Re: Re: [RFC PATCH] suspend/resume performance improvement

Re: Re: [RFC PATCH] suspend/resume performance improvement

[EunTaik Lee]

>> When a task that calls state_store() to suspend
>> the device has used up most of its time slice,
>> suspend sometimes take too long. (User noticeable)
>> 
>> Suspend/resume is a system wide operation.
>> So, instead of depending on a userspace task's time
>> slice, let kworker do the work to avoid a long wait
>> on the runqueue.

> All of that seems to duplicate the existing autosleep.c functionality.

Yes, the effect of using the kworker in autosleep will be same performance-wise.
But if a platform doesn't use the autosleep functionality then we still have the problem.
So doesn't it make sense to implement it for the autosleep-disabled use cases?

Re: Re: [RFC PATCH] suspend/resume performance improvement

[EunTaik Lee]

> Instead of replicating code in state_store here, does it make sense to
> move into a common part, so that fixes/optimization can be done in one
> place?
> Also why another config? If this has proven benefit then it can be
> default path.
I agree. I will resend the patch if this patch is agreed to be useful.

Re: Re: [RFC PATCH] suspend/resume performance improvement

[EunTaik Lee]

On Tue 2015-05-27 16:50 (GMT+09:00), Pavel Machek wrote:

>> So, instead of depending on a userspace task's time
>> slice, let kworker do the work to avoid a long wait
>> on the runqueue.
>
>...so if you really want high priority for that operation, just renice
>yourself to higher priority or something...

The problem occured on Android L version. KK had autosuspend enabled
but it was disabled in L to allow userspace to collect stats.

The delay was inside try_to_freeze_tasks function (at system_server thread of system_server process).
system_server got preempted after unlocking the tasklist_lock.
 46         while (true) {
 47                 todo = 0;
 48                 read_lock(&tasklist_lock);
 49                 for_each_process_thread(g, p) {
 50                         if (p == current || !freeze_task(p))
 51                                 continue;
 52 
 53                         if (!freezer_should_skip(p))
 54                                 todo++;
 55                 }
 56                 read_unlock(&tasklist_lock);

All of the threads of system_server process that was on the runqueue did not run during this period.
Freezing of tasks resumed after other tasks went out of the runqueue.

I have tested with the lowest nice value but the problem was still, although less, reproducible.

Best Regards,
Euntaik Lee