[PATCH 2/2] powernv/kdump: Fix cases where the kdump kernel can get HMI's

[Nicholas Piggin]

On Thu, 14 Dec 2017 11:12:13 +1100
Balbir Singh <bsingharora at gmail.com> wrote:

> On Wed, 13 Dec 2017 20:51:01 +1000
> Nicholas Piggin <npiggin at gmail.com> wrote:
> 
> > This is looking pretty nice now...
> > 
> > On Wed, 13 Dec 2017 19:08:28 +1100
> > Balbir Singh <bsingharora at gmail.com> wrote:
> >   
> > > @@ -543,7 +543,25 @@ void smp_send_debugger_break(void)
> > >  #ifdef CONFIG_KEXEC_CORE
> > >  void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
> > >  {
> > > +	int cpu;
> > > +
> > >  	smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, crash_ipi_callback, 1000000);
> > > +	if (kdump_in_progress() && crash_wake_offline) {
> > > +		for_each_present_cpu(cpu) {
> > > +			if (cpu_online(cpu))
> > > +				continue;
> > > +			/*
> > > +			 * crash_ipi_callback will wait for
> > > +			 * all cpus, including offline CPUs.
> > > +			 * We don't care about nmi_ipi_function.
> > > +			 * Offline cpus will jump straight into
> > > +			 * crash_ipi_callback, we can skip the
> > > +			 * entire NMI dance and waiting for
> > > +			 * cpus to clear pending mask, etc.
> > > +			 */
> > > +			do_smp_send_nmi_ipi(cpu);    
> > 
> > Still a little bit concerned about using NMI IPI for this.
> >  
> 
> OK -- for offline CPUs you mean?

Yes.

> > If you take an NMI IPI from stop, the idle code should do the
> > right thing and we would just return the system reset wakeup
> > reason in SRR1 here (which does not need to be cleared).
> > 
> > If you take the system reset anywhere else in the loop, it's
> > going to go out via system_reset_exception. I guess that
> > would end up doing the right thing, it probably gets to
> > crash_ipi_callback from crash_kexec_secondary?  
> 
> You mean like if we are online at the time of NMI'ing? If so
> the original loop will NMI us back into crash_ipi_callback
> anyway. We don't expect this to occur for offline CPUs

No, if the offline CPU is executing any instruction except for
stop when the crash occurs.

> 
> > 
> > It's just going to be a very untested code path :( What we
> > gain I suppose is better ability to handle a CPU that's locked
> > up somewhere in the cpu offline path. Assuming the uncommon
> > case works...
> > 
> > Actually, if you *always* go via the system reset exception
> > handler, then code paths will be shared. That might be the
> > way to go. So I would check for system reset wakeup SRR1 reason
> > and call replay_system_reset() for it. What do you think?
> >   
> 
> We could do that, but that would call pnv_system_reset_exception
> and try to call the NMI function, but we've not used that path
> to initiate the NMI, so it should call the stale nmi_ipi_function
> which is crash_ipi_callback and not go via the crash_kexec path.

It shouldn't, if the CPU is not set in the NMI bitmask, I think
it should go back out and do the rest of the system_reset_exception
handler.

Anyway we have to get this case right, because it can already hit
as I said if the offline CPU takes the NMI when it is not stopped.
This is why I want to try to use a unified code path.

> I can't call smp_send_nmi_ipi due to the nmi_ipi_busy_count and
> I'm worried about calling a stale nmi_ipi_function via the
> system_reset_exception path, if we are OK with it, I can revisit
> the code path

You shouldn't get a stale one, that would also be a bug -- we
have to cope with NMIs coming in at any time that are triggered
externally (not by smp_send_nmi_ipi), so if you see any bugs
there those need to be fixed separately.

Thanks,
Nick