Re: workqueue cpu affinity

[Max Krasnyansky <maxk@qualcomm.com>]

Oleg Nesterov wrote:
> On 06/10, Max Krasnyansky wrote:
>> Here is some backgound on this. Full cpu isolation requires some tweaks to the
>> workqueue handling. Either the workqueue threads need to be moved (which is my
>> current approach), or work needs to be redirected when it's submitted.
> 
> _IF_ we have to do this, I think it is much better to move cwq->thread.
Ok. btw That's what I'm doing now from user-space.

>> Peter Zijlstra wrote:
>>> The advantage of creating a more flexible or fine-grained flush is that
>>> large machine also profit from it.
>> I agree, our current workqueue flush scheme is expensive because it has to
>> schedule on each online cpu. So yes improving flush makes sense in general.
> 
> Yes, it is easy to implement flush_work(struct work_struct *work) which
> only waits for that work, so it can't hang unless it was enqueued on the
> isolated cpu.
> 
> But in most cases it is enough to just do
> 
> 	if (cancel_work_sync(work))
> 		work->func(work);
Cool. That would work.
btw Somehow I thought that you already implemented flush_work(). I do not see
it 2.6.25 but I could've sworn that I saw a patch flying by. Must have been
something else. Do you mind adding that ?

> Or we can add flush_workqueue_cpus(struct workqueue_struct *wq, cpumask_t *cpu_map).
That'd be special casing. I mean something will have to know what cpus cannot
be flushed. I liked your proposal above much better.

> But I don't think we should change the behaviour of flush_workqueue().
> 
>> This will require a bit of surgery across the entire tree. There is a lot of
>> code that calls flush_scheduled_work()
> 
> Almost all of them should be changed to use cancel_work_sync().

That'd be a lot of changes.

git grep flush_scheduled_work | wc
    154     376    8674

Hmm, I guess maybe not that bad. I might actually do that :-).

Max