On Mon, Apr 29, 2019 at 06:06:04PM +0200, Oleg Nesterov wrote:
> On 04/28, Paul E. McKenney wrote:
> >
> > And it still looks good after review, so I have pushed it.
>
> Thanks!
>
> > I did add
> > READ_ONCE() and WRITE_ONCE() to unprotected uses of ->gp_state, but
> > please let me know if I messed anything up.
>
> Well, at least WRITE_ONCE()'s look certainly unneeded to me, gp_state
> is protected by rss_lock.
>
> WARN_ON_ONCE(gp_state) can read gp_state lockless, but even in this case
> I do not understand what READ_ONCE() tries to prevent...
>
> Nevermind, this won't hurt and as I already said I don't understand the
> _ONCE() magic anyway ;)
If I understand correctly, rcu_sync_is_idle() can be inline and returns
->gp_state. Without the READ_ONCE(), the compiler might fuse reads from
consecutive calls to rcu_sync_is_idle() or (under register pressure)
re-read from it, getting inconsistent results. For example, this:
tmp = rcu_sync_is_idle(rsp);
do_something(tmp);
do_something_else(tmp);
Might become this:
do_something(rcu_sync_is_idle(rsp));
do_something_else(rcu_sync_is_idle(rsp));
This might actually be harmless given current calls, but it would be at
best an accident waiting to happen.
Or am I missing something here?
Thanx, Paul
