Git over HTTP; have flexible SASL authentication

Git over HTTP; have flexible SASL authentication

[Rick van Rein]

Hello,

Git providers are inventing proprietary extensions to HTTP authentication
for Git.  It seems smarter to use SASL for this purpose, possibly allowing
the client a choice and authentication ringback to the client's own domain.

I wrote an Internet Draft to allow just that, and we implemented it for
Apache, Nginx and FireFox.  I would love to learn if this list considers
it a sensible extensions to Git.
https://datatracker.ietf.org/doc/html/draft-vanrein-httpauth-sasl

If you think this is useful, it would be wonderful to mention that on the
HTTP WorkGroup list as well, because they are now considering HTTP-SASL
for adoption, a formal requisite to send "Authorization: SASL" headers.
https://datatracker.ietf.org/wg/httpbis/about/


Thanks!

Rick van Rein
InternetWide.org


Apache  :- https://gitlab.com/arpa2/apachemod/-/tree/master/arpa2_sasl
           https://gitlab.com/arpa2/apachemod/-/tree/master/arpa2_diasasl
Nginx   :- https://github.com/stef/ngx_http_auth_sasl_module
FireFox :- https://gitlab.com/arpa2/http_sasl_client

Re: Git over HTTP; have flexible SASL authentication

[Junio C Hamano]

Rick van Rein <rick@openfortress.nl> writes:

> Git providers are inventing proprietary extensions to HTTP authentication
> for Git.  It seems smarter to use SASL for this purpose, possibly allowing
> the client a choice and authentication ringback to the client's own domain.

To adopt things like this, the work to extend how to make extensible
what is on WWW-Authenticate in the thread that contains this recent
message https://lore.kernel.org/git/Y9LvFMzriAWUsS58@coredump.intra.peff.net/
may be relevant, perhaps?



    

Re: Git over HTTP; have flexible SASL authentication

[Jeff King]

On Fri, Jan 27, 2023 at 09:06:36AM -0800, Junio C Hamano wrote:

> Rick van Rein <rick@openfortress.nl> writes:
> 
> > Git providers are inventing proprietary extensions to HTTP authentication
> > for Git.  It seems smarter to use SASL for this purpose, possibly allowing
> > the client a choice and authentication ringback to the client's own domain.
> 
> To adopt things like this, the work to extend how to make extensible
> what is on WWW-Authenticate in the thread that contains this recent
> message https://lore.kernel.org/git/Y9LvFMzriAWUsS58@coredump.intra.peff.net/
> may be relevant, perhaps?

It's relevant, but I think there's a ways to go. That is just about
passing WWW-Authenticate headers to helpers, which can then try to make
sense of them. But Git would still only understand getting back a
username/password from the helper, and passing it along to curl. And
hopefully we'd do it all through curl's SASL support, and not invent our
own handling.

I'm not sure what all that might might look like. I'm sure Matthew has
probably thought about it, so I'll let him say something more
intelligent. :)

-Peff

Re: Git over HTTP; have flexible SASL authentication

[Matthew John Cheetham]

On 2023-02-01 03:24, Jeff King wrote:

> On Fri, Jan 27, 2023 at 09:06:36AM -0800, Junio C Hamano wrote:
> 
>> Rick van Rein <rick@openfortress.nl> writes:
>>
>>> Git providers are inventing proprietary extensions to HTTP authentication
>>> for Git.  It seems smarter to use SASL for this purpose, possibly allowing
>>> the client a choice and authentication ringback to the client's own domain.
>>
>> To adopt things like this, the work to extend how to make extensible
>> what is on WWW-Authenticate in the thread that contains this recent
>> message https://lore.kernel.org/git/Y9LvFMzriAWUsS58@coredump.intra.peff.net/
>> may be relevant, perhaps?
> 
> It's relevant, but I think there's a ways to go. That is just about
> passing WWW-Authenticate headers to helpers, which can then try to make
> sense of them. But Git would still only understand getting back a
> username/password from the helper, and passing it along to curl. And
> hopefully we'd do it all through curl's SASL support, and not invent our
> own handling.
> 
> I'm not sure what all that might might look like. I'm sure Matthew has
> probably thought about it, so I'll let him say something more
> intelligent. :)
> 
> -Peff

These are the same thoughts I have on this. My patches only add support for the
Git -> helper information flow, but don't make an attempt to change how helpers
can change Git's (curl's) response or behaviour.

I had earlier iterations [1] of the patch that included the ability to change
the auth type/headers that curl would respond with based on the helper's output.

For example:

protocol=https
host=example.com
password=<TOKEN>
authtype=bearer

..would have Git configure curl to set CURLOPT_XOAUTH2_BEARER. However I pulled
these patches to keep the scope of that series smaller. It's still my plan to
reintroduce such a patch series in the future.

I'd imagine that Git could advertise to helpers that its version of curl supports
SASL and a helper could enable or select this mechanism. Alternatively it could
just be a Git config to enable it outright `credential.useSasl` or something.
I haven't had chance to review SASL yet.

Thanks,
Matthew


[1] https://lore.kernel.org/git/230118.86k01kxfr7.gmgdl@evledraar.gmail.com/T/#m37fffe327593ca4f7bf32a205b7ee1d1ecd1ed46

Re: Git over HTTP; have flexible SASL authentication

[Rick van Rein]

Hi Junio,

(Sorry for the late reply)

> > Git providers are inventing proprietary extensions to HTTP authentication
> > for Git.  It seems smarter to use SASL for this purpose, possibly allowing
> > the client a choice and authentication ringback to the client's own domain.
> 
> To adopt things like this, the work to extend how to make extensible
> what is on WWW-Authenticate in the thread that contains this recent
> message https://lore.kernel.org/git/Y9LvFMzriAWUsS58@coredump.intra.peff.net/
> may be relevant, perhaps?

This framework tries to get away from just Basic / Digest auth, but
if I read correctly, it still assumes a single authentication step,
which greatly reduces the strength of authentication.

SASL assumes that the endpoints can hold state and progress through any
challenge/response exchange, in multiple steps if needed.  This is not
trivial with stateless HTTP, but we fixed it with a "s2s" argument which
can hold the (datestamped, signed, encrypted) state on the server side.

A few mechanisms that spring to mind as useful with Git over HTTP-SASL are

 - Kerberos / GSS-API, desirable for companies using its single sign-on
 _ FIDO, currently being added to SASL to benefit all protocols
 - OPAQUE, a very modern crypto, developed in IETF now
 - SXOVER-PLUS, our own work that calls back to a user's domain for login
	(so no server-stored credentials needed, only access control)

I'm curious if this would indeed be considered an improvement to Git.

Cheers,
 -Rick

Re: Git over HTTP; have flexible SASL authentication

[Junio C Hamano]

Rick van Rein <rick@openfortress.nl> writes:

> SASL assumes that the endpoints can hold state and progress through any
> challenge/response exchange, in multiple steps if needed.  This is not
> trivial with stateless HTTP, but we fixed it with a "s2s" argument which
> can hold the (datestamped, signed, encrypted) state on the server side.

That's nice to know, and its applications below do look useful, at
least to me.

> A few mechanisms that spring to mind as useful with Git over HTTP-SASL are
>
>  - Kerberos / GSS-API, desirable for companies using its single sign-on
>  _ FIDO, currently being added to SASL to benefit all protocols
>  - OPAQUE, a very modern crypto, developed in IETF now
>  - SXOVER-PLUS, our own work that calls back to a user's domain for login
> 	(so no server-stored credentials needed, only access control)
>
> I'm curious if this would indeed be considered an improvement to Git.
>
> Cheers,
>  -Rick