Maciej Stachowiak wrote:
There's been some discussion of what request headers, if any,
XMLHttpRequest should disallow for setREquestHeader.
I think we really need a clear idea of what we are trying to do by
restricting headers. I propose that the following are valid reasons to
forbid setting a header:
1) It would allow for a possible security hole.
2) It would allow a client to cause the UA to violate the http RFC
(besides just requirements on syntax, obviously those are possible with
any header).
3) It could seriously interfere with correct operation of the network
layer (specifically, it could break in-progress or future requests, or
cause improper responses to be added to the fache.
I would tentatively say the following are not valid reasons to restrict
a header:
4) It could result in content that the UA might not be able to parse as
text or as XML (this can happen anyway with no custom headers).
5) It could result in a response that the UA can't parse at the network
level (unless we think this could cause problems so serious that
category #3 applies).
6) It would not match expectations for information that is commonly
present (but in a way that is not a security issue or a violation of the
http standard).
If we agree with this, then I think the standard should list the 3
criteria for restricting headers above, and give a justification for any
header that is restricted.
Applying these rules, I think the following http headers should be
restricted:
* Connection (2, 3) - could break other requests on same persistent
connection; MUST be 'close' in UAs that don't support persistent
connections
* Date (2) - only allowed to be correct date
* Keep-Alive (2, 3) - could break requests on same connection or confuse
UAs that don't actually support keep-alive
* Trailer (2) - only allowed with chunked transfer-encoding, some header
field names not allowed, must accurately reflect presence of trailer
* Transfer-Encoding (2, 3) - property of whole message, no way this
could properly be set by the XHR client
* Upgrade (1, 2, 3) - could allow escape to SSL, violating cross-site
scripting security rules, could be used to violate RFC by specifying
protocol UA does not understand, the latter could actually interfere
with correct client operation
* Expect (2, maybe 3) - client MUST not send an Expect header if it does
not intend to send a request body
* Host (1) - could be used to subvert cross-site sripting rules on
servers that use virtual hosting
* Referer (possibly 1, 2) - the RFC does not appear to allow this to be
set to anything but the resource that originated the request (and in
some cases requires it to not be sent). Could be a security violation if
a site depends on Referer always being either correct or absent.
* TE (3) - seems like requesting an unknown transfer encoding or
trailers for clients that don't understand trailers
We should limit Content-Length too. Apparently some proxies gets
confused if there are multiple Content-Length headers (i.e. the one that
XHR produces and the one that's added) and splits a single request into
two. I believe this would fall under both 1 and 2.
The following are suspicious. I wasn't able to figure out a specific
problem but they seem like they might have the potential for security
holes, breaking the protocol, or breaking the client's network stack:
Via
We restrict Via very likely for good reason. I'll poke the relevant
people to see why.
/ Jonas