[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20180220093346.GA5505@andrea>
Date: Tue, 20 Feb 2018 10:33:47 +0100
From: Andrea Parri <parri.andrea@...il.com>
To: Alan Stern <stern@...land.harvard.edu>
Cc: "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
Akira Yokosawa <akiyks@...il.com>,
Kernel development list <linux-kernel@...r.kernel.org>,
mingo@...nel.org, Will Deacon <will.deacon@....com>,
peterz@...radead.org, boqun.feng@...il.com, npiggin@...il.com,
dhowells@...hat.com, Jade Alglave <j.alglave@....ac.uk>,
Luc Maranget <luc.maranget@...ia.fr>,
Patrick Bellasi <patrick.bellasi@....com>
Subject: Re: [PATCH] tools/memory-model: remove rb-dep,
smp_read_barrier_depends, and lockless_dereference
On Mon, Feb 19, 2018 at 12:14:45PM -0500, Alan Stern wrote:
> On Sat, 17 Feb 2018, Andrea Parri wrote:
>
> > > Akira's observation about READ_ONCE extends to all (annotated) loads. In
> > > fact, it also applies to loads corresponding to unsuccessful RMW operations;
> > > consider, for example, the following variation of MP+onceassign+derefonce:
> > >
> > > C T
> > >
> > > {
> > > y=z;
> > > z=0;
> > > }
> > >
> > > P0(int *x, int **y)
> > > {
> > > WRITE_ONCE(*x, 1);
> > > smp_store_release(y, x);
> > > }
> > >
> > > P1(int **y, int *z)
> > > {
> > > int *r0;
> > > int r1;
> > >
> > > r0 = cmpxchg_relaxed(y, z, z);
> > > r1 = READ_ONCE(*r0);
> > > }
> > >
> > > exists (1:r0=x /\ 1:r1=0)
> > >
> > > The final state is allowed w/o the patch, and forbidden w/ the patch.
> > >
> > > This also reminds me of
> > >
> > > 5a8897cc7631fa544d079c443800f4420d1b173f
> > > ("locking/atomics/alpha: Add smp_read_barrier_depends() to _release()/_relaxed() atomics")
> > >
> > > (that we probably want to mention in the commit message).
> >
> > Please also notice that 5a8897cc7631f only touched alpha's atomic.h:
> > I see no corresponding commit/change on {,cmp}xchg.h (where the "mb"
> > is currently conditionally executed).
>
> This leaves us with a question: Do we want to change the kernel by
> adding memory barriers after unsuccessful RMW operations on Alpha, or
> do we want to change the model by excluding such operations from
> address dependencies?
I'd like to continue to treat R[once] and R*[once] equally if possible.
Given the (unconditional) smp_read_barrier_depends in READ_ONCE and in
atomics, it seems reasonable to have it unconditionally in cmpxchg.
As with the following patch?
Andrea
---
diff --git a/arch/alpha/include/asm/xchg.h b/arch/alpha/include/asm/xchg.h
index 68dfb3cb71454..e2660866ce972 100644
--- a/arch/alpha/include/asm/xchg.h
+++ b/arch/alpha/include/asm/xchg.h
@@ -128,10 +128,9 @@ ____xchg(, volatile void *ptr, unsigned long x, int size)
* store NEW in MEM. Return the initial value in MEM. Success is
* indicated by comparing RETURN with OLD.
*
- * The memory barrier should be placed in SMP only when we actually
- * make the change. If we don't change anything (so if the returned
- * prev is equal to old) then we aren't acquiring anything new and
- * we don't need any memory barrier as far I can tell.
+ * The memory barrier is placed in SMP unconditionally, in order to
+ * guarantee that dependency ordering is preserved when a dependency
+ * is headed by an unsuccessful operation.
*/
static inline unsigned long
@@ -150,8 +149,8 @@ ____cmpxchg(_u8, volatile char *m, unsigned char old, unsigned char new)
" or %1,%2,%2\n"
" stq_c %2,0(%4)\n"
" beq %2,3f\n"
- __ASM__MB
"2:\n"
+ __ASM__MB
".subsection 2\n"
"3: br 1b\n"
".previous"
@@ -177,8 +176,8 @@ ____cmpxchg(_u16, volatile short *m, unsigned short old, unsigned short new)
" or %1,%2,%2\n"
" stq_c %2,0(%4)\n"
" beq %2,3f\n"
- __ASM__MB
"2:\n"
+ __ASM__MB
".subsection 2\n"
"3: br 1b\n"
".previous"
@@ -200,8 +199,8 @@ ____cmpxchg(_u32, volatile int *m, int old, int new)
" mov %4,%1\n"
" stl_c %1,%2\n"
" beq %1,3f\n"
- __ASM__MB
"2:\n"
+ __ASM__MB
".subsection 2\n"
"3: br 1b\n"
".previous"
@@ -223,8 +222,8 @@ ____cmpxchg(_u64, volatile long *m, unsigned long old, unsigned long new)
" mov %4,%1\n"
" stq_c %1,%2\n"
" beq %1,3f\n"
- __ASM__MB
"2:\n"
+ __ASM__MB
".subsection 2\n"
"3: br 1b\n"
".previous"
>
> Note that operations like atomic_add_unless() already include memory
> barriers.
>
> Alan
>
Powered by blists - more mailing lists