]> err.no Git - linux-2.6/commitdiff
sched clock: couple local and remote clocks
authorIngo Molnar <mingo@elte.hu>
Wed, 30 Jul 2008 08:22:07 +0000 (10:22 +0200)
committerIngo Molnar <mingo@elte.hu>
Thu, 31 Jul 2008 15:21:01 +0000 (17:21 +0200)
When taking the time of a remote CPU, use the opportunity to
couple (sync) the clocks to each other. (in a monotonic way)

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
kernel/sched_clock.c

index 857a1291fd238252c86db12e03b44d8adc7ed19f..074edc9893794aebac2e2f9dd7e3ff9b495a6896 100644 (file)
@@ -149,7 +149,7 @@ static void lock_double_clock(struct sched_clock_data *data1,
 u64 sched_clock_cpu(int cpu)
 {
        struct sched_clock_data *scd = cpu_sdc(cpu);
-       u64 now, clock;
+       u64 now, clock, this_clock, remote_clock;
 
        if (unlikely(!sched_clock_running))
                return 0ull;
@@ -158,26 +158,36 @@ u64 sched_clock_cpu(int cpu)
        now = sched_clock();
 
        if (cpu != raw_smp_processor_id()) {
-               /*
-                * in order to update a remote cpu's clock based on our
-                * unstable raw time rebase it against:
-                *   tick_raw           (offset between raw counters)
-                *   tick_gotd          (tick offset between cpus)
-                */
                struct sched_clock_data *my_scd = this_scd();
 
                lock_double_clock(scd, my_scd);
 
-               now += scd->tick_raw - my_scd->tick_raw;
-               now += my_scd->tick_gtod - scd->tick_gtod;
+               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(remote_clock < this_clock)) {
+                       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);
        }
 
-       clock = __update_sched_clock(scd, now);
-
        __raw_spin_unlock(&scd->lock);
 
        return clock;
@@ -223,7 +233,6 @@ EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 void sched_clock_idle_wakeup_event(u64 delta_ns)
 {
        struct sched_clock_data *scd = this_scd();
-       u64 now = sched_clock();
 
        /*
         * Override the previous timestamp and ignore all