Re: Fedora 32 rpc.gssd misbehavior

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

> On Jul 30, 2020, at 3:39 PM, Chuck Lever <chuck.lever@oracle.com> wrote:
> 
> 
> 
>> On Jul 30, 2020, at 3:10 PM, Simo Sorce <simo@redhat.com> wrote:
>> 
>> On Thu, 2020-07-30 at 13:59 -0400, Chuck Lever wrote:
>>>> On Jul 30, 2020, at 1:08 PM, Robbie Harwood <rharwood@redhat.com> wrote:
>>>> 
>>>> Simo Sorce <simo@redhat.com> writes:
>>>> 
>>>>> On Wed, 2020-07-29 at 14:27 -0400, Chuck Lever wrote:
>>>>>>> On Jul 29, 2020, at 1:19 PM, Chuck Lever <chuck.lever@oracle.com> wrote:
>>>>>>> 
>>>>>>> Hi!
>>>>>>> 
>>>>>>> I recently updated my test systems from EL7 to Fedora 32, and
>>>>>>> NFSv4.0 with Kerberos has stopped working.
>>>>>>> 
>>>>>>> I mount with "klimt.ib" as before. The client workload stops
>>>>>>> dead when the server tries to perform its first CB_RECALL.
>>>>>>> 
>>>>>>> I added some client instrumentation:
>>>>>>> 
>>>>>>> kernel: NFSv4: Callback principal (nfs@klimt.ib.1015granger.net) does not match acceptor (nfs@klimt.ib).
>>>>>>> kernel: NFS: NFSv4 callback contains invalid cred
>>>>>>> 
>>>>>>> I boosted gssd verbosity, and it says:
>>>>>>> 
>>>>>>> rpc.gssd[986]: doing downcall: lifetime_rec=72226 acceptor=nfs@klimt.ib
>>>>>>> 
>>>>>>> But it knows the full hostname for the server:
>>>>>>> 
>>>>>>> rpc.gssd[986]: Full hostname for 'klimt.ib' is 'klimt.ib.1015granger.net'
>>>>>>> 
>>>>>>> 
>>>>>>> The acceptor appears to come from the Kerberos library. Shouldn't
>>>>>>> it be canonicalized? If so, should the Kerberos library do it, or
>>>>>>> should gssd? Since this behavior appeared after an upgrade, I
>>>>>>> suspect a Kerberos library regression. But it could be config-
>>>>>>> related, since both systems were re-imaged from the ground up.
>>>>>>> 
>>>>>>> Also noticing some other problems on the server (missing hostname
>>>>>>> strings in debug messages, sssd_kcm infinite loops, and gssd
>>>>>>> sending garbage to the client after the NULL request that
>>>>>>> establishes the callback context).
>>>>>>> 
>>>>>>> But let's look at the client acceptor problem first.
>>>>>> 
>>>>>> I believe I found the problem.
>>>>>> 
>>>>>> 8bffe8c5ec1a ("gssd: add /etc/nfs.conf support") added a number of gssd config
>>>>>> options to /etc/nfs.conf, including "avoid-dns". The default setting of avoid-
>>>>>> dns is 1. When I set this option on my client system explicitly to 0, NFSv4.0
>>>>>> with Kerberos works again.
>>>>>> 
>>>>>> Is there a reason the default setting is 1?
>>>>>> 
>>>>> 
>>>>> Now that you mention DNS, this may be an interaction between a new
>>>>> default in Fedora 32 and how your environment is setup re DNS.
>>>>> 
>>>>> In F32 we changed the option dns_canonicalize_hostname from 'true' to
>>>>> 'fallback'.
>>>>> This is a transitional state to eventually move it to 'false' at some
>>>>> point in the future.
>>>>> 
>>>>> What it changes in practice is that it will first try the name passed
>>>>> in *as is* and only as a fallback try a CNAME if the name passed is not
>>>>> resolved as an A name. If you have principals in the KDC for both
>>>>> names, but you do not have keys in the keytab for both, you can have
>>>>> transitional issues.
>>>>> 
>>>>> Additionally we discovered a bug that causes non qualified names to
>>>>> fail resolution with the 'fallback' option.
>>>>> If your name in the principal is really not qualified it will try to
>>>>> qualify it anyway, so if your principal is literally nfs/foo@FOO
>>>>> libgssapi may try to use nfs/foo.my.domdain@FOO, where "my.domain" is
>>>>> what is defined in resolv.conf search path.
>>>>> 
>>>>> We are trying to address this regression.
>>>>> 
>>>>> So try to set dns_canonicalize_hostname to true to see if that may
>>>>> influence your issue. If so, please let me know, as we still need to
>>>>> address this where possible.
>>>> 
>>>> Also, please try setting `qualify_shortname = ""`.  (I did update the
>>>> config file we ship with Fedora, but upstream's default turns that on.
>>>> This is a temporary workaround while we merge something better
>>>> upstream.)
>>> 
>>> For completeness, I tried:
>>> 
>>> avoid-dns = 1
>>> dns_canonicalize_hostname = fallback
>>> qualify_shortname = ""
>>> 
>>> which is the default configuration out of the shrink wrap.
>>> 
>>> The workload hangs as before, and the acceptor is unqualified:
>>> 
>>> rpc.gssd[985]: doing downcall: lifetime_rec=84046 acceptor=nfs@klimt.ib
>>> 
>>> 
>>> The test is:
>>> 
>>> Configured domain name is "1015granger.net"
>>> 
>>> Fully-qualified client hostname is "manet.ib.granger.net"
>>> 
>>> Fully-qualified server hostname is "klimt.ib.granger.net"
>>> 
>>> mount command is "mount -o vers=4.0,sec=sys klimt.ib:/export /mnt"
>>> 
>>> In this case, both systems have keytabs and service principals, so
>>> the client automatically attempts to establish a GSS context for
>>> lease management and callback operations. The failure occurs because
>>> the server's principal is nfs@klimt.ib.1015granger.net but the
>>> acceptor now matches the server hostname from the mount command line,
>>> which is not always fully qualified.
>> 
>> Ok, TBH I personally consider the syntax you  are currently using as
>> working by accident and that you should really sue the FQDN on the
>> command line (I assume it works that way, right?), however I understand
>> this is also technically a regression, that said I do not think we can
>> really fix this case because your "shortname" is not short (it has a
>> dot in it) so the heuristicts won't trigger to qualify it even when you
>> set qualify_shortname="".
>> 
>> I have the feeling we'll break this case, and our answer will have to
>> be "use the fqdn on the command line".
> 
> See previous e-mail. Using the shrink wrap default settings, which
> includes qualify_shortname="", results in a hang on callback, as
> originally observed.
> 
> Users will notice this and complain: klimt.ib works for the NFSv3
> case and the NFSv4.1 case, and for NFSv4.0 when there is no keytab,
> but NFSv4.0,sec=* with a keytab eventually hangs.

I filed

https://bugzilla.redhat.com/show_bug.cgi?id=1867719

to document the behavior regression and some possible workarounds.
Further discussion about addressing the issue (possibly in nfs-utils)
can happen there.

Thanks!


--
Chuck Lever