kernel/sched/clock.c

   1 /*
   2  * sched_clock for unstable cpu clocks
   3  *
   4  *  Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra
   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 #include <linux/tick.h>
  65
  66 /*
  67  * Scheduler clock - returns current time in nanosec units.
  68  * This is default implementation.
  69  * Architectures and sub-architectures can override this.
  70  */
  71 unsigned long long __weak sched_clock(void)
  72 {
  73         return (unsigned long long)(jiffies - INITIAL_JIFFIES)
  74                                         * (NSEC_PER_SEC / HZ);
  75 }
  76 EXPORT_SYMBOL_GPL(sched_clock);
  77
  78 __read_mostly int sched_clock_running;
  79
  80 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  81 static struct static_key __sched_clock_stable = STATIC_KEY_INIT;
  82 static int __sched_clock_stable_early;
  83
  84 int sched_clock_stable(void)
  85 {
  86         return static_key_false(&__sched_clock_stable);
  87 }
  88
  89 static void __set_sched_clock_stable(void)
  90 {
  91         if (!sched_clock_stable())
  92                 static_key_slow_inc(&__sched_clock_stable);
  93
  94         tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
  95 }
  96
  97 void set_sched_clock_stable(void)
  98 {
  99         __sched_clock_stable_early = 1;
 100
 101         smp_mb(); /* matches sched_clock_init() */
 102
 103         if (!sched_clock_running)
 104                 return;
 105
 106         __set_sched_clock_stable();
 107 }
 108
 109 static void __clear_sched_clock_stable(struct work_struct *work)
 110 {
 111         /* XXX worry about clock continuity */
 112         if (sched_clock_stable())
 113                 static_key_slow_dec(&__sched_clock_stable);
 114
 115         tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
 116 }
 117
 118 static DECLARE_WORK(sched_clock_work, __clear_sched_clock_stable);
 119
 120 void clear_sched_clock_stable(void)
 121 {
 122         __sched_clock_stable_early = 0;
 123
 124         smp_mb(); /* matches sched_clock_init() */
 125
 126         if (!sched_clock_running)
 127                 return;
 128
 129         schedule_work(&sched_clock_work);
 130 }
 131
 132 struct sched_clock_data {
 133         u64                     tick_raw;
 134         u64                     tick_gtod;
 135         u64                     clock;
 136 };
 137
 138 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
 139
 140 static inline struct sched_clock_data *this_scd(void)
 141 {
 142         return this_cpu_ptr(&sched_clock_data);
 143 }
 144
 145 static inline struct sched_clock_data *cpu_sdc(int cpu)
 146 {
 147         return &per_cpu(sched_clock_data, cpu);
 148 }
 149
 150 void sched_clock_init(void)
 151 {
 152         u64 ktime_now = ktime_to_ns(ktime_get());
 153         int cpu;
 154
 155         for_each_possible_cpu(cpu) {
 156                 struct sched_clock_data *scd = cpu_sdc(cpu);
 157
 158                 scd->tick_raw = 0;
 159                 scd->tick_gtod = ktime_now;
 160                 scd->clock = ktime_now;
 161         }
 162
 163         sched_clock_running = 1;
 164
 165         /*
 166          * Ensure that it is impossible to not do a static_key update.
 167          *
 168          * Either {set,clear}_sched_clock_stable() must see sched_clock_running
 169          * and do the update, or we must see their __sched_clock_stable_early
 170          * and do the update, or both.
 171          */
 172         smp_mb(); /* matches {set,clear}_sched_clock_stable() */
 173
 174         if (__sched_clock_stable_early)
 175                 __set_sched_clock_stable();
 176         else
 177                 __clear_sched_clock_stable(NULL);
 178 }
 179
 180 /*
 181  * min, max except they take wrapping into account
 182  */
 183
 184 static inline u64 wrap_min(u64 x, u64 y)
 185 {
 186         return (s64)(x - y) < 0 ? x : y;
 187 }
 188
 189 static inline u64 wrap_max(u64 x, u64 y)
 190 {
 191         return (s64)(x - y) > 0 ? x : y;
 192 }
 193
 194 /*
 195  * update the percpu scd from the raw @now value
 196  *
 197  *  - filter out backward motion
 198  *  - use the GTOD tick value to create a window to filter crazy TSC values
 199  */
 200 static u64 sched_clock_local(struct sched_clock_data *scd)
 201 {
 202         u64 now, clock, old_clock, min_clock, max_clock;
 203         s64 delta;
 204
 205 again:
 206         now = sched_clock();
 207         delta = now - scd->tick_raw;
 208         if (unlikely(delta < 0))
 209                 delta = 0;
 210
 211         old_clock = scd->clock;
 212
 213         /*
 214          * scd->clock = clamp(scd->tick_gtod + delta,
 215          *                    max(scd->tick_gtod, scd->clock),
 216          *                    scd->tick_gtod + TICK_NSEC);
 217          */
 218
 219         clock = scd->tick_gtod + delta;
 220         min_clock = wrap_max(scd->tick_gtod, old_clock);
 221         max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
 222
 223         clock = wrap_max(clock, min_clock);
 224         clock = wrap_min(clock, max_clock);
 225
 226         if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
 227                 goto again;
 228
 229         return clock;
 230 }
 231
 232 static u64 sched_clock_remote(struct sched_clock_data *scd)
 233 {
 234         struct sched_clock_data *my_scd = this_scd();
 235         u64 this_clock, remote_clock;
 236         u64 *ptr, old_val, val;
 237
 238 #if BITS_PER_LONG != 64
 239 again:
 240         /*
 241          * Careful here: The local and the remote clock values need to
 242          * be read out atomic as we need to compare the values and
 243          * then update either the local or the remote side. So the
 244          * cmpxchg64 below only protects one readout.
 245          *
 246          * We must reread via sched_clock_local() in the retry case on
 247          * 32bit as an NMI could use sched_clock_local() via the
 248          * tracer and hit between the readout of
 249          * the low32bit and the high 32bit portion.
 250          */
 251         this_clock = sched_clock_local(my_scd);
 252         /*
 253          * We must enforce atomic readout on 32bit, otherwise the
 254          * update on the remote cpu can hit inbetween the readout of
 255          * the low32bit and the high 32bit portion.
 256          */
 257         remote_clock = cmpxchg64(&scd->clock, 0, 0);
 258 #else
 259         /*
 260          * On 64bit the read of [my]scd->clock is atomic versus the
 261          * update, so we can avoid the above 32bit dance.
 262          */
 263         sched_clock_local(my_scd);
 264 again:
 265         this_clock = my_scd->clock;
 266         remote_clock = scd->clock;
 267 #endif
 268
 269         /*
 270          * Use the opportunity that we have both locks
 271          * taken to couple the two clocks: we take the
 272          * larger time as the latest time for both
 273          * runqueues. (this creates monotonic movement)
 274          */
 275         if (likely((s64)(remote_clock - this_clock) < 0)) {
 276                 ptr = &scd->clock;
 277                 old_val = remote_clock;
 278                 val = this_clock;
 279         } else {
 280                 /*
 281                  * Should be rare, but possible:
 282                  */
 283                 ptr = &my_scd->clock;
 284                 old_val = this_clock;
 285                 val = remote_clock;
 286         }
 287
 288         if (cmpxchg64(ptr, old_val, val) != old_val)
 289                 goto again;
 290
 291         return val;
 292 }
 293
 294 /*
 295  * Similar to cpu_clock(), but requires local IRQs to be disabled.
 296  *
 297  * See cpu_clock().
 298  */
 299 u64 sched_clock_cpu(int cpu)
 300 {
 301         struct sched_clock_data *scd;
 302         u64 clock;
 303
 304         if (sched_clock_stable())
 305                 return sched_clock();
 306
 307         if (unlikely(!sched_clock_running))
 308                 return 0ull;
 309
 310         preempt_disable_notrace();
 311         scd = cpu_sdc(cpu);
 312
 313         if (cpu != smp_processor_id())
 314                 clock = sched_clock_remote(scd);
 315         else
 316                 clock = sched_clock_local(scd);
 317         preempt_enable_notrace();
 318
 319         return clock;
 320 }
 321 EXPORT_SYMBOL_GPL(sched_clock_cpu);
 322
 323 void sched_clock_tick(void)
 324 {
 325         struct sched_clock_data *scd;
 326         u64 now, now_gtod;
 327
 328         if (sched_clock_stable())
 329                 return;
 330
 331         if (unlikely(!sched_clock_running))
 332                 return;
 333
 334         WARN_ON_ONCE(!irqs_disabled());
 335
 336         scd = this_scd();
 337         now_gtod = ktime_to_ns(ktime_get());
 338         now = sched_clock();
 339
 340         scd->tick_raw = now;
 341         scd->tick_gtod = now_gtod;
 342         sched_clock_local(scd);
 343 }
 344
 345 /*
 346  * We are going deep-idle (irqs are disabled):
 347  */
 348 void sched_clock_idle_sleep_event(void)
 349 {
 350         sched_clock_cpu(smp_processor_id());
 351 }
 352 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 353
 354 /*
 355  * We just idled delta nanoseconds (called with irqs disabled):
 356  */
 357 void sched_clock_idle_wakeup_event(u64 delta_ns)
 358 {
 359         if (timekeeping_suspended)
 360                 return;
 361
 362         sched_clock_tick();
 363         touch_softlockup_watchdog_sched();
 364 }
 365 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 366
 367 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 368
 369 void sched_clock_init(void)
 370 {
 371         sched_clock_running = 1;
 372 }
 373
 374 u64 sched_clock_cpu(int cpu)
 375 {
 376         if (unlikely(!sched_clock_running))
 377                 return 0;
 378
 379         return sched_clock();
 380 }
 381 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 382
 383 /*
 384  * Running clock - returns the time that has elapsed while a guest has been
 385  * running.
 386  * On a guest this value should be local_clock minus the time the guest was
 387  * suspended by the hypervisor (for any reason).
 388  * On bare metal this function should return the same as local_clock.
 389  * Architectures and sub-architectures can override this.
 390  */
 391 u64 __weak running_clock(void)
 392 {
 393         return local_clock();
 394 }