On Tue, Sep 05, 2017 at 09:38:45AM +0900, Byungchul Park wrote:
> On Mon, Sep 04, 2017 at 01:42:48PM +0200, Peter Zijlstra wrote:
> > On Mon, Sep 04, 2017 at 10:30:32AM +0900, Byungchul Park wrote:
> > > On Fri, Sep 01, 2017 at 06:38:52PM +0200, Peter Zijlstra wrote:
> > > > And get tangled up with the workqueue annotation again, no thanks.
> > > > Having the first few works see the thread setup isn't worth it.
> > > >
> > > > And your work_id annotation had the same problem.
> > >
> > > I keep asking you for an example because I really understand you.
> > >
> > > Fix my problematic example with your patches,
> > >
> > > or,
> > >
> > > Show me a problematic scenario with my original code, you expect.
> > >
> > > Whatever, it would be helpful to understand you.
> >
> > I _really_ don't understand what you're worried about. Is it the kthread
> > create and workqueue init or the pool->lock that is released/acquired in
> > process_one_work()?
>
> s/in process_one_work()/in all worker code including setup code/
>
> Original code was already designed to handle real dependencies well. But
> you invalidated it _w/o_ any reason, that's why I don't agree with your
> patches.
The reasons:
- it avoids the interaction with the workqueue annotation
- it makes each work consistent
- its not different from what you did with work_id:
https://lkml.kernel.org/r/1489479542-27030-6-git-send-email-byungchul.p...@lge.com
crossrelease_work_start() vs same_context_xhlock() { if
(xhlock->work_id == curr->workid) ... }
> Your patches only do avoiding the wq issue now we focus on.
>
> Look at:
>
> worker thread another context
> ------------- ---------------
> wait_for_completion()
> |
> | (1)
> v
> +---------+
> | Work A | (2)
> +---------+
> |
> | (3)
> v
> +---------+
> | Work B | (4)
> +---------+
> |
> | (5)
> v
> +---------+
> | Work C | (6)
> +---------+
> |
> v
>
> We have to consider whole context of the worker to build dependencies
> with a crosslock e.g. wait_for_commplete().
>
> Only thing we have to care here is to make all works e.g. (2), (4) and
> (6) independent, because workqueue does _concurrency control_. As I said
> last year at the very beginning, for works not applied the control e.g.
> max_active == 1, we don't need that isolation. I said, it's a future work.
>
> It would have been much easier to communicate with each other if you
> *tried* to understand my examples like now or you *tried* to give me one
> example at least. You didn't even *try*. Only thing I want to ask you
> for is to *try* to understand my opinions on conflicts.
>
> Now, understand what I intended? Still unsufficient?
So you worry about max_active==1 ? Or you worry about pool->lock or
about the thread setup? I'm still not sure.
