Origin of blob URLs doesn't match what implementations do

[mkruisselbrink]

As mentioned in Origin of Blob URLs, the origin of a blob URL should be the origin of the incumbent settings object when the URL was created. The simple "parse the origin from the url in the first string of the path of the blob url" algorithm the URL spec uses doesn't accomplish that for opaque origins.

The way at least chrome implements this is by maintaining a separate blob url -> origin mapping, and using that to lookup the origin of a blob url. I assume firefox does someting similar (see also whatwg/html#1322 for an example where this behavior matters).

Should the URL spec be updated to actually match the File API spec about blob URLs, and to match what implementations seem to do?

[annevk]

I think the standard has a better approach. Otherwise looking up an aspect of the URL involves going through a global lookup table.

Note that the origin of a blob URL is determined at creation time and can therefore be stored in the blob URL (as the File API prescribes). So extracting it from the URL is fine too. And modifying it doesn't work since the entire URL is used as key.

[annevk]

Also, note that in https://bugzilla.mozilla.org/show_bug.cgi?id=1270451 Gecko fixed this.

[mkruisselbrink]

I think the standard has a better approach

Which standard. Currently the file API spec and the URL spec disagree here. According to the url spec the origin of a blob url with an opaque origin is a new unique opaque origin, since its created from the serialized url. While according to the file API spec the origin should be the same opaque origin as that of the incumbent settings object that created the URL.

Of course both would serialize to "null", so from the javascript URL API point of view this wouldn't matter, but when other parts of the spec try to determine if URLs are same origin this does matter. In particular for things like creating workers from blob URLs. But also other situations. Should a blob URL be same origin with the setting object that created them or not? I think yes, and I believe implementations agree (at least chrome does). The File API spec certainly agrees. The current URL spec seems to be the odd one out.

[mkruisselbrink]

Also, note that in https://bugzilla.mozilla.org/show_bug.cgi?id=1270451 Gecko fixed this.

Not sure what exactly gecko fixed. Are in latest gecko blob URLs no longer same origin with their creating context? I would find that very surprising and broken. And in particular "What is the problem with the model from the URL standard? It seems to work fine in Chrome." makes no sense. Chrome does not implement the model from the URL standard as it doesn't work.

[annevk]

I don't understand. The File API defines the origin of a blob and puts it in the URL it creates. The URL Standard extracts. I don't think there is a mismatch.

[mkruisselbrink]

The file API defines not only the origin of a blob, but also the origin of a blob URL. And defines the origin of that URL to be the same as that of the context that creates it. In particular when that origin is an opaque origin the origin of the blob URL needs to still be the same origin with it. That is impossible to do through serialization and deserialization via the URL itself (separately the FileAPI spec doesn't actually define how implementations should serialize an opaque origin in a blob URL, it leaves that implementation defined). But some out-of-band data (like the already existing Blob URL store) is needed to actually have blob URLs be same origin with the context that created them.

[annevk]

I see, opaque origins. I think File API should not serialize those other than as "null". Seems like we might reveal more than we want otherwise.

What does this mean in practice for opaque origins though? What scenario is broken and is that a problem or a feature?

[mkruisselbrink]

I think in practice browsers do indeed serialize opaque origins as "null" in blob URLs, even though the spec might allow other things. So that part is fine.

One of the features we'd lose if blob urls would no longer be same origin with their creator would be the feature discussed in whatwg/html#1322, the ability for a sandboxed iframe to start a worker using a blob URL.

Something else that would break without this same originness is using fetch to read blobs in opaque origins (but creating a blob url and then fetching that). That more or less is why blink/webkit originally implemented this lookup for the origin (although in chrome it is slightly worse since the entire FileReader API is implemented in terms of fetching blob URLs. So FileReader was broken in opaque origins without this origin mapping. But that's more an implementation detail).

[annevk]

@bakulf thoughts?

[annevk]

Okay, @bakulf seemed fine with reverting Firefox to the earlier behavior. I'll need to file a bug. So just to confirm, @mkruisselbrink, the idea is that we'd do some kind of global lookup when determining the origin of a blob URL.

Does that effectively mean the blob store is per unit of related-origin browsing contexts?

[domenic]

Probably per event loop, since they work in workers as well? I dunno, I just heard that phrase and am popping my head in.

[mkruisselbrink]

The blob store concept as defined in the latest version of the File API spec definitely needs work, as it doesn't actually reflect what browsers do (and also doesn't seem to match what older versions of the File API spec did). Having it per execution context as in the current spec certainly doesn't seem like it would make sense, as using blob urls in other execution contexts is definitely something that works today. I believe in chrome the thing that best reflects the "blob store" concept is actually fully global. But then the blob url -> (possibly opaque) origin map chrome also maintains is more per event loop. I really need to find the time to figure out what browsers are actually doing in this space, and then figure out how to best write that down in the spec, as what is currently in the specs about blob URLs is all over the place and certainly doesn't match reality... I should be able to find some time for that in the next few weeks.

[annevk]

I filed https://bugzilla.mozilla.org/show_bug.cgi?id=1282504 to get Firefox back to using a store again for the origin. Leaving this open until we figure out what the scope of that store is.

[annevk]

@mkruisselbrink so in Chrome new URL("blob:http://foo.com/bar").origin == "http://foo.com" which seems a little weird if you just use the blob URL store, no?

[mkruisselbrink]

In Chrome it first tries the blob URL store, and if that doesn't have an entry it parses the origin out of the URL string.