* Change the scope of the ASSERT_ATOMIC_LOAD() from a generic check to

a pointer-fetching specific operation check. Consequently, rename the
  operation ASSERT_ATOMIC_LOAD_PTR().
* Fix the implementation of ASSERT_ATOMIC_LOAD_PTR() by checking
  directly alignment on the word boundry, for all the given specific
  architectures. That's a bit too strict for some common case, but it
  assures safety.
* Add a comment explaining the scope of the macro
* Add a new stub in the lockmgr specific implementation

Tested by: marcel (initial version), marius
Reviewed by: rwatson, jhb (comment specific review)
Approved by: re (kib)

sys/kern/kern_lock.c

@@ -334,6 +334,9 @@ lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
 	int iflags;
 
 	MPASS((flags & ~LK_INIT_MASK) == 0);
+	ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
+            ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
+            &lk->lk_lock));
 
 	iflags = LO_SLEEPABLE | LO_UPGRADABLE;
 	if (flags & LK_CANRECURSE)

sys/kern/kern_mutex.c

@@ -783,8 +783,9 @@ mtx_init(struct mtx *m, const char *name, const char *type, int opts)
 
 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
 		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
-	ASSERT_ATOMIC_LOAD(m->mtx_lock, ("%s: mtx_lock not aligned for %s: %p",
-	    __func__, name, &m->mtx_lock));
+	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
+	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
+	    &m->mtx_lock));
 
 #ifdef MUTEX_DEBUG
 	/* Diagnostic and error correction */

sys/kern/kern_rwlock.c

@@ -174,8 +174,9 @@ rw_init_flags(struct rwlock *rw, const char *name, int opts)
 
 	MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 	    RW_RECURSE)) == 0);
-	ASSERT_ATOMIC_LOAD(rw->rw_lock, ("%s: rw_lock not aligned for %s: %p",
-	    __func__, name, &rw->rw_lock));
+	ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
+	    ("%s: rw_lock not aligned for %s: %p", __func__, name,
+	    &rw->rw_lock));
 
 	flags = LO_UPGRADABLE;
 	if (opts & RW_DUPOK)

sys/kern/kern_sx.c

@@ -205,8 +205,9 @@ sx_init_flags(struct sx *sx, const char *description, int opts)
 
 	MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 	    SX_NOPROFILE | SX_NOADAPTIVE)) == 0);
-	ASSERT_ATOMIC_LOAD(sx->sx_lock, ("%s: sx_lock not aligned for %s: %p",
-	    __func__, description, &sx->sx_lock));
+	ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
+	    ("%s: sx_lock not aligned for %s: %p", __func__, description,
+	    &sx->sx_lock));
 
 	flags = LO_SLEEPABLE | LO_UPGRADABLE;
 	if (opts & SX_DUPOK)

sys/sys/systm.h

@@ -89,9 +89,16 @@ extern int maxusers;		/* system tune hint */
 #define	__CTASSERT(x, y)	typedef char __assert ## y[(x) ? 1 : -1]
 #endif
 
-#define	ASSERT_ATOMIC_LOAD(var,msg)					\
-	KASSERT(sizeof(var) <= sizeof(uintptr_t) &&			\
-	    ALIGN(&(var)) == (uintptr_t)&(var), msg)
+/*
+ * Assert that a pointer can be loaded from memory atomically.
+ *
+ * This assertion enforces stronger alignment than necessary.  For example,
+ * on some architectures, atomicity for unaligned loads will depend on
+ * whether or not the load spans multiple cache lines.
+ */
+#define	ASSERT_ATOMIC_LOAD_PTR(var, msg)				\
+	KASSERT(sizeof(var) == sizeof(void *) &&			\
+	    ((uintptr_t)&(var) & (sizeof(void *) - 1)) == 0, msg)
 
 /*
  * XXX the hints declarations are even more misplaced than most declarations