kernel/sched_clock.c

   1 /*
   2  * sched_clock for unstable cpu clocks
   3  *
   4  *  Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   5  *
   6  *  Updates and enhancements:
   7  *    Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
   8  *
   9  * Based on code by:
  10  *   Ingo Molnar <mingo@redhat.com>
  11  *   Guillaume Chazarain <guichaz@gmail.com>
  12  *
  13  * Create a semi stable clock from a mixture of other events, including:
  14  *  - gtod
  15  *  - jiffies
  16  *  - sched_clock()
  17  *  - explicit idle events
  18  *
  19  * We use gtod as base and the unstable clock deltas. The deltas are filtered,
  20  * making it monotonic and keeping it within an expected window.  This window
  21  * is set up using jiffies.
  22  *
  23  * Furthermore, explicit sleep and wakeup hooks allow us to account for time
  24  * that is otherwise invisible (TSC gets stopped).
  25  *
  26  * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat
  27  * consistent between cpus (never more than 1 jiffies difference).
  28  */
  29 #include <linux/sched.h>
  30 #include <linux/percpu.h>
  31 #include <linux/spinlock.h>
  32 #include <linux/ktime.h>
  33 #include <linux/module.h>
  34
  35 /*
  36  * Scheduler clock - returns current time in nanosec units.
  37  * This is default implementation.
  38  * Architectures and sub-architectures can override this.
  39  */
  40 unsigned long long __attribute__((weak)) sched_clock(void)
  41 {
  42         return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
  43 }
  44
  45 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  46
  47 struct sched_clock_data {
  48         /*
  49          * Raw spinlock - this is a special case: this might be called
  50          * from within instrumentation code so we dont want to do any
  51          * instrumentation ourselves.
  52          */
  53         raw_spinlock_t          lock;
  54
  55         unsigned long           tick_jiffies;
  56         u64                     tick_raw;
  57         u64                     tick_gtod;
  58         u64                     clock;
  59 };
  60
  61 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  62
  63 static inline struct sched_clock_data *this_scd(void)
  64 {
  65         return &__get_cpu_var(sched_clock_data);
  66 }
  67
  68 static inline struct sched_clock_data *cpu_sdc(int cpu)
  69 {
  70         return &per_cpu(sched_clock_data, cpu);
  71 }
  72
  73 static __read_mostly int sched_clock_running;
  74
  75 void sched_clock_init(void)
  76 {
  77         u64 ktime_now = ktime_to_ns(ktime_get());
  78         unsigned long now_jiffies = jiffies;
  79         int cpu;
  80
  81         for_each_possible_cpu(cpu) {
  82                 struct sched_clock_data *scd = cpu_sdc(cpu);
  83
  84                 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  85                 scd->tick_jiffies = now_jiffies;
  86                 scd->tick_raw = 0;
  87                 scd->tick_gtod = ktime_now;
  88                 scd->clock = ktime_now;
  89         }
  90
  91         sched_clock_running = 1;
  92 }
  93
  94 /*
  95  * update the percpu scd from the raw @now value
  96  *
  97  *  - filter out backward motion
  98  *  - use jiffies to generate a min,max window to clip the raw values
  99  */
 100 static void __update_sched_clock(struct sched_clock_data *scd, u64 now)
 101 {
 102         unsigned long now_jiffies = jiffies;
 103         long delta_jiffies = now_jiffies - scd->tick_jiffies;
 104         u64 clock = scd->clock;
 105         u64 min_clock, max_clock;
 106         s64 delta = now - scd->tick_raw;
 107
 108         WARN_ON_ONCE(!irqs_disabled());
 109         min_clock = scd->tick_gtod + delta_jiffies * TICK_NSEC;
 110
 111         if (unlikely(delta < 0)) {
 112                 clock++;
 113                 goto out;
 114         }
 115
 116         max_clock = min_clock + TICK_NSEC;
 117
 118         if (unlikely(clock + delta > max_clock)) {
 119                 if (clock < max_clock)
 120                         clock = max_clock;
 121                 else
 122                         clock++;
 123         } else {
 124                 clock += delta;
 125         }
 126
 127  out:
 128         if (unlikely(clock < min_clock))
 129                 clock = min_clock;
 130
 131         scd->tick_jiffies = now_jiffies;
 132         scd->clock = clock;
 133 }
 134
 135 static void lock_double_clock(struct sched_clock_data *data1,
 136                                 struct sched_clock_data *data2)
 137 {
 138         if (data1 < data2) {
 139                 __raw_spin_lock(&data1->lock);
 140                 __raw_spin_lock(&data2->lock);
 141         } else {
 142                 __raw_spin_lock(&data2->lock);
 143                 __raw_spin_lock(&data1->lock);
 144         }
 145 }
 146
 147 u64 sched_clock_cpu(int cpu)
 148 {
 149         struct sched_clock_data *scd = cpu_sdc(cpu);
 150         u64 now, clock;
 151
 152         if (unlikely(!sched_clock_running))
 153                 return 0ull;
 154
 155         WARN_ON_ONCE(!irqs_disabled());
 156         now = sched_clock();
 157
 158         if (cpu != raw_smp_processor_id()) {
 159                 /*
 160                  * in order to update a remote cpu's clock based on our
 161                  * unstable raw time rebase it against:
 162                  *   tick_raw           (offset between raw counters)
 163                  *   tick_gotd          (tick offset between cpus)
 164                  */
 165                 struct sched_clock_data *my_scd = this_scd();
 166
 167                 lock_double_clock(scd, my_scd);
 168
 169                 now += scd->tick_raw - my_scd->tick_raw;
 170                 now += my_scd->tick_gtod - scd->tick_gtod;
 171
 172                 __raw_spin_unlock(&my_scd->lock);
 173         } else {
 174                 __raw_spin_lock(&scd->lock);
 175         }
 176
 177         __update_sched_clock(scd, now);
 178         clock = scd->clock;
 179
 180         __raw_spin_unlock(&scd->lock);
 181
 182         return clock;
 183 }
 184
 185 void sched_clock_tick(void)
 186 {
 187         struct sched_clock_data *scd = this_scd();
 188         u64 now, now_gtod;
 189
 190         if (unlikely(!sched_clock_running))
 191                 return;
 192
 193         WARN_ON_ONCE(!irqs_disabled());
 194
 195         now_gtod = ktime_to_ns(ktime_get());
 196         now = sched_clock();
 197
 198         __raw_spin_lock(&scd->lock);
 199         __update_sched_clock(scd, now);
 200         /*
 201          * update tick_gtod after __update_sched_clock() because that will
 202          * already observe 1 new jiffy; adding a new tick_gtod to that would
 203          * increase the clock 2 jiffies.
 204          */
 205         scd->tick_raw = now;
 206         scd->tick_gtod = now_gtod;
 207         __raw_spin_unlock(&scd->lock);
 208 }
 209
 210 /*
 211  * We are going deep-idle (irqs are disabled):
 212  */
 213 void sched_clock_idle_sleep_event(void)
 214 {
 215         sched_clock_cpu(smp_processor_id());
 216 }
 217 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 218
 219 /*
 220  * We just idled delta nanoseconds (called with irqs disabled):
 221  */
 222 void sched_clock_idle_wakeup_event(u64 delta_ns)
 223 {
 224         struct sched_clock_data *scd = this_scd();
 225         u64 now = sched_clock();
 226
 227         /*
 228          * Override the previous timestamp and ignore all
 229          * sched_clock() deltas that occured while we idled,
 230          * and use the PM-provided delta_ns to advance the
 231          * rq clock:
 232          */
 233         __raw_spin_lock(&scd->lock);
 234         scd->clock += delta_ns;
 235         __raw_spin_unlock(&scd->lock);
 236
 237         touch_softlockup_watchdog();
 238 }
 239 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 240
 241 #endif
 242
 243 unsigned long long cpu_clock(int cpu)
 244 {
 245         unsigned long long clock;
 246         unsigned long flags;
 247
 248         local_irq_save(flags);
 249         clock = sched_clock_cpu(cpu);
 250         local_irq_restore(flags);
 251
 252         return clock;
 253 }
 254 EXPORT_SYMBOL_GPL(cpu_clock);