Use atomic_fence_fence_rel() to ensure ordering in the

seq_write_begin(), instead of the load_rmb/rbm_load functions.  The
update does not need to be atomic due to the write lock owned.

Similarly, in seq_write_end(), update of *seqp needs not be atomic.
Only store must be atomic with release.

For seq_read(), the natural operation is the load acquire of the
sequence value, express this directly with atomic_load_acq_int()
instead of using custom partial fence implementation
atomic_load_rmb_int().

In seq_consistent, use atomic_thread_fence_acq() which provides the
desired semantic of ordering reads before fence before the re-reading
of *seqp, instead of custom atomic_rmb_load_int().

Reviewed by:	alc, bde
Sponsored by:	The FreeBSD Foundation
MFC after:	3 weeks

1 files changed, 6 insertions, 33 deletions

diff --git a/sys/sys/seq.h b/sys/sys/seq.h
index 21067cb..c5dcaeb 100644
--- a/sys/sys/seq.h
+++ b/sys/sys/seq.h
@@ -69,35 +69,6 @@ typedef uint32_t seq_t;
 
 #include <machine/cpu.h>
 
-/*
- * Stuff below is going away when we gain suitable memory barriers.
- *
- * atomic_load_acq_int at least on amd64 provides a full memory barrier,
- * in a way which affects performance.
- *
- * Hack below covers all architectures and avoids most of the penalty at least
- * on amd64 but still has unnecessary cost.
- */
-static __inline int
-atomic_load_rmb_int(volatile const u_int *p)
-{
-	volatile u_int v;
-
-	v = *p;
-	atomic_load_acq_int(&v);
-	return (v);
-}
-
-static __inline int
-atomic_rmb_load_int(volatile const u_int *p)
-{
-	volatile u_int v = 0;
-
-	atomic_load_acq_int(&v);
-	v = *p;
-	return (v);
-}
-
 static __inline bool
 seq_in_modify(seq_t seqp)
 {
@@ -110,14 +81,15 @@ seq_write_begin(seq_t *seqp)
 {
 
 	MPASS(!seq_in_modify(*seqp));
-	atomic_add_acq_int(seqp, 1);
+	*seqp += 1;
+	atomic_thread_fence_rel();
 }
 
 static __inline void
 seq_write_end(seq_t *seqp)
 {
 
-	atomic_add_rel_int(seqp, 1);
+	atomic_store_rel_int(seqp, *seqp + 1);
 	MPASS(!seq_in_modify(*seqp));
 }
 
@@ -127,7 +99,7 @@ seq_read(const seq_t *seqp)
 	seq_t ret;
 
 	for (;;) {
-		ret = atomic_load_rmb_int(seqp);
+		ret = atomic_load_acq_int(__DECONST(seq_t *, seqp));
 		if (seq_in_modify(ret)) {
 			cpu_spinwait();
 			continue;
@@ -142,7 +114,8 @@ static __inline seq_t
 seq_consistent(const seq_t *seqp, seq_t oldseq)
 {
 
-	return (atomic_rmb_load_int(seqp) == oldseq);
+	atomic_thread_fence_acq();
+	return (*seqp == oldseq);
 }
 
 static __inline seq_t