Consider improving interoperability of <iframe> throttling margins.

[emilio]

What is the issue with the HTML Standard?

My understanding is that all browsers implement some sort of throttling of offscreen / invisible iframes.

In Firefox, this is kind of messy, and differs between in-process and out-of-process iframes right now.

It seems according to @szager-chromium in Chromium there's no margin at all.

I think some things could probably be specified properly:

• visibility: hidden / display: none should always be throttled.
• Offscreen iframes should be throttled, with some UA-defined margin perhaps?

Are there other cases where UAs throttle iframes? cc @khushalsagar

[szager-chromium]

For reference, here is chromium's logic:

https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/renderer/core/frame/frame_view.cc;drc=4834462d69ddf6e0ac6b27572618e4436de4a2f3;l=317

Historically, we have also had one or two weird carve-outs for throttling logic that were driven by real-world usage; for example:

https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/renderer/core/frame/frame_view.cc;drc=4834462d69ddf6e0ac6b27572618e4436de4a2f3;l=279

A tighter spec with WPT would be great.

[khushalsagar]

Other than the visibility based throttling @szager-chromium mentioned, content-visibility is the other reason I could find for it (@vmpstr for it). The throttling done for view transitions is already carved out separately.

[vmpstr]

c-v language is in https://www.w3.org/TR/css-contain-2/#cv-notes point number 7, but it's "should" language

[annevk]

What does margin refer to here? Criteria?

[emilio]

Distance with the viewport (as in, IntersectionObserver margins etc).

[annevk]

Thanks! This would be reasonable to standardize more. Some relevant WebKit code:

• LocalFrameView::viewportContentsChanged
• LocalFrameView::updateScriptedAnimationsAndTimersThrottlingState (we currently do not throttle display:none, but if everyone else does this seems reasonable to change)
• Page::preferredRenderingUpdateFramesPerSecond
• DOMTimer::updateThrottlingStateIfNecessary

[khushalsagar]

I'm wondering if content-visibility is a good concept to build on for this throttling; since the goal is similar: minimize rendering cost of content not relevant to the user (as opposed to just visible). For example, IIUC c-v will force layout for find-in-page. I'm wondering if lifecycle updates should similarly be resumed for offscreen iframes when the user triggers find-in-page, it's possible that a DOM update (which populates the content the user is searching for) is blocked on the next rAF.

@vmpstr might have input based on edge cases like this addressed for c-v.

[vmpstr]

For content-visibility, it's not a comprehensive list but selection, focus, and accessibility (find in page specifically) are all things that cause us to do rendering. FWIW, content-visibility: auto also has a user-agent defined margin for visibility that causes it to be relevant, which is similar to the iframe one if I understand correctly.

I don't think specifying the exact extent of this margin is worthwhile, but specifying more explicitly cases where we don't want throttling is useful.

[szager-chromium]

IIRC, render throttling already has a carve-out that allows find-in-page to work correctly. I could be wrong.

We should be mindful that throttling rendering for off-screen iframes also prevents badly-behaved iframes from sabotaging battery usage and performance of the embedding page. In some situations chromium will throttle setTimeout and setInterval timers to 1s, and I wouldn't want requestAnimationFrame to become an escape hatch that allows an iframe to run lots of code at high frequency.

[zcorpan]

Also see https://github.com/w3c/IntersectionObserver/issues/508#issuecomment-1545646474

What WebKit does there seems better from a privacy POV since it avoids the Text Fragment side-channel using APIs affected by thottling.

[emilio]

This is still on my list fwiw. I want to make a change to Gecko here to make it more consistent and confirm such a thing is web compatible before proposing it more formally.