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                     prev_raw;
  57         u64                     tick_raw;
  58         u64                     tick_gtod;
  59         u64                     clock;
  60 };
  61
  62 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  63
  64 static inline struct sched_clock_data *this_scd(void)
  65 {
  66         return &__get_cpu_var(sched_clock_data);
  67 }
  68
  69 static inline struct sched_clock_data *cpu_sdc(int cpu)
  70 {
  71         return &per_cpu(sched_clock_data, cpu);
  72 }
  73
  74 static __read_mostly int sched_clock_running;
  75
  76 void sched_clock_init(void)
  77 {
  78         u64 ktime_now = ktime_to_ns(ktime_get());
  79         unsigned long now_jiffies = jiffies;
  80         int cpu;
  81
  82         for_each_possible_cpu(cpu) {
  83                 struct sched_clock_data *scd = cpu_sdc(cpu);
  84
  85                 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  86                 scd->tick_jiffies = now_jiffies;
  87                 scd->prev_raw = 0;
  88                 scd->tick_raw = 0;
  89                 scd->tick_gtod = ktime_now;
  90                 scd->clock = ktime_now;
  91         }
  92
  93         sched_clock_running = 1;
  94 }
  95
  96 /*
  97  * update the percpu scd from the raw @now value
  98  *
  99  *  - filter out backward motion
 100  *  - use jiffies to generate a min,max window to clip the raw values
 101  */
 102 static void __update_sched_clock(struct sched_clock_data *scd, u64 now)
 103 {
 104         unsigned long now_jiffies = jiffies;
 105         long delta_jiffies = now_jiffies - scd->tick_jiffies;
 106         u64 clock = scd->clock;
 107         u64 min_clock, max_clock;
 108         s64 delta = now - scd->prev_raw;
 109
 110         WARN_ON_ONCE(!irqs_disabled());
 111         min_clock = scd->tick_gtod + delta_jiffies * TICK_NSEC;
 112
 113         if (unlikely(delta < 0)) {
 114                 clock++;
 115                 goto out;
 116         }
 117
 118         max_clock = min_clock + TICK_NSEC;
 119
 120         if (unlikely(clock + delta > max_clock)) {
 121                 if (clock < max_clock)
 122                         clock = max_clock;
 123                 else
 124                         clock++;
 125         } else {
 126                 clock += delta;
 127         }
 128
 129  out:
 130         if (unlikely(clock < min_clock))
 131                 clock = min_clock;
 132
 133         scd->prev_raw = now;
 134         scd->tick_jiffies = now_jiffies;
 135         scd->clock = clock;
 136 }
 137
 138 static void lock_double_clock(struct sched_clock_data *data1,
 139                                 struct sched_clock_data *data2)
 140 {
 141         if (data1 < data2) {
 142                 __raw_spin_lock(&data1->lock);
 143                 __raw_spin_lock(&data2->lock);
 144         } else {
 145                 __raw_spin_lock(&data2->lock);
 146                 __raw_spin_lock(&data1->lock);
 147         }
 148 }
 149
 150 u64 sched_clock_cpu(int cpu)
 151 {
 152         struct sched_clock_data *scd = cpu_sdc(cpu);
 153         u64 now, clock;
 154
 155         if (unlikely(!sched_clock_running))
 156                 return 0ull;
 157
 158         WARN_ON_ONCE(!irqs_disabled());
 159         now = sched_clock();
 160
 161         if (cpu != raw_smp_processor_id()) {
 162                 /*
 163                  * in order to update a remote cpu's clock based on our
 164                  * unstable raw time rebase it against:
 165                  *   tick_raw           (offset between raw counters)
 166                  *   tick_gotd          (tick offset between cpus)
 167                  */
 168                 struct sched_clock_data *my_scd = this_scd();
 169
 170                 lock_double_clock(scd, my_scd);
 171
 172                 now += scd->tick_raw - my_scd->tick_raw;
 173                 now += my_scd->tick_gtod - scd->tick_gtod;
 174
 175                 __raw_spin_unlock(&my_scd->lock);
 176         } else {
 177                 __raw_spin_lock(&scd->lock);
 178         }
 179
 180         __update_sched_clock(scd, now);
 181         clock = scd->clock;
 182
 183         __raw_spin_unlock(&scd->lock);
 184
 185         return clock;
 186 }
 187
 188 void sched_clock_tick(void)
 189 {
 190         struct sched_clock_data *scd = this_scd();
 191         u64 now, now_gtod;
 192
 193         if (unlikely(!sched_clock_running))
 194                 return;
 195
 196         WARN_ON_ONCE(!irqs_disabled());
 197
 198         now_gtod = ktime_to_ns(ktime_get());
 199         now = sched_clock();
 200
 201         __raw_spin_lock(&scd->lock);
 202         __update_sched_clock(scd, now);
 203         /*
 204          * update tick_gtod after __update_sched_clock() because that will
 205          * already observe 1 new jiffy; adding a new tick_gtod to that would
 206          * increase the clock 2 jiffies.
 207          */
 208         scd->tick_raw = now;
 209         scd->tick_gtod = now_gtod;
 210         __raw_spin_unlock(&scd->lock);
 211 }
 212
 213 /*
 214  * We are going deep-idle (irqs are disabled):
 215  */
 216 void sched_clock_idle_sleep_event(void)
 217 {
 218         sched_clock_cpu(smp_processor_id());
 219 }
 220 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 221
 222 /*
 223  * We just idled delta nanoseconds (called with irqs disabled):
 224  */
 225 void sched_clock_idle_wakeup_event(u64 delta_ns)
 226 {
 227         struct sched_clock_data *scd = this_scd();
 228         u64 now = sched_clock();
 229
 230         /*
 231          * Override the previous timestamp and ignore all
 232          * sched_clock() deltas that occured while we idled,
 233          * and use the PM-provided delta_ns to advance the
 234          * rq clock:
 235          */
 236         __raw_spin_lock(&scd->lock);
 237         scd->prev_raw = now;
 238         scd->clock += delta_ns;
 239         __raw_spin_unlock(&scd->lock);
 240
 241         touch_softlockup_watchdog();
 242 }
 243 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 244
 245 #endif
 246
 247 unsigned long long cpu_clock(int cpu)
 248 {
 249         unsigned long long clock;
 250         unsigned long flags;
 251
 252         local_irq_save(flags);
 253         clock = sched_clock_cpu(cpu);
 254         local_irq_restore(flags);
 255
 256         return clock;
 257 }
 258 EXPORT_SYMBOL_GPL(cpu_clock);