Re: Thoughts on credential switching

[Jeremy Allison <jra@samba.org>]

On Thu, Mar 27, 2014 at 07:01:26AM -0700, Jeff Layton wrote:
> On Thu, 27 Mar 2014 14:06:32 +0100
> Florian Weimer <fweimer@redhat.com> wrote:
> 
> > On 03/27/2014 02:02 PM, Jeff Layton wrote:
> > 
> > >> This interface does not address the long-term lack of POSIX
> > >> compliance in setuid and friends, which are required to be
> > >> process-global and not thread-specific (as they are on the kernel
> > >> side).
> > >>
> > >> glibc works around this by reserving a signal and running set*id on
> > >> every thread in a special signal handler.  This is just crass, and
> > >> it is likely impossible to restore the original process state in
> > >> case of partial failure.  We really need kernel support to perform
> > >> the process-wide switch in an all-or-nothing manner.
> > >>
> > >
> > > I disagree. We're treading new ground here with this syscall. It's
> > > not defined by POSIX so we're under no obligation to follow its
> > > silly directives in this regard. Per-process cred switching doesn't
> > > really make much sense in this day and age, IMO. Wasn't part of the
> > > spec was written before threading existed
> > 
> > Okay, then we need to add a separate set of system calls.
> > 
> > I really, really want to get rid of that signal handler mess in
> > glibc, with its partial failures.
> > 
> 
> I agree, it's a hack, but hardly anyone these days really wants to
> switch creds on a per-process basis. It's just that we're saddled with
> a spec for those calls that was written before threads really existed.
> 
> The kernel syscalls already do the right thing as far as I'm concerned.
> What would be nice however is a blessed glibc interface to them
> that didn't involve all of the signal handling stuff. Then samba et. al.
> wouldn't need to call syscall() directly to get at them.

Amen to that :-).

However, after talking with Jeff and Jim at CollabSummit,
I was 'encouraged' to make my opinions known on the list.

To me, calling the creds handle a file descriptor just
feels wrong. IT *isn't* an fd, you can't read/write/poll
on it, and it's only done as a convenience to get the
close-on-exec semantics and the fact that the creds are
already hung off the fd's in kernel space.

I'd rather any creads call use a different type, even if
it's a typedef of 'int -> creds_handle_t', just to make
it really clear it's *NOT* an fd.

That way we can also make it clear this thing only has
meaning to a thread group, and SHOULD NOT (and indeed
preferably CAN NOT) be passed between processes.

Cheers,

	Jeremy.