sched_clock: Make it NMI safe
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Fri, 18 Sep 2009 18:14:01 +0000 (20:14 +0200)
committerIngo Molnar <mingo@elte.hu>
Fri, 18 Sep 2009 18:47:30 +0000 (20:47 +0200)
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>
kernel/perf_counter.c
kernel/sched_clock.c

index 6944bd55ec4e91062ea3a9c14675381b756df3ea..06d233a06da59842a81f51daf24f57c333660d85 100644 (file)
@@ -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;
index e1d16c9a76806ef945e4271a9cfcef9ac1a0a0d9..ac2e1dc708bdfb4f090aa1b6a743861bc5941c21 100644 (file)
@@ -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();
-
-               lock_double_clock(scd, my_scd);
-
-               this_clock = __update_sched_clock(my_scd, now);
-               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)) {
-                       clock = this_clock;
-                       scd->clock = clock;
-               } else {
-                       /*
-                        * Should be rare, but possible:
-                        */
-                       clock = remote_clock;
-                       my_scd->clock = remote_clock;
-               }
-
-               __raw_spin_unlock(&my_scd->lock);
-       } else {
-               __raw_spin_lock(&scd->lock);
-               clock = __update_sched_clock(scd, now);
-       }
+       scd = cpu_sdc(cpu);
 
-       __raw_spin_unlock(&scd->lock);
+       if (cpu != smp_processor_id())
+               clock = sched_clock_remote(scd);
+       else
+               clock = sched_clock_local(scd);
 
        return clock;
 }
@@ -223,11 +209,9 @@ void sched_clock_tick(void)
        now_gtod = ktime_to_ns(ktime_get());
        now = sched_clock();
 
-       __raw_spin_lock(&scd->lock);
        scd->tick_raw = now;
        scd->tick_gtod = now_gtod;
-       __update_sched_clock(scd, now);
-       __raw_spin_unlock(&scd->lock);
+       sched_clock_local(scd);
 }
 
 /*