Re: [RFC] x86, perf: high volume of events produces a flood of unknown NMIs

[Cyrill Gorcunov <gorcunov@gmail.com>]

On 04/20/2011 10:47 PM, Cyrill Gorcunov wrote:
> On 04/20/2011 10:26 PM, Don Zickus wrote:
>> Hi,
>>
>> Arnaldo pointed me at an NMI problem that happens when he tries to
>> generate a high volume of perf events.  He receives a flood of unknown
>> NMIs.
>>
>> I have been poking at the problem and came up with a patch, but it doesn't
>> always work.  I was hoping people who understood how the NMI works at a
>> low level might be able to help me.
>>
>> I was able to duplicate this on an AMD Phenom, Pentium 4, Xeon Core2quad,
>> and Nehalem.  The problem I think is the large generation of back-to-back
>> NMIs.  The perf nmi handler may accidentally handle some of those
>> extra/in-flight NMIs in its first pass, leaving the next NMI to be
>> unhandled and generating an unknown NMI message.
>>
>> Robert included logic to check for two back-to-back NMIs, but that falls
>> short when more then three are generated.  I modified his logic to account
>> for three back-to-back NMIs, but that didn't completely solve the problem.
>>
>> I took another approach at catching back-to-back NMIs that seemed to work
>> on all my machines except for the Xeon core2quad, but I am not entirely
>> sure if my approach is valid.
>>
>> The approach I took was based on the idea that if an NMI is being
>> generated while currently in an NMI handler, the current NMI when finished
>> won't continue executing the next instruction before the exception but
>> instead jump back into another NMI exception frame.
>>
>> As a result, the args passed in to the NMI handler should have the same ip
>> and sp as the previous NMI interrupt.  Otherwise one could assume that
>> some amount of time passed between interrupts (enough to return from the
>> exception and execute code).
>>
>> I thought this would allow me to trap an infinite number of back-to-back
>> NMIs.  Like I said it seemed to work in a number of machines, except for
>> my Xeon core2quad.
>>
>> Does anyone know if my approach is a valid one?  Or is there a better way
>> to catch this condition?  Or maybe some other tips or tricks I can use to
>> help come up with a solution for this?
>>
>> Or perhaps we don't care about this because in the end perf can't even
>> capture the data without spitting out a CPU Overload message.
>>
>> Thoughts?
>>
> 
> Hi Don, just a thought -- since pmi masks lvtpc we could read it and check if it's
> masked or no, though I fear it is quite time consuming operation in compare with
> frames :( (hmm, intel spec mentions only p4 and xeon as masking lvtpc)

Something like

	if (apic_read(APIC_LVTPC) & APIC_LVT_MASKED))
		handle-perf

-- 
    Cyrill