Tie event loops to agents

[annevk]

Event loops ought to be 1:1 with agents, let's make it so.

Helps with #4198.

---

/browsers.html ( diff )
/canvas.html ( diff )
/history.html ( diff )
/structured-data.html ( diff )
/timers-and-user-prompts.html ( diff )
/web-messaging.html ( diff )
/webappapis.html ( diff )
/window-object.html ( diff )
/workers.html ( diff )

[annevk]

I'm looking for feedback on approach and how to tackle specific things I called out in comments.

[domenic]

Hmm, I'm worried this might not align well with our desire to give implementations more freedom. Maybe what I'm worried about is more relevant for agent clusters, but I thought it was possible (and indeed likely, in many-tab situations) for multiple agents to share an event loop.

[annevk]

Since that difference isn't observable, I think it's better for the design to aim for maximum concurrency.

We could add a note, but it might get complicated given JavaScript's expectations around forward progress for agents.

[domenic]

Thoughts so far:

• It's not clear why this needs to be 1:1. I can't find the part of the spec that makes use of that property. A smaller delta would be to allow many agents -> 1 event loop. You can still go from agent -> event loop that way, which is what I see happening in this patch.

• I think it's useful to have terms like "window event loop" or "worker event loop" still instead of "an event loop for a similar-origin window agent". Indeed, introducing dedicated <dfn>s for them would be an improvement over the status quo where the terms "browsing context" and "event loop" are linked separately. I guess you might also want "worklet event loop".

• The new "we don't describe how to handle this" note is scarier than the old one. Upon consideration I guess it's the same? It'd be good to flesh out an example. I suspect this is a very common case, so showing it (e.g. "the user navigates from URL X to Y within a single browsing context. The spec doesn't handle that") would be helpful.

[annevk]

I guess we could allow sharing, but it seems that would just make things more complicated, as it's not as simple as allowing any two agents to share, that won't work. Sharing it across origins is rather questionable given Spectre. Sharing it within an agent cluster is not allowed by JavaScript. Sharing it across browsing context groups would not allow for making background tabs use less CPU, but could be allowed if user agents don't care for that. Is that the one you want?

[domenic]

Per https://freenode.logbot.info/whatwg/20190118#c1931283 it looks like we want to take a slightly different approach here.

[domenic]

Great stuff, looking pretty close. I've tried to be concrete about there is left to do.

[annevk]

I carefully checked all outstanding feedback and I think I got it covered with the latest commit.

[domenic]

LGTM with some remaining nits which I trust you to resolve. Thanks again for the patience.