Re: [PATCH 3/4] x86/nmi: wait for all CPUs in check_nmi_watchdog()

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 14/05/14 14:59, Jan Beulich wrote:
>>>> On 14.05.14 at 14:58, <david.vrabel@citrix.com> wrote:
>> The counting of a CPUs NMIs in check_nmi_watchdog() is only reliable
>> if all CPUs have been spinning for 5 or more ticks.  There may be
>> delays in waking other CPUs from deep power states that can mean that
>> when the counts are checked CPUs haven't run for long enough.
> 5 ticks ought to be a couple of orders of a magnitude longer than
> the worst possible wakeup time. I.e. I don't buy this argument
> without actual numbers to support it.
>
> Jan

Its not necesserily the wakeup time, although on some systems that does
appear to be a consideration.  Simple A

We have a prototype 8 socket system where the cpus on the further
sockets had progressively less nmi delta, and are mostly declared stuck
as the BSP reads the delta before the APs have completed their busy loop.

~Andrew