tools/memory-model: Model smp_mb__after_unlock_lock()
authorAndrea Parri <andrea.parri@amarulasolutions.com>
Mon, 3 Dec 2018 23:04:49 +0000 (15:04 -0800)
committerIngo Molnar <mingo@kernel.org>
Mon, 21 Jan 2019 10:06:55 +0000 (11:06 +0100)
The kernel documents smp_mb__after_unlock_lock() the following way:

  "Place this after a lock-acquisition primitive to guarantee that
   an UNLOCK+LOCK pair acts as a full barrier.  This guarantee applies
   if the UNLOCK and LOCK are executed by the same CPU or if the
   UNLOCK and LOCK operate on the same lock variable."

Formalize in LKMM the above guarantee by defining (new) mb-links according
to the law:

  ([M] ; po ; [UL] ; (co | po) ; [LKW] ;
fencerel(After-unlock-lock) ; [M])

where the component ([UL] ; co ; [LKW]) identifies "UNLOCK+LOCK pairs on
the same lock variable" and the component ([UL] ; po ; [LKW]) identifies
"UNLOCK+LOCK pairs executed by the same CPU".

In particular, the LKMM forbids the following two behaviors (the second
litmus test below is based on:

  Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.html

c.f., Section "Tree RCU Grace Period Memory Ordering Building Blocks"):

C after-unlock-lock-same-cpu

(*
 * Result: Never
 *)

{}

P0(spinlock_t *s, spinlock_t *t, int *x, int *y)
{
int r0;

spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(t);
smp_mb__after_unlock_lock();
r0 = READ_ONCE(*y);
spin_unlock(t);
}

P1(int *x, int *y)
{
int r0;

WRITE_ONCE(*y, 1);
smp_mb();
r0 = READ_ONCE(*x);
}

exists (0:r0=0 /\ 1:r0=0)

C after-unlock-lock-same-lock-variable

(*
 * Result: Never
 *)

{}

P0(spinlock_t *s, int *x, int *y)
{
int r0;

spin_lock(s);
WRITE_ONCE(*x, 1);
r0 = READ_ONCE(*y);
spin_unlock(s);
}

P1(spinlock_t *s, int *y, int *z)
{
int r0;

spin_lock(s);
smp_mb__after_unlock_lock();
WRITE_ONCE(*y, 1);
r0 = READ_ONCE(*z);
spin_unlock(s);
}

P2(int *z, int *x)
{
int r0;

WRITE_ONCE(*z, 1);
smp_mb();
r0 = READ_ONCE(*x);
}

exists (0:r0=0 /\ 1:r0=0 /\ 2:r0=0)

Signed-off-by: Andrea Parri <andrea.parri@amarulasolutions.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Cc: Akira Yokosawa <akiyks@gmail.com>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Daniel Lustig <dlustig@nvidia.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Jade Alglave <j.alglave@ucl.ac.uk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Luc Maranget <luc.maranget@inria.fr>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will.deacon@arm.com>
Cc: linux-arch@vger.kernel.org
Cc: parri.andrea@gmail.com
Link: http://lkml.kernel.org/r/20181203230451.28921-1-paulmck@linux.ibm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
tools/memory-model/linux-kernel.bell
tools/memory-model/linux-kernel.cat
tools/memory-model/linux-kernel.def

index b84fb2f..7965133 100644 (file)
@@ -29,7 +29,8 @@ enum Barriers = 'wmb (*smp_wmb*) ||
                'sync-rcu (*synchronize_rcu*) ||
                'before-atomic (*smp_mb__before_atomic*) ||
                'after-atomic (*smp_mb__after_atomic*) ||
-               'after-spinlock (*smp_mb__after_spinlock*)
+               'after-spinlock (*smp_mb__after_spinlock*) ||
+               'after-unlock-lock (*smp_mb__after_unlock_lock*)
 instructions F[Barriers]
 
 (* Compute matching pairs of nested Rcu-lock and Rcu-unlock *)
index 882fc33..8f23c74 100644 (file)
@@ -30,7 +30,9 @@ 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] ; po? ; [LKW] ; fencerel(After-spinlock) ; [M])
+       ([M] ; po? ; [LKW] ; fencerel(After-spinlock) ; [M]) |
+       ([M] ; po ; [UL] ; (co | po) ; [LKW] ;
+               fencerel(After-unlock-lock) ; [M])
 let gp = po ; [Sync-rcu] ; po?
 
 let strong-fence = mb | gp
index 6fa3eb2..b27911c 100644 (file)
@@ -23,6 +23,7 @@ smp_wmb() { __fence{wmb}; }
 smp_mb__before_atomic() { __fence{before-atomic}; }
 smp_mb__after_atomic() { __fence{after-atomic}; }
 smp_mb__after_spinlock() { __fence{after-spinlock}; }
+smp_mb__after_unlock_lock() { __fence{after-unlock-lock}; }
 
 // Exchange
 xchg(X,V)  __xchg{mb}(X,V)