Re: gssapi, crypto and afs/rxrpc

[J. Bruce Fields]

On Fri, Oct 16, 2020 at 05:18:26PM +0100, David Howells wrote:
> Hi Herbert, Dave, Trond,
> 
> I've written basic gssapi-derived security support for AF_RXRPC:
> 
> 	https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=rxrpc-rxgk
> 
> I've borrowed some bits from net/sunrpc/auth_gss/ but there's a lot in there
> that is quite specific to the sunrpc module that makes it hard to use for
> rxrpc (dprintk, struct xdr_buf).
> 
> Further, I've implemented some more enctypes that aren't supported yet by
> gssapi (AES with sha256/sha384 and Camellia), and that requires some changes
> to the handling as AES with sha384 has a 24-byte checksum size and a 24-byte
> calculated key size for Kc and Ki but a 32-byte Ke.
> 
> Should I pull the core out and try to make it common?  If so, should I move it
> to crypto/ or lib/, or perhaps put it in net/gssapi/?
> 
> There are two components to it:
> 
>  (1) Key derivation steps.
> 
>      My thought is to use xdr_netobj or something similar for to communicate
>      between the steps (though I'd prefer to change .data to be a void* rather
>      than u8*).
> 
>  (2) Encryption/checksumming.
> 
>      My thought is to make this interface use scattergather lists[*] since
>      that's what the crypto encryption API requires (though not the hash API).
> 
> If I do this, should I create a "kerberos" crypto API for the data wrapping
> functions?  I'm not sure that it quite matches the existing APIs because the
> size of the input data will likely not match the size of the output data and
> it's "one shot" as it needs to deal with a checksum.
> 
> Or I can just keep my implementation separate inside net/rxrpc/.

I'd love to share whatever we can, though I'm not really sure what's
involved.  Certainly some careful testing at least.

It's been about 15 years since I really worked on that code.  I remember
struggling with struct xdr_buf.  The client and server support
zero-copy, so requests and replies are represented by a combination of a
couple of linear buffers plus an array of pages.  My memory is that the
(undocumented) meanings of the fields of the xdr_buf were different for
request and replies and for server and client, and getting them right
took some trial and error.

--b.