Re: [PATCH] x86/watchdog: Use real timestamps for watchdog timeout

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

On 24/05/13 18:10, Tim Deegan wrote:
> At 15:29 +0100 on 24 May (1369409374), Andrew Cooper wrote:
>> On 24/05/13 14:55, Tim Deegan wrote:
>>> At 13:48 +0100 on 24 May (1369403327), Andrew Cooper wrote:
>>>> On 24/05/13 13:41, Tim Deegan wrote:
>>>>> Of those two, I prefer (1), just because it doesn't add any cost to the
>>>>> normal users of NOW().
>>>> I was not planning to make any requirement to change users of NOW(). 
>>> Well, you were planning to make NOW() slightly more expensive by needing
>>> to look up which of the banekd alternatives is valid.  In any case, I
>>> think some sort of approximate version based on tsc will do.
>>>
>>> Tim.
>> I was planning to memcpy the shadow set over the main set as part of
>> calibration, leaving no alteration whatsoever to NOW().
> Sorry, yes, I see how that works now.  And so I too prefer (2). :)
>
>> An approximation from the TSC alone would be better so long as it is a
>> reasonable approximation.  I am concerned about how accuate a dumb
>> approximation would be for non-stable TSCs etc.
> Yep.  I'm more and more convinced that we should gate on the number of
> NMIs we've taken without seeing a timer tick.  I'm more afratid of funny
> TSC edge cases (and remember we might take an NMI anywhere in the s3
> wakeup) than I am of machines with really bad NMI storms.  So even if
> the approximate time is wildly off we just print the wrong thing. 
>
> In the case where you saw this (and cpu0 was alive for a while before
> it managed a burst of enough NMIs), would detecting and warning
> about high NMI rates be enough to point out what's gone wrong?
>
> Tim.

Not directly, butb or my debugging case knowing when the NMI storm has
started is very useful so I can dump lspci -vvvxxxx to get the debug
state from whichever PCI device is the original cause of the SERR storm.

I certainly don't think there is anything useful Xen could automatically
do when discovering an NMI storm, but leaving a message on the serial or
in a crash state certainly helps someone trying to investigate why the
server reset.  It is certainly better than finding an NMI watchdog
timeout with serial timestamps proving that the watchdog didn't actually
time out :), and starting debugging wondering WTF was going on.

~Andrew