Re: draft-ietf-httpbis-jfv: what's next

["Poul-Henning Kamp" <[e-mail]>]

--------
In message <CAEK7mvoXqyX3cADJytjU+C158EULgPLbzAb5kiUN=[e-mail]>
, Matt Menke writes:
>I think the draft looks good, but have a couple comments:
>
>The token rule in RFC7230 already includes asterisks, so I don't think =
>identifier or token_or_asterix is needed.

Yes, I just fixed that.

>Would it make sense to codify behavior if a part of a 
>h1_common_structure value fails to parse, at least if it uses the 
>proposed "><" format)?  I suspect what browsers do is inconsistent here, 
>and having some official rule (ignore the entire element vs ignore the 
>entire line vs ignore the broken parameter) seems like it would be worth 
>having?  I'd go with throw away the entire header line, if it uses the 
>new format and that happens, since that's easiest to standardize on.

So this is a bit of a sticky wicket.

Today that is a per-header decision, for instance Accept-Encoding
can safely ignore anything it doesn't understand/parse, whereas
Content-Encoding has to be parsed perfect.

It is also a soft spot which has been used in a number of creative
attacks on deeper layer in HTTP/1 sandwiches.

Looking forward, if we want to be able to use CS to build H3
compression, we cannot allow CS headers with format errors.

I'm uncomfortable with a rule which says "just ignore", so I would
propose that failure to parse a the CS level should cause a 4xx
error, just like an ascii BEL in a HTTP1 header would.

But please note that this is only at the CS level, how valid CS
which is semantically invalid (ie: "Content-Length: ABCD") should
be handled is outside the scope of this ID.  I'm not even sure
we can give a meaningful "default" rule.

>I think it's unfortunate that the HTTP/1 serialization can't distinguish 
>between identifiers, numbers, and timestamps.

Yes, but we don't really get to decide where we start.

My hope is that we can build a machine-readable specification language
for HTTP headers from which the "semantic parsing" code can be generated,
but that is clearly in the "future work" column.

>It means that 
>per-specific-header logic will have to be responsible for that extra 
>round of parsing for HTTP/1 headers.

Not necessarily.  Parsing CS in HTTP/1 serialization is very trivial
and it is not obvious to me that it always would or should be a
separate step.  With a specification language as mentioned above,
you probably would generate combined CS+semantic parser code.

The big advantage of CS is that we don't need to know the semantics.

If your implementation receives "My-Private-Header: >[...]<" it can
take it apart and present it as a native datastructure, and the
application logic can apply the privatly known semantics to that.


-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
[e-mail]         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.