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  *
  14  * What:
  15  *
  16  * cpu_clock(i) provides a fast (execution time) high resolution
  17  * clock with bounded drift between CPUs. The value of cpu_clock(i)
  18  * is monotonic for constant i. The timestamp returned is in nanoseconds.
  19  *
  20  * ######################### BIG FAT WARNING ##########################
  21  * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
  22  * # go backwards !!                                                  #
  23  * ####################################################################
  24  *
  25  * There is no strict promise about the base, although it tends to start
  26  * at 0 on boot (but people really shouldn't rely on that).
  27  *
  28  * cpu_clock(i)       -- can be used from any context, including NMI.
  29  * local_clock()      -- is cpu_clock() on the current cpu.
  30  *
  31  * sched_clock_cpu(i)
  32  *
  33  * How:
  34  *
  35  * The implementation either uses sched_clock() when
  36  * !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK, which means in that case the
  37  * sched_clock() is assumed to provide these properties (mostly it means
  38  * the architecture provides a globally synchronized highres time source).
  39  *
  40  * Otherwise it tries to create a semi stable clock from a mixture of other
  41  * clocks, including:
  42  *
  43  *  - GTOD (clock monotomic)
  44  *  - sched_clock()
  45  *  - explicit idle events
  46  *
  47  * We use GTOD as base and use sched_clock() deltas to improve resolution. The
  48  * deltas are filtered to provide monotonicity and keeping it within an
  49  * expected window.
  50  *
  51  * Furthermore, explicit sleep and wakeup hooks allow us to account for time
  52  * that is otherwise invisible (TSC gets stopped).
  53  *
  54  */
  55 #include <linux/spinlock.h>
  56 #include <linux/hardirq.h>
  57 #include <linux/export.h>
  58 #include <linux/percpu.h>
  59 #include <linux/ktime.h>
  60 #include <linux/sched.h>
  61 #include <linux/static_key.h>
  62 #include <linux/workqueue.h>
  63 #include <linux/compiler.h>
  64
  65 /*
  66  * Scheduler clock - returns current time in nanosec units.
  67  * This is default implementation.
  68  * Architectures and sub-architectures can override this.
  69  */
  70 unsigned long long __weak sched_clock(void)
  71 {
  72         return (unsigned long long)(jiffies - INITIAL_JIFFIES)
  73                                         * (NSEC_PER_SEC / HZ);
  74 }
  75 EXPORT_SYMBOL_GPL(sched_clock);
  76
  77 __read_mostly int sched_clock_running;
  78
  79 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  80 static struct static_key __sched_clock_stable = STATIC_KEY_INIT;
  81 static int __sched_clock_stable_early;
  82
  83 int sched_clock_stable(void)
  84 {
  85         return static_key_false(&__sched_clock_stable);
  86 }
  87
  88 static void __set_sched_clock_stable(void)
  89 {
  90         if (!sched_clock_stable())
  91                 static_key_slow_inc(&__sched_clock_stable);
  92 }
  93
  94 void set_sched_clock_stable(void)
  95 {
  96         __sched_clock_stable_early = 1;
  97
  98         smp_mb(); /* matches sched_clock_init() */
  99
 100         if (!sched_clock_running)
 101                 return;
 102
 103         __set_sched_clock_stable();
 104 }
 105
 106 static void __clear_sched_clock_stable(struct work_struct *work)
 107 {
 108         /* XXX worry about clock continuity */
 109         if (sched_clock_stable())
 110                 static_key_slow_dec(&__sched_clock_stable);
 111 }
 112
 113 static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable);
 114
 115 void clear_sched_clock_stable(void)
 116 {
 117         __sched_clock_stable_early = 0;
 118
 119         smp_mb(); /* matches sched_clock_init() */
 120
 121         if (!sched_clock_running)
 122                 return;
 123
 124         schedule_work(&sched_clock_work);
 125 }
 126
 127 struct sched_clock_data {
 128         u64                     tick_raw;
 129         u64                     tick_gtod;
 130         u64                     clock;
 131 };
 132
 133 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
 134
 135 static inline struct sched_clock_data *this_scd(void)
 136 {
 137         return this_cpu_ptr(&sched_clock_data);
 138 }
 139
 140 static inline struct sched_clock_data *cpu_sdc(int cpu)
 141 {
 142         return &per_cpu(sched_clock_data, cpu);
 143 }
 144
 145 void sched_clock_init(void)
 146 {
 147         u64 ktime_now = ktime_to_ns(ktime_get());
 148         int cpu;
 149
 150         for_each_possible_cpu(cpu) {
 151                 struct sched_clock_data *scd = cpu_sdc(cpu);
 152
 153                 scd->tick_raw = 0;
 154                 scd->tick_gtod = ktime_now;
 155                 scd->clock = ktime_now;
 156         }
 157
 158         sched_clock_running = 1;
 159
 160         /*
 161          * Ensure that it is impossible to not do a static_key update.
 162          *
 163          * Either {set,clear}_sched_clock_stable() must see sched_clock_running
 164          * and do the update, or we must see their __sched_clock_stable_early
 165          * and do the update, or both.
 166          */
 167         smp_mb(); /* matches {set,clear}_sched_clock_stable() */
 168
 169         if (__sched_clock_stable_early)
 170                 __set_sched_clock_stable();
 171         else
 172                 __clear_sched_clock_stable(NULL);
 173 }
 174
 175 /*
 176  * min, max except they take wrapping into account
 177  */
 178
 179 static inline u64 wrap_min(u64 x, u64 y)
 180 {
 181         return (s64)(x - y) < 0 ? x : y;
 182 }
 183
 184 static inline u64 wrap_max(u64 x, u64 y)
 185 {
 186         return (s64)(x - y) > 0 ? x : y;
 187 }
 188
 189 /*
 190  * update the percpu scd from the raw @now value
 191  *
 192  *  - filter out backward motion
 193  *  - use the GTOD tick value to create a window to filter crazy TSC values
 194  */
 195 static u64 sched_clock_local(struct sched_clock_data *scd)
 196 {
 197         u64 now, clock, old_clock, min_clock, max_clock;
 198         s64 delta;
 199
 200 again:
 201         now = sched_clock();
 202         delta = now - scd->tick_raw;
 203         if (unlikely(delta < 0))
 204                 delta = 0;
 205
 206         old_clock = scd->clock;
 207
 208         /*
 209          * scd->clock = clamp(scd->tick_gtod + delta,
 210          *                    max(scd->tick_gtod, scd->clock),
 211          *                    scd->tick_gtod + TICK_NSEC);
 212          */
 213
 214         clock = scd->tick_gtod + delta;
 215         min_clock = wrap_max(scd->tick_gtod, old_clock);
 216         max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
 217
 218         clock = wrap_max(clock, min_clock);
 219         clock = wrap_min(clock, max_clock);
 220
 221         if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
 222                 goto again;
 223
 224         return clock;
 225 }
 226
 227 static u64 sched_clock_remote(struct sched_clock_data *scd)
 228 {
 229         struct sched_clock_data *my_scd = this_scd();
 230         u64 this_clock, remote_clock;
 231         u64 *ptr, old_val, val;
 232
 233 #if BITS_PER_LONG != 64
 234 again:
 235         /*
 236          * Careful here: The local and the remote clock values need to
 237          * be read out atomic as we need to compare the values and
 238          * then update either the local or the remote side. So the
 239          * cmpxchg64 below only protects one readout.
 240          *
 241          * We must reread via sched_clock_local() in the retry case on
 242          * 32bit as an NMI could use sched_clock_local() via the
 243          * tracer and hit between the readout of
 244          * the low32bit and the high 32bit portion.
 245          */
 246         this_clock = sched_clock_local(my_scd);
 247         /*
 248          * We must enforce atomic readout on 32bit, otherwise the
 249          * update on the remote cpu can hit inbetween the readout of
 250          * the low32bit and the high 32bit portion.
 251          */
 252         remote_clock = cmpxchg64(&scd->clock, 0, 0);
 253 #else
 254         /*
 255          * On 64bit the read of [my]scd->clock is atomic versus the
 256          * update, so we can avoid the above 32bit dance.
 257          */
 258         sched_clock_local(my_scd);
 259 again:
 260         this_clock = my_scd->clock;
 261         remote_clock = scd->clock;
 262 #endif
 263
 264         /*
 265          * Use the opportunity that we have both locks
 266          * taken to couple the two clocks: we take the
 267          * larger time as the latest time for both
 268          * runqueues. (this creates monotonic movement)
 269          */
 270         if (likely((s64)(remote_clock - this_clock) < 0)) {
 271                 ptr = &scd->clock;
 272                 old_val = remote_clock;
 273                 val = this_clock;
 274         } else {
 275                 /*
 276                  * Should be rare, but possible:
 277                  */
 278                 ptr = &my_scd->clock;
 279                 old_val = this_clock;
 280                 val = remote_clock;
 281         }
 282
 283         if (cmpxchg64(ptr, old_val, val) != old_val)
 284                 goto again;
 285
 286         return val;
 287 }
 288
 289 /*
 290  * Similar to cpu_clock(), but requires local IRQs to be disabled.
 291  *
 292  * See cpu_clock().
 293  */
 294 u64 sched_clock_cpu(int cpu)
 295 {
 296         struct sched_clock_data *scd;
 297         u64 clock;
 298
 299         if (sched_clock_stable())
 300                 return sched_clock();
 301
 302         if (unlikely(!sched_clock_running))
 303                 return 0ull;
 304
 305         preempt_disable_notrace();
 306         scd = cpu_sdc(cpu);
 307
 308         if (cpu != smp_processor_id())
 309                 clock = sched_clock_remote(scd);
 310         else
 311                 clock = sched_clock_local(scd);
 312         preempt_enable_notrace();
 313
 314         return clock;
 315 }
 316
 317 void sched_clock_tick(void)
 318 {
 319         struct sched_clock_data *scd;
 320         u64 now, now_gtod;
 321
 322         if (sched_clock_stable())
 323                 return;
 324
 325         if (unlikely(!sched_clock_running))
 326                 return;
 327
 328         WARN_ON_ONCE(!irqs_disabled());
 329
 330         scd = this_scd();
 331         now_gtod = ktime_to_ns(ktime_get());
 332         now = sched_clock();
 333
 334         scd->tick_raw = now;
 335         scd->tick_gtod = now_gtod;
 336         sched_clock_local(scd);
 337 }
 338
 339 /*
 340  * We are going deep-idle (irqs are disabled):
 341  */
 342 void sched_clock_idle_sleep_event(void)
 343 {
 344         sched_clock_cpu(smp_processor_id());
 345 }
 346 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 347
 348 /*
 349  * We just idled delta nanoseconds (called with irqs disabled):
 350  */
 351 void sched_clock_idle_wakeup_event(u64 delta_ns)
 352 {
 353         if (timekeeping_suspended)
 354                 return;
 355
 356         sched_clock_tick();
 357         touch_softlockup_watchdog();
 358 }
 359 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 360
 361 /*
 362  * As outlined at the top, provides a fast, high resolution, nanosecond
 363  * time source that is monotonic per cpu argument and has bounded drift
 364  * between cpus.
 365  *
 366  * ######################### BIG FAT WARNING ##########################
 367  * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
 368  * # go backwards !!                                                  #
 369  * ####################################################################
 370  */
 371 u64 cpu_clock(int cpu)
 372 {
 373         if (!sched_clock_stable())
 374                 return sched_clock_cpu(cpu);
 375
 376         return sched_clock();
 377 }
 378
 379 /*
 380  * Similar to cpu_clock() for the current cpu. Time will only be observed
 381  * to be monotonic if care is taken to only compare timestampt taken on the
 382  * same CPU.
 383  *
 384  * See cpu_clock().
 385  */
 386 u64 local_clock(void)
 387 {
 388         if (!sched_clock_stable())
 389                 return sched_clock_cpu(raw_smp_processor_id());
 390
 391         return sched_clock();
 392 }
 393
 394 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 395
 396 void sched_clock_init(void)
 397 {
 398         sched_clock_running = 1;
 399 }
 400
 401 u64 sched_clock_cpu(int cpu)
 402 {
 403         if (unlikely(!sched_clock_running))
 404                 return 0;
 405
 406         return sched_clock();
 407 }
 408
 409 u64 cpu_clock(int cpu)
 410 {
 411         return sched_clock();
 412 }
 413
 414 u64 local_clock(void)
 415 {
 416         return sched_clock();
 417 }
 418
 419 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 420
 421 EXPORT_SYMBOL_GPL(cpu_clock);
 422 EXPORT_SYMBOL_GPL(local_clock);
 423
 424 /*
 425  * Running clock - returns the time that has elapsed while a guest has been
 426  * running.
 427  * On a guest this value should be local_clock minus the time the guest was
 428  * suspended by the hypervisor (for any reason).
 429  * On bare metal this function should return the same as local_clock.
 430  * Architectures and sub-architectures can override this.
 431  */
 432 u64 __weak running_clock(void)
 433 {
 434         return local_clock();
 435 }