What happens to the responsible browsing context of a settings object when the browsing context is discarded?

[bzbarsky]

https://html.spec.whatwg.org/multipage/window-object.html#discard-a-document says:

each environment settings object has strong references to its global object, responsible browsing context

Does this reference outlive the browsing context being discarded?

[domenic]

I looked into this and waffled back and forth on the answer.

Currently I'm leaning toward "yes", because if you save let old = frames[0].frames[0] and then unload frames[0], old (a global object ~ environment settings object) probably has lots of properties that depend on the browsing context, and we shouldn't do a spec-segfault if you access them. We'd have to do an audit of places that use "responsible browsing context"; the ones I've seen so far are not guarded by null checks.

On the other hand it seems like the intent of discarding a browsing context is pretty clearly to allow freeing memory? And maybe this hurts that? Unclear to me how badly.

[domenic]

Hmm, there may be a clue lower down.

A WindowProxy does not have a strong reference to the browsing context it was created alongside. In particular, it is possible for a nested browsing context to be discarded even if JavaScript code holds a reference to its WindowProxy object.

(I vaguely remember you and me working through adding that in the past.)

Right now I see a strong reference graph of WindowProxy -> Window -> environment settings object -> responsible browsing context, which "discard" does not sever, so the above statement seems false. Maybe if we severed the last link in "discard" it would become true.

Edit: diagram of the current situation in #3863.

[bzbarsky]

Right now I see a strong reference graph of

Yeah, I agree.

I wonder what UAs do in practice. I am pretty sure that Gecko severs the equivalent of the "environment settings object -> responsible browsing context" link and has null-checks as needed.

[annevk]

Note that elsewhere we concluded that there should indeed be null checks on these browsing context getters throughout. #4363 tracks adding them.