Bug 220637 – REGRESSION (r271472): [ Mac WK2 ] intersection-observer/target-deleted.html is timing out

• Status: RESOLVED
• Resolution: FIXED
• Component: New Bugs

[Truitt Savell]

intersection-observer/target-deleted.html

This test is timing out constantly sense around r271473

History:
https://results.webkit.org/?suite=layout-tests&test=intersection-observer%2Ftarget-deleted.html


I can reproduce this with command:
run-webkit-tests --iterations 2000 --exit-after-n-failures 1 --exit-after-n-crashes-or-timeouts 1 --debug-rwt-logging --no-retry --force --no-build -f intersection-observer/target-deleted.html

[Radar WebKit Bug Importer]

<rdar://problem/73220571>

[Truitt Savell]

I was able to bisect this to r271471

[Truitt Savell]

(In reply to Truitt Savell from comment #2)
> I was able to bisect this to r271471

r2714712 fails while r271471 passes actually. misstyped

[Simon Fraser (smfr)]

You mean https://trac.webkit.org/changeset/271472/webkit.

Seems pretty unlikely that's the cause.

[cathiechen]

Hmm, I can't reproduce it on Catalina debug wk2.
It seems specific to the release versions.
I'll try to build release locally.

[cathiechen]

(In reply to cathiechen from comment #5)
> Hmm, I can't reproduce it on Catalina debug wk2.
> It seems specific to the release versions.
> I'll try to build release locally.

Tried the release version on macOS Catalina and Big Sur, but can't reproduce it.

[Tim Horton]

I can reproduce with Truitt's command.

[Tim Horton]

Even a single iteration is sufficient (which is consistent with the bots, and will make debugging easier).

[Tim Horton]

The observerShouldBeRemoved interval is firing forever, with 1 outstanding intersection observer live (i.e., the test is catching exactly the failure it was intended to fix...)

[Tim Horton]

Nothing even remotely near the regression point makes any sense, so I'm just going to see if I can debug.

[Tim Horton]

Test was introduced in https://trac.webkit.org/changeset/237880/webkit

[Tim Horton]

I think this is just a test depending on GC and the GC not happening (despite the attempt to force it). If I run the test with --no-timeout, it spins forever, UNTIL the instant I try to remote-web-inspect it, then suddenly the element is collected, the IntersectionObserver goes away, and the test completes...

[Ryosuke Niwa]

Oh, try increasing the number of observers and check if any observer is collected? I suspect the test is buggy. Since we have a conservative GC, we can't guarantee all objects will be collected.

[Simon Fraser (smfr)]

Maybe the test needs to do something non-trivial in the loop to make gc() do something

[Tim Horton]

Wait, also, m_intersectionObservers is a Vector of WeakPtrs, and we're just getting its length... which won't decrease if one of the WeakPtrs gets zeroed, right? This is all a bit questionable.

[Ryosuke Niwa]

(In reply to Simon Fraser (smfr) from comment #14)
> Maybe the test needs to do something non-trivial in the loop to make gc() do
> something

It's never guaranteed that gc() will collect all garbage because we have a conservative GC. Some objects can be kept alive even if they're logically dead.

[Tim Horton]

Created attachment 418108 [details]
Patch

[EWS]

Committed r271736: <https://trac.webkit.org/changeset/271736>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 418108 [details].