On Mon, Nov 11 2024 at 17:23, Peter Zijlstra wrote:
> On Fri, Nov 08, 2024 at 08:49:31AM -0500, Len Brown wrote:
>> diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
>> index 766f092dab80..910cb2d72c13 100644
>> --- a/arch/x86/kernel/smpboot.c
>> +++ b/arch/x86/kernel/smpboot.c
>> @@ -1377,6 +1377,9 @@ void smp_kick_mwait_play_dead(void)
>> for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
>> /* Bring it out of mwait */
>> WRITE_ONCE(md->control, newstate);
>> + /* If MONITOR unreliable, send IPI */
>> + if (boot_cpu_has_bug(X86_BUG_MONITOR))
>> + __apic_send_IPI(cpu, RESCHEDULE_VECTOR);
>> udelay(5);
>> }
>
> Going over that code again, mwait_play_dead() is doing __mwait(.exc=0)
> with IRQs disabled.
And the APIC is shut down. So it won't react on the IPI either.
> So that IPI you're trying to send there won't do no nothing :-/
>
> Now that comment there says MCE/NMI/SMI are still open (non-maskable
> etc.) so perhaps prod it on the NMI vector?
>
> This does seem to suggest the above code path wasn't actually tested.
I'm not sure whether that's just a suggestion :)
> Perhaps mark your local machine with BUG_MONITOR, remove the md->control
> WRITE_ONCE() and try kexec to test it?
>
> Thomas, any other thoughts?
NMI should work. See exc_nmi():
if (arch_cpu_is_offline(smp_processor_id())) {
if (microcode_nmi_handler_enabled())
microcode_offline_nmi_handler();
return;
}
Thanks,
tglx
