1455491 - Crash in static <T> core::result::unwrap_failed<T> | static bool webrender::scene_builder::SceneBuilder::process_message

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect, P1)

Description

[Darkspirit]

New on Socorro.

Only this one:
bp-80395013-b358-4659-8d88-1fb740180420 build 20180419100148 Win10
> called `Result::unwrap()` on an `Err` value: "SendError(..)"

Comment 1

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

Async scene building?

Comment 2

[Kartikaya Gupta (email:[e-mail])]

Yeah, I think so.

Comment 3

[Marcia Knous [:marcia]]

static <T> core::result::unwrap_failed<T> | static bool webrender::scene_builder::SceneBuilder::process_message signature has 157 crashes/107 installs so far.

Comment 4

[Kartikaya Gupta (email:[e-mail])]

So it looks like what's happening here is that the unwrap at [1] is failing, presumably because the RB thread has shut down and scene_rx is dead. I guess this could happen if the scene builder thread is processing a scene build concurrently with the RB thread getting a shutdown message.

[1] https://searchfox.org/mozilla-central/rev/fcd5cb3515d0e06592bba42e378f1b9518497e3d/gfx/webrender/src/scene_builder.rs#163

Comment 5

[Kartikaya Gupta (email:[e-mail])]

I have a tentative fix in upstream WR. However I haven't been able to reproduce this so I can't confirm it'll fix the problem. Based on the failure mechanism I expect that this would reproduce if somebody closes a window while we're in the middle of a long scene build (i.e. content that's expensive to deal with in WR).

Comment 6

[Darkspirit]

Open https://jrmuizel.github.io/implementation-tests/dl-test-transform.html and click X to close the window.
bp-7a708151-242b-4cb5-82e7-e45360180423

Comment 7

[Darkspirit]

mozregression --good 2018-04-18 --bad 2018-04-21 --pref gfx.webrender.all:true startup.homepage_welcome_url:'https://jrmuizel.github.io/implementation-tests/dl-test-transform.html'
> 6:48.44 INFO: Last good revision: 69008e17de668cca67e7a27c73245961fe285d23
> 6:48.44 INFO: First bad revision: c3da365d2916344fac340eeab98319089915e46f
> 6:48.44 INFO: Pushlog:
> https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=69008e17de668cca67e7a27c73245961fe285d23&tochange=c3da365d2916344fac340eeab98319089915e46f

> c3da365d2916	Kartikaya Gupta — Bug 1391318 - Automatically enable async scene building with webrender.all. r=jrmuizel

Comment 8

[Kartikaya Gupta (email:[e-mail])]

(In reply to Jan Andre Ikenmeyer [:darkspirit] from comment #6)
> Open https://jrmuizel.github.io/implementation-tests/dl-test-transform.html
> and click X to close the window.

Thanks! I can reproduce intermittently with this, and I can't using the try build at 
https://treeherder.mozilla.org/#/jobs?repo=try&revision=11b4ee7a033bebafea1f2dc1326dea8ae9c7fbc5
 which has the upstream WR fix. So hopefully the patch worked :)

Comment 9

[Darkspirit]

(In reply to Kartikaya Gupta (email:[e-mail]) from comment #8)
> So hopefully the patch worked :)

Tested 9 times and didn't crash.

Comment 11

[Kartikaya Gupta (email:[e-mail])]

This should be fixed in the Apr 27 nightly onwards. I'll leave this open for a few days to verify via crash-stats.

Comment 12

[Kartikaya Gupta (email:[e-mail])]

No crashes since Apr 26.