[PATCH v3 14/18] x86: Add rdtsc_ordered() and use it in trivial call sites

[Andy Lutomirski <luto@kernel.org>]

rdtsc_barrier(); rdtsc() is an unnecessary mouthful and requires
more thought than should be necessary.  Add an rdtsc_ordered()
helper and replace the trivial call sites with it.

This should not change generated code.  The duplication of the fence
asm is temporary.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
---
 arch/x86/entry/vdso/vclock_gettime.c | 16 ++--------------
 arch/x86/include/asm/msr.h           | 26 ++++++++++++++++++++++++++
 arch/x86/kernel/trace_clock.c        |  7 +------
 arch/x86/kvm/x86.c                   | 16 ++--------------
 arch/x86/lib/delay.c                 |  9 +++------
 5 files changed, 34 insertions(+), 40 deletions(-)

diff --git a/arch/x86/entry/vdso/vclock_gettime.c b/arch/x86/entry/vdso/vclock_gettime.c
index 0340d93c18ca..ca94fa649251 100644
--- a/arch/x86/entry/vdso/vclock_gettime.c
+++ b/arch/x86/entry/vdso/vclock_gettime.c
@@ -175,20 +175,8 @@ static notrace cycle_t vread_pvclock(int *mode)
 
 notrace static cycle_t vread_tsc(void)
 {
-	cycle_t ret;
-	u64 last;
-
-	/*
-	 * Empirically, a fence (of type that depends on the CPU)
-	 * before rdtsc is enough to ensure that rdtsc is ordered
-	 * with respect to loads.  The various CPU manuals are unclear
-	 * as to whether rdtsc can be reordered with later loads,
-	 * but no one has ever seen it happen.
-	 */
-	rdtsc_barrier();
-	ret = (cycle_t)rdtsc();
-
-	last = gtod->cycle_last;
+	cycle_t ret = (cycle_t)rdtsc_ordered();
+	u64 last = gtod->cycle_last;
 
 	if (likely(ret >= last))
 		return ret;
diff --git a/arch/x86/include/asm/msr.h b/arch/x86/include/asm/msr.h
index ff0c120dafe5..02bdd6c65017 100644
--- a/arch/x86/include/asm/msr.h
+++ b/arch/x86/include/asm/msr.h
@@ -127,6 +127,32 @@ static __always_inline unsigned long long rdtsc(void)
 	return EAX_EDX_VAL(val, low, high);
 }
 
+/**
+ * rdtsc_ordered() - read the current TSC in program order
+ *
+ * rdtsc_ordered() returns the result of RDTSC as a 64-bit integer.
+ * It is ordered like a load to a global in-memory counter.  It should
+ * be impossible to observe non-monotonic rdtsc_unordered() behavior
+ * across multiple CPUs as long as the TSC is synced.
+ */
+static __always_inline unsigned long long rdtsc_ordered(void)
+{
+	/*
+	 * The RDTSC instruction is not ordered relative to memory
+	 * access.  The Intel SDM and the AMD APM are both vague on this
+	 * point, but empirically an RDTSC instruction can be
+	 * speculatively executed before prior loads.  An RDTSC
+	 * immediately after an appropriate barrier appears to be
+	 * ordered as a normal load, that is, it provides the same
+	 * ordering guarantees as reading from a global memory location
+	 * that some other imaginary CPU is updating continuously with a
+	 * time stamp.
+	 */
+	alternative_2("", "mfence", X86_FEATURE_MFENCE_RDTSC,
+			  "lfence", X86_FEATURE_LFENCE_RDTSC);
+	return rdtsc();
+}
+
 static inline unsigned long long native_read_pmc(int counter)
 {
 	DECLARE_ARGS(val, low, high);
diff --git a/arch/x86/kernel/trace_clock.c b/arch/x86/kernel/trace_clock.c
index 67efb8c96fc4..80bb24d9b880 100644
--- a/arch/x86/kernel/trace_clock.c
+++ b/arch/x86/kernel/trace_clock.c
@@ -12,10 +12,5 @@
  */
 u64 notrace trace_clock_x86_tsc(void)
 {
-	u64 ret;
-
-	rdtsc_barrier();
-	ret = rdtsc();
-
-	return ret;
+	return rdtsc_ordered();
 }
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index b0afdc74c28a..dfccaf2f2e00 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -1419,20 +1419,8 @@ EXPORT_SYMBOL_GPL(kvm_write_tsc);
 
 static cycle_t read_tsc(void)
 {
-	cycle_t ret;
-	u64 last;
-
-	/*
-	 * Empirically, a fence (of type that depends on the CPU)
-	 * before rdtsc is enough to ensure that rdtsc is ordered
-	 * with respect to loads.  The various CPU manuals are unclear
-	 * as to whether rdtsc can be reordered with later loads,
-	 * but no one has ever seen it happen.
-	 */
-	rdtsc_barrier();
-	ret = (cycle_t)rdtsc();
-
-	last = pvclock_gtod_data.clock.cycle_last;
+	cycle_t ret = (cycle_t)rdtsc_ordered();
+	u64 last = pvclock_gtod_data.clock.cycle_last;
 
 	if (likely(ret >= last))
 		return ret;
diff --git a/arch/x86/lib/delay.c b/arch/x86/lib/delay.c
index f24bc59ab0a0..4453d52a143d 100644
--- a/arch/x86/lib/delay.c
+++ b/arch/x86/lib/delay.c
@@ -54,11 +54,9 @@ static void delay_tsc(unsigned long __loops)
 
 	preempt_disable();
 	cpu = smp_processor_id();
-	rdtsc_barrier();
-	bclock = rdtsc();
+	bclock = rdtsc_ordered();
 	for (;;) {
-		rdtsc_barrier();
-		now = rdtsc();
+		now = rdtsc_ordered();
 		if ((now - bclock) >= loops)
 			break;
 
@@ -79,8 +77,7 @@ static void delay_tsc(unsigned long __loops)
 		if (unlikely(cpu != smp_processor_id())) {
 			loops -= (now - bclock);
 			cpu = smp_processor_id();
-			rdtsc_barrier();
-			bclock = rdtsc();
+			bclock = rdtsc_ordered();
 		}
 	}
 	preempt_enable();
-- 
2.4.2