Re: [patch 24/52] fs: dcache reduce d_parent locking

[Paul E. McKenney]

On Fri, Jun 25, 2010 at 02:05:24AM +1000, Nick Piggin wrote:
> On Thu, Jun 24, 2010 at 08:32:18AM -0700, Paul E. McKenney wrote:
> > On Fri, Jun 25, 2010 at 01:07:06AM +1000, Nick Piggin wrote:
> > > On Thu, Jun 24, 2010 at 10:44:22AM +0200, Peter Zijlstra wrote:
> > > > On Thu, 2010-06-24 at 13:02 +1000, npiggin@xxxxxxx wrote:
> > > > > Use RCU property of dcache to simplify locking in some places where we
> > > > > take d_parent and d_lock.
> > > > > 
> > > > > Comment: don't need rcu_deref because we take the spinlock and recheck it.
> > > > 
> > > > But does the LOCK barrier imply a DATA DEPENDENCY barrier? (It does on
> > > > x86, and the compiler barrier implied by spin_lock() suffices to replace
> > > > ACCESS_ONCE()).
> > > 
> > > Well the dependency we care about is from loading the parent pointer
> > > to acquiring its spinlock. But we can't possibly have stale data given
> > > to the spin lock operation itself because it is a RMW.
> > 
> > As long as you check for the structure being valid after acquiring the
> > lock, I agree.  Otherwise, I would be concerned about the following
> > sequence of events:
> > 
> > 1.	CPU 0 picks up a pointer to a given data element.
> > 
> > 2.	CPU 1 removes this element from the list, drops any locks that
> > 	it might have, and starts waiting for a grace period to
> > 	elapse.
> > 
> > 3.	CPU 0 acquires the lock, does some operation that would
> > 	be appropriate had the element not been removed, then
> > 	releases the lock.
> > 
> > 4.	After the grace period, CPU 1 frees the element, negating
> > 	CPU 0's hard work.
> > 
> > The usual approach is to have a "deleted" flag or some such in the
> > element that CPU 0 would set when removing the element and that CPU 1
> > would check after acquiring the lock.  Which you might well already
> > be doing!  ;-)
> 
> Thanks, yep it's done under RCU, and after taking the lock it rechecks
> to see that it is still reachable by the same pointer (and if not,
> unlocks and retries) so it should be fine.

Very good!!!  ;-)

							Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/