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