[Webkit-unassigned] [Bug 237160] REGRESSION(r290356-r290351?): [ iOS EWS ] 3 imported/w3c/web-platform-tests/service-workers/service-worker/* tests are constant text failures.

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=237160

--- Comment #14 from Chris Dumez <cdumez at apple.com> ---
Ok, so here is what seems to be happening (I looked at partitioned-service-worker-getRegistrations.tentative.https.html but I am almost certain the issue is the same for the other one).

The test relies on an ID js variable that lives inside the service worker. The test expects this ID to not change during the execution of the test.

The test initially launches a service worker for its origin. Thanks to my change in r290352, we now run this service worker in the same WebProcess as the test (better for perf and memory usage).
The issue is that the test then loads another origin (in an iframe and in a new window). As a result, WebProcessProxy::didStartProvisionalLoadForMainFrame(), we determine that the WebContent process is no longer "origin-clean" (it has loaded more than one eTLD+1). When this happens, we have logic to terminate the service worker that's inside the current process and relaunch it is a new "origin-clean" process.

Because we relaunch the service worker, the ID variable changes and it confuses the test.

Technically, we could say that the test is wrong to assume the service worker will not exit/relaunch during the execution of the test, since service workers can famously be killed at almost any point.
However, the test is relying on FetchEvent.waitUntil() (with a promise that it doesn't resolve until the very end of the test). Per the specification, I believe we're not supposed to exit service workers that have pending events.

So while r290352 made the issue obvious, this is not really a regression from r290352. We were getting lucky before and the service worker was launching in another WebContent process (likely a dedicated one) from the beginning. As a result, the service worker would not relaunch when the test would start loading different origins.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-unassigned/attachments/20220302/d324a05e/attachment-0001.htm>