On Mon, Jul 20, 2020 at 06:48:50PM +0200, pet...@infradead.org wrote:
> On Mon, Jul 20, 2020 at 09:04:34AM -0700, Paul E. McKenney wrote:
> > 2. If we were to say "unlock" instead of "release", consistency
> > would demand that we also say "lock" instead of "acquire".
> > But "lock" is subtlely different than "acquire", and there is
> > a history of people requesting further divergence.
>
> This, acquire/release are RCpc, while (with the exception of Power)
> LOCK/UNLOCK are RCsc.
>
> ( Or did we settle on RCtso for our release/acquire order? I have vague
> memories of a long-ish thread, but seem to have forgotten the outcome,
> if any. )
>
> Lots of subtlety and head-aches right about there. Anyway, it would be
> awesome if we can get Power into the RCsc locking camp :-)
I will let you take that one up with the Power folks.
But I should give an example of a current difference between lock and
acquire, and just to spread the blame, I will pick on an architecture
other than Power. ;-)
With lock acquisition, the following orders the access to X and Z:
WRITE_ONCE(X, 1);
spin_lock(&my_lock);
smp_mb__after_lock();
r1 = READ_ONCE(Z);
But if we replace the lock acquisition with a load acquire, there are
no ordering guarantees for the accesses to X and Z:
WRITE_ONCE(X, 1);
r2 = smp_load_acquire(&Y);
smp_mb__after_lock(); // Yeah, there is no lock. ;-)
r3 = READ_ONCE(Z);
There -is- ordering between the accesses to Y and Z, but not to X and Z.
This is not a theoretical issue. The x86 platform really can reorder
the access to X to follow that of both Y and Z.
So the memory-model divergence between lock acquisition and acquire
loads is very real in the here and now.
Thanx, Paul
