[PATCH] perf tools: Unset perf_event_attr::freq when period term is set

From: Jiri Olsa
Date: Sat Aug 08 2015 - 13:12:24 EST


On Fri, Aug 07, 2015 at 12:38:43PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Thu, Aug 06, 2015 at 03:44:53PM -0400, kan.liang@xxxxxxxxx escreveu:
> > From: Kan Liang <kan.liang@xxxxxxxxx>
> > Here are some examples and test results.
> >
> > 1. Comparing the elapsed time and perf.data size from "kernbench -M -H".
> >
> > The test command for FULL callgraph and time support.
> > "perf record -e
> > '{cpu/cpu-cycles,period=100000/,cpu/instructions,period=20000/p}'
> > --call-graph fp --time"
>
> Jiri, while testing this I noticed that the message for EINVAL when
> using the cpu// syntax (per-event settings) is cryptic:
>
> [root@zoo ~]# perf record -e 'cpu/cpu-cycles,call-graph=fp,time,period=100000/p' ls
> Error:
> The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cpu/cpu-cycles,call-graph=fp,time,period=100000/p).
> /bin/dmesg may provide additional information.
> No CONFIG_PERF_EVENTS=y kernel support configured?
>
> Whereas if we use -F, it is much, much clearer, telling the user exactly
> what is failing and what needs to be done to make it work:
>
> [root@zoo ~]# perf record -F 100000 -e cpu/cpu-cycles/ usleep 1
> Maximum frequency rate (25000) reached.
> Please use -F freq option with lower value or consider
> tweaking /proc/sys/kernel/perf_event_max_sample_rate.
> [root@zoo ~]#
>
> Hope this is something easy to wire up, given your event parsing kung foo
> skillz...

my kungu foo found there was actually another issue ;-)

we did not clear up attr->freq bit.. so the sample_period
was handled as sample_freq value.. please check the patch
below

anyfoo, now period setup can raise sky high (which is correct)
and once it meets the God it shows:

[jolsa@krava perf]$ ./perf record -e 'cpu/cpu-cycles,call-graph=fp,time,period=1000000000000000000000000/' ls
event syntax error: '..time,period=1000000000000000000000000/'
\___ parser error

I'll check if we could add some default error message when value
crosses the type maximum.. which now falls to parser error

jirka


---
We need to unset 'perf_event_attr::freq' bit (default 1) when
'period' term is specified within event definition like:

-e 'cpu/cpu-cycles,call-graph=fp,time,period=100000'

otherwise it will handle the period value as frequency
(and fail if it crossed the maximum allowed frequency value).

Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
Link: http://lkml.kernel.org/n/tip-7f3gjwbekakhyxr47wvin4zb@xxxxxxxxxxxxxx
---
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index d902cad4ce10..4c779d23b1d7 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -623,6 +623,7 @@ static void apply_config_terms(struct perf_evsel *evsel,
switch (term->type) {
case PERF_EVSEL__CONFIG_TERM_PERIOD:
attr->sample_period = term->val.period;
+ attr->freq = 0;
break;
case PERF_EVSEL__CONFIG_TERM_TIME:
if (term->val.time)
--
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/