Hi Steven,
Thanks for sharing your opinion. :)
As, this went out to be a long thread of discussion (thanks Paul), i will try to answer everything here.
On 26 November 2012 22:10, Steven Rostedt rostedt@goodmis.org wrote:
This is a really bad time of year to post new patches :-/ A lot of people are trying to get their own work done by year end and then there's holidays and such that are also distractions. Not to mention that a new merge window will be opening soon.
Patches are there since End of September and it was just a ping now (actually with bad timing - merge window).
As workqueues are set off by the CPU that queued it, what real benefit does this give? A CPU was active when it queued the work and the work should be done before it gets back to sleep.
OK, an interrupt happens on an idle CPU and queues some work. That work should execute before the CPU gets back to sleep, right? I fail to see the benefit of trying to move that work elsewhere. The CPU had to wake up to execute the interrupt. It's no longer in a deep sleep (or any sleep for that matter).
To me it seems best to avoid waking up an idle CPU in the first place.
I'm still working off a turkey overdose, so maybe I'm missing something obvious.
Ok, here is the story behind these patches. The idea was first discussed by Vincent in LPC this year:
http://www.linuxplumbersconf.org/2012/wp-content/uploads/2012/08/lpc2012-sch...
Specifically slides: 12 & 19.
cpu idleness here means idleness from scheduler's perspective, i.e. For the scheduler CPU is idle, if all below are true: - current task is idle task - nr_running == 0 - wake_list is empty
There are two use cases of workqueue patch 3/4:
- queue-work from timer interrupt, which re-arms itself (slide 12): Whether timer is deferred or not, it will queue the work on current cpu once cpu wakes up. The cpu could have immediately gone back to idle state again, if the work is not queued on it. So, because this cpu is idle (from schedulers point of view), we can move this work to other cpu.
- delayed-work, which rearm's itself (slide 19): Again the same thing, we could have kept the cpu in idle state for some more time.
There might not be many users with this behavior, but a single user can actually have significant impact.
For now, it doesn't take care of big LITTLE stuff that Paul pointed out, but yes that is in plan. Some work is going on in that direction too:
http://linux.kernel.narkive.com/mCyvFVUX/rfc-0-6-sched-packing-small-tasks
The very first reason of having this patchset was to have a single preferred_cpu() routine, which can be used by all frameworks. Timers being the first user (already), workqueues tried to be the second one.
I tested it more from functionality point of view rather than with power figures :( And i understood, that it is very much required.
Having said that, I believe all the questions raised are on PATCH 3/4 (workqueue). And other 3 patches should be fine. Can you share your opinion on those patches, I will then split this patchset and send workqueue stuff after doing some power measurements later.
-- viresh