Re: [perfmon2] [PATCH] Uncore support

[stephane e. <er...@go...>]

On Sat, Feb 26, 2011 at 12:23 AM, Arun Sharma <as...@fb...> wrote:
> This is for supporting uncore events.
>
> I tested this diff with Lin Ming's uncore patch against the kernel
> http://thread.gmane.org/gmane.linux.kernel/1071064
>
> diff --git a/include/perfmon/perf_event.h b/include/perfmon/perf_event.h
> index defc88e..de5691e 100644
> --- a/include/perfmon/perf_event.h
> +++ b/include/perfmon/perf_event.h
> @@ -29,6 +29,7 @@ enum perf_type_id {
>        PERF_TYPE_HW_CACHE              = 3,
>        PERF_TYPE_RAW                   = 4,
>        PERF_TYPE_BREAKPOINT            = 5,
> +       PERF_TYPE_UNCORE                = 6,
>        PERF_TYPE_MAX,                  /* non-ABI */
>  };
>
> diff --git a/lib/pfmlib_intel_nhm_unc.c b/lib/pfmlib_intel_nhm_unc.c
> index 7fcdf28..1221190 100644
> --- a/lib/pfmlib_intel_nhm_unc.c
> +++ b/lib/pfmlib_intel_nhm_unc.c
> @@ -100,17 +100,23 @@ pfm_wsm_unc_detect(void *this)
>        return PFM_SUCCESS;
>  }
>
> +#ifdef __linux__
> +#include "pfmlib_perf_event_priv.h"
>  static int
>  intel_nhm_unc_perf_encode(void *this, pfmlib_event_desc_t *e)
>  {
> -#if 0
>        struct perf_event_attr *attr = e->os_data;
> -       /* not yet supported by perf_event */
> -       attr->type = PERF_TYPE_RAW;
> +       attr->type = PERF_TYPE_UNCORE;
>        attr->config = e->codes[0];
> -#endif
> -       return PFM_ERR_NOTSUPP;
> +       /* Can't exclude anything */
> +       attr->exclude_user = 0;
> +       attr->exclude_kernel = 0;
> +       attr->exclude_hv = 0;
> +       attr->exclude_idle = 0;
> +
> +       return PFM_SUCCESS;
>  }
> +#endif
>
>  static int
>  intel_nhm_unc_os_encode(void *this, pfmlib_event_desc_t *e)
> diff --git a/lib/pfmlib_perf_event.c b/lib/pfmlib_perf_event.c
> index 879efb6..1c7a90b 100644
> --- a/lib/pfmlib_perf_event.c
> +++ b/lib/pfmlib_perf_event.c
> @@ -83,6 +83,7 @@ pfmlib_perf_event_encode(void *this, const char *str, int dfl_plm, void *data)
>
>        e.osid = os->id;
>        e.os_data = attr;
> +       e.dfl_plm = dfl_plm;
>
That bit could be passed but I think it is useless in most cases,
perf_event controls priv levels.

Now the one thing that I need to handle better is that for uncore-type
of PMU, dfl_plm better be
PLM0|PLM3. Otherwise, libpfm4 needs to return an error. I want
something explicit so that
people understand that uncore PMU ALWAYS monitor at both user +
kernel. Silent errors
are nasty, so I want to try and avoid that as much as possible.


>        ret = pfmlib_parse_event(str, &e);
>        if (ret != PFM_SUCCESS)
>