Normative spec doesn't necessarily terminate service workers

[asakusuma]

I don't think the normative behavior of the storage directive provides the guarantees stated in the non-normative language, specifically around service workers.

Service Workers registered for an origin are terminated and deregistered.

From what I can tell, the spec says to simply call unregister() on every service worker registration for the matching origin. unregister() doesn't actually take effect until no client is using the registration. Even if reloading() did create a point in time in which the client was not using the registration, I'm not sure that iterating over each client and calling reload() will guarantee that there was a point in time in which no client was using the registration. Perhaps the Clear Registration algorithm should be invoked directly?

Original discussion: w3c/ServiceWorker#614
Related ticket: #1

[asakusuma]

FWIW, I've tested Clear-Site-Data: storage in Chrome 73 and Firefox 65, they both kick the service worker immediately, rather than waiting until no client is using the registration.

[jungkees]

For the spec language, I think we should use:

1. Remove scope to registration map[registration's scope url].
2. Invoke Clear Registration with registration.

Note that w3c/ServiceWorker#1415 moves the map entry removal step out of Clear Registration. So we need it to be merged before using the above steps.

/cc @jakearchibald, @matto, @asutherland, @SteveBeckerMSFT

[asutherland]

We noticed that as of WebKit build 171 Safari now passes all of the tests in https://github.com/web-platform-tests/wpt/blob/master/clear-site-data/storage.https.html which includes causing controlled pages to become uncontrolled, consistent with the cluster of issues associated with this one.

I can confirm for Firefox/Gecko that we believe the test behavior (and what's been proposed in this cluster) seems appropriate and consistent with previous discussions in the ServiceWorkers WG about CSD being an emergency kill-switch for ServiceWorkers, etc. and that we should go ahead with the spec changes to normalize this.

@youennf @cdumez From the reference to this issue in WebKit/WebKit#14032 and the decision to land the underlying WebKit change, I assume WebKit also believes we should specify the behavior of making controlled clients become uncontrolled?