Re: linux-next: build failure after merge of the nfsd tree

[Chuck Lever <chuck.lever@oracle.com>]

On Apr 29, 2013, at 2:57 PM, J. Bruce Fields <bfields@fieldses.org> wrote:

> On Mon, Apr 29, 2013 at 02:30:33PM -0400, Chuck Lever wrote:
>> 
>> On Apr 29, 2013, at 1:59 PM, "J. Bruce Fields" <bfields@fieldses.org>
>> wrote:
>> 
>>> On Mon, Apr 29, 2013 at 01:47:16PM -0400, Chuck Lever wrote:
>>>> 
>>>> On Apr 29, 2013, at 1:38 PM, "J. Bruce Fields"
>>>> <bfields@fieldses.org> wrote:
>>>> 
>>>>> On Mon, Apr 29, 2013 at 01:04:01PM -0400, Chuck Lever wrote:
>>>>>> Trond's nfs-for-next now has the new rpcauth_get_gssinfo() and
>>>>>> rpcauth_get_pseudoflavor() APIs, which are replacements for
>>>>>> direct calls into the GSS mech switch.  These APIs are a little
>>>>>> more generic, and more robust in the face of unloaded GSS kernel
>>>>>> modules.
>>>>>> 
>>>>>> Instead of gss_mech_get_by_OID(), I suspect you want
>>>>>> rpcauth_get_pseudoflavor(), but I haven't looked at the gssproxy
>>>>>> code.
>>>>> 
>>>>> It's doing
>>>>> 
>>>>> 		status = -EOPNOTSUPP; gm =
>>>>> 		gss_mech_get_by_OID(&ud->mech_oid); if (!gm) goto out;
>>>>> 		status = -EINVAL; status =
>>>>> 		gss_import_sec_context(ud->out_handle.data,
>>>>> 		ud->out_handle.len, gm, &rsci.mechctx, &expiry,
>>>>> 		GFP_KERNEL); if (status) goto out;
>>>>> 
>>>>> So we need a way to get from an OID and some mechanism-specific
>>>>> data to a context.
>>>>> 
>>>>> Looks to me like we just want to re-export gss_mech_get_by_OID().
>>>> 
>>>> The reason for the new wrappers is to load the kernel modules
>>>> properly before trying to discover the OID -> mechanism mapping.
>>>> 
>>>> Where are you calling it from?  If it's from outside of the GSS
>>>> module, how do you guarantee the rpc_gss_auth module is loaded?
>>>> What if the GSS mechanism for that OID isn't loaded?
>>> 
>>> Sorry, I should have said just "remove static from", not
>>> "re-export"--it's in the same module.  So there should be no problem
>>> here, right?
>> 
>> OK, gotcha.  Architecturally outside of the mech switch I'd like to
>> see OIDs passed around embedded in GSS tuples rather than by
>> themselves.
> 
> I'm not sure what you mean.  When I accept a gss context I don't yet
> know what service or qop it's going to be used with, I only know the
> mechanism OID.

RPC server GSS support didn't need the gss_mech_get_by_OID() interface before gssproxy, so I'm trying to understand why it is needed now.

But it sounds like you do need it now, so go ahead and make gss_mech_get_by_OID() global within the AUTH_GSS module. 

> 
>> An alternative would be to use gss_mech_get_by_name(), which is
>> already visible, loads GSS mechanism modules automatically, and
>> returns struct gss_api_mech *.  For NFS, we should already have a
>> clean mapping of mechanism name to pseudoflavor to GSS tuple.  Looks
>> like rsc_parse() already uses this API.
> 
> We don't have a name here, only an OID.
> 
>> Do you have gssproxy patches published in a git tree somewhere I could
>> review?
> 
> It's in my for-3.10 branch.
> 
> Which is more or less what I plan to submit for 3.10, so I'd prefer not
> to rebase it at this point.
> 
> It looks like just removing "static" would resolve the immediate
> conflict, is that right?  And then I'd happily help deal with cleaning
> up the API as followup work.



-- 
Chuck Lever
chuck[dot]lever[at]oracle[dot]com