Re: [PATCH] cred: Use RCU primitives to access RCU pointers

[Jann Horn]

On Tue, Jan 28, 2020 at 12:48 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> On 01/28, Jann Horn wrote:
> > On Tue, Jan 28, 2020 at 8:28 AM Amol Grover <frextrite@xxxxxxxxx> wrote:
> > > task_struct.cred and task_struct.real_cred are annotated by __rcu,
> >
> > task_struct.cred doesn't actually have RCU semantics though, see
> > commit d7852fbd0f0423937fa287a598bfde188bb68c22. For task_struct.cred,
> > it would probably be more correct to remove the __rcu annotation?
>
> Yes, but get_task_cred() assumes it has RCU semantics...

Oh, huh. AFAICS get_task_cred() makes no sense semantically, and I
think it ought to be deleted.


There are the following users at the moment:

rdtgroup_task_write_permission() - uses it for acting on a task as
object, should be using objective credentials instead.

__cgroup1_procs_write() - same thing.

task_state() - should also be using objective credentials instead, you
wouldn't want a task to show up as belonging to root in there just
because it's in the middle of some overlayfs filesystem method or
something like that.

apparmor_getprocattr() - same thing as for task_state()

rpcauth_bind_root_cred() - should also be using objective credentials
instead, if init is in overlayfs or whatever, you probably wouldn't
want that to have an effect here

prepare_kernel_cred() - most callers pass NULL and don't hit this
codepath, and those that don't pass NULL use `current`, so there can
be no concurrent modification. Maybe this could be rewritten to
something like this:

if (daemon) {
  BUG_ON(daemon != current);
  old = get_current_cred();
} else {
 ...
}

or maybe it could just use the objective creds, it shouldn't matter
here, I think.

> To be honest I didn't know we have this helper.

Same here.

> Can't it race with revert_creds() in
> do_faccessat() ?

Yeah, I think you can probably trigger use-after-free reads with that.
I also remember seeing something fishy in Smack at some point that had
similar issues... smack_file_send_sigiotask() does
smk_of_task(smack_cred(tsk->cred)), which looks very wrong.