freebsd/x86: Allow and override fence instructions to match kernel environment.

This is mostly based off of Olivier's (cognet@) work. I was considering
importing in __mbk but realistically, doesn't warrant the refactor.

include/freebsd/ck_md.h.in

@@ -100,19 +100,12 @@
  */
 #if defined(__i386__) && !defined(__x86__)
 #define __x86__
-#define CK_MD_TSO
-
 /*
- * On i386, we resort to building the fence instructions on top of
+ * If x86 becomes more relevant, we may want to consider importing in
- * the kernel-provided facilities.
+ * __mbk() to avoid potential issues around false sharing.
  */
-#include <sys/cdefs.h>
+#define CK_MD_TSO
-#include <sys/atomic.h>
+#define CK_MD_SSE_DISABLE	1
-
-#define CK_MD_X86_MFENCE()	mb()
-#define CK_MD_X86_SFENCE()	wmb()
-#define CK_MD_X86_LFENCE()	rmb()
-
 #elif defined(__amd64__)
 #define CK_MD_TSO
 #elif defined(__sparc64__) && !defined(__sparcv9__)

include/gcc/x86/ck_pr.h

@@ -45,15 +45,9 @@
 /* Minimum requirements for the CK_PR interface are met. */
 #define CK_F_PR
 
-#ifdef CK_MD_UMP
-#define CK_PR_LOCK_PREFIX
-#else
-#define CK_PR_LOCK_PREFIX "lock "
-#endif /* CK_MD_UMP */
-
 /*
- * Prevent speculative execution in busy-wait loops (P4 <=)
+ * Prevent speculative execution in busy-wait loops (P4 <=) or "predefined
- * or "predefined delay".
+ * delay".
  */
 CK_CC_INLINE static void
 ck_pr_stall(void)
@@ -62,28 +56,52 @@ ck_pr_stall(void)
 	return;
 }
 
+#ifdef CK_MD_UMP
+#define CK_PR_LOCK_PREFIX
+#define CK_PR_FENCE(T, I)				\
+	CK_CC_INLINE static void			\
+	ck_pr_fence_strict_##T(void)			\
+	{						\
+		__asm__ __volatile__("" ::: "memory");	\
+		return;					\
+	}
+#else
+#define CK_PR_LOCK_PREFIX "lock "
 #define CK_PR_FENCE(T, I)				\
 	CK_CC_INLINE static void			\
 	ck_pr_fence_strict_##T(void)			\
 	{						\
 		__asm__ __volatile__(I ::: "memory");	\
+		return;					\
 	}
+#endif /* CK_MD_UMP */
 
-CK_PR_FENCE(atomic, "sfence")
+#if defined(CK_MD_SSE_DISABLE)
-CK_PR_FENCE(atomic_store, "sfence")
+/* If SSE is disabled, then use atomic operations for serialization. */
-CK_PR_FENCE(atomic_load, "mfence")
+#define CK_MD_X86_MFENCE "lock addl $0, (%%esp)"
-CK_PR_FENCE(store_atomic, "sfence")
+#define CK_MD_X86_SFENCE CK_MD_X86_MFENCE
-CK_PR_FENCE(load_atomic, "mfence")
+#define CK_MD_X86_LFENCE CK_MD_X86_MFENCE
-CK_PR_FENCE(load, "lfence")
+#else
-CK_PR_FENCE(load_store, "mfence")
+#define CK_MD_X86_SFENCE "sfence"
-CK_PR_FENCE(store, "sfence")
+#define CK_MD_X86_LFENCE "lfence"
-CK_PR_FENCE(store_load, "mfence")
+#define CK_MD_X86_MFENCE "mfence"
-CK_PR_FENCE(memory, "mfence")
+#endif /* !CK_MD_SSE_DISABLE */
-CK_PR_FENCE(release, "mfence")
+
-CK_PR_FENCE(acquire, "mfence")
+CK_PR_FENCE(atomic, "")
-CK_PR_FENCE(acqrel, "mfence")
+CK_PR_FENCE(atomic_store, "")
-CK_PR_FENCE(lock, "mfence")
+CK_PR_FENCE(atomic_load, "")
-CK_PR_FENCE(unlock, "mfence")
+CK_PR_FENCE(store_atomic, "")
+CK_PR_FENCE(load_atomic, "")
+CK_PR_FENCE(load, CK_MD_X86_LFENCE)
+CK_PR_FENCE(load_store, CK_MD_X86_MFENCE)
+CK_PR_FENCE(store, CK_MD_X86_SFENCE)
+CK_PR_FENCE(store_load, CK_MD_X86_MFENCE)
+CK_PR_FENCE(memory, CK_MD_X86_MFENCE)
+CK_PR_FENCE(release, CK_MD_X86_MFENCE)
+CK_PR_FENCE(acquire, CK_MD_X86_MFENCE)
+CK_PR_FENCE(acqrel, CK_MD_X86_MFENCE)
+CK_PR_FENCE(lock, CK_MD_X86_MFENCE)
+CK_PR_FENCE(unlock, CK_MD_X86_MFENCE)
 
 #undef CK_PR_FENCE