Re: Some help needed about operf

[Maynard J. <may...@us...>]

On 05/07/2013 05:20 AM, Gilles Allard wrote:
> Hello.
> 
> To analyze the behaviour of an application, I tried to use `operf' instead of 
> `opcontrol' that is deprecated ( operf-0.9.8 )
> Building Oprofile from source, with kernel headers package installed ( kernel 
> 3.1.5  running on a P4-Celeron processor ) was perfectly right and `opcontrol' 
> gave usable results for this application.
> 
> But, in exactly the same conditions ( kernel-3.1.5, P4-Celeron processor, 
> profiling for the default event ( GLOBAL_POWER_EVENTS ) ) `operf' failed to 
> start.
> My analisys of `operf' setup shows that the system call to `perf_event_open` 
> returns EINVAL; this return code  comes from the kernel function 
> `p4_validate_raw_event' that rejected the value of `attr.config' field in the 
> event structure : the value provided by `operf' is 285 ( 1 << 8 | 0x1d ) based 
> on the code of GLOBAL_POWER_EVENTS ( 29 i.e. 0x1d ) according to the content 
> of `/usr/share/events/i386/p4/events'
> There seems to be a mismatch between the value provided by operf and this 
> expected by the kernel.

For most architectures (including x86*), operf has been using the existing event codes as a basis for what we pass to the perf_events kernel subsystem.  Looking at the "legacy" oprofile kernel driver, I see special "p4" handling. There's a map of event codes to ESCR/CCCR values.  Unfortunately, the event codes used by the oprofile kernel driver don't match what perf_events kernel code expects.

We'll need to have some special event handling code for operf for P4 and the like -- potentially use libpfm to get the code.  *Andi and Suravee* are our Intel/x86_64 experts in the community, so I will let one of them drive this problem, but I can certainly help as needed.

-Maynard
> 
> Switching to kernel-3.6.11 then to kernel-3.8.10 didn't fix the problem ( 
> Oprofle package was rebuilt with the appropriate kernel headers package 
> installed ). Neither does the loading of the kernel module prior to starting 
> `operf'.
> And I didn't find any reference to such a problem when scanning the messages 
> posted to the Oprofile mailing list during the last 2 years.
> 
> So I dont think it's a real bug but rather some sort of a operf configuration 
> problem ( ? ).
> 
> Does anyone has a suggestion to help solving that problem ( operf usage, 
> configuration ... ) ?
> 
> Many thanks in advance for any answer.
> 
> G. Allard
> 
> 
> ------------------------------------------------------------------------------
> Learn Graph Databases - Download FREE O'Reilly Book
> "Graph Databases" is the definitive new guide to graph databases and 
> their applications. This 200-page book is written by three acclaimed 
> leaders in the field. The early access version is available now. 
> Download your free book today! http://p.sf.net/sfu/neotech_d2d_may
> _______________________________________________
> oprofile-list mailing list
> opr...@li...
> https://lists.sourceforge.net/lists/listinfo/oprofile-list
>