Re: Adding plain accesses and detecting data races in the LKMM

[Paul E. McKenney]

On Fri, Apr 19, 2019 at 02:53:02AM +0200, Andrea Parri wrote:
> > Are you saying that on x86, atomic_inc() acts as a full memory barrier 
> > but not as a compiler barrier, and vice versa for 
> > smp_mb__after_atomic()?  Or that neither atomic_inc() nor 
> > smp_mb__after_atomic() implements a full memory barrier?
> 
> I'd say the former; AFAICT, these boil down to:
> 
>   https://elixir.bootlin.com/linux/v5.1-rc5/source/arch/x86/include/asm/atomic.h#L95
>   https://elixir.bootlin.com/linux/v5.1-rc5/source/arch/x86/include/asm/barrier.h#L84

OK, how about the following?

							Thanx, Paul

------------------------------------------------------------------------

commit 19d166dadc4e1bba4b248fb46d32ca4f2d10896b
Author: Paul E. McKenney <paulmck@xxxxxxxxxxxxx>
Date:   Fri Apr 19 05:20:30 2019 -0700

    tools/memory-model: Make smp_mb__{before,after}_atomic() match x86
    
    Read-modify-write atomic operations that do not return values need not
    provide any ordering guarantees, and this means that both the compiler
    and the CPU are free to reorder accesses across things like atomic_inc()
    and atomic_dec().  The stronger systems such as x86 allow the compiler
    to do the reordering, but prevent the CPU from so doing, and these
    systems implement smp_mb__{before,after}_atomic() as compiler barriers.
    The weaker systems such as Power allow both the compiler and the CPU
    to reorder accesses across things like atomic_inc() and atomic_dec(),
    and implement smp_mb__{before,after}_atomic() as full memory barriers.
    
    This means that smp_mb__before_atomic() only orders the atomic operation
    itself with accesses preceding the smp_mb__before_atomic(), and does
    not necessarily provide any ordering whatsoever against accesses
    folowing the atomic operation.  Similarly, smp_mb__after_atomic()
    only orders the atomic operation itself with accesses following the
    smp_mb__after_atomic(), and does not necessarily provide any ordering
    whatsoever against accesses preceding the atomic operation.  Full ordering
    therefore requires both an smp_mb__before_atomic() before the atomic
    operation and an smp_mb__after_atomic() after the atomic operation.
    
    Therefore, linux-kernel.cat's current model of Before-atomic
    and After-atomic is too strong, as it guarantees ordering of
    accesses on the other side of the atomic operation from the
    smp_mb__{before,after}_atomic().  This commit therefore weakens
    the guarantee to match the semantics called out above.
    
    Reported-by: Andrea Parri <andrea.parri@xxxxxxxxxxxxxxxxxxxx>
    Suggested-by: Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>
    Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxx>

diff --git a/Documentation/memory-barriers.txt b/Documentation/memory-barriers.txt
index 169d938c0b53..e5b97c3e8e39 100644
--- a/Documentation/memory-barriers.txt
+++ b/Documentation/memory-barriers.txt
@@ -1888,7 +1888,37 @@ There are some more advanced barrier functions:
 	atomic_dec(&obj->ref_count);
 
      This makes sure that the death mark on the object is perceived to be set
-     *before* the reference counter is decremented.
+     *before* the reference counter is decremented.  However, please note
+     that smp_mb__before_atomic()'s ordering guarantee does not necessarily
+     extend beyond the atomic operation.  For example:
+
+	obj->dead = 1;
+	smp_mb__before_atomic();
+	atomic_dec(&obj->ref_count);
+	r1 = a;
+
+     Here the store to obj->dead is not guaranteed to be ordered with
+     with the load from a.  This reordering can happen on x86 as follows:
+     (1) The compiler can reorder the load from a to precede the
+     atomic_dec(), (2) Because x86 smp_mb__before_atomic() is only a
+     compiler barrier, the CPU can reorder the preceding store to
+     obj->dead with the later load from a.
+
+     This could be avoided by using READ_ONCE(), which would prevent the
+     compiler from reordering due to both atomic_dec() and READ_ONCE()
+     being volatile accesses, and is usually preferable for loads from
+     shared variables.  However, weakly ordered CPUs would still be
+     free to reorder the atomic_dec() with the load from a, so a more
+     readable option is to also use smp_mb__after_atomic() as follows:
+
+	WRITE_ONCE(obj->dead, 1);
+	smp_mb__before_atomic();
+	atomic_dec(&obj->ref_count);
+	smp_mb__after_atomic();
+	r1 = READ_ONCE(a);
+
+     This orders all three accesses against each other, and also makes
+     the intent quite clear.
 
      See Documentation/atomic_{t,bitops}.txt for more information.
 
diff --git a/tools/memory-model/linux-kernel.cat b/tools/memory-model/linux-kernel.cat
index 8dcb37835b61..b6866f93abb8 100644
--- a/tools/memory-model/linux-kernel.cat
+++ b/tools/memory-model/linux-kernel.cat
@@ -28,8 +28,8 @@ include "lock.cat"
 let rmb = [R \ Noreturn] ; fencerel(Rmb) ; [R \ Noreturn]
 let wmb = [W] ; fencerel(Wmb) ; [W]
 let mb = ([M] ; fencerel(Mb) ; [M]) |
-	([M] ; fencerel(Before-atomic) ; [RMW] ; po? ; [M]) |
-	([M] ; po? ; [RMW] ; fencerel(After-atomic) ; [M]) |
+	([M] ; fencerel(Before-atomic) ; [RMW]) |
+	([RMW] ; fencerel(After-atomic) ; [M]) |
 	([M] ; po? ; [LKW] ; fencerel(After-spinlock) ; [M]) |
 	([M] ; po ; [UL] ; (co | po) ; [LKW] ;
 		fencerel(After-unlock-lock) ; [M])