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