Skip to contentSkip to conversation

Boot hangs at clocksource_done_booting on large configs

BUG Alex Thorlton 3 participants 7 messages Aug 31, 2015 – Sep 1, 2015

lore
Share & tools
lore
#0 lore raw

Hey guys,

We've been hitting issues with machines with 4096 or more logical cores hanging up at clocksource_done_booting, seemingly forever. The problem occurs somewhat intermittently, though it hits more consistently as you increase the core count. We've observed this problem on kernels starting at 3.10, all the way up through 4.2.

We initially thought there was some issue with the stop_machine code, but Thomas G. cleared that up for us with a thorough explanation of the stop_machine mechanism. After ruling out issues in stop_machine, we started investigating possible hardware related scaling issues.

I was able to hit this issue on 4.2-rc1 with our RTC disabled, to rule out any scaling issues related to multiple concurrent reads to our RTC's MMR. We start getting hung task timeouts during clocksource_done_booting:

8<----
calling clocksource_done_booting+0x0/0x42 @ 1

 INFO: task swapper/0:1 blocked for more than 480 seconds.
       Not tainted 4.2.0-rc1-nortc+ #384
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
 swapper/0       D ffff881efa933b48     0     1      0 0x00000000
  ffff881efa933b48 ffff881ef45e0000 ffffb5de7a130000 000000fdfa933b38
  ffff881efa934000 7fffffffffffffff ffff881efa933ca0 ffffb5de7a130000
  0000000000000000 ffff881efa933b68 ffffffff8158eae7 ffff881efa933b68
 Call Trace:
  [<ffffffff8158eae7>] schedule+0x37/0x80
  [<ffffffff81591707>] schedule_timeout+0x1e7/0x270
  [<ffffffff812ee9b9>] ? _find_next_bit.part.0+0x19/0x70
  [<ffffffff8158fdb0>] wait_for_completion+0xa0/0x100
  [<ffffffff8108d490>] ? wake_up_q+0x70/0x70
  [<ffffffff810f9f00>] ? cpu_stop_should_run+0x50/0x50
  [<ffffffff810fa2a0>] __stop_cpus+0x50/0x70
  [<ffffffff810b8f7d>] ? console_unlock+0x2dd/0x4a0
  [<ffffffff810f9f00>] ? cpu_stop_should_run+0x50/0x50
  [<ffffffff810fa525>] stop_cpus+0x35/0x50
  [<ffffffff810d18f0>] ? tk_setup_internals.constprop.10+0x140/0x140
  [<ffffffff810fa652>] __stop_machine+0xb2/0xe0
  [<ffffffff810d18f0>] ? tk_setup_internals.constprop.10+0x140/0x140
  [<ffffffff810d18f0>] ? tk_setup_internals.constprop.10+0x140/0x140
  [<ffffffff810fa6ae>] stop_machine+0x2e/0x50
  [<ffffffff810d1fad>] timekeeping_notify+0x2d/0x50
  [<ffffffff810d3d4b>] __clocksource_select+0xcb/0x120
  [<ffffffff81b77cfa>] ? boot_override_clock+0x4d/0x4d
  [<ffffffff81b77d2c>] clocksource_done_booting+0x32/0x42
  [<ffffffff810003a9>] do_one_initcall+0x169/0x1d0
  [<ffffffff81b5613c>] kernel_init_freeable+0x1bb/0x248
  [<ffffffff81582560>] ? rest_init+0x80/0x80
  [<ffffffff8158256e>] kernel_init+0xe/0xe0
  [<ffffffff81592c5f>] ret_from_fork+0x3f/0x70
  [<ffffffff81582560>] ? rest_init+0x80/0x80

---->8

After this I NMId the machine to get backtraces for the other CPUs. Most are stuck around here:

8<----

 CPU: 4867 PID: 19730 Comm: migration/4867 Tainted: G      D         4.2.0-rc1-nortc+ #384
 Hardware name: SGI UV2000/ROMLEY, BIOS SGI UV 2000/3000 series BIOS 01/15/2013
 task: ffffad5e798cc500 ti: ffffad5e798d4000 task.ti: ffffad5e798d4000
 RIP: 0010:[<ffffffff8105e879>]  [<ffffffff8105e879>] uv_handle_nmi+0x19/0x790
 RSP: 0000:ffffc90020156da8  EFLAGS: 00010092
 RAX: 000002fe65b95384 RBX: 000000000000002d RCX: ffffc90020162a98
 RDX: 000002fe65b95384 RSI: ffffc90020156ef8 RDI: 0000000000000001
 RBP: ffffc90020156e18 R08: 0000000000000001 R09: 0000000000000000
 R10: 0000000000000003 R11: 0000000000000020 R12: 000002fe65b95384
 R13: 0000000000000000 R14: ffffffff81a38880 R15: ffffc90020156ef8
 FS:  0000000000000000(0000) GS:ffffc90020151000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 000000000000da80 CR3: 0000000001a0b000 CR4: 00000000001406e0
 Stack:
  0000000000000000 0000000000000000 0000000000000000 0000000000000000
  0000000000000000 0000000000000000 0000000000000000 0000000000000000
  0000000000000000 000000000000002d 000002fe65b95384 0000000000000000
 Call Trace:
  <NMI>
  [<ffffffff810f9f00>] ? cpu_stop_should_run+0x50/0x50
  [<ffffffff81007869>] nmi_handle+0x79/0x100
  [<ffffffff81007873>] ? nmi_handle+0x83/0x100
  [<ffffffff810f9f00>] ? cpu_stop_should_run+0x50/0x50
  [<ffffffff81007c6c>] unknown_nmi_error+0x1c/0x90
  [<ffffffff81007e48>] default_do_nmi+0xd8/0xf0
  [<ffffffff81007ee3>] do_nmi+0x83/0xb0
  [<ffffffff815948bf>] end_repeat_nmi+0x1e/0x2e
  [<ffffffff810f9f00>] ? cpu_stop_should_run+0x50/0x50
  [<ffffffff810486c6>] ? read_hpet+0x16/0x20
  [<ffffffff810486c6>] ? read_hpet+0x16/0x20
  [<ffffffff810486c6>] ? read_hpet+0x16/0x20
  <<EOE>>
  <IRQ>
  [<ffffffff810d1b97>] ktime_get+0x37/0xb0
  [<ffffffff810d819b>] clockevents_program_event+0x4b/0x130
  [<ffffffff810d8737>] tick_handle_periodic+0x57/0x70
  [<ffffffff8103b46c>] local_apic_timer_interrupt+0x3c/0x70
  [<ffffffff815952c1>] smp_apic_timer_interrupt+0x41/0x60
  [<ffffffff815936ab>] apic_timer_interrupt+0x6b/0x70
  <EOI>
  [<ffffffff81097b41>] ? update_blocked_averages+0x381/0x860
  [<ffffffff810f9f98>] ? multi_cpu_stop+0x98/0xf0
  [<ffffffff8100272a>] ? __switch_to+0x26a/0x590
  [<ffffffff810f9f00>] ? cpu_stop_should_run+0x50/0x50
  [<ffffffff810fa1c8>] cpu_stopper_thread+0x88/0x110
  [<ffffffff8158e4e8>] ? __schedule+0x338/0x900
  [<ffffffff81085575>] smpboot_thread_fn+0x135/0x190
  [<ffffffff81085440>] ? sort_range+0x30/0x30
  [<ffffffff810820cb>] kthread+0xdb/0x100
  [<ffffffff81081ff0>] ? kthread_create_on_node+0x180/0x180
  [<ffffffff81592c5f>] ret_from_fork+0x3f/0x70
  [<ffffffff81081ff0>] ? kthread_create_on_node+0x180/0x180
 Code: 31 c0 e8 a0 c2 52 00 e9 39 ff ff ff 66 0f 1f 44 00 00 0f 1f 44 00 00 55 48 89 e5 41 57 41 56 41 55 41 54 49 89 f7 53 48 83 ec 48 <48> 8b 1d 00 f2 fa 7e 65 44 8b 2d 70 b8 fa 7e 9c 58 0f 1f 44 00

---->8

I can provide full logs if desired.

Anyways, based on my limited understanding of what's going on here, I tried disabling interrupts a bit further out in the cpu_stopper_thread code, to try and avoid that apic_timer_interrupt during the clocksource change. change_clocksource already disables interrupts, but, from the backtraces, it appears that we're taking an interrupt after the stopper thread wakes up, but before change_clocksource actually gets called.

I used this patch to try and avoid taking a apic_timer_interrupt at any point during the clocksource change.:

8<---

kernel/stop_machine.c +5 -0
index fd643d8..8502521 100644
@@ -417,8 +417,11 @@ static void cpu_stopper_thread(unsigned int cpu)
 {
        struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
        struct cpu_stop_work *work;
+       unsigned long flags;
        int ret;
 
+       local_irq_save(flags);
+
 repeat:
        work = NULL;
        spin_lock_irq(&stopper->lock);
@@ -452,6 +455,8 @@ repeat:
                cpu_stop_signal_done(done, true);
                goto repeat;
        }
+
+       local_irq_restore(flags);
 }
 
 extern void sched_set_stop_task(int cpu, struct task_struct *stop);

--->8

This was a bit of a shot in the dark, but it has worked on every kernel and every machine we've tried it on so far.

I'm hoping to get some input from the experts in this area, first of all, on whether the problem I'm seeing is actually what I think it is, and, if so, if I've solved it in the correct way.

Any input is greatly appreciated!

- Alex

#2 lore raw
In reply to Alex Thorlton

On Mon, Aug 31, 2015 at 01:04:33PM -0500, Alex Thorlton wrote:
q

Quoted 17 lines
kernel/stop_machine.c +3 -0
index fd643d8..8502521 100644
@@ -417,8 +417,11 @@ static void cpu_stopper_thread(unsigned int cpu)
 {
        struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
        struct cpu_stop_work *work;
+       unsigned long flags;
        int ret;
 
+       local_irq_save(flags);
+
 repeat:
        work = NULL;
        spin_lock_irq(&stopper->lock);

kernel/stop_machine.c repeat: · line 461

Quoted 7 lines
kernel/stop_machine.c +2 -0
@@ -452,6 +455,8 @@ repeat:
                cpu_stop_signal_done(done, true);
                goto repeat;
        }
+
+       local_irq_restore(flags);
 }
 

So I should probably just go sleep and not say anything.. _but_ *confused*.

That local_irq_save() will disable IRQs over:

	work = NULL;

But that is _all_. The spin_unlock_irq() will re-enable IRQs, after which things will run as usual.

That local_irq_restore() is a total NOP, IRQs are guaranteed enabled at the irq_local_save() (otherwise lockdep would've complained about spin_unlock_irq() unconditionally enabling them) and by the time we get to the restore that same unlock_irq will have enabled them already.

#3 lore raw
Quoted 34 lines · Peter Zijlstra

On Mon, Aug 31, 2015 at 01:04:33PM -0500, Alex Thorlton wrote:
q

Quoted 17 lines
kernel/stop_machine.c +3 -0
index fd643d8..8502521 100644
@@ -417,8 +417,11 @@ static void cpu_stopper_thread(unsigned int cpu)
 {
        struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
        struct cpu_stop_work *work;
+       unsigned long flags;
        int ret;
 
+       local_irq_save(flags);
+
 repeat:
        work = NULL;
        spin_lock_irq(&stopper->lock);

kernel/stop_machine.c repeat: · line 461

Quoted 7 lines
kernel/stop_machine.c +2 -0
@@ -452,6 +455,8 @@ repeat:
                cpu_stop_signal_done(done, true);
                goto repeat;
        }
+
+       local_irq_restore(flags);
 }
 

So I should probably just go sleep and not say anything.. _but_ *confused*.

That local_irq_save() will disable IRQs over:

	work = NULL;
Quoted 7 lines

But that is _all_. The spin_unlock_irq() will re-enable IRQs, after which things will run as usual.

That local_irq_restore() is a total NOP, IRQs are guaranteed enabled at the irq_local_save() (otherwise lockdep would've complained about spin_unlock_irq() unconditionally enabling them) and by the time we get to the restore that same unlock_irq will have enabled them already.

Ahh, right. Well that code is worthless then :)

Either way though, I guess that means that slight change just fudged the timing enough in that area to avoid the lockup we're seeing. Ignoring my useless code change, does anything else jump out at you as interesting, here?

#4 lore raw
In reply to Alex Thorlton
Quoted 3 lines · Alex Thorlton

I was able to hit this issue on 4.2-rc1 with our RTC disabled, to rule out any scaling issues related to multiple concurrent reads to our RTC's MMR.

And to rule out scaling issues you replaced the RTC MMR with HPET. Not a very good choice:

HPET does not scale either. It's uncached memory mapped I/O. See below.

Quoted 3 lines

I'm hoping to get some input from the experts in this area, first of all, on whether the problem I'm seeing is actually what I think it is, and, if so, if I've solved it in the correct way.

I fear both the analysis and the solution is wrong.

Up to the point where the actual clocksource change happens there is no reason why timer interrupts should not happen. And the code which actually changes the clocksource is definitely called with interrupts disabled. When that function returns the new clocksource is fully functional and interrupts can happen again.

Now looking at your backtraces. Most CPUs are in the migration thread and a few (3073,3078,3079,3082) are in the idle task.

Quoted 1 line

From the trace artifacts (? read_hpet) it looks like the clock source

change has been done and the cpus are on the way back from stop machine.

But they are obviously held off by something. And that something looks like the timekeeper sequence lock. Too bad, that we don't have a backtrace for CPU0 in the log.

I really wonder how a machine that large works with HPET as clocksource at all. hpet_read() is uncached memory mapped IO which takes thousands of CPU cycles. Last time I looked it was around 1us. Let's take that number to do some math.

If all CPUs do that access at the same time, then it takes NCPUS microseconds to complete if the memory mapped I/O scheduling is completely fair, which I doubt. So with 4k CPUs thats whopping 4.096ms and it gets worse if you go larger. That's more than a tick with HZ=250.

I'm quite sure that you are staring at the HPET scalability bottleneck and not at some actual kernel bug.

Your patch shifts some timing around so the issue does not happen, but that's certainly not a solution.

Thanks,

tglx

#5 lore raw
Quoted 10 lines · Thomas Gleixner
Quoted 3 lines · Alex Thorlton

I was able to hit this issue on 4.2-rc1 with our RTC disabled, to rule out any scaling issues related to multiple concurrent reads to our RTC's MMR.

And to rule out scaling issues you replaced the RTC MMR with HPET. Not a very good choice:

HPET does not scale either. It's uncached memory mapped I/O. See below.

Ahhh. I wondered if there might be an issue there. I can say that I've seen it hit a very similar issue with some other possible solutions that we've tried, but this was the most recent, fairly unmodified kernel that I've played with, so I went with this.

Quoted 12 lines
Quoted 3 lines

I'm hoping to get some input from the experts in this area, first of all, on whether the problem I'm seeing is actually what I think it is, and, if so, if I've solved it in the correct way.

I fear both the analysis and the solution is wrong.

Up to the point where the actual clocksource change happens there is no reason why timer interrupts should not happen. And the code which actually changes the clocksource is definitely called with interrupts disabled. When that function returns the new clocksource is fully functional and interrupts can happen again.

Quoted 6 lines

Now looking at your backtraces. Most CPUs are in the migration thread and a few (3073,3078,3079,3082) are in the idle task.

From the trace artifacts (? read_hpet) it looks like the clock source change has been done and the cpus are on the way back from stop machine.

That's good to note. I'll keep that in mind as I continue gathering info.

Quoted 2 lines

But they are obviously held off by something. And that something looks like the timekeeper sequence lock.

Ok. IIRC, that lock prevents reads to the clocksource until all necessary write operations have finished, correct? i.e. tk_core.seq?

I looked at that as being a possible place to get stuck in ktime_get:

        do {
                seq = read_seqcount_begin(&tk_core.seq);
                base = tk->tkr_mono.base;
                nsecs = timekeeping_get_ns(&tk->tkr_mono);

        } while (read_seqcount_retry(&tk_core.seq, seq));

But I don't have enough evidence to point there for sure.

Quoted 1 line

Too bad, that we don't have a backtrace for CPU0 in the log.

I think that may have gotten swallowed up with all the lost printks that you'll notice in there. Does the hung task backtrace not cover that though?

Quoted 5 lines

I really wonder how a machine that large works with HPET as clocksource at all. hpet_read() is uncached memory mapped IO which takes thousands of CPU cycles. Last time I looked it was around 1us. Let's take that number to do some math.

Quoted 8 lines

If all CPUs do that access at the same time, then it takes NCPUS microseconds to complete if the memory mapped I/O scheduling is completely fair, which I doubt. So with 4k CPUs thats whopping 4.096ms and it gets worse if you go larger. That's more than a tick with HZ=250.

I'm quite sure that you are staring at the HPET scalability bottleneck and not at some actual kernel bug.

That could well be the case.

Quoted 2 lines

Your patch shifts some timing around so the issue does not happen, but that's certainly not a solution.

I was fairly sure of that, but it's good to know that you agree.

I'll try and get a backtrace from a completely unmodified kernel here in the next few days. Unfortunately, as mentioned, we only hit the issue intermittently. I was able to reproduce pretty regularly on kernels from 4.0 and before, but it seems to be less common on newer kernels, though we definitely still manage to hit it from time to time.

Thanks for all the help!

- Alex

#6 lore raw
Quoted 9 lines · Alex Thorlton
Quoted 2 lines · Thomas Gleixner

But they are obviously held off by something. And that something looks like the timekeeper sequence lock.

Ok. IIRC, that lock prevents reads to the clocksource until all necessary write operations have finished, correct? i.e. tk_core.seq?

I looked at that as being a possible place to get stuck in ktime_get:

Quoted 8 lines
        do {
                seq = read_seqcount_begin(&tk_core.seq);
                base = tk->tkr_mono.base;
                nsecs = timekeeping_get_ns(&tk->tkr_mono);

        } while (read_seqcount_retry(&tk_core.seq, seq));

But I don't have enough evidence to point there for sure.

Right and it might be a red herring. If the access to HPET takes long enough for each CPU then you get into a situation where you never get out of those timer interrupts anymore.

Quoted 10 lines
Quoted 2 lines

Your patch shifts some timing around so the issue does not happen, but that's certainly not a solution.

I was fairly sure of that, but it's good to know that you agree.

I'll try and get a backtrace from a completely unmodified kernel here in the next few days. Unfortunately, as mentioned, we only hit the issue intermittently. I was able to reproduce pretty regularly on kernels from 4.0 and before, but it seems to be less common on newer kernels, though we definitely still manage to hit it from time to time.

Well, I doubt that backtraces will tell us much. Some moderate form of tracing (a few tracepoints enabled) and a trace_dump(DUMP_ALL); call before or instead of the NMI backtrace stuff might give us the really interesting information. i.e. what the cpus have been doing for the last second or such. The backtraces are just momentary snapshots and not necessarily helpful.

You might have to hack the trace dump code to limit the output to a few cpus if you don't want to wait a day or two until the information has spilled out over serial :)

Thanks,

tglx