Re: Q: check_unsafe_exec() races (Was: [PATCH 2/4] fix setuid sometimes doesn't)

[Oleg Nesterov <oleg@redhat.com>]

On 04/19, Hugh Dickins wrote:
>
> On Mon, 6 Apr 2009, Oleg Nesterov wrote:
> >
> > Sorry for delay!
>
> Please don't suppose that you can ever beat me at the slowness game!

I am still trying to compete...

> > check_unsafe_exec() doesn't need ->siglock, we can iterate over sub-threads
> > under rcu_read_lock(). Note that with RCU or ->siglock we can set the "wrong"
> > LSM_UNSAFE_SHARE if we race with copy_process(CLONE_THREAD | CLONE_FS), but
> > as it was already discussed we don't care. This means it is OK to miss the
> > freshly cloned thread which has already passed copy_fs().
>
> Yes, I agree.
> And preferable not to have IRQs disabled over that next_thread() loop.

Yes sure, we don't need local_irq_disable(), only rcu_read_lock().

> > T1 does clone(CLONE_FS /* without CLONE_THREAD */).
> >
> > T1 continues without LSM_UNSAFE_SHARE while ->fs is shared with another
> > process.
>
> If I follow you correctly, you meant to say T2 not T1 in the last step.

Yes,

> Yes, I think your clear_in_exec change is a necessary one,
> and your rcu_read_lock well worth while.

OK, I'll send 2 simple patches, the first one kills lock_task_sighand(),
another adds clear_in_exec.

But,

> One tiny change (aside from extending to compat_do_execve):
> Al originally had check_unsafe_exec()'s write_lock(&p->fs->lock)
> after the lock_task_sighand(p, &flags), but was forced to invert
> that by the IRQ issue lockdep flagged.  I think we'd all prefer
> to think of fs->lock as an innermost lock, and would like it
> now to go after your rcu_read_lock().

Since we are not going to disable IRQs, perhaps the above does
not matter? It is always safe to take rcu_read_lock(), no matter
which locks we already hold.

> (You do rcu_read_unlock() earlier, but that's okay.)

Yes, but unless we have a "strong" reason, it is better to take
fs->lock first. rcu_read_lock() is free, but disables preemption.

Oleg.