bashi@: Would you take a look at this?

Some tips to help debugging this: when you open the record dialog, the browser process under the hoods starts and stop tracing very quickly. This is to warmup the categories in all processes and discover the category names to populate the checkboxes.

The "Stop" is asynchronous: the browser process sends a TracingMsg_EndTracing IPC message and then waits all the child processes to reply with a TracingHostMsg_EndTracingAck IPC. 
So I suspect that what's happening is that MC causes the child's  ChildTraceMessageFilter::OnEndTracing() to somehow never complete and never send back the EndTracingAck, causing the browser to wait forever (or maybe until some long timeout is hit).
Now the question is: what is not responding and why?

Maybe the fact that TraceLog::FlushInternal in turn posts a message to all task runners of the local process (well, on any task runner which has at least an event, which realistically means all task runners) and expects them to reply. So if any of the taskrunners are kept frozen, the child it will never ack to the browser, and in turn it will keep the tracing UI in that suspended state.

Thanks Primiano for the tips. It's really helpful.

+tasak@

I'm not familiar with actual suspend logic in renderer (MC just call SuspendRenderer() and ResumeRenderer()) but it seem that we should resume scheduler when the renderer receives TracingMsg_EndTracing. There will be other messages that we should resume renderers.

Ironically, TraceLog::FlushInternal has a timeout mechanism [1] to prevent what I described in #2 (waiting forever for a thread task runner to flush).

However the timeout is rendered itself as a PostDelayedTask. So if you are preventing delayed tasks to run, you are also killing the timeout mechanism that prevent this from happening in the first place :)

[1] https://cs.chromium.org/chromium/src/base/trace_event/trace_log.cc?rcl=0&l=905

Suspending renderers with MC seems broken and we need to make some changes in MC messaging between renderers and browser. I'll disable suspending tentatively.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/b1536fd7e1fa4dd7559e2c12e290e79cf949e9fe

commit b1536fd7e1fa4dd7559e2c12e290e79cf949e9fe
Author: bashi <bashi@chromium.org>
Date: Tue Dec 20 04:49:30 2016

Disable suspending renderer when memory coordinator is enabled

Just calling SuspendRenderer() in OnMemoryStateChange() breaks
many things. We need a reliable and consistent way to suspend
renderers. Until we come up with a solution, disable renderer
suspending when memory coordinator is enabled. In follow-up
CLs I'll add some tests to prevent this kind of regressions.

BUG= 675735 , 675811 

Review-Url: https://codereview.chromium.org/2590073002
Cr-Commit-Position: refs/heads/master@{#439708}

[modify] https://crrev.com/b1536fd7e1fa4dd7559e2c12e290e79cf949e9fe/content/renderer/render_thread_impl.cc

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/ff1e6242e014db17e8bb736741595705c81dc774

commit ff1e6242e014db17e8bb736741595705c81dc774
Author: tasak <tasak@google.com>
Date: Thu Dec 22 04:53:55 2016

Disable PurgeAndSuspend when MemoryCoordinator is enabled.

PurgeAndSuspend depends on OnMemoryStateChange(). However,
always suspending renderer in OnMemoryStateChange(), it breaks
many things. So only enabling such suspend / resume feature when
only PurgeAndSuspend is enabled.

BUG= 675735 ,  675811 

Review-Url: https://codereview.chromium.org/2595813002
Cr-Commit-Position: refs/heads/master@{#440340}

[modify] https://crrev.com/ff1e6242e014db17e8bb736741595705c81dc774/content/renderer/render_thread_impl.cc

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/d224a9dddebc82b819515e4d9c35737c7a92e9d0

commit d224a9dddebc82b819515e4d9c35737c7a92e9d0
Author: tasak <tasak@google.com>
Date: Wed Jan 18 00:44:29 2017

Stop suspending renderer and changing purge interval to 20min

- Since SuspendRenderer sometimes causes very long style recalc and layout (c.f. youtube.com/), disable tab suspension and use default background task throttle.
- also changed time-to-resuspension to 20min. This means, purge background tab's memory every 20min(to be precise 20min+10sec).

- the original document about purge+suspend is
https://docs.google.com/document/d/1EgLimgxWK5DGhptnNVbEGSvVn6Q609ZJaBkLjEPRJvI/edit?usp=sharing
- the difference between the original purge+suspend and purge+suspend with this patch is
https://docs.google.com/document/d/1qIrXsi9BuoAgNu6lVGTcGhbAm2TLPRb_JZ0OhynLjR0/edit?usp=sharing

BUG= 675735 , 607077

Review-Url: https://codereview.chromium.org/2624063002
Cr-Commit-Position: refs/heads/master@{#444209}

[modify] https://crrev.com/d224a9dddebc82b819515e4d9c35737c7a92e9d0/chrome/browser/memory/tab_manager.cc
[modify] https://crrev.com/d224a9dddebc82b819515e4d9c35737c7a92e9d0/chrome/browser/memory/tab_manager_unittest.cc
[modify] https://crrev.com/d224a9dddebc82b819515e4d9c35737c7a92e9d0/content/renderer/render_thread_impl.cc
[modify] https://crrev.com/d224a9dddebc82b819515e4d9c35737c7a92e9d0/content/renderer/render_thread_impl.h

Tested on windows 7 using chrome M57 #57.0.2986.0 and followed the steps:

(1) Enabled the flag MemoryCoordinator.
(2) Opened lots of tabs, none of the tabs crashed.
(3) Opened chrome://tracing.
(4) Hit Record and pop up displayed named "record a new trace "

Attached screencast for reference.

@bashi--- Could you please check the attached screencast and let us know if we had missed out any steps in verifying the issue .

Thanks!

Deleted: CL.mp4 2.5 MB

	CL.mp4 2.5 MB View Download

FWIW  issue 675269  covers the MemoryCoordinator tab-suspend behaviour that triggered the hung renderer in this case, but that particular cause is incidental to this bug.

This bug is for chrome://tracing not coping correctly with "hung" renderer processes.

Not sure if I did fully catchup the situation. Are you saying that even after bashi reverted the suspend, you still experience tracing getting stuck?

#10: Sorry missed comments. Could you try to repro with ToT if you have time?

FWIW, I've been using tracing on a custom chromium w/tab suspending for a while and it works well. Not sure the root cause of this bug though.

We gave up suspending tabs for now and this shouldn't occur anymore. Let me close this. Please re-open this if this is still happening on ToT.