Handle sched_wakeup in Linux perf importer about trace-viewer HOT 14 CLOSED

From [email protected] on May 24, 2012 16:38:09

Labels: GoodStartBug

from trace-viewer.

From [email protected] on May 24, 2012 16:38:19

Labels: -GoodStartBug GoodFirstBug

from trace-viewer.

From [email protected] on June 09, 2012 02:25:08

Cc: [email protected] [email protected]
Labels: -OS-All -Area-Internals -Pri-2 -Feature-DevTools-Tracing Tool-All Area-Import Pri-1

from trace-viewer.

From [email protected] on February 02, 2013 21:13:39

Is there an example file with sched_wakeup entries?

from trace-viewer.

From [email protected] on February 02, 2013 21:28:56

(this looks like a bug entered after the hackathon this summer, so not sure if still a desired feature)

The difference between the two is that sched_wakeup tells when work (a process) has been woken up. This means its ready to be run, but not necessarily running yet (as opposed to sleeping on some event).

I'm not sure sched_wakeup would be helpful then. Right now we just show what's actively running on the CPU. If we just switched over, we'd see all processes that are ready to run (including the one currently running) at once.

from trace-viewer.

From [email protected] on February 02, 2013 21:29:43

I should have said the difference between sched_wakeup and sched_switch, which I think we use now.

from trace-viewer.

From [email protected] on February 03, 2013 15:25:38

test_data/android_systrace.html has sched_wakeup. I think that trace is from a JOP40D build, though my memory may be wrong.

We definitely want to fix this: sched_wakeup happens when a thread goes from sleeping to runnable state. Right now, we incorrectly report that the thread is still sleeping, which is really bad.

This shoudl be a matter of, in the sched parser, going and telling the thread state tracker that the state did change to runnable.

from trace-viewer.

From [email protected] on February 07, 2013 19:59:48

linux_perf_sched_parser.js has handlers registered for both sched_switch and sched_wakeup. It looks like _wakeup is doing the marking as runnable (this.importer.markPidRunnable(ts, pid, comm, prio)).

Don't think there is anything to do for this bug. Closing, please re-open if I'm missing something.

Status: Invalid
Owner: [email protected]

from trace-viewer.

From [email protected] on February 08, 2013 10:21:53

The markPidRunnable method is not implemented. I implemented it last November, but unfortunately I accidentally deleted the source tree it was in before uploading it. It wasn't too difficult, but it did require some changes to buildPerThreadCpuSlicesFromCpuState in linux_perf_importer.js IIRC.

Status: Assigned

from trace-viewer.

From [email protected] on February 09, 2013 06:46:38

Summary: Implement markPidRunnable to complete sched_wakeup handling (was: Handle sched_wakeup in Linux perf importer)

from trace-viewer.

From [email protected] on February 14, 2013 19:34:08

Today I found the change I did a while back for this bug. It's awaiting review at https://codereview.appspot.com/7303094/

from trace-viewer.

From [email protected] on February 15, 2013 16:33:33

Fixed by https://codereview.appspot.com/7303094/

Owner: [email protected]

from trace-viewer.

From [email protected] on February 15, 2013 16:33:51

Status: Fixed

from trace-viewer.

From [email protected] on February 15, 2013 16:34:49

Awesome. Thanks a lot for digging this stuff up.

from trace-viewer.