Re: [PATCH] ptrace: restore smp_rmb() in __ptrace_may_access()

[Jann Horn]

On Wed, May 29, 2019 at 6:21 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> On 05/29, Jann Horn wrote:
> > (I have no clue whatsoever what the relevant tree for this is, but I
> > guess Oleg is the relevant maintainer?)
>
> we usually route ptrace changes via -mm tree, plus I lost my account on korg.
>
> > --- a/kernel/ptrace.c
> > +++ b/kernel/ptrace.c
> > @@ -324,6 +324,16 @@ static int __ptrace_may_access(struct task_struct *task, unsigned int mode)
> >       return -EPERM;
> >  ok:
> >       rcu_read_unlock();
> > +     /*
> > +      * If a task drops privileges and becomes nondumpable (through a syscall
> > +      * like setresuid()) while we are trying to access it, we must ensure
> > +      * that the dumpability is read after the credentials; otherwise,
> > +      * we may be able to attach to a task that we shouldn't be able to
> > +      * attach to (as if the task had dropped privileges without becoming
> > +      * nondumpable).
> > +      * Pairs with a write barrier in commit_creds().
> > +      */
> > +     smp_rmb();
>
> (I am wondering if smp_acquire__after_ctrl_dep() could be used instead, just to
>  make this code look more confusing)

Uuh, I had no idea that that barrier type exists. The helper isn't
even explicitly mentioned in Documentation/memory-barriers.rst. I
don't really want to use dark magic in the middle of ptrace access
logic...

Anyway, looking at it, I think smp_acquire__after_ctrl_dep() doesn't
make sense here; quoting the documentation: "A load-load control
dependency requires a full read memory barrier, not simply a data
dependency barrier to make it work correctly". IIUC
smp_acquire__after_ctrl_dep() is for cases in which you would
otherwise need a full memory barrier - smp_mb() - and you want to be
able to reduce it to a read barrier.

> >       mm = task->mm;
>
> while at it, could you also change this into mm = READ_ONCE(task->mm) ?

I'm actually trying to get rid of the ->mm access in
__ptrace_may_access() entirely by moving the dumpability and the
user_ns into the signal_struct, but I don't have patches for that
ready (yet).