Re: BUG: using smp_processor_id() during suspend with 2.6.25-rc8

[Andi Kleen <andi@firstfloor.org>]

On Tue, Apr 08, 2008 at 12:29:30AM +0200, Rafael J. Wysocki wrote:
> On Tuesday, 8 of April 2008, Andi Kleen wrote:
> > On Tue, Apr 08, 2008 at 12:11:17AM +0200, Jiri Kosina wrote:
> > > On Tue, 8 Apr 2008, Andi Kleen wrote:
> > > 
> > > > > I know. However preempt_count is a little bit inconsistent in such cases 
> > > > > though.
> > > > And? interrupts off beats preempt count anyways. Why did you write the 
> > > > patch? Was there a (incorrect) warning triggered?
> > > 
> > > Reported at http://lkml.org/lkml/2008/4/7/130
> > > 
> > > BTW is also mce_init() (called from mce_resume()) guaranteed to run with 
> > > IRQs off?
> > 
> > [cc rafael]
> > 
> > The mce resume is a sysdev.
> > 
> > sysdevs were always supposed to run completely with interrupts off. If they 
> > don't anymore that's some kind of higher level resume code bug which you need 
> > to fix there, not hack around in the low level code.
> 
> They are executed with interrupts disabled, on one CPU.

Well then someone enables them incorrectly, see the report above.

> 
> > If it does that it likely broke more code too.
> > 
> > Obviously turning on preemption anywhere around the machine check is
> > fatal because it touches CPU state and if you reschedule you could
> > switch to another CPU and change or access the wrong CPU's state.
> 
> FWIW, at the point when sysdevs are resumed we are single-threaded.

You mean single CPUed? Even a single thread could switch to another CPU.

-Andi