1314838 - intermittent blank reftest test or reference screenshots on Linux e10s.

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Hiroyuki Ikezoe (:hiro)]

I am not sure Core:Graphics is the right component, but a similar bug (bug 1300355) belongs to it.

Comment 1

[Hiroyuki Ikezoe (:hiro)]

Oops! I accidentally hit entry key...

There are bunch of reftest failures that did not draw anything on the images intermittently.  This failure seems to start recently.  (The first occurrence is bug 1312963?)

Comment 2

[Hiroyuki Ikezoe (:hiro)]

Happened on opt build too.

Comment 3

[Bas Schouten (:bas.schouten)]

Lee, you've done a lot of linux work, have you ever seen this?

Comment 4

[Lee Salzman [:lsalzman]]

Bug 1300355 is a similar meta-bug for a Windows manifestation of this problem. The suspected reasoning in that bug thread is an OOM situation, which also puts [GFX] errors in the logs.

However, none of the Linux ones listed here have any sort of error. There are no timeouts evident in any of these failures either. The timeline also seems different, in that the Windows issues seems to start at the beginning of September, whereas the Linux bugs are starting to roll in as of the end of October.

That said, I wonder if this is in any way related to the GL compositor on Linux? That is, it may be giving us silent failures in some strange way that the basic compositor would not, as one hypothetical guess at why no errors show up in the logs when these reftest failures occur.

I guess the first coveted goal to narrowing this down would be to either get some smoking gun evidence in the logs, or better, an STR.

Karl, Nical, any thoughts?

Comment 5

[Hiroyuki Ikezoe (:hiro)]

FWIW, happened on OS X debug build too. Bug 1321087.

Comment 6

[Karl Tomlinson (:karlt)]

I'm not familiar with how reftest screenshots are captured and so I don't know whether it is affected by the kind of compositor or not.

This is happening in 64-bit builds, and so an OOM situation seems unlikely.

Is this frequent enough to narrow down a regression range through retriggers?

If it is unlikely that GL layers are going to be shipped by default in the current format, then I wonder whether the basic compositor should be the default on Nightly also.

Comment 8

[Kartikaya Gupta (email:[e-mail])]

FWIW now that bug 1364626 has merged to m-c I expect this bug to be fixed. If anybody sees instances of this still happening please point them out.

Comment 9

[Kartikaya Gupta (email:[e-mail])]

I'm going to close this bug. I did a search of all the bugs with "max difference: 255" in the summary (as of right now there's 1191 of them) and looked at the ~20 or so that were opened in the last 7 days. Of those, the vast majority were on windows (which is tracked by bug 1300355). The few that were on linux were not instances of this bug (where one of the images is entirely white) - they looked like "legitimate" reftest intermittents (both images had something painted but they didn't match).

Closing as fixed by bug 1364626.