706023 - Horizontal tearing with canvas

Status: RESOLVED WORKSFORME

(Firefox for Android Graveyard :: General, defect, P3)

Description

[Matt Woodrow (:mattwoodrow)]

When running canvas demos on my phone (Galaxy S), I'm getting bad vertical tearing that looks like a race condition between uploading textures and compositing.

The asteroids demo is particularly bad: http://www.kevs3d.co.uk/dev/asteroidsbench/

Comment 1

[Martijn Wargers (dead)]

Yes, I can reproduce. It actually locks up my LG Optimus Black phone too.
That bad vertical tearing, I've seen on more sites. But on this site it seems to happen really easily.

Comment 2

[Chris Lord [:cwiiis]]

Hmm, I see this on my Flyer, though the tears are extremely infrequent. It looks like possibly we aren't guarding against Gecko drawing into the buffer while we upload it, but I'm pretty sure we are...

The infrequency of it makes it a bit hard to track down, so I'm not treating this as top priority for now. If I think of anything, I'll report back.

Comment 3

[Chris Lord [:cwiiis]]

Oh, ok, actually, if you pan around horizontally a lot, you can see white blocks in the areas of the texture that aren't updated by the canvas once the bounce animation finishes. I think I may have an idea what causes that, looking into it.

Comment 4

[Chris Lord [:cwiiis]]

My idea was wrong (I thought something about the way we force redraws after animations may be having a negative impact, but I can't see how it would). I would still consider this low priority anyway, hopefully we'll fix it when we fix other issues, or it'll go away when we switch to our more long-term drawing plan.

Comment 5

[Chris Lord [:cwiiis]]

Ok, very sorry for the spam, but I've just noticed that we don't take the tile lock when we draw, so this is almost certainly a problem caused by updating in the middle of drawing... Cooking up a patch, as this may manifest in more serious/annoying ways.

Comment 6

[Chris Lord [:cwiiis]]

Actually, I'm wrong - I forgot that texture updates are tied to drawing and only happen synchronously before a draw, with the LayerController lock held - invalidations are only called with the LayerController lock held too, so this shouldn't be a problem. Back to square one.

Comment 7

[Brad Lassey [:blassey] (use needinfo?)]

Can we retest this now that maple has landed?

Comment 8

[Martijn Wargers (dead)]

This seems to be wfm in current trunk Fennec build on the HTC Desire HD and the Samsung Galaxy Nexus and the Samsung Galaxy SII.

Mat, could you also perhaps retest this?

Comment 9

[Brad Lassey [:blassey] (use needinfo?)]

resolving based comment 8