Re: [PATCH 2/2] perf, tools: Add documentation for perf user event parameters

From: Jiri Olsa
Date: Thu Oct 13 2016 - 07:43:22 EST


On Wed, Oct 12, 2016 at 02:02:07PM -0700, Andi Kleen wrote:
> From: Andi Kleen <ak@xxxxxxxxxxxxxxx>
>
> There was no proper documention for the user parameters for
> events that the perf parser implements. Document them all
> in the perf list manpage.
>
> Signed-off-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> ---
> tools/perf/Documentation/perf-list.txt | 42 ++++++++++++++++++++++++++++++++++
> 1 file changed, 42 insertions(+)
>
> diff --git a/tools/perf/Documentation/perf-list.txt b/tools/perf/Documentation/perf-list.txt
> index 41857cce5e86..511adc684148 100644
> --- a/tools/perf/Documentation/perf-list.txt
> +++ b/tools/perf/Documentation/perf-list.txt
> @@ -119,6 +119,48 @@ be specified as
>
> perf stat -e cpu/event=0xa8,umask=0x1,name=LSD.UOPS_CYCLES,cmask=1/ ...
>
> +EVENT PARAMETERS
> +----------------
> +
> +The pmu// syntax in perf uses parameters to configure the event for the
> +PMU.
> +
> +For a list of available parameters for each PMU please use
> +
> + ls /sys/devices/*/format
> +
> +Common parameters are
> +
> +. event=NUM: Event code
> +. umask=NUM: Event umask
> +
> +In addition perf supports some extra parameters to further configure then
> +event
> +
> +. name=name: Use name to display the event in perf
> +. period=NUM: Use sampling period NUM for the event. Similar to the -c parameter,
> +but only applies to this event. Only valid for perf record and top.
> +. freq=NUM: Use sampling frequency NUM for the event. Similar to the -c parameter,
> +but only applies to this event. Only valid for perf record and top.
> +. branch_type=TYPE: Enable last branch sampling for event. TYPE is the same
> +as the argument to perf record's -j option. Use any to be equivalent to -b.
> +when no is specified branch sampling is disabled for this event. Only valid
> +for perf record and top.
> +. time=NUM: When NUM is non zero collect a time stamp for the event.
> +. call-graph=type: Enable call graph mode for the event. When no is specified call
> +graph is disabled. The parameters are the as to the --call-graph option of
> +perf record. Only valid for perf record and top.
> +. stack-size=BYTES: For dwarf unwinding call graph mode limit the number of stack
> +bytes collected to BYTES.
> +. max-stack=NUM: For call graph mode limit the number of entries to NUM.
> +. inherit: Inherit the event to child processes.
> +. no-inherit: Do not inherit the event to child processes.
> +. overwrite: Enable overwrite mode in the perf buffer.
> +. no-overwrite: Do not enable overwrite mode in the perf buffer.
> +. config=NUM: Configure config parameter in perf event attribute.
> +. config1=NUM: Configure config1 parameter in perf event attribute.
> +. config2=NUM: Configure config2 parameter in perf event attribute.

we already have some description of these in perf-record.. we might want
to keep just one and have the other one point to it

thanks,
jirka