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 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  37
  38 #define MULTI_SHIFT 15
  39 /* Max is double, Min is 1/2 */
  40 #define MAX_MULTI (2LL << MULTI_SHIFT)
  41 #define MIN_MULTI (1LL << (MULTI_SHIFT-1))
  42
  43 struct sched_clock_data {
  44         /*
  45          * Raw spinlock - this is a special case: this might be called
  46          * from within instrumentation code so we dont want to do any
  47          * instrumentation ourselves.
  48          */
  49         raw_spinlock_t          lock;
  50
  51         unsigned long           tick_jiffies;
  52         u64                     prev_raw;
  53         u64                     tick_raw;
  54         u64                     tick_gtod;
  55         u64                     clock;
  56         s64                     multi;
  57 #ifdef CONFIG_NO_HZ
  58         int                     check_max;
  59 #endif
  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                 scd->multi = 1 << MULTI_SHIFT;
  92 #ifdef CONFIG_NO_HZ
  93                 scd->check_max = 1;
  94 #endif
  95         }
  96
  97         sched_clock_running = 1;
  98 }
  99
 100 #ifdef CONFIG_NO_HZ
 101 /*
 102  * The dynamic ticks makes the delta jiffies inaccurate. This
 103  * prevents us from checking the maximum time update.
 104  * Disable the maximum check during stopped ticks.
 105  */
 106 void sched_clock_tick_stop(int cpu)
 107 {
 108         struct sched_clock_data *scd = cpu_sdc(cpu);
 109
 110         scd->check_max = 0;
 111 }
 112
 113 void sched_clock_tick_start(int cpu)
 114 {
 115         struct sched_clock_data *scd = cpu_sdc(cpu);
 116
 117         scd->check_max = 1;
 118 }
 119
 120 static int check_max(struct sched_clock_data *scd)
 121 {
 122         return scd->check_max;
 123 }
 124 #else
 125 static int check_max(struct sched_clock_data *scd)
 126 {
 127         return 1;
 128 }
 129 #endif /* CONFIG_NO_HZ */
 130
 131 /*
 132  * update the percpu scd from the raw @now value
 133  *
 134  *  - filter out backward motion
 135  *  - use jiffies to generate a min,max window to clip the raw values
 136  */
 137 static void __update_sched_clock(struct sched_clock_data *scd, u64 now, u64 *time)
 138 {
 139         unsigned long now_jiffies = jiffies;
 140         long delta_jiffies = now_jiffies - scd->tick_jiffies;
 141         u64 clock = scd->clock;
 142         u64 min_clock, max_clock;
 143         s64 delta = now - scd->prev_raw;
 144
 145         WARN_ON_ONCE(!irqs_disabled());
 146
 147         /*
 148          * At schedule tick the clock can be just under the gtod. We don't
 149          * want to push it too prematurely.
 150          */
 151         min_clock = scd->tick_gtod + (delta_jiffies * TICK_NSEC);
 152         if (min_clock > TICK_NSEC)
 153                 min_clock -= TICK_NSEC / 2;
 154
 155         if (unlikely(delta < 0)) {
 156                 clock++;
 157                 goto out;
 158         }
 159
 160         /*
 161          * The clock must stay within a jiffie of the gtod.
 162          * But since we may be at the start of a jiffy or the end of one
 163          * we add another jiffy buffer.
 164          */
 165         max_clock = scd->tick_gtod + (2 + delta_jiffies) * TICK_NSEC;
 166
 167         delta *= scd->multi;
 168         delta >>= MULTI_SHIFT;
 169
 170         if (unlikely(clock + delta > max_clock) && check_max(scd)) {
 171                 if (clock < max_clock)
 172                         clock = max_clock;
 173                 else
 174                         clock++;
 175         } else {
 176                 clock += delta;
 177         }
 178
 179  out:
 180         if (unlikely(clock < min_clock))
 181                 clock = min_clock;
 182
 183         if (time)
 184                 *time = clock;
 185         else {
 186                 scd->prev_raw = now;
 187                 scd->clock = clock;
 188         }
 189 }
 190
 191 static void lock_double_clock(struct sched_clock_data *data1,
 192                                 struct sched_clock_data *data2)
 193 {
 194         if (data1 < data2) {
 195                 __raw_spin_lock(&data1->lock);
 196                 __raw_spin_lock(&data2->lock);
 197         } else {
 198                 __raw_spin_lock(&data2->lock);
 199                 __raw_spin_lock(&data1->lock);
 200         }
 201 }
 202
 203 u64 sched_clock_cpu(int cpu)
 204 {
 205         struct sched_clock_data *scd = cpu_sdc(cpu);
 206         u64 now, clock;
 207
 208         if (unlikely(!sched_clock_running))
 209                 return 0ull;
 210
 211         WARN_ON_ONCE(!irqs_disabled());
 212         now = sched_clock();
 213
 214         if (cpu != raw_smp_processor_id()) {
 215                 /*
 216                  * in order to update a remote cpu's clock based on our
 217                  * unstable raw time rebase it against:
 218                  *   tick_raw           (offset between raw counters)
 219                  *   tick_gotd          (tick offset between cpus)
 220                  */
 221                 struct sched_clock_data *my_scd = this_scd();
 222
 223                 lock_double_clock(scd, my_scd);
 224
 225                 now -= my_scd->tick_raw;
 226                 now += scd->tick_raw;
 227
 228                 now += my_scd->tick_gtod;
 229                 now -= scd->tick_gtod;
 230
 231                 __raw_spin_unlock(&my_scd->lock);
 232
 233                 __update_sched_clock(scd, now, &clock);
 234
 235                 __raw_spin_unlock(&scd->lock);
 236
 237         } else {
 238                 __raw_spin_lock(&scd->lock);
 239                 __update_sched_clock(scd, now, NULL);
 240                 clock = scd->clock;
 241                 __raw_spin_unlock(&scd->lock);
 242         }
 243
 244         return clock;
 245 }
 246
 247 void sched_clock_tick(void)
 248 {
 249         struct sched_clock_data *scd = this_scd();
 250         unsigned long now_jiffies = jiffies;
 251         s64 mult, delta_gtod, delta_raw;
 252         u64 now, now_gtod;
 253
 254         if (unlikely(!sched_clock_running))
 255                 return;
 256
 257         WARN_ON_ONCE(!irqs_disabled());
 258
 259         now_gtod = ktime_to_ns(ktime_get());
 260         now = sched_clock();
 261
 262         __raw_spin_lock(&scd->lock);
 263         __update_sched_clock(scd, now, NULL);
 264         /*
 265          * update tick_gtod after __update_sched_clock() because that will
 266          * already observe 1 new jiffy; adding a new tick_gtod to that would
 267          * increase the clock 2 jiffies.
 268          */
 269         delta_gtod = now_gtod - scd->tick_gtod;
 270         delta_raw = now - scd->tick_raw;
 271
 272         if ((long)delta_raw > 0) {
 273                 mult = delta_gtod << MULTI_SHIFT;
 274                 do_div(mult, delta_raw);
 275                 scd->multi = mult;
 276                 if (scd->multi > MAX_MULTI)
 277                         scd->multi = MAX_MULTI;
 278                 else if (scd->multi < MIN_MULTI)
 279                         scd->multi = MIN_MULTI;
 280         } else
 281                 scd->multi = 1 << MULTI_SHIFT;
 282
 283         scd->tick_raw = now;
 284         scd->tick_gtod = now_gtod;
 285         scd->tick_jiffies = now_jiffies;
 286         __raw_spin_unlock(&scd->lock);
 287 }
 288
 289 /*
 290  * We are going deep-idle (irqs are disabled):
 291  */
 292 void sched_clock_idle_sleep_event(void)
 293 {
 294         sched_clock_cpu(smp_processor_id());
 295 }
 296 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 297
 298 /*
 299  * We just idled delta nanoseconds (called with irqs disabled):
 300  */
 301 void sched_clock_idle_wakeup_event(u64 delta_ns)
 302 {
 303         struct sched_clock_data *scd = this_scd();
 304         u64 now = sched_clock();
 305
 306         /*
 307          * Override the previous timestamp and ignore all
 308          * sched_clock() deltas that occured while we idled,
 309          * and use the PM-provided delta_ns to advance the
 310          * rq clock:
 311          */
 312         __raw_spin_lock(&scd->lock);
 313         scd->prev_raw = now;
 314         scd->clock += delta_ns;
 315         scd->multi = 1 << MULTI_SHIFT;
 316         __raw_spin_unlock(&scd->lock);
 317
 318         touch_softlockup_watchdog();
 319 }
 320 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 321
 322 #endif
 323
 324 /*
 325  * Scheduler clock - returns current time in nanosec units.
 326  * This is default implementation.
 327  * Architectures and sub-architectures can override this.
 328  */
 329 unsigned long long __attribute__((weak)) sched_clock(void)
 330 {
 331         return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
 332 }
 333
 334 unsigned long long cpu_clock(int cpu)
 335 {
 336         unsigned long long clock;
 337         unsigned long flags;
 338
 339         local_irq_save(flags);
 340         clock = sched_clock_cpu(cpu);
 341         local_irq_restore(flags);
 342
 343         return clock;
 344 }
 345 EXPORT_SYMBOL_GPL(cpu_clock);