[OpenAFS] Authentication without aklog
Thu, 31 Jul 2014 15:29:47 -0500
I've had a few users and administrators complain to me from time to time
about the existence of 'aklog'. (By 'aklog' I really mean any mechanism
to convert krb5 tickets to AFS tokens, but I'm referring to them all as
'aklog' for simplicity.) The need for an AFS-specific authentication
step is an annoyance for some, and for some others, prohibits the use of
AFS in some applications.
The first time I heard this I was a bit surprised, but that may be just
because I'm very used to the 'aklog' approach and find it intuitive. You
need to tell the kernel what credentials you want it to use for AFS
access; makes sense to me.
The alternative is to effectively "guess" what credentials we should be
using, which is what NFSv4 does (rpc.gssd). That is, all you need to do
to authenticate is to run a plain 'kinit' or equivalent (with no
knowledge of AFS/NFS), and the kernel tries to find the ccache you used
and turn it into a token itself. This approach has a noticeable number
of cases where it does the wrong thing, and so you hear complaints about
it from time to time. But when it works correctly, it's invisible, so I
expect the only time you hear about it (from users) are the complaints.
But, at least for some environments, the downsides of rpc.gssd are
smaller than the downsides of needing to run 'aklog' at all. I don't
expect openafs to completely get rid of aklog and move to an rpc.gssd
approach (I personally don't think I would like that), but I think there
may be some compromises that would be helpful to people.
Some possible approaches:
- We could have a client option to make rpc.gssd-like behavior a
fallback, if no other credentials were set with e.g. 'aklog'.
- We could have an option to turn on rpc.gssd-like behavior as a
fallback for a specific PAG. That is, within a pag you say something
like 'pagctl pick-my-creds --enable'. ('pagctl' doesn't exist yet, of
course; it's just an idea in my head)
- We could have an option to aklog that would automatically renew
credentials using the information available to aklog at the time. For
example, if you run, say, 'aklog -autorenew', aklog would tell the
kernel its KRB5CCNAME and any other relevant information, and the
kernel would later on run an equivalent aklog command to obtain
credentials in exactly the same way for that PAG. For example, if
KRB5CCNAME was set to FILE:/tmp/foo1234, the kernel would later on
try to use the ccache file /tmp/foo1234 to obtain creds before the
existing creds expire.
With those last two, you still need to run some afs-specific command for
authentication, but you only need to run it once for the entire life of
the "session"/PAG. That's still annoying, but it removes the need for
renewing credentials within the actual session, which is not always
practical. (k5start/krenew works for many cases, but sometimes you don't
have the ability to run your own command)
Anyway, I'm sending this to the -info list to try to get some feedback
from other users. I've mentioned these ideas briefly to a few others,
who seemed to want something in this direction, but I'd like to get
opinions and feedback from as many sites as possible. The actual details
of implementing these ideas is a discussion for developers, but just
seeing what behavior people want is a discussion for here.
So, please speak up. Does this sound helpful do you? Or a bad idea? Or
anything else? Feedback is appreciated, even if it's just a "yeah, that
would be nice to have", or even "I don't really care".
Also, while I would prefer that any feedback goes to the list, if you
don't want to send to the list for any reason, even just sending
something directly to me is helpful, so I at least know your opinion.