Beta 94 GPU process OOM crashes increased as compared to 93
Categories
(Core :: Graphics, defect, P2)
Tracking
()
Tracking | Status | |
---|---|---|
firefox-esr78 | --- | unaffected |
firefox-esr91 | --- | unaffected |
firefox93 | --- | unaffected |
firefox94 | + | wontfix |
firefox95 | --- | affected |
People
(Reporter: jimm, Unassigned)
References
(Depends on 1 open bug, Blocks 1 open bug)
Details
We picked this up through crash ping telemetry - the increase is around 20%.
https://sql.telemetry.mozilla.org/queries/79062/source?p_channel=beta#196450
The change occurred around September 24th.
Suspect landings -
https://bugzil.la/1731748 — RESOLVED (nical) — Handle GL_OUT_OF_MEMORY in WebRender
https://bugzil.la/1731929 — RESOLVED (tnikkel) — make cumulative and extra resolution make sense
Reporter | ||
Comment 1•3 years ago
|
||
[Tracking Requested - why for this release]:
Worth tracking. Graphics team is looking into this.
Reporter | ||
Updated•3 years ago
|
Comment 2•3 years ago
|
||
Timothy, bug 1731929 is closer to where we saw the regression on nightly:
https://sql.telemetry.mozilla.org/queries/82430/source?p_channel=nightly#204202
Here's a rough pushlog from around that time: https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=e7f3e1f5b625&tochange=a3f0791a87fd3d2419768eff6da663361c9a3171
Updated•3 years ago
|
Reporter | ||
Updated•3 years ago
|
Comment 3•3 years ago
|
||
I can triple check the patch but bug 1731929 should not have changed anything, the numerical values we see before and after that bug should be the same: we changed from dividing by x to multiplying by 1/x. The change was purely to make the values stored with names that make logical sense to humans and so that the calculations we do with those values make sense. This was done so that further changes to that code would actually make sense (which was done in other bugs).
Reporter | ||
Updated•3 years ago
|
Comment 4•3 years ago
|
||
We talked about this in our triage meeting yesterday. Our best guess as to what's happening here is that some crashes that we're not being classified as OOMs are now. Overall crash volume hasn't appreciably changed. We'll continue to investigate but are not planning to do anything for 94.
Comment 5•3 years ago
|
||
This bug might also explain these new content process OOMs with webrender signatures that started around September 23 and then increased around October 9:
https://bugzilla.mozilla.org/show_bug.cgi?id=1738169
https://bugzilla.mozilla.org/show_bug.cgi?id=1705430
https://bugzilla.mozilla.org/show_bug.cgi?id=1738167
https://bugzilla.mozilla.org/show_bug.cgi?id=1738171
Reporter | ||
Comment 6•3 years ago
|
||
I think this signature is the source -
Bug 1692514 - Crash in [@ OOM | large | mozalloc_abort | webrender::renderer::upload::upload_to_texture_cache]
Reporter | ||
Updated•3 years ago
|
Updated•2 years ago
|
Description
•