Re: [PATCH] kernel/async.c:introduce async_schedule*_atomic

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, May 13, 2009 at 08:28:15AM +0800, Ming Lei wrote:
> 2009/5/13 Cornelia Huck <cornelia.huck@de.ibm.com>:
> > On Tue, 12 May 2009 18:52:29 +0200,
> > Frederic Weisbecker <fweisbec@gmail.com> wrote:
> >
> >> This division would make more sense indeed.
> >>
> >> - async_schedule_inatomic() would be nosync() and would use
> >>   GFP_ATOMIC. I guess the case where we want to run
> >>   a job synchronously from atomic in case of async failure is too rare
> >>   (non-existent?).
> >
> > It would add complexity for those callers providing a function that is
> > safe to be called in both contexts.
> >
> >> - async_schedule_nosync() would be only nosync() and would use
> >>   GFP_KERNEL
> >>
> >> I'm not sure the second case will ever be used though.
> >
> > It might make sense for the "just fail if we cannot get memory" case.
> >
> >>
> >> Another alternative would be to define a single async_schedule_nosync()
> >> which also takes a gfp flag.
> >
> > Wouldn't async_schedule() then need a gfp flag as well?
> >
> 
> IMHO, we should call async_schedule*() from non-atomic contexts and
> async_schedule_inatomic*() from atomic contexts explicitly, so
> async_schedule*()
> use GFP_KERNEL and async_schedule_inatomic*() use GFP_ATOMIC
> always. This can simplify the problem much more.



I think Cornelia is right about the complex case of a job
launched from atomic context that could either be run
synchronously. I have troubles to imagine such case though
but I guess it's possible.

 
> Also we still allow async_schedule*()  to run a job synchronously if
> out of memory
> or other failure. This can keep consistency with before.


Yes, but also most of the current users of async_schedule() could call
it with GFP_KERNEL. For now it's not an issue because it is not widely
used, but who knows how that will evolve...


> Any sugesstions or objections?


I have shared feelings. I don't know if the dual sense of
this new helper deserves enough disambiguation and granularity
to be split up in two parts:

- adding an async_schedule_nosync() helper
- add a new gfpflag_t parameter


Or should we just do:

- adding async_schedule_inatomic() which is a merge of nosync + GFP_ATOMIC
- use GFP_KERNEL in async_schedule()


It depends on the future users. Will someone ever accept to schedule a job
that could end up running synchronously in the worst case?

Frederic.