Re: Q: selinux_bprm_committed_creds() && signals/do_wait

From: David Howells <dhowells@redhat.com>
To: Roland McGrath <roland@redhat.com>
Cc: dhowells@redhat.com, Oleg Nesterov <oleg@redhat.com>,
	Eric Paris <eparis@parisplace.org>,
	James Morris <jmorris@namei.org>,
	Stephen Smalley <sds@tycho.nsa.gov>,
	linux-kernel@vger.kernel.org
Subject: Re: Q: selinux_bprm_committed_creds() && signals/do_wait
Date: Fri, 01 May 2009 01:44:34 +0100	[thread overview]
Message-ID: <18312.1241138674@redhat.com> (raw)
In-Reply-To: <20090501000222.8BACEFC3BF@magilla.sf.frob.com>

Roland McGrath <roland@redhat.com> wrote:

> There is a wrinkle here I don't like.  The fatal signal is "committed to"
> (sender approved by security modules, etc.) "before" the exec, but gets
> delivered "after" the exec.

That's because the signal is delivered between the system call function
returning to entry.S and userspace resuming.  We could put a lot of
check...abort clauses in exec.c and the binfmts, but is it worth the hassle?

> I don't understand why install_exec_creds() is called as late as it is.
> Can't we do that in flush_old_exec()--you know, where it says:
> 	/* install the new credentials */
> ?

I believe it's something to do with the binfmt driver needing to access files
in the old security context between calling flush_old_exec() and calling
install_exec_creds() [compute_creds() as was].  It can't do some of the
accesses before calling flush_old_exec() because it has to do funky things
with mmap().

Actually, that comment should probably be removed.  IIRC, at one time I was
trying to set all the credentials there, but was told I couldn't do that.

David