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 #ifdef CONFIG_NO_HZ
  49         int                     check_max;
  50 #endif
  51 };
  52
  53 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  54
  55 static inline struct sched_clock_data *this_scd(void)
  56 {
  57         return &__get_cpu_var(sched_clock_data);
  58 }
  59
  60 static inline struct sched_clock_data *cpu_sdc(int cpu)
  61 {
  62         return &per_cpu(sched_clock_data, cpu);
  63 }
  64
  65 static __read_mostly int sched_clock_running;
  66
  67 void sched_clock_init(void)
  68 {
  69         u64 ktime_now = ktime_to_ns(ktime_get());
  70         unsigned long now_jiffies = jiffies;
  71         int cpu;
  72
  73         for_each_possible_cpu(cpu) {
  74                 struct sched_clock_data *scd = cpu_sdc(cpu);
  75
  76                 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  77                 scd->tick_jiffies = now_jiffies;
  78                 scd->prev_raw = 0;
  79                 scd->tick_raw = 0;
  80                 scd->tick_gtod = ktime_now;
  81                 scd->clock = ktime_now;
  82 #ifdef CONFIG_NO_HZ
  83                 scd->check_max = 1;
  84 #endif
  85         }
  86
  87         sched_clock_running = 1;
  88 }
  89
  90 #ifdef CONFIG_NO_HZ
  91 /*
  92  * The dynamic ticks makes the delta jiffies inaccurate. This
  93  * prevents us from checking the maximum time update.
  94  * Disable the maximum check during stopped ticks.
  95  */
  96 void sched_clock_tick_stop(int cpu)
  97 {
  98         struct sched_clock_data *scd = cpu_sdc(cpu);
  99
 100         scd->check_max = 0;
 101 }
 102
 103 void sched_clock_tick_start(int cpu)
 104 {
 105         struct sched_clock_data *scd = cpu_sdc(cpu);
 106
 107         scd->check_max = 1;
 108 }
 109
 110 static int check_max(struct sched_clock_data *scd)
 111 {
 112         return scd->check_max;
 113 }
 114 #else
 115 static int check_max(struct sched_clock_data *scd)
 116 {
 117         return 1;
 118 }
 119 #endif /* CONFIG_NO_HZ */
 120
 121 /*
 122  * update the percpu scd from the raw @now value
 123  *
 124  *  - filter out backward motion
 125  *  - use jiffies to generate a min,max window to clip the raw values
 126  */
 127 static void __update_sched_clock(struct sched_clock_data *scd, u64 now, u64 *time)
 128 {
 129         unsigned long now_jiffies = jiffies;
 130         long delta_jiffies = now_jiffies - scd->tick_jiffies;
 131         u64 clock = scd->clock;
 132         u64 min_clock, max_clock;
 133         s64 delta = now - scd->prev_raw;
 134
 135         WARN_ON_ONCE(!irqs_disabled());
 136
 137         min_clock = scd->tick_gtod +
 138                 (delta_jiffies ? delta_jiffies - 1 : 0) * TICK_NSEC;
 139
 140         if (unlikely(delta < 0)) {
 141                 clock++;
 142                 goto out;
 143         }
 144
 145         /*
 146          * The clock must stay within a jiffie of the gtod.
 147          * But since we may be at the start of a jiffy or the end of one
 148          * we add another jiffy buffer.
 149          */
 150         max_clock = scd->tick_gtod + (2 + delta_jiffies) * TICK_NSEC;
 151
 152         if (unlikely(clock + delta > max_clock) && check_max(scd)) {
 153                 if (clock < max_clock)
 154                         clock = max_clock;
 155                 else
 156                         clock++;
 157         } else {
 158                 clock += delta;
 159         }
 160
 161  out:
 162         if (unlikely(clock < min_clock))
 163                 clock = min_clock;
 164
 165         if (time)
 166                 *time = clock;
 167         else {
 168                 scd->prev_raw = now;
 169                 scd->clock = clock;
 170         }
 171 }
 172
 173 static void lock_double_clock(struct sched_clock_data *data1,
 174                                 struct sched_clock_data *data2)
 175 {
 176         if (data1 < data2) {
 177                 __raw_spin_lock(&data1->lock);
 178                 __raw_spin_lock(&data2->lock);
 179         } else {
 180                 __raw_spin_lock(&data2->lock);
 181                 __raw_spin_lock(&data1->lock);
 182         }
 183 }
 184
 185 u64 sched_clock_cpu(int cpu)
 186 {
 187         struct sched_clock_data *scd = cpu_sdc(cpu);
 188         u64 now, clock;
 189
 190         if (unlikely(!sched_clock_running))
 191                 return 0ull;
 192
 193         WARN_ON_ONCE(!irqs_disabled());
 194         now = sched_clock();
 195
 196         if (cpu != raw_smp_processor_id()) {
 197                 /*
 198                  * in order to update a remote cpu's clock based on our
 199                  * unstable raw time rebase it against:
 200                  *   tick_raw           (offset between raw counters)
 201                  *   tick_gotd          (tick offset between cpus)
 202                  */
 203                 struct sched_clock_data *my_scd = this_scd();
 204
 205                 lock_double_clock(scd, my_scd);
 206
 207                 now -= my_scd->tick_raw;
 208                 now += scd->tick_raw;
 209
 210                 now += my_scd->tick_gtod;
 211                 now -= scd->tick_gtod;
 212
 213                 __raw_spin_unlock(&my_scd->lock);
 214
 215                 __update_sched_clock(scd, now, &clock);
 216
 217                 __raw_spin_unlock(&scd->lock);
 218
 219         } else {
 220                 __raw_spin_lock(&scd->lock);
 221                 __update_sched_clock(scd, now, NULL);
 222                 clock = scd->clock;
 223                 __raw_spin_unlock(&scd->lock);
 224         }
 225
 226         return clock;
 227 }
 228
 229 void sched_clock_tick(void)
 230 {
 231         struct sched_clock_data *scd = this_scd();
 232         unsigned long now_jiffies = jiffies;
 233         u64 now, now_gtod;
 234
 235         if (unlikely(!sched_clock_running))
 236                 return;
 237
 238         WARN_ON_ONCE(!irqs_disabled());
 239
 240         now = sched_clock();
 241         now_gtod = ktime_to_ns(ktime_get());
 242
 243         __raw_spin_lock(&scd->lock);
 244         __update_sched_clock(scd, now, NULL);
 245         /*
 246          * update tick_gtod after __update_sched_clock() because that will
 247          * already observe 1 new jiffy; adding a new tick_gtod to that would
 248          * increase the clock 2 jiffies.
 249          */
 250         scd->tick_jiffies = now_jiffies;
 251         scd->tick_raw = now;
 252         scd->tick_gtod = now_gtod;
 253         __raw_spin_unlock(&scd->lock);
 254 }
 255
 256 /*
 257  * We are going deep-idle (irqs are disabled):
 258  */
 259 void sched_clock_idle_sleep_event(void)
 260 {
 261         sched_clock_cpu(smp_processor_id());
 262 }
 263 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 264
 265 /*
 266  * We just idled delta nanoseconds (called with irqs disabled):
 267  */
 268 void sched_clock_idle_wakeup_event(u64 delta_ns)
 269 {
 270         struct sched_clock_data *scd = this_scd();
 271         u64 now = sched_clock();
 272
 273         /*
 274          * Override the previous timestamp and ignore all
 275          * sched_clock() deltas that occured while we idled,
 276          * and use the PM-provided delta_ns to advance the
 277          * rq clock:
 278          */
 279         __raw_spin_lock(&scd->lock);
 280         scd->prev_raw = now;
 281         scd->clock += delta_ns;
 282         __raw_spin_unlock(&scd->lock);
 283
 284         touch_softlockup_watchdog();
 285 }
 286 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 287
 288 #endif
 289
 290 /*
 291  * Scheduler clock - returns current time in nanosec units.
 292  * This is default implementation.
 293  * Architectures and sub-architectures can override this.
 294  */
 295 unsigned long long __attribute__((weak)) sched_clock(void)
 296 {
 297         return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
 298 }