Re: [PATCH v2] tools/memory-model: Add extra ordering for locks and remove it for ordinary release/acquire

From: Peter Zijlstra
Date: Tue Jul 17 2018 - 14:43:23 EST


On Tue, Jul 17, 2018 at 11:33:41AM -0700, Paul E. McKenney wrote:
> On Tue, Jul 17, 2018 at 09:19:15AM -0700, Linus Torvalds wrote:
> >
> > In particular, I find:
> >
> > "isync is not a memory barrier instruction, but the
> > load-compare-conditional branch-isync sequence can provide this
> > ordering property"
> >
> > so why are you doing "sync/lwsync", when it sounds like "isync/lwsync"
> > (for lock/unlock) is the right thing and would already give memory
> > barrier semantics?
>
> The PowerPC guys will correct me if I miss something here...
>
> The isync provides ordering roughly similar to lwsync, but nowhere near
> as strong as sync, and it is sync that would be needed to cause lock
> acquisition to provide full ordering. The reason for using lwsync instead
> of isync is that the former proved to be faster on recent hardware.
> The reason that the kernel still has the ability to instead generate
> isync instructions is that some older PowerPC hardware does not provide
> the lwsync instruction. If the hardware does support lwsync, the isync
> instructions are overwritten with lwsync at boot time.

Isn't ISYNC the instruction-sync pipeline flush instruction? That is
used as an smp_rmb() here to, together with the control dependency from the
LL/SC, to form a LOAD->{LOAD,STORE} (aka LOAD-ACQUIRE) ordering?

Where LWSYNC provides a TSO like ordering and SYNC provides a full
transitive barrier aka. smp_mb() (althgouh I think it is strictly
stronger than smp_mb() since it also implies completion, which smp_mb()
does not).

And since both LL/SC-CTRL + ISYNC / LWSYNC are strictly CPU local, they
cannot be used to create RCsc ordering.