Messages in this thread | | | From | Barry Song <> | Date | Mon, 4 Oct 2021 10:06:08 +1300 | Subject | Re: wakeup_affine_weight() is b0rked - was Re: [PATCH 2/2] sched/fair: Scale wakeup granularity relative to nr_running |
| |
On Mon, Oct 4, 2021 at 3:52 AM Mike Galbraith <efault@gmx.de> wrote: > > On Sun, 2021-10-03 at 20:34 +1300, Barry Song wrote: > > > > > > I looked into that crazy stacking depth... > > > > > > static int > > > wake_affine_weight(struct sched_domain *sd, struct task_struct *p, > > > int this_cpu, int prev_cpu, int sync) > > > { > > > s64 this_eff_load, prev_eff_load; > > > unsigned long task_load; > > > > > > this_eff_load = cpu_load(cpu_rq(this_cpu)); > > > ^^^^^^^^^^^^^^^^^^^^^^^^^^^ the butler didit! > > > > > > That's pretty darn busted as it sits. Between load updates, X, or any > > > other waker of many, can stack wakees to a ludicrous depth. Tracing > > > kbuild vs firefox playing a youtube clip, I watched X stack 20 of the > > > zillion firefox minions while their previous CPUs all had 1 lousy task > > > running but a cpu_load() higher than the cpu_load() of X's CPU. Most > > > of those prev_cpus were where X had left them when it migrated. Each > > > and every crazy depth migration was wake_affine_weight() deciding we > > > should pull based on crappy data. As instantaneous load on the waker > > > CPU blew through the roof in my trace snapshot, its cpu_load() did > > > finally budge.. a tiny bit.. downward. No idea where the stack would > > > have topped out, my tracing_off() limit was 20. > > > > Mike, not quite sure I caught your point. It seems you mean x wakes up > > many firefoxes within a short period, so it pulls them to the CPU where x > > is running. Technically those pulling should increase cpu_load of x' CPU. > > But due to some reason, the cpu_load is not increased in time on x' CPU, > > So this makes a lot of firefoxes piled on x' CPU, but at that time, the load > > of the cpu which firefox was running on is still larger than x' cpu with a lot > > of firefoxes? > > It looked like this. > > X-2211 [007] d...211 2327.810997: select_task_rq_fair: this_run/load:4:373 prev_run/load:4:373 waking firefox:4971 CPU7 ==> CPU7 > X-2211 [007] d...211 2327.811004: select_task_rq_fair: this_run/load:5:373 prev_run/load:1:1029 waking QXcbEventQueue:4952 CPU0 ==> CPU7 > X-2211 [007] d...211 2327.811010: select_task_rq_fair: this_run/load:6:373 prev_run/load:1:1528 waking QXcbEventQueue:3969 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811015: select_task_rq_fair: this_run/load:7:373 prev_run/load:1:1029 waking evolution-alarm:3833 CPU0 ==> CPU7 > X-2211 [007] d...211 2327.811021: select_task_rq_fair: this_run/load:8:373 prev_run/load:1:1528 waking QXcbEventQueue:3860 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811026: select_task_rq_fair: this_run/load:8:373 prev_run/load:1:1528 waking QXcbEventQueue:3800 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811032: select_task_rq_fair: this_run/load:9:373 prev_run/load:1:1528 waking xdg-desktop-por:3341 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811037: select_task_rq_fair: this_run/load:10:373 prev_run/load:1:289 waking at-spi2-registr:3165 CPU4 ==> CPU7 > X-2211 [007] d...211 2327.811042: select_task_rq_fair: this_run/load:11:373 prev_run/load:1:1029 waking ibus-ui-gtk3:2865 CPU0 ==> CPU0 > X-2211 [007] d...211 2327.811049: select_task_rq_fair: this_run/load:11:373 prev_run/load:1:226 waking ibus-x11:2868 CPU2 ==> CPU2 > X-2211 [007] d...211 2327.811054: select_task_rq_fair: this_run/load:11:373 prev_run/load:11:373 waking ibus-extension-:2866 CPU7 ==> CPU7 > X-2211 [007] d...211 2327.811059: select_task_rq_fair: this_run/load:12:373 prev_run/load:1:289 waking QXcbEventQueue:2804 CPU4 ==> CPU7 > X-2211 [007] d...211 2327.811063: select_task_rq_fair: this_run/load:13:373 prev_run/load:1:935 waking QXcbEventQueue:2756 CPU1 ==> CPU7 > X-2211 [007] d...211 2327.811068: select_task_rq_fair: this_run/load:14:373 prev_run/load:1:1528 waking QXcbEventQueue:2753 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811074: select_task_rq_fair: this_run/load:15:373 prev_run/load:1:1528 waking QXcbEventQueue:2741 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811079: select_task_rq_fair: this_run/load:16:373 prev_run/load:1:1528 waking QXcbEventQueue:2730 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811085: select_task_rq_fair: this_run/load:17:373 prev_run/load:1:5 waking QXcbEventQueue:2724 CPU0 ==> CPU0 > X-2211 [007] d...211 2327.811090: select_task_rq_fair: this_run/load:17:373 prev_run/load:1:1010 waking QXcbEventQueue:2721 CPU6 ==> CPU7 > X-2211 [007] d...211 2327.811096: select_task_rq_fair: this_run/load:18:373 prev_run/load:1:1528 waking QXcbEventQueue:2720 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811101: select_task_rq_fair: this_run/load:19:373 prev_run/load:1:1528 waking QXcbEventQueue:2704 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811105: select_task_rq_fair: this_run/load:20:373 prev_run/load:0:226 waking QXcbEventQueue:2705 CPU2 ==> CPU2 > X-2211 [007] d...211 2327.811110: select_task_rq_fair: this_run/load:19:342 prev_run/load:1:1528 waking QXcbEventQueue:2695 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811115: select_task_rq_fair: this_run/load:20:342 prev_run/load:1:1528 waking QXcbEventQueue:2694 CPU5 ==> CPU7 > X-2211 [007] d...211 2327.811120: select_task_rq_fair: this_run/load:21:342 prev_run/load:1:1528 waking QXcbEventQueue:2679 CPU5 ==> CPU7
What is the topology of your hardware? shouldn't select_idle_sibling find some other idle CPUs in CPU7's LLC domain? Why are you always getting CPU7?
one thing bothering me is that we are using the load of a single CPU in wake_affine_weight(), but we are actually scanning the whole LLC afterwards.
> > Legend: foo_run/load:foo->nr_running:cpu_load(foo) > > Every migration to CPU7 in the above was due to wake_affine_weight() > seeing more or less static effective load numbers (the trace was wider, > showing which path was taken). > > > I am wondering if this should be the responsibility of wake_wide()? > > That's a good point. I'm not so sure that would absolve use of what > appears to be stagnant state though. If we hadn't gotten there, this > stack obviously wouldn't have happened.. but we did get there, and > state that was used did not reflect reality. wake_wide() deflecting > this particular gaggle wouldn't improved state accuracy one whit for a > subsequent wakeup, or? > > -Mike
Thanks barry
| |