1805347 - On a Js1k demo, showing devtools can make parts of the Browser UI black/blank

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect)

Description

[Mayank Bansal]

Create new profile.
Go to https://js1k.com/2019-x/demo/4027

Press and hold the F12 key - this will rapidly open and close the Developer panel

AR: Elements of the browser UI (Menu bar, close buttons) will become blank, and the scrollbar area will become black.

Regression:
https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=ab7decc3020866de71069e01557f9d3c5fe7da59&tochange=b85e871f6a8d57cdc744634dadb2b0af88422504

If you cant repro, try changing the width of the dev pane, or try moving the dev pane to the bottom of the screen.

Comment 1

[Mayank Bansal]

Attachment: about:support

Comment 2

[Mayank Bansal]

Attachment: bandicam 2022-12-13 11-49-34-303.mp4

Comment 3

[Nicolas Chevobbe [:nchevobbe]]

I guess people from Graphic would have a better view of what could be happening here.
Mayank, could you try https://mozilla.github.io/mozregression/ to get a regression window?

Comment 4

[Timothy Nikkel (:tnikkel)]

Regression range is in the first comment.

Comment 5

[BugBot [:suhaib / :marco/ :calixte]]

This bug has been marked as a regression. Setting status flag for Nightly to affected.

Comment 6

[Darkspirit]

Please re-run mozregression with --pref gfx.webrender.all:true.
bug 1702989 enabled WebRender for this GPU on Windows.

Description: AMD Radeon(TM) Graphics
Vendor ID: 0x1002
Device ID: 0x1638

• 2021-04-06 in regression range: bug 1702989 switched WebRender-on-Windows allowlist from AtiRolloutWebRender to AtiAll

  d7f59acda4f9ff703581669bb072fabbb690254c Jeff Muizelaar — Bug 1702989 - Allow all AMD to be qualified for WebRender. r=aosmond

• 2021-09-30: Bug 1733412 added 0x1638 to AtiRolloutWebRender for WebRender on Linux:
  https://hg.mozilla.org/mozilla-central/file/4d047cfb39ea7a8d29b076cb7ba57cd40e63888e/widget/GfxDriverInfo.cpp

• 2022-10-28: bug 1794722 removed AtiRolloutWebRender.

Comment 7

[Timothy Nikkel (:tnikkel)]

I get this regression range with webrender on

https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=b422dd886e036907f35269f381f43785873c5296&tochange=69be3221f49a2acb92b734baf725e19bdf9a4fd9

On mac it's pretty easy to trigger the bug, don't need to hold F12, just open dev tools and close it. And it screws up the browser UI, not just inside the content window.

Comment 8

[Timothy Nikkel (:tnikkel)]

Maybe bug 1694059, bug 1694305, bug 1694284, bug 1690216 from that range?

Comment 9

[Timothy Nikkel (:tnikkel)]

And I get the same regression range on Windows.

Comment 10

[Darkspirit]

Tested with Nightly and ESR 102.5.0 on Debian Testing:
Tab text and window buttons disappear if I zoom out (at 80% or lower). The header is partially corrupted as if there is a stale tile.
It can further corrupt and Firefox becomes "transparent" in that place, I see what's behind Firefox: It might flicker between stale tile(s) of the header bar and of what was on screen at that place (desktop background image, if I minimize and restore the window. Or the color picker if I open and close it).
I made a screencast with OBS of this, but the video doesn't contain what I saw, the affected tiles of the header are just black in the video.
The header's background gradient is missing at some zoom levels.

Comment 11

[Darkspirit]

Enforcing SW WR and disabling partial present, accelerated canvas, backdrop filter, etc. doesn't help,
everything including "transparent" tile is still reproducible:
mozregression --launch 2022-12-15 -a https://js1k.com/2019-x/demo/4027 --pref gfx.webrender.allow-partial-present-buffer-age:false gfx.webrender.max-partial-present-rects:0 gfx.webrender.pbo-uploads:false gfx.webrender.debug.texture-cache.clear-evicted:false gfx.webrender.enable-item-cache:false gfx.webrender.multithreading:false gfx.webrender.svg-images:false gfx.webrender.use-optimized-shaders:false layout.css.backdrop-filter.enabled:false gfx.canvas.accelerated:false gfx.webrender.software:true

Comment 12

[Timothy Nikkel (:tnikkel)]

Hmm, it doesn't seem to be that complicated of a page. I wonder what the reduced testcase would look like. Maybe we just have too many display items and that is tripping some problem? No warnings or errors in a debug build in the console or about:support.

Comment 13

[Sotaro Ikeda [:sotaro]]

When I set pref gfx.webrender.debug.force-picture-invalidation = true, I could reproduce the similar symptom easily. Then I wonder if it is related to Webrender picture cache. "pref gfx.webrender.debug.force-picture-invalidation" was added by Bug 1678326.

With "pref gfx.webrender.debug.force-picture-invalidation = true", I got the following regression range.

https://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=b422dd886e036907f35269f381f43785873c5296&tochange=69be3221f49a2acb92b734baf725e19bdf9a4fd9

I wonder if Bug 1692250 and Bug 1694305 might be related to the bug.

Comment 14

[Sotaro Ikeda [:sotaro]]

:nical, :gw, can you comment to comment 13?

Comment 15

[Mayank Bansal]

Attachment: testcase.htm

Standalone testcase that repros part of the bug. On pressing F12, the text on the title bar becomes blank.

Comment 16

[Nicolas Silva [:nical]]

Unfortunately I suspect that the regression range points to https://hg.mozilla.org/mozilla-central/rev/d7f59acda4f9ff703581669bb072fabbb690254c which enabled webrender on AMD GPUs.

Mayank, could you do another regression range with webrender manually enabled (I think that the "gfx.webrender.all" pref would do it)?

Comment 17

[Darkspirit]

See comment 7 + comment 13.

Comment 18

[Darkspirit]

(I got the same regression range on Gnome Wayland/Debian Testing/Intel.)

Comment 19

[Glenn Watson [:gw]]

I can repro this (or similar) on current release but not on nightly. Is anyone still able to repro on nightly?

Comment 20

[Mayank Bansal]

(In reply to Glenn Watson [:gw] from comment #19)

I can repro this (or similar) on current release but not on nightly. Is anyone still able to repro on nightly?

I cant repro this bug on nightly.
I can still repro bug 1857943 .
I cant repro the testcase from bug 1840988 , but can repro the original STR from its first comment (bug 1840988 comment #0)

Comment 21

[Glenn Watson [:gw]]

I believe this is the same underlying cause as 1840988 - which is that we create too many render tasks for the shader indexing method to handle. I have a couple of ideas to remove this limitation, I'll prototype them this week and see which is most suitable.

Comment 22

[Mayank Bansal]

This is fixed on latest Nightly.