Re: c/r: broken locking when executing map_files

[Cyrill Gorcunov <gorcunov@openvz.org>]

On Mon, May 07, 2012 at 12:21:32AM +0400, Vasiliy Kulikov wrote:
> On Sat, May 05, 2012 at 23:32 +0400, Cyrill Gorcunov wrote:
> > On Sat, May 05, 2012 at 10:53:06PM +0400, Cyrill Gorcunov wrote:
> > > On Sat, May 05, 2012 at 10:20:51PM +0400, Vasiliy Kulikov wrote:
> > > > > -	if (lock_trace(task))
> > > > > +	if (!ptrace_may_access(task, PTRACE_MODE_READ))
> > > > 
> > > > Probably it will be better to change mutex_lock_killable() to
> > > > mutex_lock_killable_nested() inside of lock_trace() instead of this change?
> > > > It would keep the race-free check.
> > > 
> > > Yup, if I'm not missing something SINGLE_DEPTH_NESTING should do the trick
> > > for us. I'll test and report.
> > 
> > Hmm, this doesn't work well, the mutex remanins killable so when one does
> 
> Does it show circular locking?  It shouldn't block if it uses
> mutex_lock+mutex_lock_nested.

Nope of course, once mutex_lock_nested called the kernel doesn't complain
anymore, but it rather sleep forever, until task killed, and that is wrong
i think. moreover, since executing from inside of map_files is special one
I think changing the whole lock_trace for this sake is a wrong approach.

> 
> >  | [root@neptune ~]# /proc/self/map_files/400000-419000
> > 
> > it sleeps forever until killed, which is not good I think. Vasiliy, could
> > you remind me what exactly is problem if we use unlocked ptrace_may_access
> > here?
> 
> There is a race between ptrace_may_access() and dname_to_vma_addr().
> The target task may do exec() between these calls and
> dname_to_vma_addr() will be called against a privileged task.  This may
> lead to a leak of task maps.

Wait, the proc_map_files_lookup requires the caller to be cap-sysadmin
granted, would not this be enough?

	Cyrill