Add the Priority Hints changes to the html spec

[pmeenan]

This applies the specified changes from the Priority Hints spec to html for #7150.

The main changes are:

• Add a new fetchpriority attribute and fetchPriority IDL attribute to the img, link, script and iframe elements.
• Plumb the attribute through to the underlying fetch for the resource associated with a given element.

• At least two implementers are interested (and none opposed):
  • Chrome (Shipped in 103)
  • Mozilla
• Tests for all of the interfaces are already being run as part of the Priority Hints spec:
  • https:/web-platform-tests/wpt/blob/master/priority-hints/
• Implementation bugs are filed:
  • Chromium: https://bugs.chromium.org/p/chromium/issues/detail?id=821464
  • Gecko: https://bugzilla.mozilla.org/show_bug.cgi?id=1797715
  • WebKit: https://bugs.webkit.org/show_bug.cgi?id=247291
• MDN:
  • Already documented as experimental from the existing spec, can be moved to non-experimental after merged:
  • img
  • link
  • script
  • iframe

---

/embedded-content.html ( diff )
/images.html ( diff )
/index.html ( diff )
/indices.html ( diff )
/infrastructure.html ( diff )
/links.html ( diff )
/scripting.html ( diff )
/semantics.html ( diff )
/timers-and-user-prompts.html ( diff )
/urls-and-fetching.html ( diff )
/webappapis.html ( diff )

[pmeenan]

One thing I didn't see working when building locally was that the attribute description did not appear to populate for fetchpriority like it does for all of the other attributes. Best as I can tell, it gets it from the glossary table but it didn't populate in any of the elements where the attribute was added.

[domenic]

Tests

Can you be sure to add tests for script and iframe as well?

Can you add tests that when a service worker intercepts requests (from all four elements), the request has the correct .priority?

One thing I didn't see working when building locally was that the attribute description did not appear to populate for fetchpriority like it does for all of the other attributes.

Were you running the full build.sh? It's done via a somewhat wonky setup, but it should work. The PR preview does not work because it's not running the full build.sh...

[pmeenan]

Thanks for the thorough review. First pass of the easy cleanups are done, going to look closer at the ones that are going to require a bit more thinking.

I'll submit a set of WPT's through the Chromium flow and link them here after they have landed.

Were you running the full build.sh? It's done via a somewhat wonky setup, but it should work. The PR preview does not work because it's not running the full build.sh...

Yes, with the full build.sh, the generated local file doesn't have a description attached to any of the fetchpriority content attributes. They all look like this:

The glossary table at the end which is where it looks like the content attribute descriptions for the other attributes are pulled from looks like it is populated correctly:

[annevk]

I noticed that you defined this for iframe, but you haven't actually defined how it's passed through to the navigate algorithm and ultimately ends up in Fetch.

"process the iframe attributes" initializes a variable that is then thrown away. (I see now that @domenic also noted this and apparently this isn't limited to the iframe element.)

[pmeenan]

Sorry about the delay, just getting back to this now that fetch has merged. I think I cleaned up most of the dangling variables, formatting and references.

The big gaping hole still to fill is plumbing the iframe attribute through the navigation logic to the underlying fetch. Will follow-up shortly when I get that bit sorted out.

[pmeenan]

OK, I think it is plumbed through the iFrame case now and I don't think I missed any of the linkages along the way. It mirrors the path of referrer policy and adds a fetch priority to the document state that is used on document navigation. It defaults to auto but can be explicitly set in the iFrame case.

[pmeenan]

@domfarolino Thanks for taking a look. Looks like there was also another use of script options in the timers section. Both are now explicitly setting it to "auto".

[domfarolino]

Per the second paragraph of #8470 (review), I just want to confirm that the module script inheritance of priority hints is indeed correct and intended?

[pmeenan]

@domfarolino Ah, no. Fixed, thanks.

In the spec discussions we determined that the fetch priority wouldn't cascade so it is now wiped and set to "auto".

[annevk]

I found editorial issues. It looks like the iframe propagation is done correctly now, but it would be good to double check once all the other things are in order.

[pmeenan]

Thanks. I cleaned up the whitespace and wrapping (and tweaked where some paragraphs broke to make it much easier to spot errors). I think I got them all.

I tweaked the language around setting the priority when processing the iframe attributes (remaining comments above) but there's one part of it I'm wavering on.

I switched fetchpriority to priority in the spec text is to set the priority used when processing the iframe attributes with a link to the fetch standard which matches how referer policy links out.

I'm wondering if it would be better to say that it is used to set the request fetch priority in the document state (which then sets the priority at fetch time) like the note below mentions. It's technically more accurate about the direct effect of what it sets but setting the priority is the dev-facing end-state.

[annevk]

Thanks! I suppose it would be good to combine those sentences in some fashion. Only "The fetchpriority attribute is a fetch priority attribute." has normative meaning. Perhaps adding "also" in the note would help or perhaps the "purpose" bits should generally be in notes, though that would be a larger change.

Maybe @domenic can have a final look at this.

[domenic]

Made it halfway through before I have to run for the night; will try to pick this back up tomorrow. I found and fixed some nits, and one substantive question. Also, we need to fix the attribute not showing up in the element definitions; it's not clear why that's broken but I'll try poking at it.

[domenic]

Do we want to allow setting this on Link headers as well? If so we need to modify "apply link options from parsed header attributes".

[pmeenan]

Done

[domenic]

Hmm.

I can think of three possible ways that fetchpriority="" could impact iframe fetches. It could impact:

1. Changes to the src="" attribute
2. Navigations of any sort, e.g. location.href = "https://other.example.com/" inside the iframe or frames[0].location.href = "..." from outside.
3. Navigations specifically to the history entry that was loaded.

You've chosen (3), which seems surprising. I'd expect (1). For example, with your (3), you get semantics like the following:

iframe.fetchPriority = "high";
iframe.src = "https://example.com/";

await waitForLoadEvent(iframe);
iframe.fetchPriority = "low";
iframe.contentWindow.reload(); // uses "high", not "low"

await waitForLoadEvent(iframe);
iframe.contentWindow.location.href = "https://other.example.com/"; // uses "low"

await waitForLoadEvent(iframe);
iframe.contentWindow.back(); // uses "high", not "low"

However, I will note that your choice matches what we've done for referrer policy... hmm. Thoughts welcome from @domfarolino @jakearchibald.

[annevk]

I wonder if what is specified for referrer policy actually matches implementations. That does seem like rather weird behavior.

[domenic]

For referrer policy it makes sense, we spent a lot of time thinking about it, and I think Dom updated Chromium to match. If you fetched it with a given referrer (policy) the first time, you should keep that. For priority I'm less sure, since it's performance-ish and feels like it's meant more to apply to the container.

[annevk]

Fair, but presumably referrer policy also applies to 1? I think that's the thing that stood out as weird to me.

It's a good reminder though that these kind of fetch attributes don't translate well to iframe as navigation is just a much more complicated endeavor.

[jakearchibald]

We also stick srcdoc on the history entry, and seem to be leaning towards doing the same for sandbox #6809 (comment), although that isn't how it currently works in browsers.

srcdoc isn't quite the same, since setting it triggers a navigation.

I think I agree with @annevk here that there's no perfect choice.

[domenic]

Fair, but presumably referrer policy also applies to 1? I think that's the thing that stood out as weird to me.

Right, the current value of the referrerpolicy="" attribute will apply to src="" changes. I can see how I made things confusing by contrasting (3) and (1); this PR does indeed do (1) + (3). I meant that (1) alone would be more expected to me.

So I think the question is: is fetchpriority="" more like loading="", i.e. only applies to (1), or is it more like referrerpolicy="", i.e. applies to (1) + (3)? I tend toward the former, but don't insist on it. I can certainly see the case that it's more like the latter.

@pmeenan, what are your thoughts?

[pmeenan]

I'm starting to think that we should remove the attribute from iFrames entirely and leave it for subresources. Chrome hasn't implemented it yet and as I think about it more, trying to schedule the document fetch may not be the right level of scheduling for an iFrame.

We might be better off thinking of iFrame scheduling on its own and explore something like "defer" or other concepts that more align with lazy-loading but with more explicit control. That way it could work for scheduling the iFrame creation independent of document src (injected html, external url, srcdoc, etc).

It's quite possible that trying to prioritize an iFrame's document fetch against subresources on the page that contains it could be completely ineffective since it lives in a different document context and may end up partitioned into separate cache and networking pools depending on the browser-level sandboxing.

[pmeenan]

I stripped out the attribute from iFrame's (as well as all of the plumbing) and added the processing for link headers.

Since it's not implemented anywhere for iFrames yet and it likely won't work as people would expect, it's better to handle iFrame scheduling as its own thing (off to update the explainer and WPT tests now).

[domenic]

Woohoo, merging! Please remember to remove the iframe WPTs and update MDN, if you haven't already.