riscv/atomic: Strengthen implementations with fences
authorAndrea Parri <parri.andrea@gmail.com>
Fri, 9 Mar 2018 12:13:40 +0000 (13:13 +0100)
committerPalmer Dabbelt <palmer@sifive.com>
Tue, 3 Apr 2018 02:59:44 +0000 (19:59 -0700)
Atomics present the same issue with locking: release and acquire
variants need to be strengthened to meet the constraints defined
by the Linux-kernel memory consistency model [1].

Atomics present a further issue: implementations of atomics such
as atomic_cmpxchg() and atomic_add_unless() rely on LR/SC pairs,
which do not give full-ordering with .aqrl; for example, current
implementations allow the "lr-sc-aqrl-pair-vs-full-barrier" test
below to end up with the state indicated in the "exists" clause.

In order to "synchronize" LKMM and RISC-V's implementation, this
commit strengthens the implementations of the atomics operations
by replacing .rl and .aq with the use of ("lightweigth") fences,
and by replacing .aqrl LR/SC pairs in sequences such as:

  0:      lr.w.aqrl  %0, %addr
          bne        %0, %old, 1f
          ...
          sc.w.aqrl  %1, %new, %addr
          bnez       %1, 0b
  1:

with sequences of the form:

  0:      lr.w       %0, %addr
          bne        %0, %old, 1f
          ...
          sc.w.rl    %1, %new, %addr   /* SC-release   */
          bnez       %1, 0b
          fence      rw, rw            /* "full" fence */
  1:

following Daniel's suggestion.

These modifications were validated with simulation of the RISC-V
memory consistency model.

C lr-sc-aqrl-pair-vs-full-barrier

{}

P0(int *x, int *y, atomic_t *u)
{
int r0;
int r1;

WRITE_ONCE(*x, 1);
r0 = atomic_cmpxchg(u, 0, 1);
r1 = READ_ONCE(*y);
}

P1(int *x, int *y, atomic_t *v)
{
int r0;
int r1;

WRITE_ONCE(*y, 1);
r0 = atomic_cmpxchg(v, 0, 1);
r1 = READ_ONCE(*x);
}

exists (u=1 /\ v=1 /\ 0:r1=0 /\ 1:r1=0)

[1] https://marc.info/?l=linux-kernel&m=151930201102853&w=2
    https://groups.google.com/a/groups.riscv.org/forum/#!topic/isa-dev/hKywNHBkAXM
    https://marc.info/?l=linux-kernel&m=151633436614259&w=2

Suggested-by: Daniel Lustig <dlustig@nvidia.com>
Signed-off-by: Andrea Parri <parri.andrea@gmail.com>
Cc: Palmer Dabbelt <palmer@sifive.com>
Cc: Albert Ou <albert@sifive.com>
Cc: Daniel Lustig <dlustig@nvidia.com>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Jade Alglave <j.alglave@ucl.ac.uk>
Cc: Luc Maranget <luc.maranget@inria.fr>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Akira Yokosawa <akiyks@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: linux-riscv@lists.infradead.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Palmer Dabbelt <palmer@sifive.com>
arch/riscv/include/asm/atomic.h
arch/riscv/include/asm/cmpxchg.h

index e65d1cd89e28bb5ae52291e8fa2d13937f28e1a9..855115ace98c8cdccb7083ca4aacb3a2b20e3259 100644 (file)
 #include <asm/barrier.h>
 
 #define ATOMIC_INIT(i) { (i) }
+
+#define __atomic_op_acquire(op, args...)                               \
+({                                                                     \
+       typeof(op##_relaxed(args)) __ret  = op##_relaxed(args);         \
+       __asm__ __volatile__(RISCV_ACQUIRE_BARRIER "" ::: "memory");    \
+       __ret;                                                          \
+})
+
+#define __atomic_op_release(op, args...)                               \
+({                                                                     \
+       __asm__ __volatile__(RISCV_RELEASE_BARRIER "" ::: "memory");    \
+       op##_relaxed(args);                                             \
+})
+
 static __always_inline int atomic_read(const atomic_t *v)
 {
        return READ_ONCE(v->counter);
@@ -50,22 +64,23 @@ static __always_inline void atomic64_set(atomic64_t *v, long i)
  * have the AQ or RL bits set.  These don't return anything, so there's only
  * one version to worry about.
  */
-#define ATOMIC_OP(op, asm_op, I, asm_type, c_type, prefix)                             \
-static __always_inline void atomic##prefix##_##op(c_type i, atomic##prefix##_t *v)     \
-{                                                                                      \
-       __asm__ __volatile__ (                                                          \
-               "amo" #asm_op "." #asm_type " zero, %1, %0"                             \
-               : "+A" (v->counter)                                                     \
-               : "r" (I)                                                               \
-               : "memory");                                                            \
-}
+#define ATOMIC_OP(op, asm_op, I, asm_type, c_type, prefix)             \
+static __always_inline                                                 \
+void atomic##prefix##_##op(c_type i, atomic##prefix##_t *v)            \
+{                                                                      \
+       __asm__ __volatile__ (                                          \
+               "       amo" #asm_op "." #asm_type " zero, %1, %0"      \
+               : "+A" (v->counter)                                     \
+               : "r" (I)                                               \
+               : "memory");                                            \
+}                                                                      \
 
 #ifdef CONFIG_GENERIC_ATOMIC64
-#define ATOMIC_OPS(op, asm_op, I)                      \
+#define ATOMIC_OPS(op, asm_op, I)                                      \
         ATOMIC_OP (op, asm_op, I, w,  int,   )
 #else
-#define ATOMIC_OPS(op, asm_op, I)                      \
-        ATOMIC_OP (op, asm_op, I, w,  int,   ) \
+#define ATOMIC_OPS(op, asm_op, I)                                      \
+        ATOMIC_OP (op, asm_op, I, w,  int,   )                         \
         ATOMIC_OP (op, asm_op, I, d, long, 64)
 #endif
 
@@ -79,75 +94,115 @@ ATOMIC_OPS(xor, xor,  i)
 #undef ATOMIC_OPS
 
 /*
- * Atomic ops that have ordered, relaxed, acquire, and relese variants.
+ * Atomic ops that have ordered, relaxed, acquire, and release variants.
  * There's two flavors of these: the arithmatic ops have both fetch and return
  * versions, while the logical ops only have fetch versions.
  */
-#define ATOMIC_FETCH_OP(op, asm_op, I, asm_or, c_or, asm_type, c_type, prefix)                         \
-static __always_inline c_type atomic##prefix##_fetch_##op##c_or(c_type i, atomic##prefix##_t *v)       \
-{                                                                                                      \
-       register c_type ret;                                                                            \
-       __asm__ __volatile__ (                                                                          \
-               "amo" #asm_op "." #asm_type #asm_or " %1, %2, %0"                                       \
-               : "+A" (v->counter), "=r" (ret)                                                         \
-               : "r" (I)                                                                               \
-               : "memory");                                                                            \
-       return ret;                                                                                     \
+#define ATOMIC_FETCH_OP(op, asm_op, I, asm_type, c_type, prefix)       \
+static __always_inline                                                 \
+c_type atomic##prefix##_fetch_##op##_relaxed(c_type i,                 \
+                                            atomic##prefix##_t *v)     \
+{                                                                      \
+       register c_type ret;                                            \
+       __asm__ __volatile__ (                                          \
+               "       amo" #asm_op "." #asm_type " %1, %2, %0"        \
+               : "+A" (v->counter), "=r" (ret)                         \
+               : "r" (I)                                               \
+               : "memory");                                            \
+       return ret;                                                     \
+}                                                                      \
+static __always_inline                                                 \
+c_type atomic##prefix##_fetch_##op(c_type i, atomic##prefix##_t *v)    \
+{                                                                      \
+       register c_type ret;                                            \
+       __asm__ __volatile__ (                                          \
+               "       amo" #asm_op "." #asm_type ".aqrl  %1, %2, %0"  \
+               : "+A" (v->counter), "=r" (ret)                         \
+               : "r" (I)                                               \
+               : "memory");                                            \
+       return ret;                                                     \
 }
 
-#define ATOMIC_OP_RETURN(op, asm_op, c_op, I, asm_or, c_or, asm_type, c_type, prefix)                  \
-static __always_inline c_type atomic##prefix##_##op##_return##c_or(c_type i, atomic##prefix##_t *v)    \
-{                                                                                                      \
-        return atomic##prefix##_fetch_##op##c_or(i, v) c_op I;                                         \
+#define ATOMIC_OP_RETURN(op, asm_op, c_op, I, asm_type, c_type, prefix)        \
+static __always_inline                                                 \
+c_type atomic##prefix##_##op##_return_relaxed(c_type i,                        \
+                                             atomic##prefix##_t *v)    \
+{                                                                      \
+        return atomic##prefix##_fetch_##op##_relaxed(i, v) c_op I;     \
+}                                                                      \
+static __always_inline                                                 \
+c_type atomic##prefix##_##op##_return(c_type i, atomic##prefix##_t *v) \
+{                                                                      \
+        return atomic##prefix##_fetch_##op(i, v) c_op I;               \
 }
 
 #ifdef CONFIG_GENERIC_ATOMIC64
-#define ATOMIC_OPS(op, asm_op, c_op, I, asm_or, c_or)                          \
-        ATOMIC_FETCH_OP (op, asm_op,       I, asm_or, c_or, w,  int,   )       \
-        ATOMIC_OP_RETURN(op, asm_op, c_op, I, asm_or, c_or, w,  int,   )
+#define ATOMIC_OPS(op, asm_op, c_op, I)                                        \
+        ATOMIC_FETCH_OP( op, asm_op,       I, w,  int,   )             \
+        ATOMIC_OP_RETURN(op, asm_op, c_op, I, w,  int,   )
 #else
-#define ATOMIC_OPS(op, asm_op, c_op, I, asm_or, c_or)                          \
-        ATOMIC_FETCH_OP (op, asm_op,       I, asm_or, c_or, w,  int,   )       \
-        ATOMIC_OP_RETURN(op, asm_op, c_op, I, asm_or, c_or, w,  int,   )       \
-        ATOMIC_FETCH_OP (op, asm_op,       I, asm_or, c_or, d, long, 64)       \
-        ATOMIC_OP_RETURN(op, asm_op, c_op, I, asm_or, c_or, d, long, 64)
+#define ATOMIC_OPS(op, asm_op, c_op, I)                                        \
+        ATOMIC_FETCH_OP( op, asm_op,       I, w,  int,   )             \
+        ATOMIC_OP_RETURN(op, asm_op, c_op, I, w,  int,   )             \
+        ATOMIC_FETCH_OP( op, asm_op,       I, d, long, 64)             \
+        ATOMIC_OP_RETURN(op, asm_op, c_op, I, d, long, 64)
 #endif
 
-ATOMIC_OPS(add, add, +,  i,      , _relaxed)
-ATOMIC_OPS(add, add, +,  i, .aq  , _acquire)
-ATOMIC_OPS(add, add, +,  i, .rl  , _release)
-ATOMIC_OPS(add, add, +,  i, .aqrl,         )
+ATOMIC_OPS(add, add, +,  i)
+ATOMIC_OPS(sub, add, +, -i)
+
+#define atomic_add_return_relaxed      atomic_add_return_relaxed
+#define atomic_sub_return_relaxed      atomic_sub_return_relaxed
+#define atomic_add_return              atomic_add_return
+#define atomic_sub_return              atomic_sub_return
 
-ATOMIC_OPS(sub, add, +, -i,      , _relaxed)
-ATOMIC_OPS(sub, add, +, -i, .aq  , _acquire)
-ATOMIC_OPS(sub, add, +, -i, .rl  , _release)
-ATOMIC_OPS(sub, add, +, -i, .aqrl,         )
+#define atomic_fetch_add_relaxed       atomic_fetch_add_relaxed
+#define atomic_fetch_sub_relaxed       atomic_fetch_sub_relaxed
+#define atomic_fetch_add               atomic_fetch_add
+#define atomic_fetch_sub               atomic_fetch_sub
+
+#ifndef CONFIG_GENERIC_ATOMIC64
+#define atomic64_add_return_relaxed    atomic64_add_return_relaxed
+#define atomic64_sub_return_relaxed    atomic64_sub_return_relaxed
+#define atomic64_add_return            atomic64_add_return
+#define atomic64_sub_return            atomic64_sub_return
+
+#define atomic64_fetch_add_relaxed     atomic64_fetch_add_relaxed
+#define atomic64_fetch_sub_relaxed     atomic64_fetch_sub_relaxed
+#define atomic64_fetch_add             atomic64_fetch_add
+#define atomic64_fetch_sub             atomic64_fetch_sub
+#endif
 
 #undef ATOMIC_OPS
 
 #ifdef CONFIG_GENERIC_ATOMIC64
-#define ATOMIC_OPS(op, asm_op, I, asm_or, c_or)                                \
-        ATOMIC_FETCH_OP(op, asm_op, I, asm_or, c_or, w,  int,   )
+#define ATOMIC_OPS(op, asm_op, I)                                      \
+        ATOMIC_FETCH_OP(op, asm_op, I, w,  int,   )
 #else
-#define ATOMIC_OPS(op, asm_op, I, asm_or, c_or)                                \
-        ATOMIC_FETCH_OP(op, asm_op, I, asm_or, c_or, w,  int,   )      \
-        ATOMIC_FETCH_OP(op, asm_op, I, asm_or, c_or, d, long, 64)
+#define ATOMIC_OPS(op, asm_op, I)                                      \
+        ATOMIC_FETCH_OP(op, asm_op, I, w,  int,   )                    \
+        ATOMIC_FETCH_OP(op, asm_op, I, d, long, 64)
 #endif
 
-ATOMIC_OPS(and, and, i,      , _relaxed)
-ATOMIC_OPS(and, and, i, .aq  , _acquire)
-ATOMIC_OPS(and, and, i, .rl  , _release)
-ATOMIC_OPS(and, and, i, .aqrl,         )
+ATOMIC_OPS(and, and, i)
+ATOMIC_OPS( or,  or, i)
+ATOMIC_OPS(xor, xor, i)
 
-ATOMIC_OPS( or,  or, i,      , _relaxed)
-ATOMIC_OPS( or,  or, i, .aq  , _acquire)
-ATOMIC_OPS( or,  or, i, .rl  , _release)
-ATOMIC_OPS( or,  or, i, .aqrl,         )
+#define atomic_fetch_and_relaxed       atomic_fetch_and_relaxed
+#define atomic_fetch_or_relaxed                atomic_fetch_or_relaxed
+#define atomic_fetch_xor_relaxed       atomic_fetch_xor_relaxed
+#define atomic_fetch_and               atomic_fetch_and
+#define atomic_fetch_or                        atomic_fetch_or
+#define atomic_fetch_xor               atomic_fetch_xor
 
-ATOMIC_OPS(xor, xor, i,      , _relaxed)
-ATOMIC_OPS(xor, xor, i, .aq  , _acquire)
-ATOMIC_OPS(xor, xor, i, .rl  , _release)
-ATOMIC_OPS(xor, xor, i, .aqrl,         )
+#ifndef CONFIG_GENERIC_ATOMIC64
+#define atomic64_fetch_and_relaxed     atomic64_fetch_and_relaxed
+#define atomic64_fetch_or_relaxed      atomic64_fetch_or_relaxed
+#define atomic64_fetch_xor_relaxed     atomic64_fetch_xor_relaxed
+#define atomic64_fetch_and             atomic64_fetch_and
+#define atomic64_fetch_or              atomic64_fetch_or
+#define atomic64_fetch_xor             atomic64_fetch_xor
+#endif
 
 #undef ATOMIC_OPS
 
@@ -157,22 +212,24 @@ ATOMIC_OPS(xor, xor, i, .aqrl,         )
 /*
  * The extra atomic operations that are constructed from one of the core
  * AMO-based operations above (aside from sub, which is easier to fit above).
- * These are required to perform a barrier, but they're OK this way because
- * atomic_*_return is also required to perform a barrier.
+ * These are required to perform a full barrier, but they're OK this way
+ * because atomic_*_return is also required to perform a full barrier.
+ *
  */
-#define ATOMIC_OP(op, func_op, comp_op, I, c_type, prefix)                     \
-static __always_inline bool atomic##prefix##_##op(c_type i, atomic##prefix##_t *v) \
-{                                                                              \
-       return atomic##prefix##_##func_op##_return(i, v) comp_op I;             \
+#define ATOMIC_OP(op, func_op, comp_op, I, c_type, prefix)             \
+static __always_inline                                                 \
+bool atomic##prefix##_##op(c_type i, atomic##prefix##_t *v)            \
+{                                                                      \
+       return atomic##prefix##_##func_op##_return(i, v) comp_op I;     \
 }
 
 #ifdef CONFIG_GENERIC_ATOMIC64
-#define ATOMIC_OPS(op, func_op, comp_op, I)                    \
-        ATOMIC_OP (op, func_op, comp_op, I,  int,   )
+#define ATOMIC_OPS(op, func_op, comp_op, I)                            \
+        ATOMIC_OP(op, func_op, comp_op, I,  int,   )
 #else
-#define ATOMIC_OPS(op, func_op, comp_op, I)                    \
-        ATOMIC_OP (op, func_op, comp_op, I,  int,   )          \
-        ATOMIC_OP (op, func_op, comp_op, I, long, 64)
+#define ATOMIC_OPS(op, func_op, comp_op, I)                            \
+        ATOMIC_OP(op, func_op, comp_op, I,  int,   )                   \
+        ATOMIC_OP(op, func_op, comp_op, I, long, 64)
 #endif
 
 ATOMIC_OPS(add_and_test, add, ==, 0)
@@ -182,51 +239,87 @@ ATOMIC_OPS(add_negative, add,  <, 0)
 #undef ATOMIC_OP
 #undef ATOMIC_OPS
 
-#define ATOMIC_OP(op, func_op, I, c_type, prefix)                              \
-static __always_inline void atomic##prefix##_##op(atomic##prefix##_t *v)       \
-{                                                                              \
-       atomic##prefix##_##func_op(I, v);                                       \
+#define ATOMIC_OP(op, func_op, I, c_type, prefix)                      \
+static __always_inline                                                 \
+void atomic##prefix##_##op(atomic##prefix##_t *v)                      \
+{                                                                      \
+       atomic##prefix##_##func_op(I, v);                               \
 }
 
-#define ATOMIC_FETCH_OP(op, func_op, I, c_type, prefix)                                        \
-static __always_inline c_type atomic##prefix##_fetch_##op(atomic##prefix##_t *v)       \
-{                                                                                      \
-       return atomic##prefix##_fetch_##func_op(I, v);                                  \
+#define ATOMIC_FETCH_OP(op, func_op, I, c_type, prefix)                        \
+static __always_inline                                                 \
+c_type atomic##prefix##_fetch_##op##_relaxed(atomic##prefix##_t *v)    \
+{                                                                      \
+       return atomic##prefix##_fetch_##func_op##_relaxed(I, v);        \
+}                                                                      \
+static __always_inline                                                 \
+c_type atomic##prefix##_fetch_##op(atomic##prefix##_t *v)              \
+{                                                                      \
+       return atomic##prefix##_fetch_##func_op(I, v);                  \
 }
 
-#define ATOMIC_OP_RETURN(op, asm_op, c_op, I, c_type, prefix)                          \
-static __always_inline c_type atomic##prefix##_##op##_return(atomic##prefix##_t *v)    \
-{                                                                                      \
-        return atomic##prefix##_fetch_##op(v) c_op I;                                  \
+#define ATOMIC_OP_RETURN(op, asm_op, c_op, I, c_type, prefix)          \
+static __always_inline                                                 \
+c_type atomic##prefix##_##op##_return_relaxed(atomic##prefix##_t *v)   \
+{                                                                      \
+        return atomic##prefix##_fetch_##op##_relaxed(v) c_op I;                \
+}                                                                      \
+static __always_inline                                                 \
+c_type atomic##prefix##_##op##_return(atomic##prefix##_t *v)           \
+{                                                                      \
+        return atomic##prefix##_fetch_##op(v) c_op I;                  \
 }
 
 #ifdef CONFIG_GENERIC_ATOMIC64
-#define ATOMIC_OPS(op, asm_op, c_op, I)                                                \
-        ATOMIC_OP       (op, asm_op,       I,  int,   )                                \
-        ATOMIC_FETCH_OP (op, asm_op,       I,  int,   )                                \
+#define ATOMIC_OPS(op, asm_op, c_op, I)                                        \
+        ATOMIC_OP(       op, asm_op,       I,  int,   )                        \
+        ATOMIC_FETCH_OP( op, asm_op,       I,  int,   )                        \
         ATOMIC_OP_RETURN(op, asm_op, c_op, I,  int,   )
 #else
-#define ATOMIC_OPS(op, asm_op, c_op, I)                                                \
-        ATOMIC_OP       (op, asm_op,       I,  int,   )                                \
-        ATOMIC_FETCH_OP (op, asm_op,       I,  int,   )                                \
-        ATOMIC_OP_RETURN(op, asm_op, c_op, I,  int,   )                                \
-        ATOMIC_OP       (op, asm_op,       I, long, 64)                                \
-        ATOMIC_FETCH_OP (op, asm_op,       I, long, 64)                                \
+#define ATOMIC_OPS(op, asm_op, c_op, I)                                        \
+        ATOMIC_OP(       op, asm_op,       I,  int,   )                        \
+        ATOMIC_FETCH_OP( op, asm_op,       I,  int,   )                        \
+        ATOMIC_OP_RETURN(op, asm_op, c_op, I,  int,   )                        \
+        ATOMIC_OP(       op, asm_op,       I, long, 64)                        \
+        ATOMIC_FETCH_OP( op, asm_op,       I, long, 64)                        \
         ATOMIC_OP_RETURN(op, asm_op, c_op, I, long, 64)
 #endif
 
 ATOMIC_OPS(inc, add, +,  1)
 ATOMIC_OPS(dec, add, +, -1)
 
+#define atomic_inc_return_relaxed      atomic_inc_return_relaxed
+#define atomic_dec_return_relaxed      atomic_dec_return_relaxed
+#define atomic_inc_return              atomic_inc_return
+#define atomic_dec_return              atomic_dec_return
+
+#define atomic_fetch_inc_relaxed       atomic_fetch_inc_relaxed
+#define atomic_fetch_dec_relaxed       atomic_fetch_dec_relaxed
+#define atomic_fetch_inc               atomic_fetch_inc
+#define atomic_fetch_dec               atomic_fetch_dec
+
+#ifndef CONFIG_GENERIC_ATOMIC64
+#define atomic64_inc_return_relaxed    atomic64_inc_return_relaxed
+#define atomic64_dec_return_relaxed    atomic64_dec_return_relaxed
+#define atomic64_inc_return            atomic64_inc_return
+#define atomic64_dec_return            atomic64_dec_return
+
+#define atomic64_fetch_inc_relaxed     atomic64_fetch_inc_relaxed
+#define atomic64_fetch_dec_relaxed     atomic64_fetch_dec_relaxed
+#define atomic64_fetch_inc             atomic64_fetch_inc
+#define atomic64_fetch_dec             atomic64_fetch_dec
+#endif
+
 #undef ATOMIC_OPS
 #undef ATOMIC_OP
 #undef ATOMIC_FETCH_OP
 #undef ATOMIC_OP_RETURN
 
-#define ATOMIC_OP(op, func_op, comp_op, I, prefix)                             \
-static __always_inline bool atomic##prefix##_##op(atomic##prefix##_t *v)       \
-{                                                                              \
-       return atomic##prefix##_##func_op##_return(v) comp_op I;                \
+#define ATOMIC_OP(op, func_op, comp_op, I, prefix)                     \
+static __always_inline                                                 \
+bool atomic##prefix##_##op(atomic##prefix##_t *v)                      \
+{                                                                      \
+       return atomic##prefix##_##func_op##_return(v) comp_op I;        \
 }
 
 ATOMIC_OP(inc_and_test, inc, ==, 0,   )
@@ -238,19 +331,19 @@ ATOMIC_OP(dec_and_test, dec, ==, 0, 64)
 
 #undef ATOMIC_OP
 
-/* This is required to provide a barrier on success. */
+/* This is required to provide a full barrier on success. */
 static __always_inline int __atomic_add_unless(atomic_t *v, int a, int u)
 {
        int prev, rc;
 
        __asm__ __volatile__ (
-               "0:\n\t"
-               "lr.w.aqrl  %[p],  %[c]\n\t"
-               "beq        %[p],  %[u], 1f\n\t"
-               "add       %[rc],  %[p], %[a]\n\t"
-               "sc.w.aqrl %[rc], %[rc], %[c]\n\t"
-               "bnez      %[rc], 0b\n\t"
-               "1:"
+               "0:     lr.w     %[p],  %[c]\n"
+               "       beq      %[p],  %[u], 1f\n"
+               "       add      %[rc], %[p], %[a]\n"
+               "       sc.w.rl  %[rc], %[rc], %[c]\n"
+               "       bnez     %[rc], 0b\n"
+               "       fence    rw, rw\n"
+               "1:\n"
                : [p]"=&r" (prev), [rc]"=&r" (rc), [c]"+A" (v->counter)
                : [a]"r" (a), [u]"r" (u)
                : "memory");
@@ -263,13 +356,13 @@ static __always_inline long __atomic64_add_unless(atomic64_t *v, long a, long u)
        long prev, rc;
 
        __asm__ __volatile__ (
-               "0:\n\t"
-               "lr.d.aqrl  %[p],  %[c]\n\t"
-               "beq        %[p],  %[u], 1f\n\t"
-               "add       %[rc],  %[p], %[a]\n\t"
-               "sc.d.aqrl %[rc], %[rc], %[c]\n\t"
-               "bnez      %[rc], 0b\n\t"
-               "1:"
+               "0:     lr.d     %[p],  %[c]\n"
+               "       beq      %[p],  %[u], 1f\n"
+               "       add      %[rc], %[p], %[a]\n"
+               "       sc.d.rl  %[rc], %[rc], %[c]\n"
+               "       bnez     %[rc], 0b\n"
+               "       fence    rw, rw\n"
+               "1:\n"
                : [p]"=&r" (prev), [rc]"=&r" (rc), [c]"+A" (v->counter)
                : [a]"r" (a), [u]"r" (u)
                : "memory");
@@ -300,37 +393,63 @@ static __always_inline long atomic64_inc_not_zero(atomic64_t *v)
 
 /*
  * atomic_{cmp,}xchg is required to have exactly the same ordering semantics as
- * {cmp,}xchg and the operations that return, so they need a barrier.
- */
-/*
- * FIXME: atomic_cmpxchg_{acquire,release,relaxed} are all implemented by
- * assigning the same barrier to both the LR and SC operations, but that might
- * not make any sense.  We're waiting on a memory model specification to
- * determine exactly what the right thing to do is here.
+ * {cmp,}xchg and the operations that return, so they need a full barrier.
  */
-#define ATOMIC_OP(c_t, prefix, c_or, size, asm_or)                                             \
-static __always_inline c_t atomic##prefix##_cmpxchg##c_or(atomic##prefix##_t *v, c_t o, c_t n)         \
-{                                                                                              \
-       return __cmpxchg(&(v->counter), o, n, size, asm_or, asm_or);                            \
-}                                                                                              \
-static __always_inline c_t atomic##prefix##_xchg##c_or(atomic##prefix##_t *v, c_t n)           \
-{                                                                                              \
-       return __xchg(n, &(v->counter), size, asm_or);                                          \
+#define ATOMIC_OP(c_t, prefix, size)                                   \
+static __always_inline                                                 \
+c_t atomic##prefix##_xchg_relaxed(atomic##prefix##_t *v, c_t n)                \
+{                                                                      \
+       return __xchg_relaxed(&(v->counter), n, size);                  \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_xchg_acquire(atomic##prefix##_t *v, c_t n)                \
+{                                                                      \
+       return __xchg_acquire(&(v->counter), n, size);                  \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_xchg_release(atomic##prefix##_t *v, c_t n)                \
+{                                                                      \
+       return __xchg_release(&(v->counter), n, size);                  \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_xchg(atomic##prefix##_t *v, c_t n)                        \
+{                                                                      \
+       return __xchg(&(v->counter), n, size);                          \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_cmpxchg_relaxed(atomic##prefix##_t *v,            \
+                                    c_t o, c_t n)                      \
+{                                                                      \
+       return __cmpxchg_relaxed(&(v->counter), o, n, size);            \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_cmpxchg_acquire(atomic##prefix##_t *v,            \
+                                    c_t o, c_t n)                      \
+{                                                                      \
+       return __cmpxchg_acquire(&(v->counter), o, n, size);            \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_cmpxchg_release(atomic##prefix##_t *v,            \
+                                    c_t o, c_t n)                      \
+{                                                                      \
+       return __cmpxchg_release(&(v->counter), o, n, size);            \
+}                                                                      \
+static __always_inline                                                 \
+c_t atomic##prefix##_cmpxchg(atomic##prefix##_t *v, c_t o, c_t n)      \
+{                                                                      \
+       return __cmpxchg(&(v->counter), o, n, size);                    \
 }
 
 #ifdef CONFIG_GENERIC_ATOMIC64
-#define ATOMIC_OPS(c_or, asm_or)                       \
-       ATOMIC_OP( int,   , c_or, 4, asm_or)
+#define ATOMIC_OPS()                                                   \
+       ATOMIC_OP( int,   , 4)
 #else
-#define ATOMIC_OPS(c_or, asm_or)                       \
-       ATOMIC_OP( int,   , c_or, 4, asm_or)            \
-       ATOMIC_OP(long, 64, c_or, 8, asm_or)
+#define ATOMIC_OPS()                                                   \
+       ATOMIC_OP( int,   , 4)                                          \
+       ATOMIC_OP(long, 64, 8)
 #endif
 
-ATOMIC_OPS(        , .aqrl)
-ATOMIC_OPS(_acquire,   .aq)
-ATOMIC_OPS(_release,   .rl)
-ATOMIC_OPS(_relaxed,      )
+ATOMIC_OPS()
 
 #undef ATOMIC_OPS
 #undef ATOMIC_OP
@@ -340,13 +459,13 @@ static __always_inline int atomic_sub_if_positive(atomic_t *v, int offset)
        int prev, rc;
 
        __asm__ __volatile__ (
-               "0:\n\t"
-               "lr.w.aqrl  %[p],  %[c]\n\t"
-               "sub       %[rc],  %[p], %[o]\n\t"
-               "bltz      %[rc],    1f\n\t"
-               "sc.w.aqrl %[rc], %[rc], %[c]\n\t"
-               "bnez      %[rc],    0b\n\t"
-               "1:"
+               "0:     lr.w     %[p],  %[c]\n"
+               "       sub      %[rc], %[p], %[o]\n"
+               "       bltz     %[rc], 1f\n"
+               "       sc.w.rl  %[rc], %[rc], %[c]\n"
+               "       bnez     %[rc], 0b\n"
+               "       fence    rw, rw\n"
+               "1:\n"
                : [p]"=&r" (prev), [rc]"=&r" (rc), [c]"+A" (v->counter)
                : [o]"r" (offset)
                : "memory");
@@ -361,13 +480,13 @@ static __always_inline long atomic64_sub_if_positive(atomic64_t *v, int offset)
        long prev, rc;
 
        __asm__ __volatile__ (
-               "0:\n\t"
-               "lr.d.aqrl  %[p],  %[c]\n\t"
-               "sub       %[rc],  %[p], %[o]\n\t"
-               "bltz      %[rc],    1f\n\t"
-               "sc.d.aqrl %[rc], %[rc], %[c]\n\t"
-               "bnez      %[rc],    0b\n\t"
-               "1:"
+               "0:     lr.d     %[p],  %[c]\n"
+               "       sub      %[rc], %[p], %[o]\n"
+               "       bltz     %[rc], 1f\n"
+               "       sc.d.rl  %[rc], %[rc], %[c]\n"
+               "       bnez     %[rc], 0b\n"
+               "       fence    rw, rw\n"
+               "1:\n"
                : [p]"=&r" (prev), [rc]"=&r" (rc), [c]"+A" (v->counter)
                : [o]"r" (offset)
                : "memory");
index db249dbc7b9760c319e96e441d2f2ebc18ada3d4..c12833f7b6bd1505f407e59c94bac0804fa7eefd 100644 (file)
 #include <linux/bug.h>
 
 #include <asm/barrier.h>
+#include <asm/fence.h>
 
-#define __xchg(new, ptr, size, asm_or)                         \
-({                                                             \
-       __typeof__(ptr) __ptr = (ptr);                          \
-       __typeof__(new) __new = (new);                          \
-       __typeof__(*(ptr)) __ret;                               \
-       switch (size) {                                         \
-       case 4:                                                 \
-               __asm__ __volatile__ (                          \
-                       "amoswap.w" #asm_or " %0, %2, %1"       \
-                       : "=r" (__ret), "+A" (*__ptr)           \
-                       : "r" (__new)                           \
-                       : "memory");                            \
-               break;                                          \
-       case 8:                                                 \
-               __asm__ __volatile__ (                          \
-                       "amoswap.d" #asm_or " %0, %2, %1"       \
-                       : "=r" (__ret), "+A" (*__ptr)           \
-                       : "r" (__new)                           \
-                       : "memory");                            \
-               break;                                          \
-       default:                                                \
-               BUILD_BUG();                                    \
-       }                                                       \
-       __ret;                                                  \
-})
-
-#define xchg(ptr, x)    (__xchg((x), (ptr), sizeof(*(ptr)), .aqrl))
-
-#define xchg32(ptr, x)                         \
-({                                             \
-       BUILD_BUG_ON(sizeof(*(ptr)) != 4);      \
-       xchg((ptr), (x));                       \
-})
-
-#define xchg64(ptr, x)                         \
-({                                             \
-       BUILD_BUG_ON(sizeof(*(ptr)) != 8);      \
-       xchg((ptr), (x));                       \
+#define __xchg_relaxed(ptr, new, size)                                 \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(new) __new = (new);                                  \
+       __typeof__(*(ptr)) __ret;                                       \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "       amoswap.w %0, %2, %1\n"                 \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "       amoswap.d %0, %2, %1\n"                 \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
+
+#define xchg_relaxed(ptr, x)                                           \
+({                                                                     \
+       __typeof__(*(ptr)) _x_ = (x);                                   \
+       (__typeof__(*(ptr))) __xchg_relaxed((ptr),                      \
+                                           _x_, sizeof(*(ptr)));       \
+})
+
+#define __xchg_acquire(ptr, new, size)                                 \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(new) __new = (new);                                  \
+       __typeof__(*(ptr)) __ret;                                       \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "       amoswap.w %0, %2, %1\n"                 \
+                       RISCV_ACQUIRE_BARRIER                           \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "       amoswap.d %0, %2, %1\n"                 \
+                       RISCV_ACQUIRE_BARRIER                           \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
+
+#define xchg_acquire(ptr, x)                                           \
+({                                                                     \
+       __typeof__(*(ptr)) _x_ = (x);                                   \
+       (__typeof__(*(ptr))) __xchg_acquire((ptr),                      \
+                                           _x_, sizeof(*(ptr)));       \
+})
+
+#define __xchg_release(ptr, new, size)                                 \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(new) __new = (new);                                  \
+       __typeof__(*(ptr)) __ret;                                       \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       RISCV_RELEASE_BARRIER                           \
+                       "       amoswap.w %0, %2, %1\n"                 \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       RISCV_RELEASE_BARRIER                           \
+                       "       amoswap.d %0, %2, %1\n"                 \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
+
+#define xchg_release(ptr, x)                                           \
+({                                                                     \
+       __typeof__(*(ptr)) _x_ = (x);                                   \
+       (__typeof__(*(ptr))) __xchg_release((ptr),                      \
+                                           _x_, sizeof(*(ptr)));       \
+})
+
+#define __xchg(ptr, new, size)                                         \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(new) __new = (new);                                  \
+       __typeof__(*(ptr)) __ret;                                       \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "       amoswap.w.aqrl %0, %2, %1\n"            \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "       amoswap.d.aqrl %0, %2, %1\n"            \
+                       : "=r" (__ret), "+A" (*__ptr)                   \
+                       : "r" (__new)                                   \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
+
+#define xchg(ptr, x)                                                   \
+({                                                                     \
+       __typeof__(*(ptr)) _x_ = (x);                                   \
+       (__typeof__(*(ptr))) __xchg((ptr), _x_, sizeof(*(ptr)));        \
+})
+
+#define xchg32(ptr, x)                                                 \
+({                                                                     \
+       BUILD_BUG_ON(sizeof(*(ptr)) != 4);                              \
+       xchg((ptr), (x));                                               \
+})
+
+#define xchg64(ptr, x)                                                 \
+({                                                                     \
+       BUILD_BUG_ON(sizeof(*(ptr)) != 8);                              \
+       xchg((ptr), (x));                                               \
 })
 
 /*
  * store NEW in MEM.  Return the initial value in MEM.  Success is
  * indicated by comparing RETURN with OLD.
  */
-#define __cmpxchg(ptr, old, new, size, lrb, scb)                       \
+#define __cmpxchg_relaxed(ptr, old, new, size)                         \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(*(ptr)) __old = (old);                               \
+       __typeof__(*(ptr)) __new = (new);                               \
+       __typeof__(*(ptr)) __ret;                                       \
+       register unsigned int __rc;                                     \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "0:     lr.w %0, %2\n"                          \
+                       "       bne  %0, %z3, 1f\n"                     \
+                       "       sc.w %1, %z4, %2\n"                     \
+                       "       bnez %1, 0b\n"                          \
+                       "1:\n"                                          \
+                       : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
+                       : "rJ" (__old), "rJ" (__new)                    \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "0:     lr.d %0, %2\n"                          \
+                       "       bne %0, %z3, 1f\n"                      \
+                       "       sc.d %1, %z4, %2\n"                     \
+                       "       bnez %1, 0b\n"                          \
+                       "1:\n"                                          \
+                       : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
+                       : "rJ" (__old), "rJ" (__new)                    \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
+
+#define cmpxchg_relaxed(ptr, o, n)                                     \
+({                                                                     \
+       __typeof__(*(ptr)) _o_ = (o);                                   \
+       __typeof__(*(ptr)) _n_ = (n);                                   \
+       (__typeof__(*(ptr))) __cmpxchg_relaxed((ptr),                   \
+                                       _o_, _n_, sizeof(*(ptr)));      \
+})
+
+#define __cmpxchg_acquire(ptr, old, new, size)                         \
 ({                                                                     \
        __typeof__(ptr) __ptr = (ptr);                                  \
        __typeof__(*(ptr)) __old = (old);                               \
        switch (size) {                                                 \
        case 4:                                                         \
                __asm__ __volatile__ (                                  \
-               "0:"                                                    \
-                       "lr.w" #scb " %0, %2\n"                         \
-                       "bne         %0, %z3, 1f\n"                     \
-                       "sc.w" #lrb " %1, %z4, %2\n"                    \
-                       "bnez        %1, 0b\n"                          \
-               "1:"                                                    \
+                       "0:     lr.w %0, %2\n"                          \
+                       "       bne  %0, %z3, 1f\n"                     \
+                       "       sc.w %1, %z4, %2\n"                     \
+                       "       bnez %1, 0b\n"                          \
+                       RISCV_ACQUIRE_BARRIER                           \
+                       "1:\n"                                          \
                        : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
                        : "rJ" (__old), "rJ" (__new)                    \
                        : "memory");                                    \
                break;                                                  \
        case 8:                                                         \
                __asm__ __volatile__ (                                  \
-               "0:"                                                    \
-                       "lr.d" #scb " %0, %2\n"                         \
-                       "bne         %0, %z3, 1f\n"                     \
-                       "sc.d" #lrb " %1, %z4, %2\n"                    \
-                       "bnez        %1, 0b\n"                          \
-               "1:"                                                    \
+                       "0:     lr.d %0, %2\n"                          \
+                       "       bne %0, %z3, 1f\n"                      \
+                       "       sc.d %1, %z4, %2\n"                     \
+                       "       bnez %1, 0b\n"                          \
+                       RISCV_ACQUIRE_BARRIER                           \
+                       "1:\n"                                          \
                        : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
                        : "rJ" (__old), "rJ" (__new)                    \
                        : "memory");                                    \
        __ret;                                                          \
 })
 
-#define cmpxchg(ptr, o, n) \
-       (__cmpxchg((ptr), (o), (n), sizeof(*(ptr)), .aqrl, .aqrl))
+#define cmpxchg_acquire(ptr, o, n)                                     \
+({                                                                     \
+       __typeof__(*(ptr)) _o_ = (o);                                   \
+       __typeof__(*(ptr)) _n_ = (n);                                   \
+       (__typeof__(*(ptr))) __cmpxchg_acquire((ptr),                   \
+                                       _o_, _n_, sizeof(*(ptr)));      \
+})
 
-#define cmpxchg_local(ptr, o, n) \
-       (__cmpxchg((ptr), (o), (n), sizeof(*(ptr)), , ))
+#define __cmpxchg_release(ptr, old, new, size)                         \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(*(ptr)) __old = (old);                               \
+       __typeof__(*(ptr)) __new = (new);                               \
+       __typeof__(*(ptr)) __ret;                                       \
+       register unsigned int __rc;                                     \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       RISCV_RELEASE_BARRIER                           \
+                       "0:     lr.w %0, %2\n"                          \
+                       "       bne  %0, %z3, 1f\n"                     \
+                       "       sc.w %1, %z4, %2\n"                     \
+                       "       bnez %1, 0b\n"                          \
+                       "1:\n"                                          \
+                       : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
+                       : "rJ" (__old), "rJ" (__new)                    \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       RISCV_RELEASE_BARRIER                           \
+                       "0:     lr.d %0, %2\n"                          \
+                       "       bne %0, %z3, 1f\n"                      \
+                       "       sc.d %1, %z4, %2\n"                     \
+                       "       bnez %1, 0b\n"                          \
+                       "1:\n"                                          \
+                       : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
+                       : "rJ" (__old), "rJ" (__new)                    \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
+
+#define cmpxchg_release(ptr, o, n)                                     \
+({                                                                     \
+       __typeof__(*(ptr)) _o_ = (o);                                   \
+       __typeof__(*(ptr)) _n_ = (n);                                   \
+       (__typeof__(*(ptr))) __cmpxchg_release((ptr),                   \
+                                       _o_, _n_, sizeof(*(ptr)));      \
+})
+
+#define __cmpxchg(ptr, old, new, size)                                 \
+({                                                                     \
+       __typeof__(ptr) __ptr = (ptr);                                  \
+       __typeof__(*(ptr)) __old = (old);                               \
+       __typeof__(*(ptr)) __new = (new);                               \
+       __typeof__(*(ptr)) __ret;                                       \
+       register unsigned int __rc;                                     \
+       switch (size) {                                                 \
+       case 4:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "0:     lr.w %0, %2\n"                          \
+                       "       bne  %0, %z3, 1f\n"                     \
+                       "       sc.w.rl %1, %z4, %2\n"                  \
+                       "       bnez %1, 0b\n"                          \
+                       "       fence rw, rw\n"                         \
+                       "1:\n"                                          \
+                       : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
+                       : "rJ" (__old), "rJ" (__new)                    \
+                       : "memory");                                    \
+               break;                                                  \
+       case 8:                                                         \
+               __asm__ __volatile__ (                                  \
+                       "0:     lr.d %0, %2\n"                          \
+                       "       bne %0, %z3, 1f\n"                      \
+                       "       sc.d.rl %1, %z4, %2\n"                  \
+                       "       bnez %1, 0b\n"                          \
+                       "       fence rw, rw\n"                         \
+                       "1:\n"                                          \
+                       : "=&r" (__ret), "=&r" (__rc), "+A" (*__ptr)    \
+                       : "rJ" (__old), "rJ" (__new)                    \
+                       : "memory");                                    \
+               break;                                                  \
+       default:                                                        \
+               BUILD_BUG();                                            \
+       }                                                               \
+       __ret;                                                          \
+})
 
-#define cmpxchg32(ptr, o, n)                   \
-({                                             \
-       BUILD_BUG_ON(sizeof(*(ptr)) != 4);      \
-       cmpxchg((ptr), (o), (n));               \
+#define cmpxchg(ptr, o, n)                                             \
+({                                                                     \
+       __typeof__(*(ptr)) _o_ = (o);                                   \
+       __typeof__(*(ptr)) _n_ = (n);                                   \
+       (__typeof__(*(ptr))) __cmpxchg((ptr),                           \
+                                      _o_, _n_, sizeof(*(ptr)));       \
 })
 
-#define cmpxchg32_local(ptr, o, n)             \
-({                                             \
-       BUILD_BUG_ON(sizeof(*(ptr)) != 4);      \
-       cmpxchg_local((ptr), (o), (n));         \
+#define cmpxchg_local(ptr, o, n)                                       \
+       (__cmpxchg_relaxed((ptr), (o), (n), sizeof(*(ptr))))
+
+#define cmpxchg32(ptr, o, n)                                           \
+({                                                                     \
+       BUILD_BUG_ON(sizeof(*(ptr)) != 4);                              \
+       cmpxchg((ptr), (o), (n));                                       \
 })
 
-#define cmpxchg64(ptr, o, n)                   \
-({                                             \
-       BUILD_BUG_ON(sizeof(*(ptr)) != 8);      \
-       cmpxchg((ptr), (o), (n));               \
+#define cmpxchg32_local(ptr, o, n)                                     \
+({                                                                     \
+       BUILD_BUG_ON(sizeof(*(ptr)) != 4);                              \
+       cmpxchg_relaxed((ptr), (o), (n))                                \
 })
 
-#define cmpxchg64_local(ptr, o, n)             \
-({                                             \
-       BUILD_BUG_ON(sizeof(*(ptr)) != 8);      \
-       cmpxchg_local((ptr), (o), (n));         \
+#define cmpxchg64(ptr, o, n)                                           \
+({                                                                     \
+       BUILD_BUG_ON(sizeof(*(ptr)) != 8);                              \
+       cmpxchg((ptr), (o), (n));                                       \
+})
+
+#define cmpxchg64_local(ptr, o, n)                                     \
+({                                                                     \
+       BUILD_BUG_ON(sizeof(*(ptr)) != 8);                              \
+       cmpxchg_relaxed((ptr), (o), (n));                               \
 })
 
 #endif /* _ASM_RISCV_CMPXCHG_H */