- From: Poul-Henning Kamp <phk@phk.freebsd.dk>
- Date: Sat, 15 Oct 2016 15:49:04 +0000
- To: Martin Thomson <martin.thomson@gmail.com>
- cc: Matt Menke <mmenke@google.com>, HTTP Working Group <ietf-http-wg@w3.org>
-------- In message <CABkgnnXw7WacnMf4Nsx-drktn__V4afK61G67A5bT5SSdqaucQ@mail.gmail.com> , Martin Thomson writes: >On 15 October 2016 at 20:41, Poul-Henning Kamp <phk@phk.freebsd.dk> wrote: >> Looking forward, if we want to be able to use CS to build H3 >> compression, we cannot allow CS headers with format errors. > >I tend to agree with this, though there are levels of format errors. >For instance, if you use the >< notation and the < is absent, that's a >flat parse error (I would argue that the < is redundant actually, save >an octet). It is redundant, but it might still be a good idea. Truncation of headers happens a lot more than it should in the wild, so apart from the recursive role of the '<' I do like that it also tells you that you are not missing half the header. >But what I think that Matt is looking for is a grammar that supports >an in-band signal about type so that syntax checking can be done by >the parser (and not by the semantics layer). That - to me - seems >like a pretty reasonable request. Yes, I agree, but it runs into the very inclusive definition of RFC7230::token. We need three markers: '(h1_)number', 'h1_timestamp' and 'h1_blob', which are all valid 'identifier' (= RFC7230::token) today. We have three options: 1. Keep using RFC7230::token for 'identifier' Then only '<', '>', '{', '}', '[' and ']' are still available. We have already given '<', '>' meaning and even though we could disambiguate them, I'd really like to avoid overloading. So one proposal could be: h1_number = '}' number h1_timestamp = '{' number h1_blob = '[' base64 Leaving us with only ']' if we forgot something. (It rattles my OCD to use "precious" balanced glyphs this way...) (For reasons of transmission efficiency I am intentionally not proposing formats such as "{#" number "}") 2. Restrict 'identifier' If we use a restricted RFC7230::token for 'identifier', we can shake some special characters free for type marking duty. Maybe: identifier = ALPHA *(ALPHA / DIGIT / '_' / '-' / '*') h1_number = '#' number h1_timestamp = '$' number h1_blob = "'" base64 "'" 3. Let the semantic layer sort it out. As the draft does today. This has best H1/H2/HPACK transmission efficiency. This also enforces only the minimum necessary restriction on HTTP-header inventors. For instance: h1_blob is a valid identifier and thus a valid name of a dictionary. On the contra side, exploiting such "loopholes" is almost guaranteed to hurt H3 compression for that header later on. I picked 3 based on 'minimum intrusiveness', but I can live with all three. Given that a H3 compression likely will serialize the type, the two first alternatives are probably the most forward-compatible. -- Poul-Henning Kamp | UNIX since Zilog Zeus 3.20 phk@FreeBSD.ORG | TCP/IP since RFC 956 FreeBSD committer | BSD since 4.3-tahoe Never attribute to malice what can adequately be explained by incompetence.
Received on Saturday, 15 October 2016 15:49:31 UTC