Bug 168033 – Nwtr ignores ImageDiff's errors for ref test

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

[Fujii Hironori]

nrwt: Reftests pass unexpectedly in case of ImageDiff's crash

>  fast/css/cascade/box-shadow-and-webkit-box-shadow-cascade-order.html -> ref test hashes didn't match but diff passed

On the other hand, pixel test fails and reports "ImageDiff crashed" as expected.

>  fast/css/background-clip-radius-values.html : ImageDiff crashed
>[1/1] fast/css/background-clip-radius-values.html failed unexpectedly (image diff)

[Fujii Hironori]

Created attachment 301004 [details]
Patch

[Build Bot]

Comment on attachment 301004 [details]
Patch

Attachment 301004 [details] did not pass mac-debug-ews (mac):
Output: http://webkit-queues.webkit.org/results/3030141

New failing tests:
compositing/transitions/transform-on-large-layer.html
fast/css/sticky/sticky-left-percentage.html

[Build Bot]

Created attachment 301013 [details]
Archive of layout-test-results from ews117 for mac-elcapitan

The attached test failures were seen while running run-webkit-tests on the mac-debug-ews.
Bot: ews117  Port: mac-elcapitan  Platform: Mac OS X 10.11.6

[Build Bot]

Comment on attachment 301004 [details]
Patch

Attachment 301004 [details] did not pass mac-wk2-ews (mac-wk2):
Output: http://webkit-queues.webkit.org/results/3030161

New failing tests:
compositing/transitions/transform-on-large-layer.html
fast/css/sticky/sticky-left-percentage.html

[Build Bot]

Created attachment 301015 [details]
Archive of layout-test-results from ews105 for mac-elcapitan-wk2

The attached test failures were seen while running run-webkit-tests on the mac-wk2-ews.
Bot: ews105  Port: mac-elcapitan-wk2  Platform: Mac OS X 10.11.6

[Alexey Proskuryakov]

Comment on attachment 301004 [details]
Patch

Looks like this breaks tests.

But also, why does ImageDiff crash? In a crazy situation like this, it's probably appropriate to terminate run-webkit-tests with an exception.

[Fujii Hironori]

(In reply to comment #4)
> New failing tests:
> compositing/transitions/transform-on-large-layer.html
> fast/css/sticky/sticky-left-percentage.html

This change discovers problems which have been hidden by false negative.

>   compositing/transitions/transform-on-large-layer.html : ImageDiff produced stderr output:
> Error: test and reference images have different sizes. Test image is 800x600, reference image is 1600x977
> [1/10] compositing/transitions/transform-on-large-layer.html failed unexpectedly (reference didn't generate pixel results.)
>   fast/css/sticky/sticky-left-percentage.html : ImageDiff produced stderr output:
> Error: test and reference images have different sizes. Test image is 800x700, reference image is 800x600
> [2/10] fast/css/sticky/sticky-left-percentage.html failed unexpectedly (reference didn't generate pixel results.)

[Fujii Hironori]

(In reply to comment #6)
> But also, why does ImageDiff crash? In a crazy situation like this, it's
> probably appropriate to terminate run-webkit-tests with an exception.

My ImageDiff crashed just because mis-configuration of my system (Bug 168036 comment 1).

But, this change isn't only for ImageDiff's crash.
Nwtr hides tree types of other problems at the moment:
* ImageDiff produced stderr output
* ImageDiff timed out
* ImageDiff crashed

https://trac.webkit.org/browser/trunk/Tools/Scripts/webkitpy/port/image_diff.py?rev=211931#L97

[Fujii Hironori]

Nwtr ignores ImageDiff's errors only for ref test.
Nwtr checks the err_str of pixel test.

[Fujii Hironori]

Created attachment 301022 [details]
Patch

[Alexey Proskuryakov]

Comment on attachment 301022 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=301022&action=review

> Tools/ChangeLog:3
> +        Nwtr ignores ImageDiff's errors for ref test

Looking at the patch, but it's taking me some effort to understand all that it changes.

In particular, I'm still not sure what the problems with these two tests were. What exactly makes them fail now? They clearly don't crash or hang. When I run them locally, I get a warning saying "ref test hashes didn't match but diff passed", but many more tests produce this warning, so what is the difference with those?

  compositing/masks/compositing-clip-path.html -> ref test hashes didn't match but diff passed
  compositing/overlap-blending/children-opacity-huge.html -> ref test hashes didn't match but diff passed
  compositing/overlap-blending/children-opacity-no-overlap.html -> ref test hashes didn't match but diff passed
  compositing/overlap-blending/nested-non-overlap-clipping.html -> ref test hashes didn't match but diff passed
  compositing/transitions/transform-on-large-layer.html -> ref test hashes didn't match but diff passed
  css3/filters/backdrop/backdrop-filter-with-border-radius-and-reflection-add.html -> ref test hashes didn't match but diff passed
  css3/filters/backdrop/backdrop-filter-with-border-radius-and-reflection.html -> ref test hashes didn't match but diff passed
  css3/filters/filters-on-svg-root.html -> ref test hashes didn't match but diff passed

> LayoutTests/ChangeLog:13
> +        Set a tiny margin-top to #sandbox not to overwrap with a preceding
> +        text.

Why do we need it to overlap?

[Alexey Proskuryakov]

Comment on attachment 301022 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=301022&action=review

>> LayoutTests/ChangeLog:13
>> +        text.
> 
> Why do we need it to overlap?

I meant, why do we need it to not overlap?

[Ryosuke Niwa]

Comment on attachment 301022 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=301022&action=review

> Tools/ChangeLog:9
> +        expected.  Bacause _compare_image checks returned err_str of

We use a single space after period.

> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:339
>                  diff_result = self._port.diff_image(reference_driver_output.image, actual_driver_output.image, tolerance=0)
> -                if not diff_result[0]:
> -                    failures.append(test_failures.FailureReftestMismatchDidNotOccur(reference_filename))
> +                err_str = diff_result[2]
> +                if err_str:
> +                    _log.warning('  %s : %s' % (self._test_name, err_str))
> +                    failures.append(test_failures.FailureReftestNoImagesGenerated(reference_filename))

Can we extract this as a helper method with _ prefix so that we don't have to repeat the same code twice?

> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:343
>                  else:
> -                    _log.warning("  %s -> ref test hashes matched but diff failed" % self._test_name)
> +                    if not diff_result[0]:
> +                        failures.append(test_failures.FailureReftestMismatchDidNotOccur(reference_filename))

Why don't we use elif here?

> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:355
>              else:
> -                _log.warning("  %s -> ref test hashes didn't match but diff passed" % self._test_name)
> +                if diff_result[0]:

Ditto.

> LayoutTests/ChangeLog:9
> +        ImageDiff reports errors for different image sizes of expected and
> +        actual image of ref tests.  They should have same window size.

Can we make this change in a separate patch?

[Fujii Hironori]

Thank you for reviewing.

(In reply to comment #11)
> Comment on attachment 301022 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=301022&action=review
> 
> > Tools/ChangeLog:3
> > +        Nwtr ignores ImageDiff's errors for ref test
> 
> Looking at the patch, but it's taking me some effort to understand all that
> it changes.
> 
> In particular, I'm still not sure what the problems with these two tests
> were. What exactly makes them fail now? They clearly don't crash or hang.
> When I run them locally, I get a warning saying "ref test hashes didn't
> match but diff passed", but many more tests produce this warning, so what is
> the difference with those?
> 
>   compositing/masks/compositing-clip-path.html -> ref test hashes didn't
> match but diff passed
>   compositing/overlap-blending/children-opacity-huge.html -> ref test hashes
> didn't match but diff passed
>   compositing/overlap-blending/children-opacity-no-overlap.html -> ref test
> hashes didn't match but diff passed
>   compositing/overlap-blending/nested-non-overlap-clipping.html -> ref test
> hashes didn't match but diff passed
>   compositing/transitions/transform-on-large-layer.html -> ref test hashes
> didn't match but diff passed
>  
> css3/filters/backdrop/backdrop-filter-with-border-radius-and-reflection-add.
> html -> ref test hashes didn't match but diff passed
>  
> css3/filters/backdrop/backdrop-filter-with-border-radius-and-reflection.html
> -> ref test hashes didn't match but diff passed
>   css3/filters/filters-on-svg-root.html -> ref test hashes didn't match but
> diff passed

You didn't apply my patch. It's difficult to know what's an
actual failure without my patch because nwtr doesn't report
ImageDiff's error.

You can see the ImageDiff's error message by applying out patch (Comment 7):

> Error: test and reference images have different sizes. Test image is 800x600, reference image is 1600x977

And, ImageDiff doen't check any pixels if the comparing two image have different sizes.
So, it can not fail the test by modifying the text in -expected.html.

[Fujii Hironori]

(In reply to comment #12)
> Comment on attachment 301022 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=301022&action=review
> 
> >> LayoutTests/ChangeLog:13
> >> +        text.
> > 
> > Why do we need it to overlap?
> 
> I meant, why do we need it to not overlap?

It caused pixel diffs while I was runnning nwtr of GTK port.
I thought it was caused by z-index specifed in the test.
But, I check the test again, there is no such z-index specifed, but just translate3d.
This is my fault. I need to investigate the failure of GTK port.
Thank you for pointing out.

[Fujii Hironori]

Thank you for reviewing.

(In reply to comment #13)
> Comment on attachment 301022 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=301022&action=review
> 
> > Tools/ChangeLog:9
> > +        expected.  Bacause _compare_image checks returned err_str of
> 
> We use a single space after period.

Oh, this is Emacs default behavior of fill-paragraph. I changed
my config (sentence-end-double-space) now.

> > Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:339
> >                  diff_result = self._port.diff_image(reference_driver_output.image, actual_driver_output.image, tolerance=0)
> > -                if not diff_result[0]:
> > -                    failures.append(test_failures.FailureReftestMismatchDidNotOccur(reference_filename))
> > +                err_str = diff_result[2]
> > +                if err_str:
> > +                    _log.warning('  %s : %s' % (self._test_name, err_str))
> > +                    failures.append(test_failures.FailureReftestNoImagesGenerated(reference_filename))
> 
> Can we extract this as a helper method with _ prefix so that we don't have
> to repeat the same code twice?

I found another problem (Bug 168221).
The login of match and mismatch ref tests are similar but same.
I think it's difficult.
I'll revisit this bug after finishing Bug 168221.

> > Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:343
> >                  else:
> > -                    _log.warning("  %s -> ref test hashes matched but diff failed" % self._test_name)
> > +                    if not diff_result[0]:
> > +                        failures.append(test_failures.FailureReftestMismatchDidNotOccur(reference_filename))
> 
> Why don't we use elif here?
> 
> > Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:355
> >              else:
> > -                _log.warning("  %s -> ref test hashes didn't match but diff passed" % self._test_name)
> > +                if diff_result[0]:
> 
> Ditto.

I just copied the code of pixel tests. You're right. I'll revise.

> > LayoutTests/ChangeLog:9
> > +        ImageDiff reports errors for different image sizes of expected and
> > +        actual image of ref tests.  They should have same window size.
> 
> Can we make this change in a separate patch?

That's good idea. Filed (Bug 168217 and Bug 168218).

[Alexey Proskuryakov]

> You didn't apply my patch. It's difficult to know what's an
> actual failure without my patch because nwtr doesn't report
> ImageDiff's error.

I don't think that this answers my question. Can you explain what is different about the two tests that you modified in this patch, in contrast with the others that currently have the same output?

> And, ImageDiff doen't check any pixels if the comparing two image have different sizes.

Are you sure? My recollection is that is does, at least on the Mac.

[Fujii Hironori]

(In reply to comment #17)
> > You didn't apply my patch. It's difficult to know what's an
> > actual failure without my patch because nwtr doesn't report
> > ImageDiff's error.
> 
> I don't think that this answers my question. Can you explain what is
> different about the two tests that you modified in this patch, in contrast
> with the others that currently have the same output?

The two tests call window.resize, but others don't.
Is this an answer to your question? I think I don't understand your question.


> > And, ImageDiff doen't check any pixels if the comparing two image have different sizes.
> 
> Are you sure? My recollection is that is does, at least on the Mac.

I tested in my mac mini now. My result is different with yours.

> fujii@fujihiro-mac-mini $ ./Tools/Scripts/run-webkit-tests compositing/transitions/transform-on-large-layer.html
> Using port 'mac-elcapitan-wk2'
> Test configuration: <elcapitan, x86_64, release>
> Placing test results in /Users/fujii/work/webkit/g1/WebKitBuild/Release/layout-test-results
> Baseline search path: platform/mac-wk2 -> platform/wk2 -> platform/mac-elcapitan -> platform/mac -> generic
> Using Release build
> Pixel tests disabled
> Regular timeout: 30000, slow test timeout: 150000
> Command line: /Users/fujii/work/webkit/g1/WebKitBuild/Release/WebKitTestRunner -
> 
> --lint-test-files warnings:
> LayoutTests/TestExpectations:384 Path does not exist. imported/w3c/web-platform-tests/tools
> 
> Found 1 test; running 1, skipping 0.
> 
> Running 1 test
> 
> Starting helper ...2017-02-13 00:00:03.439 LayoutTestHelper[28724:1359551] Could not determine current color profile, so it will not be reset after running the tests.     Running 1 WebKitTestRunner.
> 
>   compositing/transitions/transform-on-large-layer.html -> ref test hashes didn't match but diff passed
> The test ran as expected.
> 
> fujii@fujihiro-mac-mini $ sed -ie 's/turns green/XXX/' LayoutTests/compositing/transitions/transform-on-large-layer-expected.html
> fujii@fujihiro-mac-mini $ ./Tools/Scripts/run-webkit-tests compositing/transitions/transform-on-large-layer.html
> Using port 'mac-elcapitan-wk2'
> Test configuration: <elcapitan, x86_64, release>
> Placing test results in /Users/fujii/work/webkit/g1/WebKitBuild/Release/layout-test-results
> Baseline search path: platform/mac-wk2 -> platform/wk2 -> platform/mac-elcapitan -> platform/mac -> generic
> Using Release build
> Pixel tests disabled
> Regular timeout: 30000, slow test timeout: 150000
> Command line: /Users/fujii/work/webkit/g1/WebKitBuild/Release/WebKitTestRunner -
> 
> --lint-test-files warnings:
> LayoutTests/TestExpectations:384 Path does not exist. imported/w3c/web-platform-tests/tools
> 
> Found 1 test; running 1, skipping 0.
> 
> Running 1 test
> 
> Starting helper ...2017-02-13 00:02:06.395 LayoutTestHelper[28882:1360579] Could not determine current color profile, so it will not be reset after running the tests.     Running 1 WebKitTestRunner.
> 
>   compositing/transitions/transform-on-large-layer.html -> ref test hashes didn't match but diff passed
> The test ran as expected.
> 
> fujii@fujihiro-mac-mini $


Here is a source code of ImageDiff of Mac port.
https://trac.webkit.org/browser/trunk/Tools/DumpRenderTree/cg/ImageDiffCG.cpp?rev=198559#L216

It doesn't check any pixels if the sizes are different.

[Fujii Hironori]

I discovered old discussions.
This problem was discussed two years ago.

  [webkit-dev] run-webkit-tests question; hashes when comparing ref test output
  https://lists.webkit.org/pipermail/webkit-dev/2015-January/027166.html

[Fujii Hironori]

Created attachment 301589 [details]
Patch

[Fujii Hironori]

Created attachment 301593 [details]
Patch

[Fujii Hironori]

Created attachment 301594 [details]
Patch

[Fujii Hironori]

Created attachment 301694 [details]
Patch

[Fujii Hironori]

Comment on attachment 301593 [details]
Patch

I've finished fixing depending bugs, please review again.

[Ryosuke Niwa]

Comment on attachment 301593 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=301593&action=review

> Tools/ChangeLog:8
> +        On the other hand, pixel test fails and reports the error as

It's strange to start your log by saying "on the other hand" given you haven't stated what on one hand happens.

> Tools/ChangeLog:9
> +        expected. Bacause _compare_image checks returned err_str of

Typo: B"a"cause.

> Tools/ChangeLog:16
> +

Please also explain the change you're making to expected match case.

> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:340
> +            err_str = diff_result[2]

Please spell out "error" instead of abbreviating it as "err".

[Fujii Hironori]

Created attachment 303635 [details]
Patch

[Alexey Proskuryakov]

Comment on attachment 303635 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=303635&action=review

Looks good to me. Can automated tests be added for this fix?

> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:276
> +            error = diff_result[2]

I think that error_string would be a better name, there are too many things that an "error" can contain.

> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:344
> +                actual_driver_output.error = (actual_driver_output.error or '') + error

I don't see any other place where we set actual_driver_output.error. Why is this needed?

[Fujii Hironori]

Comment on attachment 303635 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=303635&action=review

>> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:276
>> +            error = diff_result[2]
> 
> I think that error_string would be a better name, there are too many things that an "error" can contain.

I agree. I'll revise.

>> Tools/Scripts/webkitpy/layout_tests/controllers/single_test_runner.py:344
>> +                actual_driver_output.error = (actual_driver_output.error or '') + error
> 
> I don't see any other place where we set actual_driver_output.error. Why is this needed?

actual_driver_output is the return value of self._driver.run_test.
Then, actual_driver_output.error may contain a error of executing WTR for actual case.

[Fujii Hironori]

(In reply to comment #27)
> Looks good to me. Can automated tests be added for this fix?

I don't think so. I have no idea.

[Fujii Hironori]

Created attachment 304227 [details]
Patch

[Alexey Proskuryakov]

CC'ing some folks who may have an idea for how to make a regression test for this fix.

[WebKit Commit Bot]

Comment on attachment 304227 [details]
Patch

Clearing flags on attachment: 304227

Committed r213918: <http://trac.webkit.org/changeset/213918>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.