Running talos tests with --geckoProfile is very slow or crashes
Categories
(Core :: Gecko Profiler, defect, P1)
Tracking
()
Tracking | Status | |
---|---|---|
firefox71 | --- | fixed |
People
(Reporter: jdescottes, Assigned: mozbugz)
References
(Regression)
Details
(Keywords: regression)
Attachments
(1 file)
It seems like running talos tests locally with the --geckoProfile option is now either very slow (for small tests) or crashes (for bigger tests).
For instance on my machine (macos):
./mach talos-test --activeTests damp --subtest custom.debugger --cycles 1 --tppagecycles 1 --geckoProfile
finishes with a tab crash (no issue without --geckoProfile)
It crashes with the following exception in the commandline:
05:52:06 INFO - PID 86064 | dyld: Library not loaded: @executable_path/libmozglue.dylib
05:52:06 INFO - PID 86064 | Referenced from: /[...]/Nightly.app/Contents/MacOS/crashreporter.app/Contents/MacOS/minidump-analyzer
05:52:06 INFO - PID 86064 | Reason: no suitable image found. Did find:
05:52:06 INFO - PID 86064 | /Users/jdescottes/lib/libmozglue.dylib: stat() failed with errno=20
Also running ./mach talos-test --activeTests damp --subtest simple.debugger --cycles 1 --tppagecycles 1 --geckoProfile
finishes and opens the profiler in a new window but only after ~20 seconds
According to a bisect, it was regressed by Bug 1581049, more accurately: https://hg.mozilla.org/integration/autoland/rev/4001594f15dd
Reporter | ||
Comment 1•5 years ago
|
||
:gerald, any idea why we are getting this crash?
Assignee | ||
Comment 2•5 years ago
|
||
Thank you Julian.
I'm able to reproduce it locally, so hopefully I'll be able to find the cause...
Here it seems to regress in the previous revision https://hg.mozilla.org/integration/autoland/rev/1ceccc3c86a49e41 -- a more likely culprit to me than the simpler nearby revision you found (but as we discussed you were using artifact builds, so maybe that was not as precise.)
Assignee | ||
Comment 3•5 years ago
|
||
In some situations, entries may in fact take more than half the buffer size
(e.g., when duplicating a stack into a small temporary buffer).
So we now allow blocks to take the full buffer size -- but not more, as they
would start overwriting themselves!
Pushed by gtatum@mozilla.com: https://hg.mozilla.org/integration/autoland/rev/318c6aa896db Allow profiler entries up to the size of the buffer - r=gregtatum
Comment 5•5 years ago
|
||
bugherder |
Assignee | ||
Comment 6•5 years ago
|
||
Julian, the crashes should not happen anymore, starting with Nightly 71.0a1 (2019-09-20).
But I'm guessing tests being slow may still be an issue. Could you please try again, and if there is still a problem, please open a new bug with your experience of how slow tests are. Thank you.
Reporter | ||
Comment 7•5 years ago
|
||
Thanks a lot for fixing this!
I quickly measured the time needed to get profiles on my machine for:
- custom.debugger: 12s
- simple.debugger: 10s
I compared with FIREFOX_NIGHTLY_70_END, and I get the same timings.
So, for me it's back to normal and definitely usable.
Updated•5 years ago
|
Updated•2 years ago
|
Description
•