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

[Oleg Nesterov <oleg@redhat.com>]

On 04/21, Linus Torvalds wrote:
>
> On Tue, 21 Apr 2009, Oleg Nesterov wrote:
> >
> > > (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.
>
> .. but so does taking a spinlock. So it shouldn't matter.
>
> We could play games with that (the same way I think we have some games for
> large-system irq latency with '__raw_spin_lock_flags()' on ia64), but that
> makes sense only when you have lots of CPU's and expect irq latency to
> suffer.
>
> And it doesn't tend to make sense for preemption latency, because if you
> have so many CPU's that you have lots of spinning on locks, you would
> normally not really care deeply about preemption (sure, in theory it's a
> real-time thing, in practice I doubt you'll find anybody who cares).

OK, I agree, it doesn't really matter from latency/etc pov.

But still I can't understand why it is better to take fs->lock under
RCU lock. I mean, "fs->lock is the innermost lock" should not apply
to rcu_read_lock(). Because the latter is a bit special, no?

Oleg.