Unable to use perf in VM

[marc.zyngier@arm.com (Marc Zyngier)]

On 30/11/16 19:17, Wei Huang wrote:
> 
> 
> On 11/30/2016 07:37 AM, Marc Zyngier wrote:
>> On 30/11/16 11:48, Marc Zyngier wrote:
>>> + Shannon
>>>
>>> On 29/11/16 22:04, Itaru Kitayama wrote:
>>>> Hi,
>>>>
>>>> In a VM (virsh controlled, KVM acceleration enabled) on a recent
>>>> kvmarm kernel host, I find I am unable to use perf to obtain
>>>> performance statistics for a complex task like kernel build.
>>>> (I've verified this is seen with a Fedora 25 VM and host combination
>>>> as well)
>>>> APM folks CC'ed think this might be caused by a bug in the core PMU 
>>>> framework code, thus I'd like to have experts opinion on this issue.
>>>>
>>>> [root at localhost linux]# perf stat -B make
>>>>    CHK     include/config/kernel.release
>>>> [  119.617684] git[1144]: undefined instruction: pc=fffffc000808ff30
>>>> [  119.623040] Code: 51000442 92401042 d51b9ca2 d5033fdf (d53b9d40)
>>>> [  119.627607] Internal error: undefined instruction: 0 [#1] SMP
>>>
>>> [...]
>>>
>>> In a VM running mainline hosted on an AMD Seattle box:
>>>
>>>  Performance counter stats for 'make':
>>>
>>>     1526089.499304      task-clock:u (msec)       #    0.932 CPUs utilized          
>>>                  0      context-switches:u        #    0.000 K/sec                  
>>>                  0      cpu-migrations:u          #    0.000 K/sec                  
>>>           29527793      page-faults:u             #    0.019 M/sec                  
>>>      2913174122673      cycles:u                  #    1.909 GHz                    
>>>      2365040892322      instructions:u            #    0.81  insn per cycle         
>>>    <not supported>      branches:u                                                  
>>>        32049215378      branch-misses:u           #    0.00% of all branches        
>>>
>>>     1637.531444837 seconds time elapsed
>>>
>>> Running the same host kernel on a Mustang system, the guest explodes
>>> in the way you reported. The failing instruction always seems to be
>>> an access to pmxevcntr_el0 (I've seen both reads and writes).
>>>
>>> Funnily enough, it dies if you try any HW event other than cycles
>>> ("perf stat -e cycles ls" works, and "perf stat -e instructions ls"
>>> explodes). Which would tend to indicate that we're screwing up
>>> the counter selection, but I have no proof of that (specially that
>>> the Seattle guest is working just as expected).
>>
>> It turns out that we *don't* inject an undef. It seems to be generated
>> locally at EL1.
>>
>> Still digging.
> 
> Just FYI: I saw it on Mustang before. My initial thought was HW related,
> but without proof. I am interested to see your findings...

It would have been good to report it earlier. Anyway, I've identified
the root issue, which seems to boil down to how you interpret a small
corner of the PMU architecture. I've raised it with the architecture
team here, and I should have a workaround/fix shortly.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...