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 static __read_mostly int sched_clock_running;
  46
  47 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  48
  49 struct sched_clock_data {
  50         /*
  51          * Raw spinlock - this is a special case: this might be called
  52          * from within instrumentation code so we dont want to do any
  53          * instrumentation ourselves.
  54          */
  55         raw_spinlock_t          lock;
  56
  57         unsigned long           tick_jiffies;
  58         u64                     tick_raw;
  59         u64                     tick_gtod;
  60         u64                     clock;
  61 };
  62
  63 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  64
  65 static inline struct sched_clock_data *this_scd(void)
  66 {
  67         return &__get_cpu_var(sched_clock_data);
  68 }
  69
  70 static inline struct sched_clock_data *cpu_sdc(int cpu)
  71 {
  72         return &per_cpu(sched_clock_data, cpu);
  73 }
  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 u64 __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         return clock;
 135 }
 136
 137 static void lock_double_clock(struct sched_clock_data *data1,
 138                                 struct sched_clock_data *data2)
 139 {
 140         if (data1 < data2) {
 141                 __raw_spin_lock(&data1->lock);
 142                 __raw_spin_lock(&data2->lock);
 143         } else {
 144                 __raw_spin_lock(&data2->lock);
 145                 __raw_spin_lock(&data1->lock);
 146         }
 147 }
 148
 149 u64 sched_clock_cpu(int cpu)
 150 {
 151         struct sched_clock_data *scd = cpu_sdc(cpu);
 152         u64 now, clock, this_clock, remote_clock;
 153
 154         if (unlikely(!sched_clock_running))
 155                 return 0ull;
 156
 157         WARN_ON_ONCE(!irqs_disabled());
 158         now = sched_clock();
 159
 160         if (cpu != raw_smp_processor_id()) {
 161                 struct sched_clock_data *my_scd = this_scd();
 162
 163                 lock_double_clock(scd, my_scd);
 164
 165                 this_clock = __update_sched_clock(my_scd, now);
 166                 remote_clock = scd->clock;
 167
 168                 /*
 169                  * Use the opportunity that we have both locks
 170                  * taken to couple the two clocks: we take the
 171                  * larger time as the latest time for both
 172                  * runqueues. (this creates monotonic movement)
 173                  */
 174                 if (likely(remote_clock < this_clock)) {
 175                         clock = this_clock;
 176                         scd->clock = clock;
 177                 } else {
 178                         /*
 179                          * Should be rare, but possible:
 180                          */
 181                         clock = remote_clock;
 182                         my_scd->clock = remote_clock;
 183                 }
 184
 185                 __raw_spin_unlock(&my_scd->lock);
 186         } else {
 187                 __raw_spin_lock(&scd->lock);
 188                 clock = __update_sched_clock(scd, now);
 189         }
 190
 191         __raw_spin_unlock(&scd->lock);
 192
 193         return clock;
 194 }
 195
 196 void sched_clock_tick(void)
 197 {
 198         struct sched_clock_data *scd = this_scd();
 199         u64 now, now_gtod;
 200
 201         if (unlikely(!sched_clock_running))
 202                 return;
 203
 204         WARN_ON_ONCE(!irqs_disabled());
 205
 206         now_gtod = ktime_to_ns(ktime_get());
 207         now = sched_clock();
 208
 209         __raw_spin_lock(&scd->lock);
 210         __update_sched_clock(scd, now);
 211         /*
 212          * update tick_gtod after __update_sched_clock() because that will
 213          * already observe 1 new jiffy; adding a new tick_gtod to that would
 214          * increase the clock 2 jiffies.
 215          */
 216         scd->tick_raw = now;
 217         scd->tick_gtod = now_gtod;
 218         __raw_spin_unlock(&scd->lock);
 219 }
 220
 221 /*
 222  * We are going deep-idle (irqs are disabled):
 223  */
 224 void sched_clock_idle_sleep_event(void)
 225 {
 226         sched_clock_cpu(smp_processor_id());
 227 }
 228 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 229
 230 /*
 231  * We just idled delta nanoseconds (called with irqs disabled):
 232  */
 233 void sched_clock_idle_wakeup_event(u64 delta_ns)
 234 {
 235         struct sched_clock_data *scd = this_scd();
 236
 237         /*
 238          * Override the previous timestamp and ignore all
 239          * sched_clock() deltas that occured while we idled,
 240          * and use the PM-provided delta_ns to advance the
 241          * rq clock:
 242          */
 243         __raw_spin_lock(&scd->lock);
 244         scd->clock += delta_ns;
 245         __raw_spin_unlock(&scd->lock);
 246
 247         touch_softlockup_watchdog();
 248 }
 249 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 250
 251 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 252
 253 void sched_clock_init(void)
 254 {
 255         sched_clock_running = 1;
 256 }
 257
 258 u64 sched_clock_cpu(int cpu)
 259 {
 260         if (unlikely(!sched_clock_running))
 261                 return 0;
 262
 263         return sched_clock();
 264 }
 265
 266 #endif
 267
 268 unsigned long long cpu_clock(int cpu)
 269 {
 270         unsigned long long clock;
 271         unsigned long flags;
 272
 273         local_irq_save(flags);
 274         clock = sched_clock_cpu(cpu);
 275         local_irq_restore(flags);
 276
 277         return clock;
 278 }
 279 EXPORT_SYMBOL_GPL(cpu_clock);