On Mon, Nov 20, 2017 at 08:28:06PM +0100, Peter Zijlstra wrote:

On Mon, Nov 20, 2017 at 06:05:55PM +0000, Will Deacon wrote:

quoted

This is a thorny issue, but RCU (specifically rcu_dereference but probably
also some READ_ONCEs) relies on being able to utilise syntactic dependency
chains to order local accesses to shared variables.

Well, we used to have READ_ONCE() and smp_read_barrier_depends(), but
we recently munged them together, in the process getting rid of
lockless_dereference().

So for sure, READ_ONCE() must be able to do the address dependency
thing, otherwise tons of code comes apart.

I do hope that they track the volatile accesses produced by READ_ONCE().
Otherwise, it would not be good to apply this to anything touching MMIO.

							Thanx, Paul