Re: [PATCH 2/3] work_on_cpu: Use our own workqueue.

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 10 Feb 2009 19:24:07 +1030 Rusty Russell <rusty@rustcorp.com.au> wrote:

> On Thursday 05 February 2009 02:06:36 Andrew Morton wrote:
> > On Wed, 4 Feb 2009 21:11:35 +1030 Rusty Russell <rusty@rustcorp.com.au> wrote:
> > 
> > > On Wednesday 04 February 2009 13:31:11 Andrew Morton wrote:
> > > > On Wed, 4 Feb 2009 13:14:31 +1030 Rusty Russell <rusty@rustcorp.com.au> wrote:
> > > > > I think you're right though: smp_call_function_single (or neat wrappers)
> > > > > where possible, work_on_cpu which can fail for the others, and we'll just
> > > > > have to plumb in the error returns.
> > > > 
> > > > I bet a lot of those can use plain old schedule_work_on().
> > > 
> > > Which is where work_on_cpu started: a little wrapper around schedule_work_on.
> > > 
> > > We're going in circles, no?
> > 
> > No, we've made some progress.  We have a better understanding of what
> > the restrictions, shortcomings and traps are in this stuff.  We've
> > learned (surprise!) that a one-size-fits-all big hammer wasn't such a
> > great idea.
> > 
> > Proposed schedule_work_on() rule: either the flush_work() caller or the
> > callback should not hold any explicit or implicit sleeping locks.
> 
> But as you found out looking through these, it's really hard to tell.  I can
> guess, but that's a little fraught...

yup.

> How about we make work_on_cpu spawn a temp thread; if you care, use
> something cleverer?  Spawning a thread just isn't that slow.

That's what
work_on_cpu-rewrite-it-to-create-a-kernel-thread-on-demand.patch does?

> Meanwhile, I'll prepare patches to convert all the non-controversial cases
> (ie. smp_call_function-style ones).

arch-x86-kernel-acpi-cstatec-avoid-using-work_on_cpu.patch
arch-x86-kernel-cpu-cpufreq-acpi-cpufreqc-avoid-using-work_on_cpu.patch
arch-x86-kernel-cpu-mcheck-mce_amd_64c-avoid-using-work_on_cpu.patch

convert three work_on_cpu() callers.  The drivers/pci/pci-driver.c one
is a bit problematic.

I guess as long as we don't find a high frequency set_cpus_allowed()
callsite which can't be converted to smp_call_function_single() we'll
be OK.