Load latency sampling on Sandy Bridge

Load latency sampling on Sandy Bridge

[Manuel Selva]

Hi,

I am trying to setup load latency sampling on a Sandy Bridge processor 
(Cpu familly 6, model 42, model name  Intel(R) Core(TM) i5-2520M CPU @ 
2.50GHz) running a 3.13 linux kernel.

After deep investigations made of reading of the Intel SDM, examples 
provided by the pfm library and the source code of perf, I am still not 
able to get sample. I think for now that my problem is related to the 
precise_ip parameter of the  perf_event_attr structure.

Any time I change this value from 0 to 1 or from 0 to 2 (2 is the value 
to enable PEBS isn't it ?) the perf_event_open system call failed with a 
"Not supported operation" error. Running perf mem record on this Sandy 
Bridge platform also result in the same error.

So my question is, which way should I follow to identify why I am not 
able to do load latency sampling on this hardware ?

Thanks for any advice,

----
Manu

Re: Load latency sampling on Sandy Bridge

[Vince Weaver]

On Wed, 29 Oct 2014, Manuel Selva wrote:

> I am trying to setup load latency sampling on a Sandy Bridge processor (Cpu
> familly 6, model 42, model name  Intel(R) Core(TM) i5-2520M CPU @ 2.50GHz)
> running a 3.13 linux kernel.
> 
> After deep investigations made of reading of the Intel SDM, examples provided
> by the pfm library and the source code of perf, I am still not able to get
> sample. I think for now that my problem is related to the precise_ip parameter
> of the  perf_event_attr structure.
> 
> Any time I change this value from 0 to 1 or from 0 to 2 (2 is the value to
> enable PEBS isn't it ?) the perf_event_open system call failed with a "Not
> supported operation" error. Running perf mem record on this Sandy Bridge
> platform also result in the same error.

Which event are you trying to use?

On my Ivybrige machine using 3.14 I am able to get latency values with
	pe.config=0x1cd;        
        pe.config1=3;           
        pe.precise_ip=2;

Vince

Re: Load latency sampling on Sandy Bridge

[Manuel Selva]

The same event as the one you mentionned (I also checked that it's the same event than the  one used by perf mem record)

Setting precise-ip to 0 allows me to open the event and then mmap the result, but I don't have any sample.

Manu



> Le 29 oct. 2014 à 17:09, Vince Weaver <vincent.weaver@maine.edu> a écrit :
> 
>> On Wed, 29 Oct 2014, Manuel Selva wrote:
>> 
>> I am trying to setup load latency sampling on a Sandy Bridge processor (Cpu
>> familly 6, model 42, model name  Intel(R) Core(TM) i5-2520M CPU @ 2.50GHz)
>> running a 3.13 linux kernel.
>> 
>> After deep investigations made of reading of the Intel SDM, examples provided
>> by the pfm library and the source code of perf, I am still not able to get
>> sample. I think for now that my problem is related to the precise_ip parameter
>> of the  perf_event_attr structure.
>> 
>> Any time I change this value from 0 to 1 or from 0 to 2 (2 is the value to
>> enable PEBS isn't it ?) the perf_event_open system call failed with a "Not
>> supported operation" error. Running perf mem record on this Sandy Bridge
>> platform also result in the same error.
> 
> Which event are you trying to use?
> 
> On my Ivybrige machine using 3.14 I am able to get latency values with
>    pe.config=0x1cd;        
>        pe.config1=3;           
>        pe.precise_ip=2;
> 
> Vince

Re: Load latency sampling on Sandy Bridge

[Manuel Selva]

Yes, perf mem doesn't work on my Sandy Bridge. I should be missing 
something, maybe a kernel parameter somewhere, or something like that, 
but I don't know where searching ...

Manu

On 10/29/2014 07:53 PM, Vince Weaver wrote:
> On Wed, 29 Oct 2014, Manuel Selva wrote:
>
>> The same event as the one you mentionned (I also checked that it's the
>> same event than the one used by perf mem record)
>
> so perf mem record doesn't work for you either?
>
> I checked on my sandybridge-ep machine running Linux 3.14 and latency
> measurement seems to be working fine there too.
>
> Vince
>

Re: Load latency sampling on Sandy Bridge

[Vince Weaver]

On Wed, 29 Oct 2014, Manuel Selva wrote:

> The same event as the one you mentionned (I also checked that it's the 
> same event than the one used by perf mem record)

so perf mem record doesn't work for you either?

I checked on my sandybridge-ep machine running Linux 3.14 and latency 
measurement seems to be working fine there too.

Vince

Re: Load latency sampling on Sandy Bridge

[Vince Weaver]

On Wed, 29 Oct 2014, Manuel Selva wrote:

> Yes, perf mem doesn't work on my Sandy Bridge. I should be missing something,
> maybe a kernel parameter somewhere, or something like that, but I don't know
> where searching ...

well if you're getting operation not supported you are probably running 
into this chunk of code in arch/x86/kernel/cpu/perf_event.c


int x86_pmu_hw_config(struct perf_event *event)
{
        if (event->attr.precise_ip) {
                int precise = 0;

                /* Support for constant skid */
                if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
                        precise++;

                        /* Support for IP fixup */
                        if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2) 
                                precise++;
                }

                if (event->attr.precise_ip > precise)
                        return -EOPNOTSUPP;



and if you look, pebs_broken can be enabled on Sandybridge CPUs due to 
firmware errata.  Did you check your dmesg for messages starting with 
PEBS?

Vince

Re: Load latency sampling on Sandy Bridge

[Manuel Selva]

Yes the problem is there ! Thanks I am not enough "fluent" in kernel 
code to identify this. Here is my dmesg:

perf_event_intel: PEBS disabled due to CPU errata, please upgrade microcode

Does it mean that I must "upgrade my hardware" through the upgrade of 
microcode ? I never did such a thing.

Manu

On 10/29/2014 08:12 PM, Vince Weaver wrote:
> On Wed, 29 Oct 2014, Manuel Selva wrote:
>
>> Yes, perf mem doesn't work on my Sandy Bridge. I should be missing something,
>> maybe a kernel parameter somewhere, or something like that, but I don't know
>> where searching ...
>
> well if you're getting operation not supported you are probably running
> into this chunk of code in arch/x86/kernel/cpu/perf_event.c
>
>
> int x86_pmu_hw_config(struct perf_event *event)
> {
>          if (event->attr.precise_ip) {
>                  int precise = 0;
>
>                  /* Support for constant skid */
>                  if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
>                          precise++;
>
>                          /* Support for IP fixup */
>                          if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
>                                  precise++;
>                  }
>
>                  if (event->attr.precise_ip > precise)
>                          return -EOPNOTSUPP;
>
>
>
> and if you look, pebs_broken can be enabled on Sandybridge CPUs due to
> firmware errata.  Did you check your dmesg for messages starting with
> PEBS?
>
> Vince
>

Re: Load latency sampling on Sandy Bridge

[Vince Weaver]

On Wed, 29 Oct 2014, Manuel Selva wrote:

> perf_event_intel: PEBS disabled due to CPU errata, please upgrade microcode
> 
> Does it mean that I must "upgrade my hardware" through the upgrade of
> microcode ? I never did such a thing.

Usually the microcode is updated at boot time.  Usually your distribution 
will have some sort of microcode update package you can install that will 
do this for you automatically.

Vince

Re: Load latency sampling on Sandy Bridge

[Manuel Selva]

many thanks, it works !!!

On ubuntu: sudo apt-get install intel-microcode

-----
Manu

On 10/29/2014 09:02 PM, Vince Weaver wrote:
> On Wed, 29 Oct 2014, Manuel Selva wrote:
>
>> perf_event_intel: PEBS disabled due to CPU errata, please upgrade microcode
>>
>> Does it mean that I must "upgrade my hardware" through the upgrade of
>> microcode ? I never did such a thing.
>
> Usually the microcode is updated at boot time.  Usually your distribution
> will have some sort of microcode update package you can install that will
> do this for you automatically.
>
> Vince
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
>