Re: [PATCH v2] x86/mce: Don't participate in rendezvous process once nmi_shootdown_cpus() was made

[Xunlei Pang <xpang@redhat.com>]

On 02/20/2017 at 09:29 PM, Xunlei Pang wrote:
> On 02/20/2017 at 07:09 PM, Borislav Petkov wrote:
>> On Mon, Feb 20, 2017 at 02:10:37PM +0800, Xunlei Pang wrote:
>>> @@ -1128,8 +1129,9 @@ void do_machine_check(struct pt_regs *regs, long error_code)
>>>  	 */
>>>  	int lmce = 1;
>>>  
>>> -	/* If this CPU is offline, just bail out. */
>>> -	if (cpu_is_offline(smp_processor_id())) {
>>> +	/* If nmi shootdown happened or this CPU is offline, just bail out. */
>>> +	if (cpus_shotdown() ||
>> I don't like "cpus_shotdown" - it doesn't hint at all that this is
>> special-handling crash/kdump.
>>
>> And more importantly, I want it to be obvious that we do let the
>> crashing CPU into the MCE handler.
> Ok, I will export crashing_cpu and use it directly in mce handler.

Forget to mention, one reason I introduced cpus_shotdown() is that "crashing_cpu"
is defined with CONFIG_SMP=y, so we have to export it unconditionally if we don't want
to add the conditional code(i.e. with #ifdef CONFIG_SMP quoted) in mce.c.

Regards,
Xunlei

>
>> Why?
>>
>> If we didn't, you will not handle *any* MCE, even a fatal one, during
>> dumping memory so if that dump is corrupted from the MCE, you won't
>> know. And I don't want to be the one staring at the corrupted dump and
>> wondering why I'm seeing what I'm seeing.
>>
>> IOW, if we get a fatal MCE during dumping then we should go and die.
>> This is much better than silently corrupting the dump and not even
>> saying anything about it.
>>
> My thought is that it doesn't matter after kdump boots as new mce handler
> will be installed. If we get a fatal MCE during kdumping, the new handler will
> handle the cpus running kdump kernel correctly.
>
> There is a small window between crash and kdump kernel boot, so if a SRAO comes
> within this window it will also cause the mce synchronization problem on the crashing
> cpu if we don't bail out the crashing cpu.
>
> Regards,
> Xunlei