Re: proc_flush_task oops

[ebiederm@xmission.com (Eric W. Biederman)]

Dave Jones <davej@codemonkey.org.uk> writes:

> On Thu, Dec 21, 2017 at 12:38:12PM +0200, Alexey Dobriyan wrote:
>  > On 12/21/17, Eric W. Biederman <ebiederm@xmission.com> wrote:
>  > > I have stared at this code, and written some test programs and I can't
>  > > see what is going on.  alloc_pid by design and in implementation (as far
>  > > as I can see) is always single threaded when allocating the first pid
>  > > in a pid namespace.  idr_init always initialized idr_next to 0.
>  > >
>  > > So how we can get past:
>  > >
>  > > 	if (unlikely(is_child_reaper(pid))) {
>  > > 		if (pid_ns_prepare_proc(ns)) {
>  > > 			disable_pid_allocation(ns);
>  > > 			goto out_free;
>  > > 		}
>  > > 	}
>  > >
>  > > with proc_mnt still set to NULL is a mystery to me.
>  > >
>  > > Is there any chance the idr code doesn't always return the lowest valid
>  > > free number?  So init gets assigned something other than 1?
>  > 
>  > Well, this theory is easy to test (attached).
>
> I'll give this a shot and report back when I get to the office.
>
>  > There is a "valid" way to break the code via kernel.ns_last_pid:
>  > unshare+write+fork but the reproducer doesn't seem to use it (or it does?)
>
> that sysctl is root only, so that isn't at play here.

ns_capable(CAP_SYS_ADMIN) will allow root in a user namespace.   So the
sysctl should be fuzzable.

The ns_last_pid sysctl is still not in play because it changes
task_active_pid_ns (aka the pid namespace of the callers pid) not
pid_ns_for_children.  So it still is not in play.

Every time I think of a "valid" way to break the code, I double check
myself and find there are already checks in place to prevent that.

Eric