Re: [PATCH] x86, UV: Fix NMI handler for UV platforms

[Cyrill Gorcunov <gorcunov@gmail.com>]

On 03/23/2011 12:25 AM, Jack Steiner wrote:
> On Tue, Mar 22, 2011 at 02:44:50PM -0400, Don Zickus wrote:
>> On Tue, Mar 22, 2011 at 12:11:18PM -0500, Jack Steiner wrote:
>>> How certain are you that multiple NMIs triggered at about the same time will
>>> deliver discrete NMI events? I updated the patch so that I'm running with:
>>
>> I think as long as there isn't more than two (1 active, 1 latched), you
>> would be ok.  A third one looks like it would get dropped.
>>
>>>
>>> 	- no special code in traps.c (I removed the traps.c code that was
>>> 	  in the patch I posted)
>>> 	- used die_notifier for calling the UV nmi handler
>>> 	- UV priority is higher than the hw_perf priority
>>>
>>> Both hw_perf (perf top) & UV NMIs work correctly under light loads. However, if I
>>> run for 10 - 15 minutes injecting UV NMIs at a rate of about 30/min, "perf top"
>>> stops generating output. Strace shows that it continues to poll() but no data
>>> is received.
>>
>> That's a low frequency and it still gets stuck?
>>
>>>
>>> While "perf top" is hung, if I inject an NMI into the system in a way that will NOT
>>> be consumed by the UV nmi handler, "perf top" resumes output but will stop again after
>>> a few minutes.
>>
>> So that means the PMU set its interrupt bit but the cpu failed to get the
>> NMI.
>>
>>>
>>>
>>> AFAICT, the UV nmi handler is not consuming extra NMI interrupts. I can't
>>> rule out that I'm missing something but I don't see it.
>>
>> What happens if you put the UV nmi handler below the hw_perf handler in
>> priority?  I assume the DIE_NMIUNKNOWN snippet in the hw_perf handler will
>> swallow some of the UV NMIs, but more importantly does it still generate
>> the hang you see?
> 
> I verified that the failures ("perf top" stops) are the same on both RHEL6.1 & the
> latest x86 2.6.38+ tree.
> 
> I switched priorities & as expected, "perf top" no longer hangs. I see an occassional
> missed UV NMI - about 1 every minute. I also see a few "dazed" messages as
> well - 3 in a 5 minute period. This testing was done on a 2.6.38+ kernel.
> 
> I'm running on a 48p system.
> 
> Ideas?
> 

  I fear there is always a probability for eaten nmi (due to inflight nmi logic
we have) or missed nmi (due to non-instant deliery of nmi).  Say the following
scenario may happen:

1) perf-nmi-0 (from counter 0) issued
2) uv-nmi issued
3) perf-nmi-0 latched
4) perf-nmi-1 (from counter 1) not yet issued but couter overflowed
5) nmi-handler
6) uv-nmi-latched
7) nmi-handler eats both nmis from perf-nmi-0 and uv-nmi because of in-flight
   nmi logic we have
8) finally perf-nmi-1 should appear on line but counter already pulled down so
   no nmi

and here you get missed nmi you expect from uv. I *guess*, not sure if it's possible.
If you disable nmi-watchdog on boot line, does it help?
-- 
    Cyrill