiffy forward, taking last_jiffies_update as a base. But last_jiffies_update hasn't been updated for 2 jiffies since the timekeeper has interrupts disabled. 5) clockevents_program_event(), which relies on ktime_get(), observes that the expiration is in the past and therefore programs the min delta event on the clock. 6) The tick fires immediately, goto 3) 7) Tick storm, the nohz_full CPU is drown and takes ages to reach MULTI_STOP_DISABLE_IRQ, which is the only way out of this situation. Solve this with unconditionally updating jiffies if the value is stale on nohz_full IRQ entry. IRQs and other disturbances are expected to be rare enough on nohz_full for the unconditional call to ktime_get() to actually matter. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Paul E. McKenney <paulmck@kernel.org> Link: https://lore.kernel.org/r/20211026141055.57358-2-frederic@kernel.org
diff --git a/kernel/softirq.c b/kernel/softirq.c index 322b65d456767..41f470929e991 100644 --- a/kernel/softirq.c +++ b/kernel/softirq.c @@ -595,7 +595,8 @@ void irq_enter_rcu(void) { __irq_enter_raw(); - if (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET)) + if (tick_nohz_full_cpu(smp_processor_id()) || + (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET))) tick_irq_enter(); account_hardirq_enter(current); diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c index 6bffe5af8cb11..17a283ce2b20f 100644 --- a/kernel/time/tick-sched.c +++ b/kernel/time/tick-sched.c @@ -1375,6 +1375,13 @@ static inline void tick_nohz_irq_enter(void) now = ktime_get(); if (ts->idle_active) tick_nohz_stop_idle(ts, now); + /* + * If all CPUs are idle. We may need to update a stale jiffies value. + * Note nohz_full is a special case: a timekeeper is guaranteed to stay + * alive but it might be busy looping with interrupts disabled in some + * rare case (typically stop machine). So we must make sure we have a + * last resort. + */ if (ts->tick_stopped) tick_nohz_update_jiffies(now); }
| |