Re: [PATCH 02/18] Blackfin Serial Driver: use barrier instead of cpu_relax for Blackfin SMP like patch

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

On Wed, 11 Mar 2009 17:50:21 +0800 graff yang <graff.yang@gmail.com> wrote:

> On Wed, Mar 11, 2009 at 12:48 AM, Andrew Morton
> <akpm@linux-foundation.org> wrote:
> >
> > On Tue, 10 Mar 2009 06:25:08 -0400 Mike Frysinger <vapier.adi@gmail.com> wrote:
> >
> > > On Tue, Mar 10, 2009 at 06:07, gyang wrote:
> > > > On Fri, 2009-03-06 at 14:37 -0800, Andrew Morton wrote:
> > > >> On Fri, __6 Mar 2009 14:42:44 +0800
> > > >> Bryan Wu <cooloney@kernel.org> wrote:
> > > >>
> > > >> > From: Graf Yang <graf.yang@analog.com>
> > > >> >
> > > >> > We are making a SMP like patch to blackfin, cpu_relax() is replaced by a
> > > >> > data cache flush function which will count it to a per-cpu counter.
> > > >> > If this serial function is called too early, the per-cpu data area have
> > > >> > not been initialized, this call will cause crash.
> > > >>
> > > >> That's a bug in blackfin architecture support. __The kernel should be
> > > >> able to call cpu_relax() at any time, surely. __It's a very low-level
> > > >> and simple thing.
> > > >>
> > > >> > So we'd like to use barrier() instead of cpu_relax().
> > > >> >
> > > >>
> > > >> barrier() is purely a compiler concept. __We might as well just remove
> > > >> the cpu_relax() altogether.
> > > >
> > > > Do you mean remove cpu_relax(), and either not add barrier() here?
> > >
> > > afaik, early printk all runs before SMP is setup, so having it be a
> > > 100% busy wait is fine
> >
> > No, blackfin is busted, please fix this bug in blackfin core.
> >
> > What happens if core kernel code decides to run cpu_relax() prior to
> > initialising per-cpu data?
> 
> cpu_relax() will call smp_mb(), and it need to invalidate data-cache
> in blackfin smp-like kernel,
> the cache flush number is increased and recorded into a per-cpu data.
> 
> When cpu_relax() is called from early-printk functions, the per-cpu
> data areas have not
> been initialized.

Right.  And we've demonstrated here that this was a bad idea.

So I suggest that the blackfin maintainers delete that code.

If it's really valuable (which I doubt) then it could be reimplemented
via a statically-allocated array of longs and local_irq_save()
protection.  Or a statically allocated array of atomic_long_t's.