Re: [PATCH v3] x86/mce: Don't participate in rendezvous process once nmi_shootdown_cpus() was made

["Luck, Tony" <tony.luck@intel.com>]

On Wed, Feb 22, 2017 at 12:11:14PM +0800, Xunlei Pang wrote:
> +	/*
> +	 * Cases to bail out to avoid rendezvous process timeout:
> +	 * 1)If this CPU is offline.
> +	 * 2)If crashing_cpu was set, e.g. entering kdump,
> +	 *   we need to skip cpus remaining in 1st kernel.
> +	 */
> +	if (cpu_is_offline(cpu) ||
> +	    (crashing_cpu != -1 && crashing_cpu != cpu)) {
>  		u64 mcgstatus;
>  
>  		mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);


I think we should document the remaining race conditions. I don't
think there is any good way to eliminate them, and they are already
pretty small windows.

I think the sequence of events looks like:

     1	Panic occurs
     2	nmi_shootdown_cpus() sets crashing_cpu
     3	send NMI to everyone else
     4	wait up to a second for other CPUs to take NMI
     5	go to kexec code
     6	start new kernel
     7	new kernel establishes #MC handler

If one of the other cpus triggers a machine check while
getting to, or in, the NMI handler ... then that cpu will
skip processing (if RIPV is set).

Between '2' and '5' if crashing_cpu gets a machine check it
will execute in the old kernel handler, and do the right thing.

There's a fuzzy area between '6' and '7' where a machine check
might not end up in the right code.

>From '7' onwards the kexec kernel will handle and machine
checks caused by kdump.

-Tony