sched_clock: Make it NMI safe

Arjan complained about the suckyness of TSC on modern machines, and
asked if we could do something about that for PERF_SAMPLE_TIME.

Make cpu_clock() NMI safe by removing the spinlock and using
cmpxchg. This also makes it smaller and more robust.

Affects architectures that use HAVE_UNSTABLE_SCHED_CLOCK, i.e. IA64
and x86.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

2 files changed, 56 insertions, 75 deletions

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 6944bd55ec4e..06d233a06da5 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -2955,10 +2955,7 @@ void perf_prepare_sample(struct perf_event_header *header,
         }
 
         if (sample_type & PERF_SAMPLE_TIME) {
-                /*
-                 * Maybe do better on x86 and provide cpu_clock_nmi()
-                 */
-                data->time = sched_clock();
+                data->time = perf_clock();
 
                 header->size += sizeof(data->time);
         }
@@ -3488,7 +3485,7 @@ static void perf_log_throttle(struct perf_counter *counter, int enable)
                         .misc = 0,
                         .size = sizeof(throttle_event),
                 },
-                .time           = sched_clock(),
+                .time           = perf_clock(),
                 .id             = primary_counter_id(counter),
                 .stream_id      = counter->id,
         };
@@ -3540,7 +3537,7 @@ static int __perf_counter_overflow(struct perf_counter *counter, int nmi,
         }
 
         if (counter->attr.freq) {
-                u64 now = sched_clock();
+                u64 now = perf_clock();
                 s64 delta = now - hwc->freq_stamp;
 
                 hwc->freq_stamp = now;
diff --git a/kernel/sched_clock.c b/kernel/sched_clock.c
index e1d16c9a7680..ac2e1dc708bd 100644
--- a/kernel/sched_clock.c
+++ b/kernel/sched_clock.c
@@ -48,13 +48,6 @@ static __read_mostly int sched_clock_running;
 __read_mostly int sched_clock_stable;
 
 struct sched_clock_data {
-        /*
-         * Raw spinlock - this is a special case: this might be called
-         * from within instrumentation code so we dont want to do any
-         * instrumentation ourselves.
-         */
-        raw_spinlock_t          lock;
-
         u64                     tick_raw;
         u64                     tick_gtod;
         u64                     clock;
@@ -80,7 +73,6 @@ void sched_clock_init(void)
         for_each_possible_cpu(cpu) {
                 struct sched_clock_data *scd = cpu_sdc(cpu);
 
-                scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
                 scd->tick_raw = 0;
                 scd->tick_gtod = ktime_now;
                 scd->clock = ktime_now;
@@ -109,14 +101,19 @@ static inline u64 wrap_max(u64 x, u64 y)
  *  - filter out backward motion
  *  - use the GTOD tick value to create a window to filter crazy TSC values
  */
-static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
+static u64 sched_clock_local(struct sched_clock_data *scd)
 {
-        s64 delta = now - scd->tick_raw;
-        u64 clock, min_clock, max_clock;
+        u64 now, clock, old_clock, min_clock, max_clock;
+        s64 delta;
 
+again:
+        now = sched_clock();
+        delta = now - scd->tick_raw;
         if (unlikely(delta < 0))
                 delta = 0;
 
+        old_clock = scd->clock;
+
         /*
          * scd->clock = clamp(scd->tick_gtod + delta,
          *                    max(scd->tick_gtod, scd->clock),
@@ -124,84 +121,73 @@ static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
          */
 
         clock = scd->tick_gtod + delta;
-        min_clock = wrap_max(scd->tick_gtod, scd->clock);
-        max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC);
+        min_clock = wrap_max(scd->tick_gtod, old_clock);
+        max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
 
         clock = wrap_max(clock, min_clock);
         clock = wrap_min(clock, max_clock);
 
-        scd->clock = clock;
+        if (cmpxchg(&scd->clock, old_clock, clock) != old_clock)
+                goto again;
 
-        return scd->clock;
+        return clock;
 }
 
-static void lock_double_clock(struct sched_clock_data *data1,
-                                struct sched_clock_data *data2)
+static u64 sched_clock_remote(struct sched_clock_data *scd)
 {
-        if (data1 < data2) {
-                __raw_spin_lock(&data1->lock);
-                __raw_spin_lock(&data2->lock);
+        struct sched_clock_data *my_scd = this_scd();
+        u64 this_clock, remote_clock;
+        u64 *ptr, old_val, val;
+
+        sched_clock_local(my_scd);
+again:
+        this_clock = my_scd->clock;
+        remote_clock = scd->clock;
+
+        /*
+         * Use the opportunity that we have both locks
+         * taken to couple the two clocks: we take the
+         * larger time as the latest time for both
+         * runqueues. (this creates monotonic movement)
+         */
+        if (likely((s64)(remote_clock - this_clock) < 0)) {
+                ptr = &scd->clock;
+                old_val = remote_clock;
+                val = this_clock;
         } else {
-                __raw_spin_lock(&data2->lock);
-                __raw_spin_lock(&data1->lock);
+                /*
+                 * Should be rare, but possible:
+                 */
+                ptr = &my_scd->clock;
+                old_val = this_clock;
+                val = remote_clock;
         }
+
+        if (cmpxchg(ptr, old_val, val) != old_val)
+                goto again;
+
+        return val;
 }
 
 u64 sched_clock_cpu(int cpu)
 {
-        u64 now, clock, this_clock, remote_clock;
         struct sched_clock_data *scd;
+        u64 clock;
+
+        WARN_ON_ONCE(!irqs_disabled());
 
         if (sched_clock_stable)
                 return sched_clock();
 
-        scd = cpu_sdc(cpu);
-
-        /*
-         * Normally this is not called in NMI context - but if it is,
-         * trying to do any locking here is totally lethal.
-         */
-        if (unlikely(in_nmi()))
-                return scd->clock;
-
         if (unlikely(!sched_clock_running))
                 return 0ull;
 
-        WARN_ON_ONCE(!irqs_disabled());
-        now = sched_clock();
-
-        if (cpu != raw_smp_processor_id()) {
-                struct sched_clock_data *my_scd = this_scd();
+        scd = cpu_sdc(cpu);