Re: [Integrity] Revisit RFC6920 URIs for benefit of servers and Semantic Web applications

Alright, thanks.

I understand the CSP usage is already in the wild, otherwise I would make
the same submission to CSP. Perhaps future editions can add an option to
refer to the IANA registry (or even use ni: outright)?

It's difficult to prove that forward compatibility would be
"impossible". Looking
forward I imagine older user-agents might try to parse a URI as an
`option-expression`, and so I'd prefer the alternative that looks less like
a URI, other things being equal. Additionally, use of the equals sign has
precedent in other contexts, such as the `parameter` production for media
types (RFC2045), the HSTS `directive` production (RFC6797), the HTTP Link
header (RFC5988), and more.

As a maintainer of libraries that parse many of these productions, and
author of Web robots (that will likely, in the distant future, have to grok
CSP and SRI), I would find re-using an existing production less complex
than implementing a new one.

Since I don't find any option-expression production in CSP, would the WG be
amenable to using the equals sign character in place of colon, as follows,
or an existing similar production?

; proposed changes
option-expression  = option-name "=" option-value
option-name        = 1*option-name-char
option-value       = *option-value-char
option-name-char   = ALPHA / DIGIT / "-"
option-value-char  = ALPHA / DIGIT / "-" / "+" / "." / "/"

(I noticed the current ABNF only allows a single character, here I've
allowed multiple characters. As was probably intended before, this
modification leaves open the possibility of unambigiously expanding the
range of option-value later to allow a quoted-string or urlencoded value,
or both.)

With regards to re-using productions from CSP, I still see a significant
distinction between CSP and SRI: The former is making a series of
assertions about an HTTP request, the latter is making an assertion about a
link relation, and has implications beyond HTTP. If the WG wants to define
a subset of behavior as it applies to just HTTP, that's good, but I still
would like to be wary of future re-use and generalization. I don't believe
importing the CSP syntax satisfies this. Compared to the benefit of being
able to re-use quoted-string/urlencoded/etc, over defining a new one (as
the Editor's Draft currently does), I believe the introduced complexity of
having a different style of token, or even using ni: outright, is minimal.

Finally, do you refer to the `hash-source` production [1], but without the
single quotes? I can't find a literal `hash-src` production like you refer
to, idk if it might be unreleased somewhere, or something like that.

Cheers,

Austin.

[1] <https://w3c.github.io/webappsec/specs/CSP2/#hash_source>
(Unfortunately, I can't find a canonical URL for this version of the
editor's draft, but the most recent change appears to be
commit 7fe5ce1e2e54130702b0a678a40966f39fab1bab)

On Wed, Mar 11, 2015 at 8:35 PM, Devdatta Akhawe <dev.akhawe@gmail.com>
wrote:

> Hi Austin
>
> We should definitely fix the whitespace issue. Re the proposed changes: we
> want to reuse the deployed CSP hash-src format. So, it is unlikely that we
> can change that format. I don't see that it explicitly makes it impossible
> to use RFC6920 URIs in the future, if we so deemed necessary? a ni:// URI
> can just be one without space and we can define its semantics as ignore
> content type specified earlier if it exists in ni:// URI. While not ideal,
> there is value to reusing the CSP hash-src format.
>
> cheers
> Dev
>
> On 11 March 2015 at 11:38, Austin William Wright <aaa@bzfx.net> wrote:
>
>> Understood that it doesn't need to support everything right from the
>> first version.
>>
>> The spec, as currently defined, could make minor adjustments to enhance
>> forward compatibility, that I do not believe would add additional cost to
>> user agent implementors.
>>
>> If RFC6920, for whatever reason, is completely off the table, then I
>> would suggest changing the format of the token so it doesn't look like a
>> valid RFC3986 URI, by replacing use of the colon and dash with equals and
>> comma, respectively:
>>
>> ; proposed changes
>> option-expression  =  option-name "=" [ option-value ]
>> hash-expression    = hash-algo "," base64-value
>>
>> This should be unambiguous, because neither comma nor equal sign will
>> appear in option-name nor hash-algo. And the lack of a colon prevents
>> RFC3986 parsing.
>>
>> This enhancement would also allow the TR to give tokens to hashes
>> matching those in the IANA registry ("SHA-256" instead of "sha256"),
>> further supporting forward compatibility.
>>
>> Aside: The ABNF in the current Editor's Draft seems to allow the
>> following string:
>>
>> option-name:option-valuesha256-Cg==
>>
>> i.e. without a WS between option-expression and hash-expression. Am I
>> reading this correct? /aside
>>
>> I don't think RFC6920 needs to be off the table, however; merely define
>> an ABNF subset compatible with RFC6920 `NI-URI` production. For all intents
>> and purposes, it would be just a coincidence that it happens to be a valid
>> NI-URI production (if funny-looking). The only significant difference (not
>> formatting/parsing related) would seem to be handling of ct= (appearing in
>> every hash production, instead of just once).
>>
>> I don't see anything wrong with mixing option-expression and NI-URI
>> productions, that might be an ideal solution to explore. I'm thinking all
>> option-expressions must be matched, if any, and at least one
>> hash-expression must be matched.
>>
>> Finally, the current option-value production prohibits the expression of
>> arbitrary media types. I regularly use quoted-strings (containing spaces
>> and special characters) in media types, including in Accept headers, for
>> instance with JSON Schema, e.g.:
>>
>> application/json; charset=utf-8; profile="http://example.com/book"
>>
>> (Note charset isn't a registered parameter for application/json, utf-8 is
>> by definition, but it's not prohibited, and I've had difficulty getting
>> some Web browsers to play nicely without it.)
>>
>> The current option-value production appears to prohibit a media type like
>> this, whereas the urlencoded ct= named information parameter utilizes
>> urlencoding.
>>
>> Thanks,
>>
>> Austin.
>>
>> On Wed, Mar 11, 2015 at 9:45 AM, Devdatta Akhawe <dev.akhawe@gmail.com>
>> wrote:
>>
>>> Hi Austin
>>>
>>> These are all great points, but we are not really trying to address
>>> these in the first version of SRI. The goal in version 1 is only to be able
>>> to check the hashes of scripts and links. That said, I tend to agree with
>>> you that this should be in our radar; do you think the spec, as currently
>>> defined, makes it impossible to address these concerns in future
>>> iterations? For example, the parser is intentionally forgiving of formats
>>> it is not aware of, so as to allow such changes in the future.
>>>
>>> cheers
>>> Dev
>>>
>>>
>>> On Mar 9, 2015 11:26 PM, "Austin William Wright" <aaa@bzfx.net> wrote:
>>>
>>>> [CCing swig because I know many of you are also using ni: URIs. Let me
>>>> know if there's anything to add!]
>>>>
>>>> I understand that the ni: URI was removed in a recent update to the SRI
>>>> draft. I would like to ask this be reconsidered.
>>>>
>>>> Using the ni: URI for SRI is important to Semantic Web applications as
>>>> it allows us to treat the assertion as a link relation. This distinction
>>>> might not be significant to many user-agents (and thus many on this list),
>>>> but in Semantic Web applications, especially Web servers, this is of great
>>>> significance, allows us to make useful relationships between resources.
>>>>
>>>> It also confers benefits to servers and application designers, as
>>>> RFC6920 defines a mapping between ni: URIs and a </.well-known/> URL. In a
>>>> corporate project under development, we're already using ni: URIs to keep a
>>>> content-addressable database of files, making them accessible through this
>>>> mapping. I intend to use Subresource Integrity when serving assets from
>>>> this store.
>>>>
>>>> It also provides an intuitive abstraction: If we think of the ni: URI
>>>> as identifying a resource (the definition of the URI), the integrity=
>>>> attribute is performing an assertion: "These two URIs must identify the
>>>> same information resource, otherwise abort!"
>>>>
>>>> For additional support for this use case, I'd like to propose making
>>>> the "integrity" attribute a a link-extension for RFC5988 Web Linking,
>>>> suitable for use on any declaration of a link.
>>>>
>>>> User agents do not need to think of the ni: URI as a URL if they do not
>>>> need to, they just follow the ABNF defined in the RFC. There's many cases
>>>> where URIs are used as identifiers in Web applications; in namespaces [1],
>>>> schemas (e.g. JSON Schema), DTDs, RDFa [2], and in rel= attributes in all
>>>> sorts of tags and HTTP headers.
>>>>
>>>> Additionally, RFC6920 defines a registry of hashes [3], to ensure
>>>> forward compatibility (e.g. SHA-3, when standardized later this year). I
>>>> would like to avoid duplication of effort defining a database of hashes.
>>>>
>>>> In short, (1) signing was an explicit goal of the ni: URI, along with
>>>> other features. Not using ni: would mean servers being unable to take
>>>> advantage of these other features; and, (2) don't forget about the HTTP
>>>> Link: header.
>>>>
>>>> Thanks for your consideration,
>>>>
>>>> Austin Wright.
>>>>
>>>> [1] http://www.w3.org/TR/html5/infrastructure.html#xml
>>>> [2] http://www.w3.org/TR/html-rdfa/
>>>> [3]
>>>> http://www.iana.org/assignments/named-information/named-information.xhtml
>>>>
>>>>
>>
>