Messages in this thread | | | Date | Tue, 29 Apr 2014 08:11:00 -0700 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH v2] rwsem: Support optimistic spinning |
| |
On Mon, Apr 28, 2014 at 05:50:49PM -0700, Tim Chen wrote: > On Mon, 2014-04-28 at 16:10 -0700, Paul E. McKenney wrote: > > > > +#ifdef CONFIG_SMP > > > +static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem) > > > +{ > > > + int retval; > > > + struct task_struct *owner; > > > + > > > + rcu_read_lock(); > > > + owner = ACCESS_ONCE(sem->owner); > > > > OK, I'll bite... > > > > Why ACCESS_ONCE() instead of rcu_dereference()? > > We're using it as a speculative check on the sem->owner to see > if the owner is running on the cpu. The rcu_read_lock > is used for ensuring that the owner->on_cpu memory is > still valid.
OK, so if we read complete garbage, all that happens is that we lose a bit of performance? If so, I am OK with it as long as there is a comment (which Davidlohr suggested later in this thread).
Thanx, Paul
> > (My first question was "where is the update side", but this is covered > > by task_struct allocation and deallocation.) > > Tim >
| |