Re: [PATCH v3 2/3] x86/smp native_play_dead: Prefer cpuidle_play_dead() over mwait_play_dead()

[Thomas Gleixner <tglx@linutronix.de>]

On Fri, Nov 15 2024 at 11:07, Peter Zijlstra wrote:
> On Fri, Nov 15, 2024 at 02:21:19AM +0100, Thomas Gleixner wrote:
>> On Thu, Nov 14 2024 at 13:03, Peter Zijlstra wrote:
>> > On Tue, Nov 12, 2024 at 03:01:27PM +0100, Peter Zijlstra wrote:
>> >
>> >> No, not mwait hint. We need an instruction that:
>> >> 
>> >>  - goes to deepest C state
>> >>  - drops into WAIT-for-Start-IPI (SIPI)
>> >> 
>> >> Notably, it should not wake from:
>> >> 
>> >>  - random memory writes
>> >>  - NMI, MCE, SMI and other such non-maskable thingies
>> >>  - anything else -- the memory pointed to by RIP might no longer exist
>> >> 
>> >> Lets call the instruction: DEAD.
>> >
>> > So, turns out that when you send INIT to an AP it does the whole drop
>> > into Wait-for-SIPI and ignore non-maskable crap.
>> >
>> > The reason we don't do that is because INIT to CPU0 (BP) is somewhat
>> > fatal, but since Thomas killed all that CPU0 hotplug crap, I think we
>> > can actually go do that.
>> 
>> Instead of playing dead or to kick out CPUs from whatever dead play
>> routine they are in?
>> 
>> playimg dead is to stay because INIT will bring back the MCE broadcast
>> problem, which we try to avoid by bringing SMT siblings up just to shut
>> them down again by playing dead.
>> 
>> You need a MCE broadcast free system and/or some sensible BIOS bringup
>> code for that to work...
>
> Isn't INIT a better state to be in during kexec than HLT?

For the kexec transition I agree that INIT is better. We just need to be
careful for a crash kexec when the crashing CPU is not CPU0...

Thanks,

        tglx