Re: The elephant in the room is... (was RE: [caplet] Re: ADsafe, Take 6)
- On 23/10/2007, Larry Masinter <lmm@...> wrote:
>Maybe I'm being horribly unfair to protocol designers, but implementors do.
> On standards:
> The benefit of HTTP and XML and HTML is not that they are well-designed protocol and syntax and language, but that there are many different and (more-or-less) interoperable implementations for many operating systems and languages and well-deployed support infrastructure; with enough general agreement about them at the lower layers that you can get on with it defining the next layer up. So a "let's redesign them to be cleaner" effort isn't helpful, really. You'd have to be 10 times or 100 times better before getting traction.
> Protocol designers don't "gloss over" escaping; protocol designers are software developers (or maybe developers-gone-bad) for whom escaping is generally an ugly after-the-fact design addition or compromise between allowing everything to be encoded but letting simple cases be encoded concisely. Think of it as Huffman coding at the design level.
An example is entities in URIs embedded in HTML.
is invalid HTML, but browser implementors, faced with a bunch of
markup writers who don't understand escaping decided that they should
guess what it means, so it means something vastly different from
> <p><html>... </html></p>That's how it's done today, and that's the source of a huge number of problems.
> is how you write html in html, not because xx; and &symbolname; are wonderful quoting mechanisms, but because the &entity; syntax was already there, and inventing another one for what was seemed like an uncommon case appeared unnecessary.
And a point I was trying to make (don't remember if I made it) was
that embedding one structured language in another shouldn't require
escaping. You should be able to include the parse tree directly using
some kind of consistent (quote ...) syntax.
> On quoting:Agreed. But there's a third possibility:
> No matter what your escaping and encoding system, developers will have problems with them: you either understand the general principle or you don't.
I understand them and I've got deadlines and it'll usually work unless
some lunatic gives their kid a name with apostrophes, so I'll not
think through it now, and then come back to it once I've gotten some
sleep which'll happen never.
> I know (from ancient experience) that most programmers learning LISP had trouble with thinking about X, (QUOTE X) and (EVAL X) when learning Lisp. The problem is keeping track of the different layers of interpretation – it isn't the syntax.I think there's something fundamentally different about embedding by
escaping and embedding by injecting attaching another languages parse
tree to your own. Escaping requires clients of your language to be
able to also parse any embeddable languages, and agree on the parse
tree that any given string should produce.
> Common Lisp added many other escaping conventions: `(let ((,x "abc\"def") (y ',z)) ,w) so it's hard to claim that S-expressions have consistent delimiters.Fair enough, but that's only a concern if you are embedding using escaping.
> On layering of escaping/quoting:Quite right. If you're trying to identify a "secure" subset of any
> The multiple layered quoting systems work well enough, because each layer does its own escaping/encoding and unescaping/unencoding and tools either hide or assist with the process. It's only when you're writing a program trying to process multiple layers simultaneously that you have trouble.
system (like web applications) which use multiple languages then you
have two choices:
- identify a safe subset of each language individually
- deal with all languages at once and try to identify a safe subset of the union
The first approach is certainly easier but has
least-common-denominator problems -- you have to exclude things that
might be allowable under the second approach.
If you consider the three individually, you have to conclude that the
If you deal with all languages at once, then you can apply your
To do that you have to parse all three languages and then hope that
your parse trees agree with browsers' interpretations. My point is
that's easier if you have a unified parse tree representation for all
languages that appear in a document.
> The reason that ADsafe is having trouble is that it is trying to do filtering without actually using the normal layer software for parsing and interpretation, and skip what turns out to be necessary complexity. Try to write a regular expression that will determines whether a Lisp program might divide by zero, and you'd have similar problems.I disagree. URLs are more than a pointer into a registry of handlers.
> On checking URLs:
> I think you can't check for invalid URLs by examining a program's syntax because (a) programs can construct URLs, and you can't check for invalid (vs. valid) URLs any more than you can do all array bounds checking at compile time and (b) the rules for what constitute a "safe" URL are complicated and evolving. After all, a URL is just a reference to a registry of protocols, which requires the registering body define some syntax for how the URL syntax might identify something or invoke some protocol or process. Each URL scheme has its own syntax and story for what might be "safe" to execute in different contexts, but that depends as much on the implementation of the URL-interpreter as anything else.
Again, you'll get fewer false negatives if you do that classification
in the context of the larger document.
> If you're going to do dynamic URL safety checking, there's not much point in doing syntactic checking, because you'll get lots of false positives ("this is unsafe" when it isn't) and won't catch any more problems syntactically than would be caught by the run-time check.There's always reason to do static checking since it let's you skip
runtime checks :) But runtime checks are out of scope for ADSafe.