Skip to contentSkip to conversation

tick/broadcast: Do not set oneshot_mask except was_periodic was true

PATCH Superseded Victor Hassan 2 participants 8 messages Mar 28, 2023 – Apr 11, 2023 Nohz, dynticks support

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 tick/broadcast: Do not set oneshot_mask except was_periodic was true

If a broadcast timer is registered after the system switched to oneshot mode, a hang_task err could occur like that:

INFO: task kworker/u15:0:7 blocked for more than 120 seconds.
      Tainted: G            E     5.15.41-android13-8-00002-xxx #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:kworker/u16:0   state:D stack: 9808 pid:  7 ppid: 2 flags:0x00000008
Workqueue: events_unbound deferred_probe_work_func.cfi_jt
Call trace:
 __switch_to+0y240/0x490
 __schedule+0x620/0xafc
 schedule+0x110/0x204
 schedule_hrtimeout_range_clock+0x9c/0x118
 usleep_range_state+0x150/0x1ac
 _regulator_do_enable+0x528/0x878
 set_machine_constraints+0x6a0/0xf2c
 regulator_register+0x3ac/0x7ac
 devm_regulator_register+0xbc/0x120
 pmu_ext_regulator_probe+0xb0/0x1b4 [pmu_ext_regulator]
 platform_probe+0x70/0x194
 really_proe+0x320/0x68c
 __driver_probe_device+0x204/0x260
 driver_probe_device+0x48/0x1e0

When the new broadcast timer was registered after the system switched to oneshot mode, the broadcast timer was not used as periodic. If the oneshot mask was set incorrectly, all cores which did not enter cpu_idle state can't enter cpu_idle normally, causing the hrtimer mechanism to break.

This patch fixes the issue by moving the update action about oneshot mask to a more strict conditions. The tick_broadcast_setup_oneshot would be called in two typical condition, and they all will work.

1. tick_handle_periodic -> tick_broadcast_setup_oneshot

The origin broadcast was periodic, so it can set the oneshot_mask bits for those waiting for periodic broadcast and program the broadcast timer to fire.

2. tick_install_broadcast_device -> tick_broadcast_setup_oneshot

The origin broadcast was oneshot, so the cores which enter the cpu_idle already used the oneshot_mask bits. It is unnecessary to update the oneshot_mask.

Fixes 9c336c9935cf ("tick/broadcast: Allow late registered device to enter oneshot mode")
---
 kernel/time/tick-broadcast.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
kernel/time/tick-broadcast.c +3 -2
index 93bf2b4e47e5..fdbbba487978 100644
@@ -1041,12 +1041,13 @@ static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 		 */
 		cpumask_copy(tmpmask, tick_broadcast_mask);
 		cpumask_clear_cpu(cpu, tmpmask);
-		cpumask_or(tick_broadcast_oneshot_mask,
-			   tick_broadcast_oneshot_mask, tmpmask);
 
 		if (was_periodic && !cpumask_empty(tmpmask)) {
 			ktime_t nextevt = tick_get_next_period();
 
+			cpumask_or(tick_broadcast_oneshot_mask,
+				   tick_broadcast_oneshot_mask, tmpmask);
+
 			clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 			tick_broadcast_init_next_event(tmpmask, nextevt);
 			tick_broadcast_set_event(bc, cpu, nextevt);
Signature · 1 line
2.29.0
#1 lore raw
Quoted 63 lines · Victor Hassan

If a broadcast timer is registered after the system switched to oneshot mode, a hang_task err could occur like that:

INFO: task kworker/u15:0:7 blocked for more than 120 seconds.
       Tainted: G            E     5.15.41-android13-8-00002-xxx #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:kworker/u16:0   state:D stack: 9808 pid:  7 ppid: 2 flags:0x00000008
Workqueue: events_unbound deferred_probe_work_func.cfi_jt
Call trace:
  __switch_to+0y240/0x490
  __schedule+0x620/0xafc
  schedule+0x110/0x204
  schedule_hrtimeout_range_clock+0x9c/0x118
  usleep_range_state+0x150/0x1ac
  _regulator_do_enable+0x528/0x878
  set_machine_constraints+0x6a0/0xf2c
  regulator_register+0x3ac/0x7ac
  devm_regulator_register+0xbc/0x120
  pmu_ext_regulator_probe+0xb0/0x1b4 [pmu_ext_regulator]
  platform_probe+0x70/0x194
  really_proe+0x320/0x68c
  __driver_probe_device+0x204/0x260
  driver_probe_device+0x48/0x1e0

When the new broadcast timer was registered after the system switched to oneshot mode, the broadcast timer was not used as periodic. If the oneshot mask was set incorrectly, all cores which did not enter cpu_idle state can't enter cpu_idle normally, causing the hrtimer mechanism to break.

This patch fixes the issue by moving the update action about oneshot mask to a more strict conditions. The tick_broadcast_setup_oneshot would be called in two typical condition, and they all will work.

1. tick_handle_periodic -> tick_broadcast_setup_oneshot

The origin broadcast was periodic, so it can set the oneshot_mask bits for those waiting for periodic broadcast and program the broadcast timer to fire.

2. tick_install_broadcast_device -> tick_broadcast_setup_oneshot

The origin broadcast was oneshot, so the cores which enter the cpu_idle already used the oneshot_mask bits. It is unnecessary to update the oneshot_mask.

Fixes 9c336c9935cf ("tick/broadcast: Allow late registered device to enter oneshot mode")
---
  kernel/time/tick-broadcast.c | 5 +++--
  1 file changed, 3 insertions(+), 2 deletions(-)
kernel/time/tick-broadcast.c +0 -2
index 93bf2b4e47e5..fdbbba487978 100644
@@ -1041,12 +1041,13 @@ static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
  		 */
  		cpumask_copy(tmpmask, tick_broadcast_mask);
  		cpumask_clear_cpu(cpu, tmpmask);
-		cpumask_or(tick_broadcast_oneshot_mask,
-			   tick_broadcast_oneshot_mask, tmpmask);

kernel/time/tick-broadcast.c static void tick_broadcast_setup_oneshot(struct clock_event_device *bc) · line 1053

Quoted 10 lines
kernel/time/tick-broadcast.c +3 -0
 
  		if (was_periodic && !cpumask_empty(tmpmask)) {
  			ktime_t nextevt = tick_get_next_period();
 
+			cpumask_or(tick_broadcast_oneshot_mask,
+				   tick_broadcast_oneshot_mask, tmpmask);
+
  			clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
  			tick_broadcast_init_next_event(tmpmask, nextevt);
  			tick_broadcast_set_event(bc, cpu, nextevt);

Friendly ping.

maintainer Reviewed-by on v1 1/1
#2 lore raw
In reply to Victor Hassan
Quoted 60 lines · Victor Hassan

If a broadcast timer is registered after the system switched to oneshot mode, a hang_task err could occur like that:

INFO: task kworker/u15:0:7 blocked for more than 120 seconds.
      Tainted: G            E     5.15.41-android13-8-00002-xxx #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:kworker/u16:0   state:D stack: 9808 pid:  7 ppid: 2 flags:0x00000008
Workqueue: events_unbound deferred_probe_work_func.cfi_jt
Call trace:
 __switch_to+0y240/0x490
 __schedule+0x620/0xafc
 schedule+0x110/0x204
 schedule_hrtimeout_range_clock+0x9c/0x118
 usleep_range_state+0x150/0x1ac
 _regulator_do_enable+0x528/0x878
 set_machine_constraints+0x6a0/0xf2c
 regulator_register+0x3ac/0x7ac
 devm_regulator_register+0xbc/0x120
 pmu_ext_regulator_probe+0xb0/0x1b4 [pmu_ext_regulator]
 platform_probe+0x70/0x194
 really_proe+0x320/0x68c
 __driver_probe_device+0x204/0x260
 driver_probe_device+0x48/0x1e0

When the new broadcast timer was registered after the system switched to oneshot mode, the broadcast timer was not used as periodic. If the oneshot mask was set incorrectly, all cores which did not enter cpu_idle state can't enter cpu_idle normally, causing the hrtimer mechanism to break.

This patch fixes the issue by moving the update action about oneshot mask to a more strict conditions. The tick_broadcast_setup_oneshot would be called in two typical condition, and they all will work.

1. tick_handle_periodic -> tick_broadcast_setup_oneshot

The origin broadcast was periodic, so it can set the oneshot_mask bits for those waiting for periodic broadcast and program the broadcast timer to fire.

2. tick_install_broadcast_device -> tick_broadcast_setup_oneshot

The origin broadcast was oneshot, so the cores which enter the cpu_idle already used the oneshot_mask bits. It is unnecessary to update the oneshot_mask.

Fixes 9c336c9935cf ("tick/broadcast: Allow late registered device to enter oneshot mode")
---
 kernel/time/tick-broadcast.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
kernel/time/tick-broadcast.c +0 -0
index 93bf2b4e47e5..fdbbba487978 100644
@@ -1041,12 +1041,13 @@ static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 		 */
 		cpumask_copy(tmpmask, tick_broadcast_mask);

kernel/time/tick-broadcast.c static void tick_broadcast_setup_oneshot(struct clock_event_device *bc) · line 1050

Quoted 10 lines
kernel/time/tick-broadcast.c +3 -2
 		cpumask_clear_cpu(cpu, tmpmask);
-		cpumask_or(tick_broadcast_oneshot_mask,
-			   tick_broadcast_oneshot_mask, tmpmask);
 
 		if (was_periodic && !cpumask_empty(tmpmask)) {
 			ktime_t nextevt = tick_get_next_period();
 
+			cpumask_or(tick_broadcast_oneshot_mask,
+				   tick_broadcast_oneshot_mask, tmpmask);
+

Good catch, it looks like one issue that can trigger is due to the resulting ignored calls to tick_broadcast_exit(). Indeed if the cpu is already in tick_broadcast_oneshot_mask then cpuidle won't call the exit.

Leading to such race:

* CPU 1 stop its tick, next event is in one hour
* CPU 0 registers new broadcast and sets CPU 1 in tick_broadcast_oneshot_mask
* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask
* CPU 1 goes to sleep
* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour, program the broadcast to that deadline
* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy
* CPU 1 don't call tick_broadcast_exit and thus don't remove itself from tick_broadcast_oneshot_mask
* CPU 1 re-enters in cpuidle_enter_state(), tick_broadcast_enter() is again ignored so the new timer isn't propagated to the broadcast.
* CPU 1 goes to sleep and won't be woken before one hour.

Thanks.

Quoted 6 lines
kernel/time/tick-broadcast.c +0 -0
 			clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 			tick_broadcast_init_next_event(tmpmask, nextevt);
 			tick_broadcast_set_event(bc, cpu, nextevt);
Signature · 1 line
2.29.0
#3 lore raw
Quoted 68 lines · Frederic Weisbecker
Quoted 67 lines · Victor Hassan

If a broadcast timer is registered after the system switched to oneshot mode, a hang_task err could occur like that:

INFO: task kworker/u15:0:7 blocked for more than 120 seconds.
       Tainted: G            E     5.15.41-android13-8-00002-xxx #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:kworker/u16:0   state:D stack: 9808 pid:  7 ppid: 2 flags:0x00000008
Workqueue: events_unbound deferred_probe_work_func.cfi_jt
Call trace:
  __switch_to+0y240/0x490
  __schedule+0x620/0xafc
  schedule+0x110/0x204
  schedule_hrtimeout_range_clock+0x9c/0x118
  usleep_range_state+0x150/0x1ac
  _regulator_do_enable+0x528/0x878
  set_machine_constraints+0x6a0/0xf2c
  regulator_register+0x3ac/0x7ac
  devm_regulator_register+0xbc/0x120
  pmu_ext_regulator_probe+0xb0/0x1b4 [pmu_ext_regulator]
  platform_probe+0x70/0x194
  really_proe+0x320/0x68c
  __driver_probe_device+0x204/0x260
  driver_probe_device+0x48/0x1e0

When the new broadcast timer was registered after the system switched to oneshot mode, the broadcast timer was not used as periodic. If the oneshot mask was set incorrectly, all cores which did not enter cpu_idle state can't enter cpu_idle normally, causing the hrtimer mechanism to break.

This patch fixes the issue by moving the update action about oneshot mask to a more strict conditions. The tick_broadcast_setup_oneshot would be called in two typical condition, and they all will work.

1. tick_handle_periodic -> tick_broadcast_setup_oneshot

The origin broadcast was periodic, so it can set the oneshot_mask bits for those waiting for periodic broadcast and program the broadcast timer to fire.

2. tick_install_broadcast_device -> tick_broadcast_setup_oneshot

The origin broadcast was oneshot, so the cores which enter the cpu_idle already used the oneshot_mask bits. It is unnecessary to update the oneshot_mask.

Fixes 9c336c9935cf ("tick/broadcast: Allow late registered device to enter oneshot mode")
---
  kernel/time/tick-broadcast.c | 5 +++--
  1 file changed, 3 insertions(+), 2 deletions(-)
kernel/time/tick-broadcast.c +0 -2
index 93bf2b4e47e5..fdbbba487978 100644
@@ -1041,12 +1041,13 @@ static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
  		 */
  		cpumask_copy(tmpmask, tick_broadcast_mask);
  		cpumask_clear_cpu(cpu, tmpmask);
-		cpumask_or(tick_broadcast_oneshot_mask,
-			   tick_broadcast_oneshot_mask, tmpmask);
 
  		if (was_periodic && !cpumask_empty(tmpmask)) {
  			ktime_t nextevt = tick_get_next_period();
 
Quoted 7 lines

kernel/time/tick-broadcast.c static void tick_broadcast_setup_oneshot(struct clock_event_device *bc) · line 1050

Quoted 3 lines
kernel/time/tick-broadcast.c +3 -0
+			cpumask_or(tick_broadcast_oneshot_mask,
+				   tick_broadcast_oneshot_mask, tmpmask);
+

Good catch, it looks like one issue that can trigger is due to the resulting ignored calls to tick_broadcast_exit(). Indeed if the cpu is already in tick_broadcast_oneshot_mask then cpuidle won't call the exit

Hi Frederic,
Thanks for the review. Still I have a few questions and would like to discuss them further.

Quoted 7 lines

Leading to such race:

* CPU 1 stop its tick, next event is in one hour
* CPU 0 registers new broadcast and sets CPU 1 in tick_broadcast_oneshot_mask
* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask

Yes.

Quoted 6 lines

* CPU 1 goes to sleep
* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour, program the broadcast to that deadline
* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy
* CPU 1 don't call tick_broadcast_exit and thus don't remove itself from tick_broadcast_oneshot_mask

I'm not sure about this... Actually, I believe CPU 1 *will* call tick_broadcast_exit in this condition because I cannot find a limitation on this execution path.

Quoted 19 lines

* CPU 1 re-enters in cpuidle_enter_state(), tick_broadcast_enter() is again ignored so the new timer isn't propagated to the broadcast.
* CPU 1 goes to sleep and won't be woken before one hour.

Thanks.

Quoted 6 lines
kernel/time/tick-broadcast.c +0 -0
  			clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
  			tick_broadcast_init_next_event(tmpmask, nextevt);
  			tick_broadcast_set_event(bc, cpu, nextevt);
Signature · 1 line
2.29.0
#4 lore raw
Quoted 17 lines · Victor Hassan
Quoted 7 lines

Leading to such race:

* CPU 1 stop its tick, next event is in one hour
* CPU 0 registers new broadcast and sets CPU 1 in tick_broadcast_oneshot_mask
* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask

Yes.

Quoted 6 lines

* CPU 1 goes to sleep
* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour, program the broadcast to that deadline
* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy
* CPU 1 don't call tick_broadcast_exit and thus don't remove itself from tick_broadcast_oneshot_mask

Quoted 3 lines

I'm not sure about this... Actually, I believe CPU 1 *will* call tick_broadcast_exit in this condition because I cannot find a limitation on this execution path.

You're right, what I wrote doesn't make sense. Let me try again:

* CPU 1 stop its tick, next event is in one hour. It calls tick_broadcast_enter() and goes to sleep.

* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy (note it's not yet actually programmed in the tick device)

* CPU 1 call tick_broadcast_exit().

* CPU 0 registers new broadcast device and sets CPU 1 in tick_broadcast_oneshot_mask

* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour (because the recently enqueued timer for CPU 1 hasn't been programmed yet), so it programs the broadcast to that 1 hour deadline.

* CPU 1 runs tick_nohz_idle_stop_tick() which eventually writes and program dev->next_event to next jiffy

* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask, so the dev->next_event change isn't propagated to broadcast.

* CPU 1 goes to sleep for 1 hour.

Does it make more sense? There might be more simple scenario of course.

Thanks.

#5 lore raw
Quoted 38 lines · Frederic Weisbecker
Quoted 17 lines · Victor Hassan
Quoted 7 lines

Leading to such race:

* CPU 1 stop its tick, next event is in one hour
* CPU 0 registers new broadcast and sets CPU 1 in tick_broadcast_oneshot_mask
* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask

Yes.

Quoted 6 lines

* CPU 1 goes to sleep
* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour, program the broadcast to that deadline
* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy
* CPU 1 don't call tick_broadcast_exit and thus don't remove itself from tick_broadcast_oneshot_mask

Quoted 3 lines

I'm not sure about this... Actually, I believe CPU 1 *will* call tick_broadcast_exit in this condition because I cannot find a limitation on this execution path.

You're right, what I wrote doesn't make sense. Let me try again:

* CPU 1 stop its tick, next event is in one hour. It calls tick_broadcast_enter() and goes to sleep.

* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy (note it's not yet actually programmed in the tick device)

* CPU 1 call tick_broadcast_exit().

* CPU 0 registers new broadcast device and sets CPU 1 in tick_broadcast_oneshot_mask

* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour (because the recently enqueued timer for CPU 1 hasn't been programmed yet), so it programs the broadcast to that 1 hour deadline.

Quoted 8 lines

* CPU 1 runs tick_nohz_idle_stop_tick() which eventually writes and program dev->next_event to next jiffy

* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask, so the dev->next_event change isn't propagated to broadcast.

* CPU 1 goes to sleep for 1 hour.

Hi Frederic,
Yes, I think that make sense :)

Quoted 4 lines

Does it make more sense? There might be more simple scenario of course.

Thanks.

#6 lore raw
Quoted 45 lines · Victor Hassan
Quoted 42 lines · Frederic Weisbecker
Quoted 17 lines · Victor Hassan
Quoted 7 lines

Leading to such race:

* CPU 1 stop its tick, next event is in one hour
* CPU 0 registers new broadcast and sets CPU 1 in tick_broadcast_oneshot_mask
* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask

Yes.

Quoted 6 lines

* CPU 1 goes to sleep
* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour, program the broadcast to that deadline
* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy
* CPU 1 don't call tick_broadcast_exit and thus don't remove itself from tick_broadcast_oneshot_mask

Quoted 3 lines

I'm not sure about this... Actually, I believe CPU 1 *will* call tick_broadcast_exit in this condition because I cannot find a limitation on this execution path.

You're right, what I wrote doesn't make sense. Let me try again:

* CPU 1 stop its tick, next event is in one hour. It calls tick_broadcast_enter() and goes to sleep.
* CPU 1 gets an interrupt that enqueues a new timer expiring in the next jiffy (note it's not yet actually programmed in the tick device)
* CPU 1 call tick_broadcast_exit().

* CPU 0 registers new broadcast device and sets CPU 1 in tick_broadcast_oneshot_mask

* CPU 0 runs the broadcast callback, sees that the next timer for CPU 1 is in one hour (because the recently enqueued timer for CPU 1 hasn't been programmed yet), so it programs the broadcast to that 1 hour deadline.

* CPU 1 runs tick_nohz_idle_stop_tick() which eventually writes and program dev->next_event to next jiffy
* CPU 1 runs into cpuidle_enter_state(), and tick_broadcast_enter() is ignored because the CPU is already in tick_broadcast_oneshot_mask, so the dev->next_event change isn't propagated to broadcast.

Quoted 4 lines
Quoted 1 line

* CPU 1 goes to sleep for 1 hour.

Hi Frederic,
  Yes, I think that make sense :)

Hi Frederic,
If we have reached a consensus, may I add "Reviewed-by: Frederic"
in the next patch?

Quoted 7 lines
Quoted 6 lines

Does it make more sense? There might be more simple scenario of course.

Thanks.

#7 lore raw
Quoted 3 lines · Victor Hassan

Hi Frederic,
If we have reached a consensus, may I add "Reviewed-by: Frederic" in the
next patch?

Sure and perhaps add the racing scenario I proposed to clarify the changelog.

Thanks.