Re: [RFC][PATCH v4 -next 1/4] Move kmsg_dump(KMSG_DUMP_PANIC) below smp_send_stop()

[Don Zickus <dzickus@redhat.com>]

On Fri, Feb 03, 2012 at 10:32:31PM +0000, Luck, Tony wrote:
> > What if we send the REBOOT_IPI first and let it block for up to a second.
> > Most code paths that are done with spin_locks will use
> > spin_lock_irqrestore.  As soon as the interrupts are re-enabled the
> > REBOOT_IPI comes in and takes the processor.  If after a second the cpu
> > still is blocking interrupts, just use the NMI as a big hammer to shut it
> > down.
> 
> This looks good - it certainly deals with my "if we just let them run
> a bit, they'd release the locks" quibble.  One second sounds very
> generous - but I'm not going to bikeshed that (so long as it is a total
> of one second - not one second per cpu). So the pseudo-code is:

This is how the stop_cpus is implemented on x86 and the one second comes
from there

arch/x86/kernel/smp.c::native_irq_stop_other_cpus and
native_nmi_stop_other_cpus

> 
> 	send_reboot_ipi_to_everyone_else()
> 
> 	wait_1_second()
> 
> 	for_each_cpu_that_didnt_respond_to_reboot_ipi {
> 		hit_that_cpu_with_NMI()
> 	}
> 
> Perhaps a notification printk() if we had to use the NMI hammer?

Yes.

Again this is for x86, but I guess that is our common case with pstore.

Cheers,
Don