Re: RT scheduling and a way to make a process hang, unkillable

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2009-02-16 at 15:16 -0500, Kyle Moffett wrote:
> On Mon, Feb 16, 2009 at 5:36 AM, Dhaval Giani <dhaval@linux.vnet.ibm.com> wrote:
> > On Sun, Feb 15, 2009 at 12:24:56PM +0100, Peter Zijlstra wrote:
> >> On Sat, 2009-02-14 at 16:51 -0800, Corey Hickey wrote:
> >> > The procedure is for a program to:
> >> > 1. run as root
> >> > 2. set SCHED_FIFO
> >> > 3. change UID to a user with no realtime CPU share allocated
> >>
> >> Hmm, setuid() should fail in that situation.
> >>
> >> /me goes peek at code.
> >>
> >> Can't find any code to make that happen, Dhaval didn't we fix that at
> >> one point?
> >
> > So after some searching around, I realized we did not. Does this help?
> > It fixes it on my system,
> >
> > --
> > sched: Don't allow setuid to succeed if the user does not have rt bandwidth
> 
> Erm, hrm, this reminds me of the old sendmail capabilities bug.  There
> are an awful lot of buggy binaries out there who assume that if they
> have uid 0 and they call setuid() that it cannot fail.  They then do
> all sorts of insecure operations, assuming that they have dropped to
> an unprivileged UID.  This one is especially bad because it could bite
> *any* program using setuid() which an admin happens to run with chrt.
> 
> Specifically, I personally think that:
>   *  Process is stuck and unkillable
> 
> is a much better result than:
>   *  Process runs arbitrary untrusted code with full-root privs in RT mode.

You have a point, however there are plenty of ways to fail setuid(), one
of them is severe memory pressure, another is exceeding rlimits, also
the security_*() hooks can do pretty much whatever.

So while security is important, its IMHO not a good enough reason to
preserve broken stuff.

[ dhaval, michael, it appears setuid() already returns errors outside
those specified by POSIX, so I'd rather fail with -ENOTIME, or similar,
rather than with -EAGAIN ]