Which gss calls may block on IO

[Simo Sorce]

On Fri, 2017-04-14 at 22:15 +0300, Isaac Boukris wrote:
> On Fri, Apr 14, 2017 at 6:53 PM, Greg Hudson <ghudson at mit.edu> wrote:
> > On 04/14/2017 08:26 AM, Isaac Boukris wrote:
> >> If I only acquire credentials for krb5 mech, can I assume that
> >> gss_accpet would not block on IO?
> >
> > I think that's true, assuming the keytab and profile and /etc/gss/mech
> > file aren't on a network filesystem.  In the future we might have
> > pluggable keytab types, at which time a keytab module might do blocking
> > I/O, but that would be an exotic case.
> >
> >> Is there any information available on which gss calls may block on IO
> >> and which not?
> >
> > I'm not aware of a writeup.  From memory: gss_acquire_cred() and
> > gss_init_sec_context() can block on either DNS or AS/TGS operations.
> > The per-message functions shouldn't block.  gss_inquire_cred() can be
> > less trivial that it sounds (because it can force a resolution of which
> > ccache to use which would ordinarily be delayed until init_sec_context)
> > but I don't think it does network I/O (unless there's a ccselect plugin
> > module which does so, or $HOME/.k5identity is on a network filesystem).
> > The other inquire functions shouldn't block.  The delete/release
> > functions shouldn't block.
> 
> Thanks for this information.
> fwiw I also noticed rfc 2743 has the following comment about
> gss_{init,accept} calls only: "this call may block pending network
> interactions".

RFC2743 is generic, and with mechanisms different from krb5
gss_accept_sec_context() can indeed block. For example the NTLMSSP
mechanism on a member server would have to contact the DC on accept() to
be able to perform the challenge-response with the client.

Simo.

-- 
Simo Sorce
Sr. Principal Software Engineer
Red Hat, Inc