lkml.org 
[lkml]   [2018]   [Dec]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] Linux: Implement membarrier function
On Mon, Dec 10, 2018 at 10:42:25PM -0800, David Goldblatt wrote:
> Hi Paul, thank you for thinking about all this.
>
> I think the modelling you suggest captures most of the algorithms I
> would want to write. I think it's slightly too weak, though, to
> implement the model suggested in P1202R0[1], which permits the SC
> outcome to be recovered in C-Goldblat-memb-2[2] by inserting a second
> smp_memb() after the first, which is a rather nice property (and I
> believe is supported by the underlying implementation options). I
> afraid though that I'm not familiar enough with the Linux herd
> definitions to suggest a tweak (or know how easy a tweak might be).

Actually, there has been an offlist discussion on exactly this.

What is the general rule? Is it that a given cycle have at least as
many heavy barriers as it does light ones? Either way, why?

Gah! I updated the tests to add the second "t", apologies!!!

Thanx, Paul

> - David
>
> [1] Which I think may be strengthened a little bit more even in R1.
> [2] As a nit, my name has two "t"'s in it, although I'd throw into the
> ring "memb-pairwise", "memb-nontransitive", and "memb-sequenced" if
> these get non-placeholder names.
>
> On Thu, Dec 6, 2018 at 1:54 PM Paul E. McKenney <paulmck@linux.ibm.com> wrote:
> >
> > Hello, David,
> >
> > I took a crack at extending LKMM to accommodate what I think would
> > support what you have in your paper. Please see the very end of this
> > email for a patch against the "dev" branch of my -rcu tree.
> >
> > This gives the expected result for the following three litmus tests,
> > but is probably deficient or otherwise misguided in other ways. I have
> > added the LKMM maintainers on CC for their amusement. ;-)
> >
> > Thoughts?
> >
> > Thanx, Paul
> >
> > ------------------------------------------------------------------------
> >
> > C C-Goldblat-memb-1
> > {
> > }
> >
> > P0(int *x0, int *x1)
> > {
> > WRITE_ONCE(*x0, 1);
> > r1 = READ_ONCE(*x1);
> > }
> >
> >
> > P1(int *x0, int *x1)
> > {
> > WRITE_ONCE(*x1, 1);
> > smp_memb();
> > r2 = READ_ONCE(*x0);
> > }
> >
> > exists (0:r1=0 /\ 1:r2=0)
> >
> > ------------------------------------------------------------------------
> >
> > C C-Goldblat-memb-2
> > {
> > }
> >
> > P0(int *x0, int *x1)
> > {
> > WRITE_ONCE(*x0, 1);
> > r1 = READ_ONCE(*x1);
> > }
> >
> >
> > P1(int *x1, int *x2)
> > {
> > WRITE_ONCE(*x1, 1);
> > smp_memb();
> > r1 = READ_ONCE(*x2);
> > }
> >
> > P2(int *x2, int *x0)
> > {
> > WRITE_ONCE(*x2, 1);
> > r1 = READ_ONCE(*x0);
> > }
> >
> > exists (0:r1=0 /\ 1:r1=0 /\ 2:r1=0)
> >
> > ------------------------------------------------------------------------
> >
> > C C-Goldblat-memb-3
> > {
> > }
> >
> > P0(int *x0, int *x1)
> > {
> > WRITE_ONCE(*x0, 1);
> > r1 = READ_ONCE(*x1);
> > }
> >
> >
> > P1(int *x1, int *x2)
> > {
> > WRITE_ONCE(*x1, 1);
> > smp_memb();
> > r1 = READ_ONCE(*x2);
> > }
> >
> > P2(int *x2, int *x3)
> > {
> > WRITE_ONCE(*x2, 1);
> > r1 = READ_ONCE(*x3);
> > }
> >
> > P3(int *x3, int *x0)
> > {
> > WRITE_ONCE(*x3, 1);
> > smp_memb();
> > r1 = READ_ONCE(*x0);
> > }
> >
> > exists (0:r1=0 /\ 1:r1=0 /\ 2:r1=0 /\ 3:r1=0)
> >
> > ------------------------------------------------------------------------
> >
> > On Thu, Nov 29, 2018 at 11:02:17AM -0800, David Goldblatt wrote:
> > > One note with the suggested patch is that
> > > `atomic_thread_fence(memory_order_acq_rel)` should probably be
> > > `atomic_thread_fence (memory_order_seq_cst)` (otherwise the call would
> > > be a no-op on, say, x86, which it very much isn't).
> > >
> > > The non-transitivity thing makes the resulting description arguably
> > > incorrect, but this is informal enough that it might not be a big deal
> > > to add something after "For these threads, the membarrier function
> > > call turns an existing compiler barrier (see above) executed by these
> > > threads into full memory barriers" that clarifies it. E.g. you could
> > > make it into "turns an existing compiler barrier [...] into full
> > > memory barriers, with respect to the calling thread".
> > >
> > > Since this is targeting the description of the OS call (and doesn't
> > > have to concern itself with also being implementable by other
> > > asymmetric techniques or degrading to architectural barriers), I think
> > > that the description in "approach 2" in P1202 would also make sense
> > > for a formal description of the syscall. (Of course, without the
> > > kernel itself committing to a rigorous semantics, anything specified
> > > on top of it will be on slightly shaky ground).
> > >
> > > - David
> > >
> > > On Thu, Nov 29, 2018 at 7:04 AM Paul E. McKenney <paulmck@linux.ibm.com> wrote:
> > > >
> > > > On Thu, Nov 29, 2018 at 09:44:22AM -0500, Mathieu Desnoyers wrote:
> > > > > ----- On Nov 29, 2018, at 8:50 AM, Florian Weimer fweimer@redhat.com wrote:
> > > > >
> > > > > > * Torvald Riegel:
> > > > > >
> > > > > >> On Wed, 2018-11-28 at 16:05 +0100, Florian Weimer wrote:
> > > > > >>> This is essentially a repost of last year's patch, rebased to the glibc
> > > > > >>> 2.29 symbol version and reflecting the introduction of
> > > > > >>> MEMBARRIER_CMD_GLOBAL.
> > > > > >>>
> > > > > >>> I'm not including any changes to manual/ here because the set of
> > > > > >>> supported operations is evolving rapidly, we could not get consensus for
> > > > > >>> the language I proposed the last time, and I do not want to contribute
> > > > > >>> to the manual for the time being.
> > > > > >>
> > > > > >> Fair enough. Nonetheless, can you summarize how far you're along with
> > > > > >> properly defining the semantics (eg, based on the C/C++ memory model)?
> > > > > >
> > > > > > I wrote down what you could, but no one liked it.
> > > > > >
> > > > > > <https://sourceware.org/ml/libc-alpha/2017-12/msg00796.html>
> > > > > >
> > > > > > I expect that a formalization would interact in non-trivial ways with
> > > > > > any potential formalization of usable relaxed memory order semantics,
> > > > > > and I'm not sure if anyone knows how to do the latter today.
> > > > >
> > > > > Adding Paul E. McKenney in CC.
> > > >
> > > > There is some prototype C++ memory model wording from David Goldblatt (CCed)
> > > > here (search for "Standarese"):
> > > >
> > > > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1202r0.pdf
> > > >
> > > > David's key insight is that (in Linuxese) light fences cannot pair with
> > > > each other.
> >
> > ------------------------------------------------------------------------
> >
> > commit 17e3b6b60e57d1cb791f68a1a6a36e942cb2baad
> > Author: Paul E. McKenney <paulmck@linux.ibm.com>
> > Date: Thu Dec 6 13:40:40 2018 -0800
> >
> > EXP tools/memory-model: Add semantics for sys_membarrier()
> >
> > This prototype commit extends LKMM to accommodate sys_membarrier(),
> > which is a asymmetric barrier with a limited ability to insert full
> > ordering into tasks that provide only compiler ordering. This commit
> > currently uses the "po" relation for this purpose, but something more
> > sophisticated will be required when plain accesses are added, which
> > the compiler can reorder.
> >
> > For more detail, please see David Goldblatt's C++ working paper:
> > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1202r0.pdf
> >
> > Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
> >
> > diff --git a/tools/memory-model/linux-kernel.bell b/tools/memory-model/linux-kernel.bell
> > index 9c42cd9ddcb4..4ef41453f569 100644
> > --- a/tools/memory-model/linux-kernel.bell
> > +++ b/tools/memory-model/linux-kernel.bell
> > @@ -24,6 +24,7 @@ instructions RMW[{'once,'acquire,'release}]
> > enum Barriers = 'wmb (*smp_wmb*) ||
> > 'rmb (*smp_rmb*) ||
> > 'mb (*smp_mb*) ||
> > + 'memb (*sys_membarrier*) ||
> > 'rcu-lock (*rcu_read_lock*) ||
> > 'rcu-unlock (*rcu_read_unlock*) ||
> > 'sync-rcu (*synchronize_rcu*) ||
> > diff --git a/tools/memory-model/linux-kernel.cat b/tools/memory-model/linux-kernel.cat
> > index 8dcb37835b61..837c3ee20bea 100644
> > --- a/tools/memory-model/linux-kernel.cat
> > +++ b/tools/memory-model/linux-kernel.cat
> > @@ -33,9 +33,10 @@ let mb = ([M] ; fencerel(Mb) ; [M]) |
> > ([M] ; po? ; [LKW] ; fencerel(After-spinlock) ; [M]) |
> > ([M] ; po ; [UL] ; (co | po) ; [LKW] ;
> > fencerel(After-unlock-lock) ; [M])
> > +let memb = [M] ; fencerel(Memb) ; [M]
> > let gp = po ; [Sync-rcu | Sync-srcu] ; po?
> >
> > -let strong-fence = mb | gp
> > +let strong-fence = mb | gp | memb
> >
> > (* Release Acquire *)
> > let acq-po = [Acquire] ; po ; [M]
> > @@ -86,6 +87,13 @@ acyclic hb as happens-before
> > let pb = prop ; strong-fence ; hb*
> > acyclic pb as propagation
> >
> > +(********************)
> > +(* sys_membarrier() *)
> > +(********************)
> > +
> > +let memb-step = ( prop ; po ; prop )? ; memb
> > +acyclic memb-step as memb-before
> > +
> > (*******)
> > (* RCU *)
> > (*******)
> > diff --git a/tools/memory-model/linux-kernel.def b/tools/memory-model/linux-kernel.def
> > index 1d6a120cde14..9ff0691c5f2c 100644
> > --- a/tools/memory-model/linux-kernel.def
> > +++ b/tools/memory-model/linux-kernel.def
> > @@ -17,6 +17,7 @@ rcu_dereference(X) __load{once}(X)
> > smp_store_mb(X,V) { __store{once}(X,V); __fence{mb}; }
> >
> > // Fences
> > +smp_memb() { __fence{memb}; }
> > smp_mb() { __fence{mb}; }
> > smp_rmb() { __fence{rmb}; }
> > smp_wmb() { __fence{wmb}; }
> >
>

\
 
 \ /
  Last update: 2018-12-11 15:50    [W:0.319 / U:0.008 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site