930575 - Graphic defect occurs when scrolling in FTU/Settings

Status: RESOLVED WORKSFORME

(Core :: Graphics, defect)

Description

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

Attachment: 2013-10-24-08-40-17.png

1. flash today's build ( 2013024040207 ) on Buri
2. scroll in FTU

Expected: no graphic defect
Actual: the scrolling leaves parts undrawn

Note: it will correct itself after about a minute after not scrolling.

Gaia:     62a84888f1c10556af8fbbd50e126bb1afe96276
Gecko:    http://hg.mozilla.org/mozilla-central/rev/19fd3388c372
BuildID   20131024040207
Version   27.0a1

Comment 2

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

As per Marcia's bug; this seems to occur in FTU and Settings.  Updating based upon Marcia's findings.

Comment 3

[Jason Smith [:jsmith]]

The only bug that looks relevant that might have caused this is bug 907643. Might be worth getting the regression range down to 12 hours.

Comment 4

[Jason Smith [:jsmith]]

(In reply to Jason Smith [:jsmith] from comment #3)
> The only bug that looks relevant that might have caused this is bug 907643.
> Might be worth getting the regression range down to 12 hours.

Meant to say - bug 907463.

Comment 5

[Nick Cameron [:nrc]]

(In reply to Jason Smith [:jsmith] from comment #4)
> (In reply to Jason Smith [:jsmith] from comment #3)
> > The only bug that looks relevant that might have caused this is bug 907643.
> > Might be worth getting the regression range down to 12 hours.
> 
> Meant to say - bug 907463.

I doubt it would be 907463 - it only affects the shutdown of the graphics system - so I would expect either a crash or no rendering at all - corruption like this seems unlikely (always possible though).

Comment 6

[Jason Smith [:jsmith]]

Apparently there aren't 12 hour builds available today per nkot, so we have the regression window down to a day (this was working yesterday, broken today).

Comment 7

[Nick Cameron [:nrc]]

So I tried to build and test with and without 907463, but in both cases my device crashed on startup, so I that is pretty inconclusive :-(

I asked around and no graphics people were willing to admit to landing any patches recently, so it might be worth backing out 907463 to check if that helps or not. Sorry I can't be any more useful here - I'm going to be travelling until Tuesday now so I can't easily fix my builds.

Comment 8

[Marcia Knous [:marcia]]

Testing with a new nightly which has the fix in Bug 907463 backed out, the issue still occurs:

Gaia   afbf45f26a73b7cd5e0a831bea48087331975286
SourceStamp 2f2a45f04e7c
BuildID 20131025100746
Version 27.0a1

Comment 9

[Jason Smith [:jsmith]]

(In reply to Marcia Knous [:marcia] from comment #8)
> Testing with a new nightly which has the fix in Bug 907463 backed out, the
> issue still occurs:
> 
> Gaia   afbf45f26a73b7cd5e0a831bea48087331975286
> SourceStamp 2f2a45f04e7c
> BuildID 20131025100746
> Version 27.0a1

The build that Ed spun does not contain the backout of this patch, as it hasn't landed on central.

Comment 10

[Marcia Knous [:marcia]]

(In reply to Jason Smith [:jsmith] from comment #9)
> (In reply to Marcia Knous [:marcia] from comment #8)
> > Testing with a new nightly which has the fix in Bug 907463 backed out, the
> > issue still occurs:
> > 
> > Gaia   afbf45f26a73b7cd5e0a831bea48087331975286
> > SourceStamp 2f2a45f04e7c
> > BuildID 20131025100746
> > Version 27.0a1
> 
> The build that Ed spun does not contain the backout of this patch, as it
> hasn't landed on central.

jsmith is right - apparently only one of the 2 bugs we asked for got backed out in the respin: https://tbpl.mozilla.org/?rev=2f2a45f04e7c

Comment 11

[Gregor Wagner [:gwagner]]

and backed out on mc
remote:   https://hg.mozilla.org/mozilla-central/rev/2846fd235ad0
remote:   https://hg.mozilla.org/mozilla-central/rev/3f22bd55ce6f
remote:   https://hg.mozilla.org/mozilla-central/rev/b6f5a4a737c8

Comment 12

[Marcia Knous [:marcia]]

Backout addressed the issue as this issue is not seen on Buri using:

Gaia   c59be1ad4a0bc005ff648eae59c441cb03771fbf
SourceStamp ef3f5669b53e
BuildID 20131026040201
Version 27.0a1

Confirmed backout is in the build: https://tbpl.mozilla.org/?rev=ef3f5669b53e

Resolving this as WFM.

Comment 13

[Nick Cameron [:nrc]]

I would like to try to reland part of bug 907463 (first the part which does not look suspicious, then the other part with a speculative change). Is there a way to request re-testing of the part of the smoke test which failed? Say on a try run? I could just land and see if we fail, but that seems a bit anti-social.

Comment 14

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

I suppose if you supply a patch/pull request to me somewhere, I can apply the patch in my own build and then test around the patch to see if I see the issue again or not.  Just needinfo me with the patch supplied.

Comment 15

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

Oh, I think you can also push to try, which would build for unagi.
I could be mistaken.

If you come up with a better method, please let me know.

Comment 16

[Nick Cameron [:nrc]]

(In reply to Naoki Hirata :nhirata (please use needinfo instead of cc) from comment #14)
> I suppose if you supply a patch/pull request to me somewhere, I can apply
> the patch in my own build and then test around the patch to see if I see the
> issue again or not.  Just needinfo me with the patch supplied.

Could you do one run with https://bugzilla.mozilla.org/attachment.cgi?id=805858 (the first patch from bug 907463) and a separate run with https://bugzilla.mozilla.org/attachment.cgi?id=826593 (a tweaked version of the third patch). Thank you!

Comment 17

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

I can.  Just to note, I'm in a work week so I won't be able to get to it until possibly Thursday.

Comment 18

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

patch : https://bugzilla.mozilla.org/attachment.cgi?id=805858 
This causes the scrolling defect to appear in the settings app.

I backed out of that patch and tried to apply the second item as a patch; having said that I got an error : 
error: patch failed: gfx/layers/composite/CompositableHost.h:256
error: gfx/layers/composite/CompositableHost.h: patch does not apply

Comment 19

[Nick Cameron [:nrc]]

(In reply to Naoki Hirata :nhirata (please use needinfo instead of cc) from comment #18)
> patch : https://bugzilla.mozilla.org/attachment.cgi?id=805858 
> This causes the scrolling defect to appear in the settings app.
> 

Thanks that is extremely helpful.

> I backed out of that patch and tried to apply the second item as a patch;
> having said that I got an error : 
> error: patch failed: gfx/layers/composite/CompositableHost.h:256
> error: gfx/layers/composite/CompositableHost.h: patch does not apply

I rebased the patch, it is at https://bugzilla.mozilla.org/attachment.cgi?id=829032, thanks!

Comment 20

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

Is there any other testing that you would like me to conduct?  Here's a report of my testing around the patch:

Summary : With the patch from https://bugzilla.mozilla.org/attachment.cgi?id=829032, the scrolling effect displays as expected ( no graphic defect as per comment 0).  This is on the buri device.

Overall, I think I prefer having this patch.  There are certain webpages that seems to load faster, scroll faster, better visually.  I could be mistaken.  I didn't take fps measurements.

Testing:
1. Trying to see what happens with these changes, I went to glmol.com.  the browser tab crashes upon loading that web page.  Having said that looking at the logcat, it looks like there's a lot more sharedsurface Gralloc creation and deletions happening in the logs.
2. Played with the camera app (camera shots, video shots)
3. Played with video app
4. Going to planet.mozilla.org ( zoom/panning );it seems like a very slight tearing effect that's barely noticible to the point I could be mistaken.  Tearing effect seems to only be when scrolling text.  Scrolling seems better as an overall animation.
5. 3d test from the marketplace, seems to work.
6. played with http://bit.ly/css3d-mol (from bug 862097).

- I had one experience where if the device goes to sleep some graphics doesn't render correctly when coming back on the http://bit.ly/css3d-mol.  Specifically, the molecules didn't render.

Happens on master as well:
- white screenshot for browser content when launching task manager with browser in the foreground.
- zooming can offset and chop rendering areas; panning fixes the issue.
- zooming on planet.mozilla.org blanks out.  looking at log, it's OOM, even when memory pressure is over, it doesn't recover from OOM.  Browser doesn't die, I had to kill it.
- black flashes ( bug 934602 / Bug 905882 )
- when placing the device in landscape, the browser would shrink the rendering area esp when the device is off ; this might be something
to note : tab crashes on master as well when going to glmol.com

Comment 21

[Nick Cameron [:nrc]]

Awesome, thank you! That is very useful.

(In reply to Naoki Hirata :nhirata (please use needinfo instead of cc) from comment #20)
> Is there any other testing that you would like me to conduct?  Here's a
> report of my testing around the patch:

No, that is great. I might ask again to check some patches before landing, if that is OK.

> 
> Summary : With the patch from
> https://bugzilla.mozilla.org/attachment.cgi?id=829032, the scrolling effect
> displays as expected ( no graphic defect as per comment 0).  This is on the
> buri device.
> 
> Overall, I think I prefer having this patch.  There are certain webpages
> that seems to load faster, scroll faster, better visually.  I could be
> mistaken.  I didn't take fps measurements.
> 
> Testing:
> 1. Trying to see what happens with these changes, I went to glmol.com.  the
> browser tab crashes upon loading that web page.  Having said that looking at
> the logcat, it looks like there's a lot more sharedsurface Gralloc creation
> and deletions happening in the logs.

Is this a new crash?

> 2. Played with the camera app (camera shots, video shots)
> 3. Played with video app
> 4. Going to planet.mozilla.org ( zoom/panning );it seems like a very slight
> tearing effect that's barely noticible to the point I could be mistaken. 
> Tearing effect seems to only be when scrolling text.  Scrolling seems better
> as an overall animation.

I'm pretty sure my patch cannot cause this.

> 5. 3d test from the marketplace, seems to work.
> 6. played with http://bit.ly/css3d-mol (from bug 862097).
> 
> - I had one experience where if the device goes to sleep some graphics
> doesn't render correctly when coming back on the http://bit.ly/css3d-mol. 
> Specifically, the molecules didn't render.
> 

This is something that could be due to my patch and is a bit worrying. I would like to investigate this before landing.

Is this reproducible? Does it happen every time? And have you seen this behaviour before?

> Happens on master as well:
> - white screenshot for browser content when launching task manager with
> browser in the foreground.
> - zooming can offset and chop rendering areas; panning fixes the issue.
> - zooming on planet.mozilla.org blanks out.  looking at log, it's OOM, even
> when memory pressure is over, it doesn't recover from OOM.  Browser doesn't
> die, I had to kill it.
> - black flashes ( bug 934602 / Bug 905882 )
> - when placing the device in landscape, the browser would shrink the
> rendering area esp when the device is off ; this might be something
> to note : tab crashes on master as well when going to glmol.com

So all of the above happen on master and are therefore not due to the patch?

Comment 22

[Naoki Hirata :nhirata (please use needinfo instead of cc)]

> No, that is great. I might ask again to check some patches before landing,
> if that is OK.

It's definitely OK as that is my job.  Having said that, I do have to juggle priorities, so depending on things I might not be able to get to it right away.

> > 1. Trying to see what happens with these changes, I went to glmol.com.  the
> > browser tab crashes upon loading that web page.  Having said that looking at
> > the logcat, it looks like there's a lot more sharedsurface Gralloc creation
> > and deletions happening in the logs.
> 
> Is this a new crash?

No, this is not a new crash.  It happens on the releng builds as well.  I was just mentioning it because I noticed a lot more alloc/deallocs in the logs.
 
> > 2. Played with the camera app (camera shots, video shots)
> > 3. Played with video app
> > 4. Going to planet.mozilla.org ( zoom/panning );it seems like a very slight
> > tearing effect that's barely noticible to the point I could be mistaken. 
> > Tearing effect seems to only be when scrolling text.  Scrolling seems better
> > as an overall animation.
> 
> I'm pretty sure my patch cannot cause this.

I'll investigate the tearing effect in the main build again then.  Thanks.  :)

> > 5. 3d test from the marketplace, seems to work.
> > 6. played with http://bit.ly/css3d-mol (from bug 862097).
> > 
> > - I had one experience where if the device goes to sleep some graphics
> > doesn't render correctly when coming back on the http://bit.ly/css3d-mol. 
> > Specifically, the molecules didn't render.
> > 
> 
> This is something that could be due to my patch and is a bit worrying. I
> would like to investigate this before landing.
> 
> Is this reproducible? Does it happen every time? And have you seen this
> behaviour before?

It happened one time and I can not reproduce it. I am currently unsure how to reproduce it; I tried multiple times to just have the device sleep and then wake it.  I think there's some other state I have to get into in order to reproduce it... I hadn't figured it out.  I have not seen the behavior before in the main build.
 
> > Happens on master as well:
> > - white screenshot for browser content when launching task manager with
> > browser in the foreground.
> > - zooming can offset and chop rendering areas; panning fixes the issue.
> > - zooming on planet.mozilla.org blanks out.  looking at log, it's OOM, even
> > when memory pressure is over, it doesn't recover from OOM.  Browser doesn't
> > die, I had to kill it.
> > - black flashes ( bug 934602 / Bug 905882 )
> > - when placing the device in landscape, the browser would shrink the
> > rendering area esp when the device is off ; this might be something
> > to note : tab crashes on master as well when going to glmol.com
> 
> So all of the above happen on master and are therefore not due to the patch?

yes, these are not because of the patch.  These are all issues that I have to investigate further and create bugs for if they do not exist.

Comment 23

[Nick Cameron [:nrc]]

Cool, thanks!