Re: [Patch V0] x86, mce: Ensure offline CPU's don't participate in mce rendezvous process.

[Andy Lutomirski]

On Fri, Dec 4, 2015 at 4:08 PM, Raj, Ashok <ashok.raj@xxxxxxxxx> wrote:
> On Fri, Dec 04, 2015 at 02:34:52PM -0800, Andy Lutomirski wrote:
>> On Fri, Dec 4, 2015 at 9:53 AM, Luck, Tony <tony.luck@xxxxxxxxx> wrote:
>> > ist_enter() is black magic to me. Andy? Would you be worried about executing
>> > ist_{enter,exit}() on a cpu that was once online, but is currently marked offline
>> > by Linux?
>>
>> Offline CPUs are black magic to me.  But as long as the CPU works the
>> way that the normal specs say it should, then ist_enter is fair game.
>> In any event, if context tracking blows up on an offline CPU, I'd
>> argue that's a context tracking bug and needs to be fixed.
>>
>> But maybe offlined CPUs are supposed to have all interrupts off
>> (including MCE?) and the argument goes the other way?  Dunno.
>
> MCE's are broadcast by the hardware and cannot be blocked. Offline
> is only a Linux specific state. Now if the offline was a result of an ACPI
> event (eject) that triggered the CPU removal (offline in Linux, as it would
> have in a platform that supports true hotplug) then the platform would
> remove this cpu from the broadcast list.
>
> if kernel were to set CR4.MCE=0 that would cause system shutdown when
> an MCE is broadcast and hits this cpu.

I meant "supposed" as in Linux might expect arch code to prevent the
CPU from receiving interrupts.

Anyway, I think that would be silly and we should just expect
ist_enter to work regardless of online state.

This does mean that if we plug in a new CPU and online it, then
there's a window before we set up percpu memory and enable CR4.MCE in
which an MCE on any CPU will kill the system, at least on hardware for
which MCE broadcast can't be turned off.

--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/