Re: [RFC -tip] perf, x86: Add PERF_COUNT_HW_NMI_WATCHDOG event v2

[Stephane Eranian]

On Tue, Jun 21, 2011 at 7:54 PM, Cyrill Gorcunov <gorcunov@xxxxxxxxx> wrote:
> On Tue, Jun 21, 2011 at 07:26:16PM +0200, Stephane Eranian wrote:
> ...
>> >
>> > ÂIt doesn't change things much. Of course I can put some specifics into
>> > .config but where is the guarantee some new generic event would not ever
>> > intersect with it. I know (for example) we could reserve -1ULL for this
>> > event but again where is the guarantee that it will never ever be used
>> > system wide in future for some different event?
>> >
>> I am not talking about a new generic PMU event. I am talking about
>> you hardcoding a raw event in the callback: type = PERF_TYPE_RAW,
>> config=0x003c.
>>
>
> watchdog.c is system-wide and even if it's not used by archs other
> than x86 doesn't mean it'll not in future ;) So we could encode
> this watchdog event as you mention (and I would check for this
> values and map it internally to bits I need, this btw will require me
> to check if this bits do not intersect with some already valid bits
> combination but it's a different problem) but than we _must_ be sure it
> never bring problems as new arch or consumer appear. So, no, I don't
> like this approach, it's fragile. But gimme some time -- I'll re-check
> all bits, probably there is some workaround.
>
I don't understand what possible conflict bits you're talking about

If in arch/x86/kernel/cpu/perf_event.c I add a callback which
checks:

    if (netburst_host) {
       attr.type = PERF_TYPE_RAW
       attr.config = 0x3c
   } else {
      attr.type = PERF_TYPE_HARDWARE
      attr.config = PERF_COUNT_HW__CPU_CYCLES
  }

I simply don't think PERF_COUNT_HW_NMI_WATCHDOG should
ever be exposed to users.

> Â Â Â ÂCyrill
>
--
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/