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

[Alan Stern]

On Tue, 17 Jul 2018, Linus Torvalds wrote:

> On Tue, Jul 17, 2018 at 11:31 AM Paul E. McKenney
> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > 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.
> 
> That's only true when looking at isync in isolation.
> 
> Read the part I quoted. The AIX documentation implies that the
> *sequence* of load-compare-conditional branch-isync is a memory
> barrier, even if isync on its own is now.

I'm not a huge expert on the PowerPC architecture, but I do have a
pretty good understanding of the widely accepted memory model published
by the Peter Sewell's group at Cambridge (PPCMEM).  According to that
model, load-compare-conditional branch-isync is _not_ a full memory
barrier.

> So I'm just saying that
> 
>  (a) isync-on-lock is supposed to be much cheaper than sync-on-lock
> 
>  (b) the AIX documentation at least implies that isync-on-lock (when
> used together the the whole locking sequence) is actually a memory
> barrier
> 
> Now, admittedly the powerpc barrier instructions are unfathomable
> crazy stuff, so who knows. But:
> 
>  (a) lwsync is a memory barrier for all the "easy" cases (ie
> load->store, load->load, and store->load).
> 
>  (b) lwsync is *not* a memory barrier for the store->load case.
> 
>  (c) isync *is* (when in that *sequence*) a memory barrier for a
> store->load case (and has to be: loads inside a spinlocked region MUST
> NOT be done earlier than stores outside of it!).

Why not?  Instructions are allowed to migrate _into_ critical sections,
just not _out_ of them.  So a store preceding the start of a spinlocked
region can migrate in and be executed after a load that is inside the
region.

Alan Stern

> So a unlock/lock sequence where the unlock is using lwsync, and the
> lock is using isync, should in fact be a full memory barrier (which is
> the semantics we're looking for).
> 
> So doing performance testing on sync/lwsync (for lock/unlock
> respectively) seems the wrong thing to do.  Please test the
> isync/lwsync case instead.
> 
> Hmm? What am I missing?