Re: [PATCH] perf sched: kill time stamp discrepancy between script and latency

From: Joonwoo Park
Date: Tue Oct 04 2016 - 13:11:30 EST




On 10/04/2016 05:04 AM, Namhyung Kim wrote:
On Mon, Oct 03, 2016 at 03:04:48PM -0700, Joonwoo Park wrote:


On 09/30/2016 10:15 PM, Namhyung Kim wrote:
Hi Joonwoo,

On Wed, Sep 28, 2016 at 07:25:26PM -0700, Joonwoo Park wrote:
Perf sched latency is handy to find out the maximum sched latency and
the time stamp of the event. After running sched latency, if a found
latency looks suspicious it's quite reasonable to run perf script
subsequently and search with the time stamp given by perf sched latency
to continue further debugging. However, at present, it's possible the
time stamp given by perf sched latency cannot be found in the trace
output by perf script because perf sched latency converts the event
time from ns to ms as double float and prints it with printf which
does banker's rounding as opposed to perf script doesn't.

For example:

0x750ff0 [0x80]: event: 9
<snip>
2 1858303049520 0x750ff0 [0x80]: PERF_RECORD_SAMPLE(IP, 0x1): 15281/15281: 0xffffffff8162a63a period: 1 addr: 0
... thread: hackbench:15281
<snip>

$ perf sched -i perf.data latency | grep hackbench
hackbench:(401) + 3539.283 ms | 23347 | avg: 7.286 ms | max: 829.998 ms | max at: 1858.303050 s

$ perf script -i perf.data | grep "1858\.303050"

$ perf script -i perf.data | grep "1858\.303049"
hackbench 15281 [002] 1858.303049: sched:sched_switch: prev_comm=hackbench prev_pid=15281 prev_prio=120 prev_state=D ==> next_comm=hackbench next_pid=15603 next_prio=120

Fix perf latency to print out time stamp without rounding to avoid such
discrepancy.

Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Ingo Molnar <mingo@xxxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
Cc: linux-kernel@xxxxxxxxxxxxxxx
Signed-off-by: Joonwoo Park <joonwoop@xxxxxxxxxxxxxx>
---

I was tempted to get rid of all u64 to double casting in the function
output_lat_thread but didn't because there is no data loss as of
today. Double float gives at least 15 significant decimal digits
precision while the function requires only 14 significant digits precision.

$ python -c "print(len(str(int(0xffffffffffffffff / 1e6))))"
14

tools/lib/traceevent/event-parse.h | 1 +
tools/perf/builtin-sched.c | 12 ++++++++++--
2 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/tools/lib/traceevent/event-parse.h b/tools/lib/traceevent/event-parse.h
index 9ffde37..f42703c 100644
--- a/tools/lib/traceevent/event-parse.h
+++ b/tools/lib/traceevent/event-parse.h
@@ -174,6 +174,7 @@ struct pevent_plugin_option {

#define NSECS_PER_SEC 1000000000ULL
#define NSECS_PER_USEC 1000ULL
+#define MSECS_PER_SEC 1000ULL

enum format_flags {
FIELD_IS_ARRAY = 1,
diff --git a/tools/perf/builtin-sched.c b/tools/perf/builtin-sched.c
index afa0576..e5cf51a 100644
--- a/tools/perf/builtin-sched.c
+++ b/tools/perf/builtin-sched.c
@@ -1190,6 +1190,7 @@ static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_
int i;
int ret;
u64 avg;
+ u64 max_lat_at_sec, max_lat_at_msec;

Isn't it usec rathen than msec? :)

It's to contain three decimal digits which are msecs when 'max_lat_at' is
expressed in sec.
For example when max_lat_at = 1858303049520 which is 1858.3030495199998 sec,
max_lat_at_msec is meant to be 303.

??? But didn't you want to print it as 303049?

Looking at the code, the 'max_lat' is a latency which was printed in
msec (%9.3f ms) but the 'max_lat_at' is a timestamp which was printed
in sec (%13.6f s). This is confusing but they use effectively same
unit (usec) by using different number of digit after the period.


I must admit variable's name is bit misleading. Maybe just secs, msecs are
better?
Also just noticed u64 isn't needed for msecs. Will size down.

IIUC you wanted usec for timestamp rather than msec, aren't you?

Ugh. I was a dummy. My bad. Will fix this.





if (!work_list->nb_atoms)
return;
@@ -1212,11 +1213,18 @@ static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_

avg = work_list->total_lat / work_list->nb_atoms;

- printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
+ /*
+ * Avoid round up with printf to prevent event time discrepency
+ * between sched script and latency.
+ */
+ max_lat_at_sec = work_list->max_lat_at / NSECS_PER_SEC;
+ max_lat_at_msec = (work_list->max_lat_at -
+ max_lat_at_sec * NSECS_PER_SEC) / MSECS_PER_SEC;
+ printf("+%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %6lu.%06lu s\n",

Maybe you'd better to be in sync with the script code:

if (PRINT_FIELD(TIME)) {
nsecs = sample->time;
secs = nsecs / NSECS_PER_SEC;
nsecs -= secs * NSECS_PER_SEC;
usecs = nsecs / NSECS_PER_USEC;
if (nanosecs)
printf("%5lu.%09llu: ", secs, nsecs);
else
printf("%5lu.%06lu: ", secs, usecs);
}

Apart from variable name, I'm not quite sure what to sync because sched
doesn't print in nsecs.
Maybe you just wanted for variable names in sync rather than logic?

You found inconsistency between the two, and wanted to show same time.
So wouldn't it be better factor out the code and reuse it from the
both callsite? If you don't need nsec, you could make it an option.

But this is just a general suggestion and maybe it's not worth doing
here since the code is small.. So I won't insist on it strongly.

I see what you're suggesting. Will try to factor out the common part in the next revision.

Thanks,
Joonwoo


Thanks
Namhyung



(double)work_list->total_runtime / 1e6,
work_list->nb_atoms, (double)avg / 1e6,
(double)work_list->max_lat / 1e6,
- (double)work_list->max_lat_at / 1e9);
+ max_lat_at_sec, max_lat_at_msec);
}

static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
hosted by The Linux Foundation