Re: [PATCH 0/3] perf/core, x86: unify perfctr bitmasks

[Robert Richter]

On 30.03.10 12:11:46, Stephane Eranian wrote:
> On Mon, Mar 29, 2010 at 6:36 PM, Robert Richter <robert.richter@xxxxxxx> wrote:
> > This patch set unifies performance counter bit masks for x86. All mask
> > are almost the same for all x86 models and thus can use the same macro
> > definitions in arch/x86/include/asm/perf_event.h. It removes duplicate
> > code. There is also a patch that reverts some changes of the big
> > perf_counter -> perf_event rename.
> >
> 
> But there are still fields which are unique to each vendor:
> - GUEST vs. HOST on AMD
> - ANY_THREAD on Intel.
> 
> For instance, I noticed that in
> 
> arch/x86/kernel/cpu/perf_event.c:__hw_perf_event_init():
> 
>       if (attr->type == PERF_TYPE_RAW) {
>                 hwc->config |= x86_pmu.raw_event(attr->config);
>                 if ((hwc->config & ARCH_PERFMON_EVENTSEL_ANY) &&
>                     perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
>                         return -EACCES;
>                 return 0;
>         }
> 
> Assumes ANY also exists on AMD processors. That is not the case.
> This check needs to be moved into an Intel specific function.

Generally, ARCH_PERFMON_EVENTSEL_* refers to:

 Intel 64 and IA-32 Architectures Software Developerâs Manual Volume
 3B: System Programming Guide, Part 2
 30.2 ARCHITECTURAL PERFORMANCE MONITORING
 Appendix A: Performance-Monitoring Events
 Appendix B: Model-Specific Registers (MSRs) 

and AMD64_EVENTSEL_* to:

 AMD64 Architecture Programmer's Manual Volume 2: System Programming
 13.3.1 Performance Counters

X86_* is generic.

If a feature is available from both vendors, the names shouln't be
changed. Instead the first introduced mask should be used (at least
this is my suggestion).

So, there are some ARCH_PERFMON_EVENTSEL_* masks that are Intel only,
which is true for ARCH_PERFMON_EVENTSEL_ANY. And indead, the code
should be checked for this. ARCH_PERFMON_EVENTSEL_ANY is always
cleared on AMD cpus, so this code is ok. Actually the bit is cleared
for *all* cpus in x86_pmu_raw_event(), the code was and is broken for
this.

-Robert

-- 
Advanced Micro Devices, Inc.
Operating System Research Center
email: robert.richter@xxxxxxx

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/