1402585 - webrendest: (nearly all) text got invisible today

Status: VERIFIED FIXED

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

Description

[Darkspirit]

Attachment: Screenshot_20170923_121747.png

Nightly 58 x64 20170922220129 de_DE @ Debian Testing (KDE / Radeon RX480)
layers hw accel + webrender + webrendest + layersfree

Disabling webrendest fixes this.

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Can you get a regression window using mozregression?

Comment 2

[Darkspirit]

I got different results. The last one is the best, I think.

1. run mozregression
2. set these prefs, then type "retry":
> layers.acceleration.force-enabled;true
> gfx.webrender.enabled;true
> gfx.webrender.layers-free;true
> gfx.webrendest.enabled;true
3. proceed as usual (good/bad)


> mozregression --good 2017-09-21 --bad 2017-09-23 --profile-persistence clone-first
> [...]
> 7:37.39 INFO: Last good revision: 80f6fc604e5a2cf89d55292b8fb80206fc6c87a4
> 7:37.39 INFO: First bad revision: 6ff0f49d6e25827d4615136a0100571acf4b9eac
> 7:37.40 INFO: Pushlog:
> https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=80f6fc604e5a2cf89d55292b8fb80206fc6c87a4&tochange=6ff0f49d6e25827d4615136a0100571acf4b9eac
-> doesn't make sense

If text was invisble (on tabs/about:newtab/about:support), I flipped webrendest to false and everything was fine. Then it flipped it to false and it was broken again:
> mozregression --good 2017-09-22 --bad 2017-09-23 --profile-persistence clone-first
> [...]
>  3:14.60 INFO: Last good revision: 3d72fdb0e561ea59d9e5850c3e71367dbb8a7148
>  3:14.60 INFO: First bad revision: 9d1fcbc2fdd58011903aa1b987264d58dc1a06b5
>  3:14.60 INFO: Pushlog:
> https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=3d72fdb0e561ea59d9e5850c3e71367dbb8a7148&tochange=9d1fcbc2fdd58011903aa1b987264d58dc1a06b5
-> doesn't make sense

I retried a "bad" build with layersfree disabled (=only layers force accel + webrender + webrendest). Same results.
Compositing was always WebRender, avaliable by user/enabled by pref and there were no crashes.

> 11:24.60 INFO: Running autoland build built on 2017-09-22 14:43:10.038000, revision 66f49bef
At some point I get "Nightly Safe Mode: Nightly closed unexpectedly while starting. Thight might be caused by add-ons or other problems. You can try to resolve the problem by troubleshooing in Safe Mode. [Start in Safe Mode]."
Then I tryped "retry" and this came again and again.
> https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=176e16629030869dfd6e31e240006f9ea70be2ba&tochange=43983cf4138e13ba21b1a4cdcf21ab56336cedff
"skip". "1". Got that message again. "bad". msg again. "bad".
> 14:35.85 INFO: Last good revision: 176e16629030869dfd6e31e240006f9ea70be2ba
> 14:35.85 INFO: First bad revision: 754a0d9a44ea5208b6b204e06033e008a6c17b52
> 14:35.85 INFO: Pushlog:
> https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=176e16629030869dfd6e31e240006f9ea70be2ba&tochange=754a0d9a44ea5208b6b204e06033e008a6c17b52
-> There is a webrender update.

Let's verify this. Got the "Same Mode" popup at some point again. This time, I entered "bad".
> mozregression --good 2017-09-21 --bad 2017-09-23 --profile-persistence clone-first
> [...]
> 3:55.99 INFO: Last good revision: 5a63d8457a2a69a2ad54a50765bff412580df2a4
> 3:55.99 INFO: First bad revision: 14db7c0bcf9ae86c9ec2cc9d3c249a42e459a2a9
> 3:55.99 INFO: Pushlog:
> https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=5a63d8457a2a69a2ad54a50765bff412580df2a4&tochange=14db7c0bcf9ae86c9ec2cc9d3c249a42e459a2a9

This time completely without layersfree (=only layers force accel + webrender + webrendest). This time I didn't got a "Safe Mode" popup:
> mozregression --good 2017-09-21 --bad 2017-09-23 --profile-persistence clone-first
> [...]
>  6:15.41 INFO: Narrowed inbound regression window from [176e1662, 66f49bef] (3 builds) to [176e1662, 754a0d9a] (2 builds) (~1 steps left)
> 6:15.41 INFO: No more inbound revisions, bisection finished.
> 6:15.41 INFO: Last good revision: 176e16629030869dfd6e31e240006f9ea70be2ba
> 6:15.41 INFO: First bad revision: 754a0d9a44ea5208b6b204e06033e008a6c17b52
> 6:15.41 INFO: Pushlog:
> https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=176e16629030869dfd6e31e240006f9ea70be2ba&tochange=754a0d9a44ea5208b6b204e06033e008a6c17b52

Comment 3

[Aria Beingessner [:Gankra]]

Affects macos too.

Comment 4

[Gary [:streetwolf52]]

Affects windows 10 too.  WR totally borked.

Comment 5

[Aria Beingessner [:Gankra]]

It looks like it might be font related.

On https://gankro.github.io/blah/webtests/text.html

The bold paragraph is the only thing with visible text, and the visible text is specifically that using --apple-system-font and the "basic" font I download for adobe's edge fonts.

But even those only show up if they're styled bold?

Comment 6

[Gary [:streetwolf52]]

Some pictures and other graphics don't display either.

Comment 7

[Darkspirit]

Attachment: Screenshot_20170923_195841.png

(In reply to Alexis Beingessner [:Gankro] from comment #5)
Nightly 58 x64 20170923100045 de_DE @ Debian Testing (KDE / Radeon RX480)

Comment 8

[Asif Youssuff]

> 6:15.41 INFO: No more inbound revisions, bisection finished.
> 6:15.41 INFO: Last good revision: 176e16629030869dfd6e31e240006f9ea70be2ba
> 6:15.41 INFO: First bad revision: 754a0d9a44ea5208b6b204e06033e008a6c17b52
> 6:15.41 INFO: Pushlog:
> https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=176e16629030869dfd6e31e240006f9ea70be2ba&tochange=754a0d9a44ea5208b6b204e06033e008a6c17b52

I got the same results when I tried mozregression on my profile, FWIW.

Comment 9

[Glenn Watson [:gw]]

(Copying details added to GH bug)

OK, this seems to be an issue in the Gecko code.

The regression range was https://hg.mozilla.org/integration/autoland/pushloghtml?fromchange=176e16629030869dfd6e31e240006f9ea70be2ba&tochange=754a0d9a44ea5208b6b204e06033e008a6c17b52

This line https://hg.mozilla.org/integration/autoland/rev/754a0d9a44ea#l5.12 seems to always be failing. This results in no font instances being added, which means that WR drops the text runs with an unknown font instance.

Comment 10

[Lee Salzman [:lsalzman]]

Attachment: don't error when ShmSegmentReader reads empty segments

When the OffsetRange has 0 size, this causes some small chaos because reading did not expect this. It really shouldn't be erroring here.

Comment 11

[Aria Beingessner [:Gankra]]

Comment on attachment 8911619 [details] [diff] [review]
don't error when ShmSegmentReader reads empty segments

Locally confirmed this fixes the issue, and the fix makes sense to me!

Comment 12

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/cb1fd93593b4
don't error when ShmSegmentReader reads empty segments. r=gankro

Comment 13

[Sebastian Hengst [:aryx] (needinfo me if it's about an intermittent or backout)]

https://hg.mozilla.org/mozilla-central/rev/cb1fd93593b4

Comment 14

[Darkspirit]

Verified fixed in Nightly 58 x64 20170925100307 de_DE @ Debian Testing (KDE / Radeon RX480)