Bug 239506 – Fix accelerated compositing for null ScrollingCoordinator and overflow scroll in LayerAncestorClippingStack.

• Status: NEW
• Component: Compositing

[Jigen Zhou]

With null ScrollingCoordinator and overflow scroll, a crash can occur in LayerAncestorClippingStack::updateWithClipData in the following code block:

        if (existingEntry.clipData.isOverflowScroll && !clipDataEntry.isOverflowScroll) {
            ASSERT(scrollingCoordinator);
            scrollingCoordinator->unparentChildrenAndDestroyNode(existingEntry.overflowScrollProxyNodeID);
            existingEntry.overflowScrollProxyNodeID = 0;
        }

(https://github.com/WebKit/WebKit/blob/main/Source/WebCore/rendering/LayerAncestorClippingStack.cpp#L137)

[Jigen Zhou]

Created attachment 457951 [details]
Patch

[Simon Fraser (smfr)]

Do you have a testcase that hits this crash?

[Jigen Zhou]

(In reply to Simon Fraser (smfr) from comment #2)
> Do you have a testcase that hits this crash?

Yes, we have a crash when running MiniBrowser built for Playstation with AC for youtube.com. Sometimes, just scrolling down and entering a clip, WebProcess can crash in the above code block.

[Simon Fraser (smfr)]

Can you create a reduction and include a test with the patch?

[Radar WebKit Bug Importer]

<rdar://problem/92342843>

[Jigen Zhou]

Created attachment 458416 [details]
Patch

[Jigen Zhou]

(In reply to Simon Fraser (smfr) from comment #4)
> Can you create a reduction and include a test with the patch?

Hi Simon, I just updated the patch with a reduction test that can reproduce the crash for Playstation build. This is the first time I am adding a layout test, please let me know I missed anything. Thanks.

[Jigen Zhou]

Created attachment 458418 [details]
Patch

Including the new test file.

[Jigen Zhou]

Created attachment 458561 [details]
Patch

Updated the test.

[Simon Fraser (smfr)]

Comment on attachment 458561 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=458561&action=review

> LayoutTests/compositing/overflow/overflow-scroll-hidden-swap-crash.html:64
> +                }, 100);
> +              }, 100);
> +            }, 100);
> +          }, 100);
> +        }, 100);
> +      }, 100);

This test will take 600ms to run, and seems to be trying to catch non-deterministic behavior. It needs to run quicker (imagine if all 40,000 tests took > 500ms).

> Source/WebCore/rendering/LayerAncestorClippingStack.cpp:69
> +//
> +// If scrollingCoordinator is null, overflowScrollProxyNodeID should be 0 for all existing entries
> +// of the clipping stack. The non-zero for overflowScrollProxyNodeID can only possibily be set in
> +// RenderLayerCompositor::updateScrollingNodeForScrollingProxyRole, but it won't be called for null
> +// scrollingCoordinator.
> +//

Weird for this comment to be here and not next to the relevant code.

> Source/WebCore/rendering/LayerAncestorClippingStack.cpp:147
> +        if (existingEntry.overflowScrollProxyNodeID && existingEntry.clipData.isOverflowScroll && !clipDataEntry.isOverflowScroll) {

If existingEntry.clipData.isOverflowScroll is set, I would expect a non-0 existingEntry.overflowScrollProxyNodeID. When does this happen?

[Simon Fraser (smfr)]

This test does not trigger a crash on macOS without the rest of the patch, so I suspect something specific to Playstation here.

[Jigen Zhou]

(In reply to Simon Fraser (smfr) from comment #10)
> Comment on attachment 458561 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=458561&action=review
> 
> > LayoutTests/compositing/overflow/overflow-scroll-hidden-swap-crash.html:64
> > +                }, 100);
> > +              }, 100);
> > +            }, 100);
> > +          }, 100);
> > +        }, 100);
> > +      }, 100);
> 
> This test will take 600ms to run, and seems to be trying to catch
> non-deterministic behavior. It needs to run quicker (imagine if all 40,000
> tests took > 500ms).

The multiple levels of setTimeout in this test is meant to catch a crash with higher possibilities due to the special non-deterministic nature. I tested with various shorter durations, but it seems that the current one captures the crash better.

> 
> > Source/WebCore/rendering/LayerAncestorClippingStack.cpp:69
> > +//
> > +// If scrollingCoordinator is null, overflowScrollProxyNodeID should be 0 for all existing entries
> > +// of the clipping stack. The non-zero for overflowScrollProxyNodeID can only possibily be set in
> > +// RenderLayerCompositor::updateScrollingNodeForScrollingProxyRole, but it won't be called for null
> > +// scrollingCoordinator.
> > +//
> 
> Weird for this comment to be here and not next to the relevant code.
>
 
When I read the code for LayerAncestorClippingStack, it took me a while to understand why for non-zero overflowScrollProxyNodeID, scrollingCoordinator can be assumed to be valid in multiple functions in the file. If comment like this was somewhere in the code, it would have helped me a little bit as a new developer in this part. 

> > Source/WebCore/rendering/LayerAncestorClippingStack.cpp:147
> > +        if (existingEntry.overflowScrollProxyNodeID && existingEntry.clipData.isOverflowScroll && !clipDataEntry.isOverflowScroll) {
> 
> If existingEntry.clipData.isOverflowScroll is set, I would expect a non-0
> existingEntry.overflowScrollProxyNodeID. When does this happen?

Apparently this is not the case in the crash we have experienced. We have null scrollingCoordinator, overflowScrollProxyNodeID is always zero (as the above comment shows). The above test was to simulate the scenario when swapping overflow between scroll and hidden.  Then we have isOverflowScroll set for existingEntry, isOverflowScroll is not set for clipDataEntry.

[Jigen Zhou]

(In reply to Simon Fraser (smfr) from comment #11)
> This test does not trigger a crash on macOS without the rest of the patch,
> so I suspect something specific to Playstation here.

In our case, scrollingCoordinator is always null. Otherwise, there is some randomness to capture the crash. But just analyzing the code itself, it is pretty obvious to me there is a crash scenario there.