1191040 - Erratic tab rendering on external display of retina Mac

Status: RESOLVED FIXED

(Core :: Graphics: Layers, defect)

Description

[Ryan Feeley [:rfeeley]]

Attachment: nightly-tab-bug.png

See attached. After I resized my window and dragged it around, it went away, but was happening for over an hour after reconnecting to my non-retina monitor.

Comment 1

[u279076]

Have you been able to replicate this at all or was it a one-off?

Comment 2

[Ryan Feeley [:rfeeley]]

One-off

Comment 3

[Ryan Feeley [:rfeeley]]

Attachment: nightly-tab-bug2.png

Second one just in case

Comment 4

[Mike Conley (:mconley) (:⚙️)]

Can you try to recall the exact steps you went through in order to get into this state?

You had a Firefox window on your Retina Macbook display... you plugged it into a secondary non-Retina monitor... did the window automatically move to the non-Retina display, or did you drag it there?

Comment 5

[Ryan Feeley [:rfeeley]]

Typically I use the external display when at my desk. When I go to a meeting I unplug and all windows move to the retina. This happened after I came back. I didn't do anything different than usual, but haven't reproduced it yet. I expect I will at some point when in Nightly (had to switch to release as I was having other problems).

Comment 6

[Mike Conley (:mconley) (:⚙️)]

STR that work for me:

1) Start with Retina Macbook plugged into secondary standard DPI display.
2) Start Nightly, and put it on the Retina display.
3) Unplug the display. Nightly should stay on the Retina.
4) Re-plug in the display. Nightly should stay on the Retina. If it didn't, move it to it.
5) Quickly open and close a window in Nightly.
6) Now open a new tab - alternatively, overflow the tabstrip. Regardless, behold some hilarious rendering action.

Comment 7

[Mike Conley (:mconley) (:⚙️)]

Note that my STR worked perfectly with a fresh profile on 10.8.5.

Comment 8

[Mike Conley (:mconley) (:⚙️)]

STR:

1) Start with Retina Macbook plugged into secondary standard DPI display.
2) Start Nightly, and put it on the Retina display.
3) Unplug the display.
5) Quickly open and close a window in Nightly. Now open a new tab. If it seems garbled - congratulations, you've seen the bug. If not, go to the next step.
6) Plug the display back in, and put Nightly on the standard DPI display. Now open a new tab.

Behold some hilarious rendering action.

Comment 9

[Mike Conley (:mconley) (:⚙️)]

mozregression says it's this changeset range that introduced the regression:

https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=c14a0de10f23&tochange=b0d4af6d3d2a

Comment 10

[Mike Conley (:mconley) (:⚙️)]

Anything in that range look particularly suspicious?

Comment 11

[Jeff Muizelaar [:jrmuizel]]

Yes:
https://hg.mozilla.org/integration/mozilla-inbound/rev/3a266383e726 is suspicious.

If you could be back that one out and try again that would be wonderful.

Comment 12

[Mike Conley (:mconley) (:⚙️)]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #11)
> Yes:
> https://hg.mozilla.org/integration/mozilla-inbound/rev/3a266383e726 is
> suspicious.
> 
> If you could be back that one out and try again that would be wonderful.

Yes - I can't reproduce with that changeset backed out.

Comment 13

[Mike Conley (:mconley) (:⚙️)]

Can we back out bug 1188995 until this gets resolved? The UX when the user gets into this state is the bad kind of spectacular.

Comment 14

[Matt Woodrow (:mattwoodrow)]

I'll fix this today.

Comment 15

[Mike Conley (:mconley) (:⚙️)]

Cool, thanks. Happy to verify.

Comment 16

[Matt Woodrow (:mattwoodrow)]

Attachment: compute-tile-size-once

Comment 17

[Benoit Girard (:BenWa)]

Comment on attachment 8645047 [details] [diff] [review]
compute-tile-size-once

Review of attachment 8645047 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/thebes/gfxPlatform.cpp
@@ +1003,5 @@
>  {
>    // The tile size should be picked in the parent processes
>    // and sent to the child processes over IPDL GetTileSize.
>    if (!XRE_IsParentProcess()) {
> +    return;

.... meeehhh ... ok fine :)

We've got good asserts covering the tile size getting set properly anyways.

::: gfx/thebes/gfxPlatform.h
@@ +755,5 @@
>      virtual void GetPlatformCMSOutputProfile(void *&mem, size_t &size);
>  
> +    /**
> +     * Calling this function will compute and set the ideal tile size for the
> +     * platform. This should only be called in the parent process; child processes

This comment could use an update.

Comment 18

[Mike Conley (:mconley) (:⚙️)]

I cannot reproduce the bug with this patch applied. Great job!

Comment 19

[Pulsebot]

https://hg.mozilla.org/integration/mozilla-inbound/rev/4e63305fe592

Comment 20

[Ryan VanderMeulen (PTO, back July 27)]

Backed out for Linux gtest crashes.
https://treeherder.mozilla.org/logviewer.html#?job_id=12638223&repo=mozilla-inbound

https://hg.mozilla.org/integration/mozilla-inbound/rev/e102197e3034

Comment 21

[Pulsebot]

https://hg.mozilla.org/integration/mozilla-inbound/rev/daef3729bce2

Comment 22

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/daef3729bce2