Implementing authenticated Smart HTTP - which URLs to secure

Implementing authenticated Smart HTTP - which URLs to secure

[Jonas H.]

Howdy!

I'd like to implement HTTP authentication for Git Smart HTTP using 
Dulwich (a Python binding):

1) read-only if unauthenticated and write only if authenticated
2) read/write only if authenticated

I couldn't find any documentation on which URLs need be secured and what 
response codes are expected in case the cloner/pusher is unauthenticated.

Is there any documentation on Smart HTTP workflow?  The C sources 
(vanilla Git and libgit2) didn't help me too much since I found it very 
difficult to follow the code-flow... it's probably just abstracted too 
well :-)

Thanks!
Jonas

Re: Implementing authenticated Smart HTTP - which URLs to secure

[Shawn Pearce]

On Sun, Jul 15, 2012 at 6:43 AM, Jonas H. <jonas@lophus.org> wrote:
> I'd like to implement HTTP authentication for Git Smart HTTP using Dulwich
> (a Python binding):
>
> 1) read-only if unauthenticated and write only if authenticated
> 2) read/write only if authenticated
>
> I couldn't find any documentation on which URLs need be secured and what
> response codes are expected in case the cloner/pusher is unauthenticated.

Smart HTTP uses HTTP authentication, so return 401 with a proper
WWW-Authenticate header to prompt the client for authentication, and
use the Authorization header to verify the client. Return 403 to tell
the client they cannot access the service because the Authorization
header is invalid[1].

You can tell check for a write request by looking at the service
parameter on the /info/refs request, if its git-receive-pack, you are
about to receive a write, so you want to prompt for authentication.
You should also check for authentication on the /git-receive-pack
request.

[1] This is actually a lie. The servers I have written over the years
return 200 OK with a special Git payload in this case. The payload
uses the "ERR" in the /info/refs response to print a message to the
client telling the user access is forbidden. This allows a custom
message to be sent, and stops the client from discarding the message
and falling back to the dumb protocol.

> Is there any documentation on Smart HTTP workflow?  The C sources (vanilla
> Git and libgit2) didn't help me too much since I found it very difficult to
> follow the code-flow... it's probably just abstracted too well :-)

There is no authentication/authorization in the server components in
git-core. This is delegated to the web server that runs in front of
Git, just like with the system SSH server handling authentication for
Git over SSH.

Re: Implementing authenticated Smart HTTP - which URLs to secure

[Jonas H.]

On 07/15/2012 10:49 PM, Shawn Pearce wrote:
> On Sun, Jul 15, 2012 at 6:43 AM, Jonas H. <jonas@lophus.org> wrote:
>> I'd like to implement HTTP authentication for Git Smart HTTP using Dulwich
>> (a Python binding):
>>
>> 1) read-only if unauthenticated and write only if authenticated
>> 2) read/write only if authenticated
>>
>> I couldn't find any documentation on which URLs need be secured and what
>> response codes are expected in case the cloner/pusher is unauthenticated.
>
> Smart HTTP uses HTTP authentication, so return 401 with a proper
> WWW-Authenticate header to prompt the client for authentication, and
> use the Authorization header to verify the client. Return 403 to tell
> the client they cannot access the service because the Authorization
> header is invalid[1].
>
> You can tell check for a write request by looking at the service
> parameter on the /info/refs request, if its git-receive-pack, you are
> about to receive a write, so you want to prompt for authentication.
> You should also check for authentication on the /git-receive-pack
> request.

Thanks man! That helped a lot. I figured it has something to do with the 
argument to /info/refs but didn't make the connection between "I want to 
upload" and "git-receive-pack" -- it's pretty confusing that the 
argument is "receive" if I want to *upload*. At least when you don't 
know the argument is the name of the service the client wants the server 
to start.

> [1] This is actually a lie. The servers I have written over the years
> return 200 OK with a special Git payload in this case. The payload
> uses the "ERR" in the /info/refs response to print a message to the
> client telling the user access is forbidden. This allows a custom
> message to be sent, and stops the client from discarding the message
> and falling back to the dumb protocol.

I just experienced that exact problem. What worked for me is just 
keeping responding with 401 Unauthorized.  Git will give up with a nice 
error message ("Authentication failed") after a 2 or so tries.

Do you know of any better way that does not require using the raw Git 
protocol?

> There is no authentication/authorization in the server components in
> git-core. This is delegated to the web server that runs in front of
> Git, just like with the system SSH server handling authentication for
> Git over SSH.

Well that explains why I couldn't find anything helpful in the code :-) 
Though I'm a bit confused by the lack of documentation on the subject 
since I'm probably not the first one to set up a Git server with 
authentication?!

Jonas