Re: bug on shutdown from 68-mm4 (machine_power_off returning causes problems)

[ebiederm@xmission.com (Eric W. Biederman)]

"Martin J. Bligh" <mbligh@aracnet.com> writes:

> >> Yup, backing out kexec fixes it.
> > 
> > 
> > Ok.  Thinking it through the differences is that I have machine_power_off
> > call stop_apics (which is roughly equivalent to the old smp_send_stop).
> 
> Mmmm. Not sure NUMA-Q will like disconnect_bsp_APIC() much, but I guess
> that's my problem, not yours ;-) I can't do init 6 at the moment, so I'm
> walking on thin ice as is ... if I have to fix a couple of things up for
> NUMA-Q, that's no problem.

Ah ok.  This could explain some of your problems with kexec if you
can't even reboot the box....
 
> > In the kexec patch that does 2 things.
> > 1) It shuts down the secondary cpus, and returns the bootstrap cpu to
> >    virtual wire mode. 
> > 2) It calls set_cpus_allowed to force the reboot to be on the primary
> >    cpu.
> > 
> > After returning from machine_power_off. We run into a problem
> > in flush_tlb_mm.  Because we have a cpu disabled, that is still part
> > of the mm's vm mask.
> 
> OK ... I presume that's just because you shut down the secondaries then.
> If so ... shouldn't it remove them from the relevant vm mask when it
> offlines them? (probably not your code, but still)

Right.  The code is more about shutting down the secondaries then off
lining them.  In a general purpose scenario I expect more would need
to happen.

> > Does anyone know why machine_halt, and machine_power_off return?
> 
> Nope, thats ... odd ;-)

That is what I thought.

I have sent the fixed patches to Andrew and CC'd the list so we will
see what develops.

Eric