[PATCH] Enforce that CR precede LF in chunk lines

[Ben Kallus]

> Still, there is a robustness principle which allows applications
> to parse requests with various deviations from the grammar.

Whether this is principle is good is a matter of opinion. I tend to
lean toward thinking that it is not (as you can probably tell) but
reasonable minds will differ on this point.

> You may want to be more specific what "off by one" means here.

Happy to :)

Here's an example of a payload that smuggles a request past Apache
Traffic Server to a Node.js backend:
```
POST / HTTP/1.1\r\n
Transfer-Encoding: chunked\r\n
\r\n
2\r\r
;a\r\n
02\r\n
2d\r\n
0\r\n
\r\n
DELETE / HTTP/1.1\r\n
Content-Length: 183\r\n
\r\n
0\r\n\r\nGET / HTTP/1.1\r\n\r\n
```
The exact mechanism of this payload is relatively unimportant (it has
to do with the `2\r\r;a`). The important point is that the POST is
seen by both ATS and Node, the DELETE is seen only by Node, and the
GET is seen only by ATS. Thus, the DELETE request is smuggled.

(A very similar attack worked on Google Cloud's classic application
load balancer, and on Akamai's load balancer until very recently when
companies patched the bugs. I'm still working on the writeup for those
bugs, but you can see us present the material here:
https://yewtu.be/watch?v=aKPAX00ft5s&t=2h19m0s)

You'll notice that the DELETE request has a Content-Length header.
This is because in order for the smuggling to go undetected, the
response to the DELETE request needs to be sent only after the GET
request is forwarded. One way to do this is to add a message body to
the DELETE request, so that it remains incomplete until the arrival of
the GET request. It is therefore necessary for the attacker to predict
the length of the GET request after it has passed through the reverse
proxy, so that this length can be used to compute the Content-Length
(or chunk size) in the DELETE request. Because reverse proxies often
modify requests, this is not always straightforward.

In this instance, I use a Content-Length header of 183 because with
the configuration of ATS that I was attacking, `GET /
HTTP/1.1\r\n\r\n` ends up becoming 178 bytes long due to the insertion
of X-Forwarded-For, Via, etc., +5 for `0\r\n\r\n`. If I had used a
length less than 183, then Node would send a 400 after responding to
the DELETE request, which makes the reverse proxy aware that request
smuggling has occurred. If I had used a length greater than 183, then
Node would time out waiting for the rest of the DELETE request's
message body. Thus, I need to guess the length exactly right to pull
off undetected request smuggling. Guessing correctly can be
challenging, especially when added headers have unpredictable lengths.
This is common with CDNs, which often insert random identifiers into
request headers.

If instead of using Content-Length, I had used a chunked message body
to smuggle the DELETE request, and the backend server allows bare LF
as a chunk line terminator, then my length guess could be one less
than the correct value without invalidating the message for servers
that accept bare LF in chunk lines. Thus, when developing future
request smuggling attacks, getting my length guess correct is a little
easier when the backend server allows bare LF chunk line endings.

> including manual testing such
> as with nc(1).

If you pass the -C flag, nc will translate LF to CRLF for you :)