Re: [PATCH -tip 2/3] sched/wake_q: Relax to acquire semantics

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, Sep 22, 2015 at 04:33:07PM +0200, Martin Schwidefsky wrote:
> On Tue, 22 Sep 2015 21:29:14 +0800
> Boqun Feng <boqun.feng@gmail.com> wrote:
> 
> > On Tue, Sep 22, 2015 at 02:51:36PM +0200, Martin Schwidefsky wrote:
> > > On Tue, 22 Sep 2015 20:23:26 +0800
> > > Boqun Feng <boqun.feng@gmail.com> wrote:
> > > 
> > > > Hi Martin,
> > > > 
> > > > On Tue, Sep 22, 2015 at 12:27:35PM +0200, Martin Schwidefsky wrote:
> > > > > On Mon, 21 Sep 2015 11:22:52 +0200
> > > > > Martin Schwidefsky <schwidefsky@de.ibm.com> wrote:
> > > > > 
> > > > > > On Fri, 18 Sep 2015 14:41:20 -0700
> > > > > > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > > > > > 
> > > > > > > On Tue, Sep 15, 2015 at 10:09:41AM -0700, Paul E. McKenney wrote:
> > > > > > > > On Tue, Sep 15, 2015 at 06:30:28PM +0200, Peter Zijlstra wrote:
> > > > > > > > > On Tue, Sep 15, 2015 at 08:34:48AM -0700, Paul E. McKenney wrote:
> > > > > > > > > > On Tue, Sep 15, 2015 at 04:14:39PM +0200, Peter Zijlstra wrote:
> > > > > > > > > > > On Tue, Sep 15, 2015 at 07:09:22AM -0700, Paul E. McKenney wrote:
> > > > > > > > > > > > On Tue, Sep 15, 2015 at 02:48:00PM +0200, Peter Zijlstra wrote:
> > > > > > > > > > > > > On Tue, Sep 15, 2015 at 05:41:42AM -0700, Paul E. McKenney wrote:
> > > > > > > > > > > > > > > Never mind, the PPC people will implement this with lwsync and that is
> > > > > > > > > > > > > > > very much not transitive IIRC.
> > > > > > > > > > > > > > 
> > > > > > > > > > > > > > I am probably lost on context, but...
> > > > > > > > > > > > > > 
> > > > > > > > > > > > > > It turns out that lwsync is transitive in special cases.  One of them
> > > > > > > > > > > > > > is a series of release-acquire pairs, which can extend indefinitely.
> > > > > > > > > > > > > > 
> > > > > > > > > > > > > > Does that help in this case?
> > > > > > > > > > > > > 
> > > > > > > > > > > > > Probably not, but good to know. I still don't think we want to rely on
> > > > > > > > > > > > > ACQUIRE/RELEASE being transitive in general though.
> > > > > > > > > > > > 
> > > > > > > > > > > > OK, I will bite...  Why not?
> > > > > > > > > > > 
> > > > > > > > > > > It would mean us reviewing all archs (again) and documenting it I
> > > > > > > > > > > suppose. Which is of course entirely possible.
> > > > > > > > > > > 
> > > > > > > > > > > That said, I don't think the case at hand requires it, so lets postpone
> > > > > > > > > > > this for now ;-)
> > > > > > > > > > 
> > > > > > > > > > True enough, but in my experience smp_store_release() and
> > > > > > > > > > smp_load_acquire() are a -lot- easier to use than other barriers,
> > > > > > > > > > and transitivity will help promote their use.  So...
> > > > > > > > > > 
> > > > > > > > > > All the TSO architectures (x86, s390, SPARC, HPPA, ...) support transitive
> > > > > > > > > > smp_store_release()/smp_load_acquire() via their native ordering in
> > > > > > > > > > combination with barrier() macros.  x86 with CONFIG_X86_PPRO_FENCE=y,
> > > > > > > > > > which is not TSO, uses an mfence instruction.  Power supports this via
> > > > > > > > > > lwsync's partial cumulativity.  ARM64 supports it in SMP via the new ldar
> > > > > > > > > > and stlr instructions (in non-SMP, it uses barrier(), which suffices
> > > > > > > > > > in that case).  IA64 supports this via total ordering of all release
> > > > > > > > > > instructions in theory and by the actual full-barrier implementation
> > > > > > > > > > in practice (and the fact that gcc emits st.rel and ld.acq instructions
> > > > > > > > > > for volatile stores and loads).  All other architectures use smp_mb(),
> > > > > > > > > > which is transitive.
> > > > > > > > > > 
> > > > > > > > > > Did I miss anything?
> > > > > > > > > 
> > > > > > > > > I think that about covers it.. the only odd duckling might be s390 which
> > > > > > > > > is documented as TSO but recently grew smp_mb__{before,after}_atomic(),
> > > > > > > > > which seems to confuse matters.
> > > > > > > > 
> > > > > > > > Fair point, adding Martin and Heiko on CC for their thoughts.
> > > > > > 
> > > > > > Well we always had the full memory barrier for the various versions of
> > > > > > smp_mb__xxx, they just have moved around and renamed several times.
> > > > > > 
> > > > > > After discussing this with Heiko we came to the conclusion that we can use
> > > > > > a simple barrier() for smp_mb__before_atomic() and smp_mb__after_atomic().
> > > > > > 
> > > > > > > > It looks like this applies to recent mainframes that have new atomic
> > > > > > > > instructions, which, yes, might need something to make them work with
> > > > > > > > fully transitive smp_load_acquire() and smp_store_release().
> > > > > > > > 
> > > > > > > > Martin, Heiko, the question is whether or not the current s390
> > > > > > > > smp_store_release() and smp_load_acquire() can be transitive.
> > > > > > > > For example, if all the Xi variables below are initially zero,
> > > > > > > > is it possible for all the r0, r1, r2, ... rN variables to
> > > > > > > > have the value 1 at the end of the test.
> > > > > > > 
> > > > > > > Right...  This time actually adding Martin and Heiko on CC...
> > > > > > > 
> > > > > > > 							Thanx, Paul
> > > > > > > 
> > > > > > > > CPU 0
> > > > > > > > 	r0 = smp_load_acquire(&X0);
> > > > > > > > 	smp_store_release(&X1, 1);
> > > > > > > > 
> > > > > > > > CPU 1
> > > > > > > > 	r1 = smp_load_acquire(&X1);
> > > > > > > > 	smp_store_release(&X2, 1);
> > > > > > > > 
> > > > > > > > CPU 2
> > > > > > > > 	r2 = smp_load_acquire(&X2);
> > > > > > > > 	smp_store_release(&X3, 1);
> > > > > > > > 
> > > > > > > > ...
> > > > > > > > 
> > > > > > > > CPU N
> > > > > > > > 	rN = smp_load_acquire(&XN);
> > > > > > > > 	smp_store_release(&X0, 1);
> > > > > > > > 
> > > > > > > > If smp_store_release() and smp_load_acquire() are transitive, the
> > > > > > > > answer would be "no".
> > > > > > 
> > > > > > The answer is "no". Christian recently summarized what the principles of
> > > > > > operation has to say about the CPU read / write behavior. If you consider
> > > > > > the sequential order of instructions then
> > > > > > 
> > > > > > 1) reads are in order
> > > > > > 2) writes are in order
> > > > > > 3) reads can happen earlier
> > > > > > 4) writes can happen later
> > > > > 
> > > > > Correction. The principles of operation states this:
> > > > > 
> > > > > "A storage-operand store specified by one instruction appears to precede
> > > > > all storage-operand stores specified by conceptually subsequent instructions,
> > > > > but it does not necessarily precede storage-operand fetches specified by
> > > > > conceptually subsequent instructions. However, a storage-operand store
> > > > > appears to precede a conceptually subsequent storage-operand fetch from the
> > > > > same main-storage location."
> > > > > 
> > 
> > Confused...
> 
> Yeah, seems like I'm confused as well. This stuff always make my head hurt..
>  
> > IIUC, the previous paragraph actually means that a STORE-LOAD pair can be
> > reordered. But the below reasoning is saying that a LOAD-STORE pair can
> > be reordered. Do I miss something here?
> 
> True, the above paragraph allows a store to move past a load and not the other
> way around.
> 
> > > > > As observed by other CPUs a write to one memory location can "overtake" a
> > > > > read of another memory location if there is no explicit memory-barrier
> > > > > between the load and the store instruction.
> > > > > 
> > > > > In the above example X0, X1, ... XN are different memory locations, so
> > > > > architecturally the answer is "yes", all r0, r1, ... rN variables can have
> > > > > the value of 1 after the test. I doubt that any existing machine will
> > > > > show this behavior though.
> > > > > 
> > > > 
> > > > Just be curious, how about when N == 1? The test then becomes:
> > > > 
> > > > CPU 0
> > > > 	r0 = smp_load_acquire(&X0);
> > > > 	smp_store_release(&X1,1);
> > > > 
> > > > CPU 1
> > > > 	r1 = smp_load_acquire(&X1);
> > > > 	smp_store_release(&X0,1);
> > > > 
> > > > Is it possible that r0 == 1 and r1 == 1 at the end, due to the same
> > > > reason?
> > > 
> > > Yes, that is possible for the same reason. To change that we would have
> > > to replace the barrier() in smp_load_acquire/smp_store_release with
> > > smp_mb().
> > > 
> > 
> > I thought that s390 is TSO, so this is prohibitted. If that is possible,
> > I think, that means the current implementation of smp_load_acquire and
> > smp_store_release on s390 is incorrect...
> 
> Ok, further reading of chapter 5 of the principles revealed this:
> 
> "As observed by other CPUs and by channel programs, storage-operand fetches
> associated with one instruction execution appear to precede all storage
> operand references for conceptually subsequent instructions."
> 
> So no writes before reads. Correction to the correction: all r0, r1, ...rN
> equal to one can not happen after all. Got me worried there ;-)

Whew!!!

So s390's current smp_store_release() and smp_load_acquire() provide
ordering as needed, then, right?  For example, suppose that there
was a long chain of smp_load_acquire()/smp_store_release() pairs
involving many CPUs.  Would the all-ones case still be impossible?

							Thanx, Paul