Boris Zbarsky wrote:
On 1/15/13 2:48 AM, Brendan Eich wrote:
I don't buy it. Processes are heavy enough, especially on mobile
devices, that Chrome pragmatically limits the number in the pool and
shares non-HTTPS (last I heard, someone correct me if this has changed
materially).
My point was that if they're doing process per origin, effectively,
then we can put different origins in different DOM Rust tasks.
Cool, I'm with you -- this has been a gleam in various eyes for a while.
It turns out (I'm told) that IE9 and up (possibly IE8) process-isolate
such that window.open returns a DCOM proxy, sometimes!
For finer-grained arrangements such as cross-site <script src=> evolved
to use a subordinate label for the ad code (to subject it to stricter
CSP), I don't propose inversion of control flow, however.
Alex Russell zeroed in on this in the Q&A for my AppSecUSA talk where I
proposed the cross-site script sub-label, perhaps in order to see
whether process or finer isolation was in the cards. I don't think it is
a "compatible-enough" change from today's script inclusion execution model.
But this does not mean we should keep to the original
synchronous-calling (with async-loading) design of windows and frames.
That seems to be something that broke without "breaking the web".
We don't have to use separate processes for them (though we should
perhaps design with the option of doing so) but we _can_ run them in
parallel, kill them independently of each other, and so forth. I'm
slightly more interested in that than in the crash-proofing aspects,
though obviously I like those too.
Same here.
Or put another way, you'll be able to nuke all those ad iframes
without affecting the site that embeds them.
There's a race to get advertizers to use iframes. I still see tons of
cross-site script inclusion, CSS flyovers, etc. Nuking those via the
sub-label idea has appeal. Tell me if you think
no-iframe/cross-site-script ads are a dying technique, though.
Process per iframe? How about per img?
Just because we have the option of using a process doesn't mean we
have to.
We do have the inverted control flow issue on some parts of the DOM,
though. How much of that can we break?
Note that the linked-to document has some discussion about how this
should interact with the renderer process limit in Chrome.
Good point about expanded principals. I'm not sure of the b2g details
(will ask around and report back) but I've proposed that cross-site
scripts (script src="http://ads-r-us.com/...") should be given a
distinct trust label that's subordinate in the lattice to the including
page's origin-label.
I think the point is that if we ever want WebKit to actually implement
our b2g stuff, we should ensure that any non-same-origin communication
is async in our designs.
I don't agree for cross-site scripts. Those are labeled with the
includers origin, depriving us of an important policy knob. Yet their
code must have synchronous DOM access. Either they become a non-issue,
or we do sub-labels without inverted control flow.
/be
_______________________________________________
dev-servo mailing list
[email protected]
https://lists.mozilla.org/listinfo/dev-servo