1309714 - Investigate what needs doing to submit GPU crash reports

Status: RESOLVED DUPLICATE of bug 1317972

(Core :: Graphics, defect, P3)

Comment 1

[David Anderson [:dvander] - inactive, e-mail if emergency]

Confirmed that manual submissions work: https://crash-stats.mozilla.com/report/index/48606713-b012-4139-a8fb-739612161101

Bonus, all the app notes and everything seem to be in order. So I guess this bug can be about whether we want UI, and if so, what it should look like.

Comment 2

[Milan Sreckovic [:milan] (needinfo for best results)]

Note for Mike and George to sort this out, whether we want the UI or not.  Given that we're going for a seamless experience when the GPU process crashes, it seems vaguely wrong to pop something in front of the user telling them things went bad. On the other hand, we do want to see these reports - Mike, do we have any idea as to the success of the "submit these old reports" UI that pops up?

Comment 3

[George Wright (:gw280) (needinfo me!)]

It was on my todo list for today to speak to Mike about this :) I'll loop back with the fruits of that discussion.

Comment 4

[Mike Conley (:mconley) (:⚙️)]

(In reply to Milan Sreckovic [:milan] from comment #2)
> reports - Mike, do we have any idea as to the success of the "submit these
> old reports" UI that pops up?

We're not currently collecting Telemetry on how many users have opted in to sending backlogged crash reports, so we don't currently know. I believe adding such Telemetry is more or less trivial, assuming data peer approval.

Out of curiosity, what's the most useful part of the crash report? If the stack is what's useful, it might be worth pursuing gsvelto's work in bug 1280469 et al and making it work for the compositor process. I believe the advantage here is that while we don't get the whole crash report, we get the symbolicated crash stack for free, since it's sent in the crash Telemetry ping.

If, however, the full crash report (including the more sensitive data inside it) is more valuable, then I suspect we'll have to go the same route with traditional crash reports: show UI and offer auto-submission.

Comment 5

[David Anderson [:dvander] - inactive, e-mail if emergency]

I would say 90% of the useful data is captured in the stack trace. The metadata is the last 10%, but most of that is already in the telemetry environment.

Comment 6

[George Wright (:gw280) (needinfo me!)]

OK, here's what I think we should do, having now spoken to Mike.

- Ensure the client-side stack walking functionality works in the GPU process, and we will get the full symbolicated stack sent in the telemetry ping. This doesn't contain any user sensitive data, so we're good for that.

- Ensure the more detailed crash reports are stored in the same location as existing chrome/content crash reports. Do not prompt for crash report submission at the time of crash, but hook into the existing stuff to prompt the user at next session startup to see if they're willing to send the archived crash reports. This is the same behaviour for content process crashes where the crashing content process is not backing the current tab.

Comment 7

[George Wright (:gw280) (needinfo me!)]

I think we can dupe this off bug 1317972. Probably shouldn't have filed the other bug, but most of the real discussion is happening there now.