Skip to contentSkip to conversation

multiple timer interrupts a second

nohz_full/apic Luiz Capitulino 2 participants 3 messages Nov 3, 2017 – Nov 6, 2017

lore
Share & tools
lore
#0 lore raw

[CC'ing lkml this time]

Hi,

I've observed that smp_apic_timer_interrupt() is sometimes called two or more times a second on a nohz_full core which has a single task taking 100% of the core. In one of the calls, hrtimer_interrupt() runs tick_sched_timer(), but in others it doesn't call any handler. Here's an example (Linus HEAD f34157878):

<...>-1831  [008]  1060.115578: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1060.115578: funcgraph_entry:                   |    hrtimer_interrupt() {
<...>-1831  [008]  1060.115578: hrtimer_expire_entry: hrtimer=0xffff8edfefd12d60 function=tick_sched_timer now=1060079001980
<...>-1831  [008]  1060.115578: funcgraph_entry:        1.172 us   |      tick_sched_timer();
<...>-1831  [008]  1060.115580: funcgraph_exit:         1.757 us   |    }
<...>-1831  [008]  1060.115580: hrtimer_start:        hrtimer=0xffff8edfefd12d60 function=tick_sched_timer expires=1061079000000 softexpires=1061079000000
<...>-1831  [008]  1060.115581: funcgraph_exit:         3.026 us   |  }
<...>-1831  [008]  1061.115577: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1061.115577: funcgraph_entry:        0.261 us   |    hrtimer_interrupt(); <---------- NO handler called
<...>-1831  [008]  1061.115578: funcgraph_exit:         1.349 us   |  }
<...>-1831  [008]  1061.115579: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1061.115579: funcgraph_entry:                   |    hrtimer_interrupt() {
<...>-1831  [008]  1061.115579: hrtimer_expire_entry: hrtimer=0xffff8edfefd12d60 function=tick_sched_timer now=1061079001473
<...>-1831  [008]  1061.115580: funcgraph_entry:        1.413 us   |      tick_sched_timer();
<...>-1831  [008]  1061.115581: funcgraph_exit:         2.124 us   |    }
<...>-1831  [008]  1061.115582: hrtimer_start:        hrtimer=0xffff8edfefd12d60 function=tick_sched_timer expires=1062079000000 softexpires=1062079000000
<...>-1831  [008]  1061.115582: funcgraph_exit:         3.255 us   |  }

Is this expected for some reason?

I guess what's happening is that the deadline timer is firing earlier than expected. From a few dozen to a few hundreds nanoseconds earlier. When this happens, hrtimer_interrupt() skips calling the hrtimer handler (since it's early) and the apic is programmed to fire in the next microsecond.

On further research I saw that Nicolai tried to fix a very similar problem last year:

commit 1a9e4c564ab174e53ed86def922804a5ddc63e7d

Author: Nicolai Stange <[email protected]>
Date:   Thu Jul 14 17:22:54 2016 +0200

x86/timers/apic: Fix imprecise timer interrupts by eliminating TSC clockevents frequency roundoff error

I noticed the following bug/misbehavior on certain Intel systems: with a single task running on a NOHZ CPU on an Intel Haswell, I recognized that I did not only get the one expected local_timer APIC interrupt, but two per second at minimum. (!)

Maybe this issue is still present?

#1 lore raw

Hi Luiz,

[John Stultz added to CC]

Quoted 26 lines · Luiz Capitulino

[CC'ing lkml this time]

I've observed that smp_apic_timer_interrupt() is sometimes called two or more times a second on a nohz_full core which has a single task taking 100% of the core. In one of the calls, hrtimer_interrupt() runs tick_sched_timer(), but in others it doesn't call any handler. Here's an example (Linus HEAD f34157878):

<...>-1831  [008]  1060.115578: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1060.115578: funcgraph_entry:                   |    hrtimer_interrupt() {
<...>-1831  [008]  1060.115578: hrtimer_expire_entry: hrtimer=0xffff8edfefd12d60 function=tick_sched_timer now=1060079001980
<...>-1831  [008]  1060.115578: funcgraph_entry:        1.172 us   |      tick_sched_timer();
<...>-1831  [008]  1060.115580: funcgraph_exit:         1.757 us   |    }
<...>-1831  [008]  1060.115580: hrtimer_start:        hrtimer=0xffff8edfefd12d60 function=tick_sched_timer expires=1061079000000 softexpires=1061079000000
<...>-1831  [008]  1060.115581: funcgraph_exit:         3.026 us   |  }
<...>-1831  [008]  1061.115577: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1061.115577: funcgraph_entry:        0.261 us   |    hrtimer_interrupt(); <---------- NO handler called
<...>-1831  [008]  1061.115578: funcgraph_exit:         1.349 us   |  }
<...>-1831  [008]  1061.115579: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1061.115579: funcgraph_entry:                   |    hrtimer_interrupt() {
<...>-1831  [008]  1061.115579: hrtimer_expire_entry: hrtimer=0xffff8edfefd12d60 function=tick_sched_timer now=1061079001473
<...>-1831  [008]  1061.115580: funcgraph_entry:        1.413 us   |      tick_sched_timer();
<...>-1831  [008]  1061.115581: funcgraph_exit:         2.124 us   |    }
<...>-1831  [008]  1061.115582: hrtimer_start:        hrtimer=0xffff8edfefd12d60 function=tick_sched_timer expires=1062079000000 softexpires=1062079000000
<...>-1831  [008]  1061.115582: funcgraph_exit:         3.255 us   |  }
Quoted 7 lines

Is this expected for some reason?

I guess what's happening is that the deadline timer is firing earlier than expected. From a few dozen to a few hundreds nanoseconds earlier. When this happens, hrtimer_interrupt() skips calling the hrtimer handler (since it's early) and the apic is programmed to fire in the next microsecond.

Exactly.

Quoted 7 lines

On further research I saw that Nicolai tried to fix a very similar problem last year:

commit 1a9e4c564ab174e53ed86def922804a5ddc63e7d

Author: Nicolai Stange <[email protected]>
Date:   Thu Jul 14 17:22:54 2016 +0200
Quoted 6 lines

x86/timers/apic: Fix imprecise timer interrupts by eliminating TSC clockevents frequency roundoff error

I noticed the following bug/misbehavior on certain Intel systems: with a single task running on a NOHZ CPU on an Intel Haswell, I recognized that I did not only get the one expected local_timer APIC interrupt, but two per second at minimum. (!)

Note that there's also 6731b0d611a1 ("x86/timers/apic: Inform TSC deadline clockevent device about recalibration").

Quoted 1 line

Maybe this issue is still present?

Yes it is. The reason is that NTP frequency adjustments would make it into the mono clocksource's frequency but not into the clock_event_device's.

I sent a series back then [1] addressing this.

However, in the meanwhile, I changed jobs and thus, got somehow distracted. My bad :/

The result is that this got merged only up to what corresponds to [10/28] in [1].

I'll pick this up again now and try to get the rest accepted for inclusion.

Thanks,

Nicolai

[1] last time I sent this series as a whole can be found at

    http://lkml.kernel.org/r/[email protected]
#2 lore raw
Quoted 69 lines · On Sat, 04 Nov 2017 10:14:52 +0100 Nicolai Stange <[email protected]> wrote:

Hi Luiz,

[John Stultz added to CC]

Quoted 26 lines · Luiz Capitulino

[CC'ing lkml this time]

I've observed that smp_apic_timer_interrupt() is sometimes called two or more times a second on a nohz_full core which has a single task taking 100% of the core. In one of the calls, hrtimer_interrupt() runs tick_sched_timer(), but in others it doesn't call any handler. Here's an example (Linus HEAD f34157878):

<...>-1831  [008]  1060.115578: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1060.115578: funcgraph_entry:                   |    hrtimer_interrupt() {
<...>-1831  [008]  1060.115578: hrtimer_expire_entry: hrtimer=0xffff8edfefd12d60 function=tick_sched_timer now=1060079001980
<...>-1831  [008]  1060.115578: funcgraph_entry:        1.172 us   |      tick_sched_timer();
<...>-1831  [008]  1060.115580: funcgraph_exit:         1.757 us   |    }
<...>-1831  [008]  1060.115580: hrtimer_start:        hrtimer=0xffff8edfefd12d60 function=tick_sched_timer expires=1061079000000 softexpires=1061079000000
<...>-1831  [008]  1060.115581: funcgraph_exit:         3.026 us   |  }
<...>-1831  [008]  1061.115577: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1061.115577: funcgraph_entry:        0.261 us   |    hrtimer_interrupt(); <---------- NO handler called
<...>-1831  [008]  1061.115578: funcgraph_exit:         1.349 us   |  }
<...>-1831  [008]  1061.115579: funcgraph_entry:                   |  smp_apic_timer_interrupt() {
<...>-1831  [008]  1061.115579: funcgraph_entry:                   |    hrtimer_interrupt() {
<...>-1831  [008]  1061.115579: hrtimer_expire_entry: hrtimer=0xffff8edfefd12d60 function=tick_sched_timer now=1061079001473
<...>-1831  [008]  1061.115580: funcgraph_entry:        1.413 us   |      tick_sched_timer();
<...>-1831  [008]  1061.115581: funcgraph_exit:         2.124 us   |    }
<...>-1831  [008]  1061.115582: hrtimer_start:        hrtimer=0xffff8edfefd12d60 function=tick_sched_timer expires=1062079000000 softexpires=1062079000000
<...>-1831  [008]  1061.115582: funcgraph_exit:         3.255 us   |  }
Quoted 7 lines

Is this expected for some reason?

I guess what's happening is that the deadline timer is firing earlier than expected. From a few dozen to a few hundreds nanoseconds earlier. When this happens, hrtimer_interrupt() skips calling the hrtimer handler (since it's early) and the apic is programmed to fire in the next microsecond.

Exactly.

Quoted 7 lines

On further research I saw that Nicolai tried to fix a very similar problem last year:

commit 1a9e4c564ab174e53ed86def922804a5ddc63e7d

Author: Nicolai Stange <[email protected]>
Date:   Thu Jul 14 17:22:54 2016 +0200
Quoted 6 lines

x86/timers/apic: Fix imprecise timer interrupts by eliminating TSC clockevents frequency roundoff error

I noticed the following bug/misbehavior on certain Intel systems: with a single task running on a NOHZ CPU on an Intel Haswell, I recognized that I did not only get the one expected local_timer APIC interrupt, but two per second at minimum. (!)

Note that there's also 6731b0d611a1 ("x86/timers/apic: Inform TSC deadline clockevent device about recalibration").

Quoted 1 line

Maybe this issue is still present?

Yes it is. The reason is that NTP frequency adjustments would make it into the mono clocksource's frequency but not into the clock_event_device's.

I sent a series back then [1] addressing this.

Quoted 8 lines

However, in the meanwhile, I changed jobs and thus, got somehow distracted. My bad :/

The result is that this got merged only up to what corresponds to [10/28] in [1].

I'll pick this up again now and try to get the rest accepted for inclusion.

OK. If you get a new version, please include me in the CC so that I can give it a try.

Quoted 10 lines

Thanks,

Nicolai

[1] last time I sent this series as a whole can be found at

    http://lkml.kernel.org/r/[email protected]