1133781 - (perf-tool-crashes) [meta] Many tests crash after ending in the new performance tool

Status: RESOLVED FIXED

(DevTools :: Performance Tools (Profiler/Timeline), defect)

Comment 1

[Victor Porof [:vporof][:vp]]

Take a look at this gum run: https://treeherder.mozilla.org/#/jobs?repo=gum&revision=2af669f79b9e

Almost all failures are crashes, even though all the checks pass during the test.

Comment 2

[Victor Porof [:vporof][:vp]]

Looks like browser_perf-compatibility-04.js is the worst offender, so that could be a good starting point to reproduce/debug this.

Comment 3

[Jordan Santell [:jsantell] [@jsantell]]

For the new performance developer tools, we're seeing a lot from what we believe is the profiler. Tromey has done some debugging in bug 1133791 (comment #1), and seems to be some sort of JIT error. Do any of you have any idea what could be causing this, possible solutions, next steps? Once these crashes are solved, we'll flip the switch to enable in Nightly for the new profiler/timeline tool.

Thanks!

Comment 4

[Jan de Mooij [:jandem]]

Forwarding to Kannan as this is probably fallout from the new profiler backend.

Comment 5

[Victor Porof [:vporof][:vp]]

Attachment: debugging.patch

Attaching the patch I used with djvj to debug this. We ended up concluding that a normally dispatched event on the main thread isn't being executed (via `executeSoon`), because of the beachball, which can mean one of three things: 1. the event is removed from the dispatcher, 2. the dispatcher is blocked waiting for something, 3. the dispatcher itself died. Since the debugger server is making heavy use of those, it's entirely possible that while the clients are shutting down, some packets end up being stuck.

We'll be looking more into this on Monday.

I hope there's some delirium tremens in our keg.

Comment 6

[Shu-yu Guo [:shu]]

Clearing needinfo. Found iloop in _findOptimalTickInterval in devtools/shared/timeline/markers-overview.js

Comment 7

[Victor Porof [:vporof][:vp]]

Thanks shu. Will look at this tomorrow.

Comment 8

[Victor Porof [:vporof][:vp]]

Infinite loop was fixed in bug 1137503.

Comment 9

[Victor Porof [:vporof][:vp]]

Paul, these are the current outstanding crashes on gum: https://bugzilla.mozilla.org/showdependencytree.cgi?id=1133781&hide_resolved=1

Comment 10

[Victor Porof [:vporof][:vp]]

Talked to shu and I'm not panicking anymore. Removing needinfo