Bug 32955 – REGRESSION(52554): Two layout test failures on Leopard Debug after r52554 (partial? rollout of r52536)

• Status: RESOLVED
• Resolution: FIXED
• Component: Tools / Tests

[Eric Seidel (no email)]

REGRESSION(52554): Two layout test failures on Leopard after r52554 (partial? rollout of r52536)
http://build.webkit.org/results/Leopard%20Intel%20Debug%20(Tests)/r52555%20(8652)/results.html

http://trac.webkit.org/changeset/52554

The previous build was clean.  The bots have been red since that checkin.  I suggest we rollout this rollout even if that means breaking Chromium.  Or we Skip the offending tests.  Breaking Leopard is not cool. :(  I have no clue what the problem with the SVG test is.

Tests where results did not match expected results:
svg/W3C-SVG-1.1/filters-conv-01-f.svg	expected	actual	diff	pretty diff

Tests that caused the DumpRenderTree tool to crash:
storage/open-database-while-transaction-in-progress.html	stderr

--- layout-test-results/svg/W3C-SVG-1.1/filters-conv-01-f-expected.txt	2009-12-24 22:24:25.000000000 -0800
+++ layout-test-results/svg/W3C-SVG-1.1/filters-conv-01-f-actual.txt	2009-12-24 22:24:25.000000000 -0800
@@ -39,7 +39,7 @@
             RenderSVGImage {image} at (180,163) size 50x63 [filter=convolve5]
           RenderSVGText {text} at (10,20) size 117x18 contains 1 chunk(s)
             RenderSVGInlineText {#text} at (0,-14) size 117x18
-              chunk 1 text run 1 at (10.00,20.00) startOffset 0 endOffset 19 width 117.00: "Vertical blur (1x3)"
+              chunk 1 text run 1 at (10.00,20.00) startOffset 0 endOffset 19 width 115.00: "Vertical blur (1x3)"
           RenderSVGContainer {g} at (330,163) size 50x63 [transform={m=((1.00,0.00)(0.00,1.00)) t=(300.00,93.00)}]
             RenderSVGImage {image} at (330,163) size 50x63 [filter=convolve6]
     RenderSVGText {text} at (10,340) size 264x46 contains 1 chunk(s)

[Eric Seidel (no email)]

r52553 definitely ran clean:
http://build.webkit.org/builders/Leopard%20Intel%20Debug%20%28Tests%29/builds/8650

[Eric Seidel (no email)]

I meant to link to build http://build.webkit.org/builders/Leopard%20Intel%20Debug%20%28Tests%29/builds/8651 instead of 8652 in my initial message.  8652 corresponds to r52554 and is the first build we saw failures from.

[Eric Seidel (no email)]

I'm updating my tree and then I will make a rollout of (r52554).  I've notified the #chromium channel, as according to build.chromium.org no one is sheriff today.

[Eric Seidel (no email)]

I've kicked off a rollout.  Should land shortly.

[Eric Seidel (no email)]

Committed r52569: <http://trac.webkit.org/changeset/52569>

[Eric Seidel (no email)]

Looks like this breaks 21 tests for Chromium:
http://build.chromium.org/buildbot/layout_test_results/webkit-rel-linux-webkit-org/35287/index-14120.html

However, if it passes on Leopard (and thus makes the WebKit builders green again), I plan to leave the change in for now.

[Eric Seidel (no email)]

The 21 test failures for Chromium are:
Threw exception Error: INVALID_STATE_ERR: DOM Exception 11
when doing database operations.

Dumi's commentary would be helpful here.

[Eric Seidel (no email)]

It's possible the failures before were a bogus ASSERT.  I never tried to reproduce them locally.

[Eric Seidel (no email)]

Yes, looks like the Leopard ASSERT was:
ASSERTION FAILED: m_openDatabaseSet && m_openDatabaseSet->contains(database)
(/Volumes/Big/WebKit-BuildSlave/leopard-intel-debug/build/WebCore/dom/Document.cpp:4448 void WebCore::Document::removeOpenDatabase(WebCore::Database*))

[Nikolas Zimmermann]

The svg test only fails if another test crashes before. I have _no_ clue what causes text metrics differences, if a test crashed before that filters-conv-01-f test. Indicates a DRT problem.

[Shinichiro Hamaji]

Oops... I'm so sorry about this re-regression, and thanks Eric for landing the fix. I should have checked webkit build bot (I usually check the bots after I land a patch, but in this case I couldn't imagine reverting a change can cause regression)...

Adding Pam in CC list of this bug as he is the next webkit gardener of chromium. I guess we may want to disable the 21 tests in chromium side if Dumi is on year end vacation.

[Eric Seidel (no email)]

Niko:  I expect that DRT crashing just causes DRT to restart at the right point to trigger interaction between the SVG tests.  I expect that we could cause the SVG test to fail in other circumstances (including jus running every test after storage/open-database-while-transaction-in-progress.html).  I've not yet attempted to reproduce it locally.

Counting tests, it looks like there are about 1000 tests between storage/open-database-while-transaction-in-progress.html and svg/W3C-SVG-1.1/filters-conv-01-f.svg.

Sinchiro:  Thank you.  I expect that the 21 tests all have a single root cause.  I also expect that the ASSERT seen on Leopard also has a simple resolution.  I do not feel comfortable enough with this code to guess off the top of my head, but I expect that someone like Dumi might know.

[Nikolas Zimmermann]

(In reply to comment #12)
> Niko:  I expect that DRT crashing just causes DRT to restart at the right point
> to trigger interaction between the SVG tests.  I expect that we could cause the
> SVG test to fail in other circumstances (including jus running every test after
> storage/open-database-while-transaction-in-progress.html).  I've not yet
> attempted to reproduce it locally.

The problem are not the SVG tests, I've already investigated. If I run just "run-webkit-tests svg" I can reproduce the filters-conv-01-f.svg bug on my machine, everytime! If DRT crashes ~ 1000 tests before the SVG test, and restarts, the problem with filters-conv-01-f.svg appears. So some test is modifying text metrics for all following tests, no idea why.

[Dirk Schulze]

(In reply to comment #13)
> The problem are not the SVG tests, I've already investigated. If I run just
> "run-webkit-tests svg" I can reproduce the filters-conv-01-f.svg bug on my
> machine, everytime! If DRT crashes ~ 1000 tests before the SVG test, and
> restarts, the problem with filters-conv-01-f.svg appears. So some test is
> modifying text metrics for all following tests, no idea why.

But it's strange if it allways fails on filters-conv-01-f.svg?