sched/stat: correct the task blocking state
PATCH v2 1/1 sched/stat: correct the task blocking state
From: Alex Shi <[email protected]>
The commit 80ed87c8a9ca ("sched/wait: Introduce TASK_NOLOAD and TASK_IDLE") stopped the idle kthreads from contributing to the load average. However, the idle state time still contributes to the blocked state time instead of the sleep time. As a result, we cannot determine if a task is stopped due to some reasons or if it is idle by its own initiative.
Distinguishing between these two states would make the system state clearer and provide us with an opportunity to use the 'D' state of a task as an indicator of latency issues.
Originally-from: Curu Wong <[email protected]>
To: [email protected]
To: Valentin Schneider <[email protected]>
To: Daniel Bristot de Oliveira <[email protected]>
To: Mel Gorman <[email protected]>
To: Ben Segall <[email protected]>
To: Steven Rostedt <[email protected]>
To: Dietmar Eggemann <[email protected]>
To: Vincent Guittot <[email protected]>
To: Juri Lelli <[email protected]>
To: Peter Zijlstra <[email protected]>
To: Ingo Molnar <[email protected]>
--- include/linux/sched.h | 6 ++++++ kernel/sched/deadline.c | 5 +++-- kernel/sched/fair.c | 5 +++-- kernel/sched/rt.c | 5 +++-- 4 files changed, 15 insertions(+), 6 deletions(-)
include/linux/sched.h +6 -0
@@ -140,6 +140,12 @@ struct user_event_mm;#define is_special_task_state (state) \ ((state) & (__TASK_STOPPED | __TASK_TRACED | TASK_PARKED | TASK_DEAD)) +/* blocked task is UNINTERRUPTIBLE but not NOLOAD */ +#define is_blocked_state (state) \ + ((state) & TASK_UNINTERRUPTIBLE && (!((state) & TASK_NOLOAD))) + +#define is_idle_state (state) (((state) & TASK_IDLE) == TASK_IDLE) +#ifdef CONFIG_DEBUG_ATOMIC_SLEEP # define debug_normal_state_change (state_value) \ do { \
kernel/sched/deadline.c +3 -2
@@ -1566,11 +1566,12 @@ update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,unsigned int state; state =READ_ONCE (p->__state); -if (state &TASK_INTERRUPTIBLE ) +/* idle state still accounts into sleep */ +if (state &TASK_INTERRUPTIBLE ||is_idle_state (state))__schedstat_set (p->stats.sleep_start,rq_clock (rq_of_dl_rq (dl_rq))); -if (state &TASK_UNINTERRUPTIBLE ) +if (is_blocked_state (state))__schedstat_set (p->stats.block_start,rq_clock (rq_of_dl_rq (dl_rq))); }
kernel/sched/fair.c +3 -2
@@ -1278,10 +1278,11 @@ update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int fl/* XXX racy against TTWU */ state =READ_ONCE (tsk->__state); -if (state &TASK_INTERRUPTIBLE ) +/* idle state still accounts into sleep */ +if (state &TASK_INTERRUPTIBLE ||is_idle_state (state))__schedstat_set (tsk->stats.sleep_start,rq_clock (rq_of (cfs_rq))); -if (state &TASK_UNINTERRUPTIBLE ) +if (is_blocked_state (state))__schedstat_set (tsk->stats.block_start,rq_clock (rq_of (cfs_rq))); }
kernel/sched/rt.c +3 -2
@@ -1371,11 +1371,12 @@ update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,unsigned int state; state =READ_ONCE (p->__state); -if (state &TASK_INTERRUPTIBLE ) +/* idle state still accounts into sleep */ +if (state &TASK_INTERRUPTIBLE ||is_idle_state (state))__schedstat_set (p->stats.sleep_start,rq_clock (rq_of_rt_rq (rt_rq))); -if (state &TASK_UNINTERRUPTIBLE ) +if (is_blocked_state (state))__schedstat_set (p->stats.block_start,rq_clock (rq_of_rt_rq (rt_rq))); }
Signature · 1 line
2.43.0
Valentin Schneider
reviewerQuoted 8 lines · On 03/01/24 16:10, [email protected] wrote:
From: Alex Shi <[email protected]>
The commit
80ed87c8a9ca("sched/wait: IntroduceTASK_NOLOADandTASK_IDLE") stopped the idle kthreads from contributing to the load average. However, the idle state time still contributes to the blocked state time instead of the sleep time. As a result, we cannot determine if a task is stopped due to some reasons or if it is idle by its own initiative.
Quoted 4 lines
Distinguishing between these two states would make the system state clearer and provide us with an opportunity to use the 'D' state of a task as an indicator of latency issues.
get_task_state() already reports TASK_IDLE as 'I', which should be what userspace sees (e.g. via /proc/$pid/stat). This is also the case for the sched_switch and sched_wakeup trace events.
I assume what you mean here is you first turn to schedstats to check whether there is any abnormal amount of blocking, and then if there is any you turn to tracing, in which case you'd like the schedstats to not make things look worse than they really are?
Quoted 15 lines · Valentin Schneider
Quoted 8 lines · On 03/01/24 16:10, [email protected] wrote:
From: Alex Shi <[email protected]>
The commit
80ed87c8a9ca("sched/wait: IntroduceTASK_NOLOADandTASK_IDLE") stopped the idle kthreads from contributing to the load average. However, the idle state time still contributes to the blocked state time instead of the sleep time. As a result, we cannot determine if a task is stopped due to some reasons or if it is idle by its own initiative.Quoted 4 lines
Distinguishing between these two states would make the system state clearer and provide us with an opportunity to use the 'D' state of a task as an indicator of latency issues.
Quoted 8 lines
get_task_state()already reportsTASK_IDLEas 'I', which should be what userspace sees (e.g. via/proc/$pid/stat). This is also the case for thesched_switchandsched_wakeuptrace events.I assume what you mean here is you first turn to schedstats to check whether there is any abnormal amount of blocking, and then if there is any you turn to tracing, in which case you'd like the schedstats to not make things look worse than they really are?
Yes, switch/wakeup or others could help to figure out real blocked time, but with this change, schedstats could give a neat and elegant way.
Thanks!
Quoted 1 line
Quoted 21 lines · Alex Shi <[email protected]> 于2024年1月4日周四 19:29写道:
Quoted 19 lines · Valentin Schneider
Quoted 8 lines · On 03/01/24 16:10, [email protected] wrote:
From: Alex Shi <[email protected]>
The commit
80ed87c8a9ca("sched/wait: IntroduceTASK_NOLOADandTASK_IDLE") stopped the idle kthreads from contributing to the load average. However, the idle state time still contributes to the blocked state time instead of the sleep time. As a result, we cannot determine if a task is stopped due to some reasons or if it is idle by its own initiative.Quoted 4 lines
Distinguishing between these two states would make the system state clearer and provide us with an opportunity to use the 'D' state of a task as an indicator of latency issues.
get_task_state()already reportsTASK_IDLEas 'I', which should be what userspace sees (e.g. via/proc/$pid/stat). This is also the case for thesched_switchandsched_wakeuptrace events.
Quoted 8 lines
Quoted 4 lines
I assume what you mean here is you first turn to schedstats to check whether there is any abnormal amount of blocking, and then if there is any you turn to tracing, in which case you'd like the schedstats to not make things look worse than they really are?
Yes, switch/wakeup or others could help to figure out real blocked time, but with this change, schedstats could give a neat and elegant way.
For all of the shortages I can imagine, we can't treat blocked and sleep states as before, and that may force some scripts to change on these things, but giving 2 states the correct way is better way, and make sleep/block state more meaningful and helpful in normal usage. Are there any concerns on this?
Thanks
Alex