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  * sched_clock_cpu(i) -- must be used with local IRQs disabled (implied by NMI)
  30  * local_clock()      -- is cpu_clock() on the current cpu.
  31  *
  32  * How:
  33  *
  34  * The implementation either uses sched_clock() when
  35  * !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK, which means in that case the
  36  * sched_clock() is assumed to provide these properties (mostly it means
  37  * the architecture provides a globally synchronized highres time source).
  38  *
  39  * Otherwise it tries to create a semi stable clock from a mixture of other
  40  * clocks, including:
  41  *
  42  *  - GTOD (clock monotomic)
  43  *  - sched_clock()
  44  *  - explicit idle events
  45  *
  46  * We use GTOD as base and use sched_clock() deltas to improve resolution. The
  47  * deltas are filtered to provide monotonicity and keeping it within an
  48  * expected window.
  49  *
  50  * Furthermore, explicit sleep and wakeup hooks allow us to account for time
  51  * that is otherwise invisible (TSC gets stopped).
  52  *
  53  *
  54  * Notes:
  55  *
  56  * The !IRQ-safetly of sched_clock() and sched_clock_cpu() comes from things
  57  * like cpufreq interrupts that can change the base clock (TSC) multiplier
  58  * and cause funny jumps in time -- although the filtering provided by
  59  * sched_clock_cpu() should mitigate serious artifacts we cannot rely on it
  60  * in general since for !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK we fully rely on
  61  * sched_clock().
  62  */
  63 #include <linux/spinlock.h>
  64 #include <linux/hardirq.h>
  65 #include <linux/export.h>
  66 #include <linux/percpu.h>
  67 #include <linux/ktime.h>
  68 #include <linux/sched.h>
  69
  70 /*
  71  * Scheduler clock - returns current time in nanosec units.
  72  * This is default implementation.
  73  * Architectures and sub-architectures can override this.
  74  */
  75 unsigned long long __attribute__((weak)) sched_clock(void)
  76 {
  77         return (unsigned long long)(jiffies - INITIAL_JIFFIES)
  78                                         * (NSEC_PER_SEC / HZ);
  79 }
  80 EXPORT_SYMBOL_GPL(sched_clock);
  81
  82 __read_mostly int sched_clock_running;
  83
  84 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  85 __read_mostly int sched_clock_stable;
  86
  87 struct sched_clock_data {
  88         u64                     tick_raw;
  89         u64                     tick_gtod;
  90         u64                     clock;
  91 };
  92
  93 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  94
  95 static inline struct sched_clock_data *this_scd(void)
  96 {
  97         return &__get_cpu_var(sched_clock_data);
  98 }
  99
 100 static inline struct sched_clock_data *cpu_sdc(int cpu)
 101 {
 102         return &per_cpu(sched_clock_data, cpu);
 103 }
 104
 105 void sched_clock_init(void)
 106 {
 107         u64 ktime_now = ktime_to_ns(ktime_get());
 108         int cpu;
 109
 110         for_each_possible_cpu(cpu) {
 111                 struct sched_clock_data *scd = cpu_sdc(cpu);
 112
 113                 scd->tick_raw = 0;
 114                 scd->tick_gtod = ktime_now;
 115                 scd->clock = ktime_now;
 116         }
 117
 118         sched_clock_running = 1;
 119 }
 120
 121 /*
 122  * min, max except they take wrapping into account
 123  */
 124
 125 static inline u64 wrap_min(u64 x, u64 y)
 126 {
 127         return (s64)(x - y) < 0 ? x : y;
 128 }
 129
 130 static inline u64 wrap_max(u64 x, u64 y)
 131 {
 132         return (s64)(x - y) > 0 ? x : y;
 133 }
 134
 135 /*
 136  * update the percpu scd from the raw @now value
 137  *
 138  *  - filter out backward motion
 139  *  - use the GTOD tick value to create a window to filter crazy TSC values
 140  */
 141 static u64 sched_clock_local(struct sched_clock_data *scd)
 142 {
 143         u64 now, clock, old_clock, min_clock, max_clock;
 144         s64 delta;
 145
 146 again:
 147         now = sched_clock();
 148         delta = now - scd->tick_raw;
 149         if (unlikely(delta < 0))
 150                 delta = 0;
 151
 152         old_clock = scd->clock;
 153
 154         /*
 155          * scd->clock = clamp(scd->tick_gtod + delta,
 156          *                    max(scd->tick_gtod, scd->clock),
 157          *                    scd->tick_gtod + TICK_NSEC);
 158          */
 159
 160         clock = scd->tick_gtod + delta;
 161         min_clock = wrap_max(scd->tick_gtod, old_clock);
 162         max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
 163
 164         clock = wrap_max(clock, min_clock);
 165         clock = wrap_min(clock, max_clock);
 166
 167         if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
 168                 goto again;
 169
 170         return clock;
 171 }
 172
 173 static u64 sched_clock_remote(struct sched_clock_data *scd)
 174 {
 175         struct sched_clock_data *my_scd = this_scd();
 176         u64 this_clock, remote_clock;
 177         u64 *ptr, old_val, val;
 178
 179 #if BITS_PER_LONG != 64
 180 again:
 181         /*
 182          * Careful here: The local and the remote clock values need to
 183          * be read out atomic as we need to compare the values and
 184          * then update either the local or the remote side. So the
 185          * cmpxchg64 below only protects one readout.
 186          *
 187          * We must reread via sched_clock_local() in the retry case on
 188          * 32bit as an NMI could use sched_clock_local() via the
 189          * tracer and hit between the readout of
 190          * the low32bit and the high 32bit portion.
 191          */
 192         this_clock = sched_clock_local(my_scd);
 193         /*
 194          * We must enforce atomic readout on 32bit, otherwise the
 195          * update on the remote cpu can hit inbetween the readout of
 196          * the low32bit and the high 32bit portion.
 197          */
 198         remote_clock = cmpxchg64(&scd->clock, 0, 0);
 199 #else
 200         /*
 201          * On 64bit the read of [my]scd->clock is atomic versus the
 202          * update, so we can avoid the above 32bit dance.
 203          */
 204         sched_clock_local(my_scd);
 205 again:
 206         this_clock = my_scd->clock;
 207         remote_clock = scd->clock;
 208 #endif
 209
 210         /*
 211          * Use the opportunity that we have both locks
 212          * taken to couple the two clocks: we take the
 213          * larger time as the latest time for both
 214          * runqueues. (this creates monotonic movement)
 215          */
 216         if (likely((s64)(remote_clock - this_clock) < 0)) {
 217                 ptr = &scd->clock;
 218                 old_val = remote_clock;
 219                 val = this_clock;
 220         } else {
 221                 /*
 222                  * Should be rare, but possible:
 223                  */
 224                 ptr = &my_scd->clock;
 225                 old_val = this_clock;
 226                 val = remote_clock;
 227         }
 228
 229         if (cmpxchg64(ptr, old_val, val) != old_val)
 230                 goto again;
 231
 232         return val;
 233 }
 234
 235 /*
 236  * Similar to cpu_clock(), but requires local IRQs to be disabled.
 237  *
 238  * See cpu_clock().
 239  */
 240 u64 sched_clock_cpu(int cpu)
 241 {
 242         struct sched_clock_data *scd;
 243         u64 clock;
 244
 245         WARN_ON_ONCE(!irqs_disabled());
 246
 247         if (sched_clock_stable)
 248                 return sched_clock();
 249
 250         if (unlikely(!sched_clock_running))
 251                 return 0ull;
 252
 253         scd = cpu_sdc(cpu);
 254
 255         if (cpu != smp_processor_id())
 256                 clock = sched_clock_remote(scd);
 257         else
 258                 clock = sched_clock_local(scd);
 259
 260         return clock;
 261 }
 262
 263 void sched_clock_tick(void)
 264 {
 265         struct sched_clock_data *scd;
 266         u64 now, now_gtod;
 267
 268         if (sched_clock_stable)
 269                 return;
 270
 271         if (unlikely(!sched_clock_running))
 272                 return;
 273
 274         WARN_ON_ONCE(!irqs_disabled());
 275
 276         scd = this_scd();
 277         now_gtod = ktime_to_ns(ktime_get());
 278         now = sched_clock();
 279
 280         scd->tick_raw = now;
 281         scd->tick_gtod = now_gtod;
 282         sched_clock_local(scd);
 283 }
 284
 285 /*
 286  * We are going deep-idle (irqs are disabled):
 287  */
 288 void sched_clock_idle_sleep_event(void)
 289 {
 290         sched_clock_cpu(smp_processor_id());
 291 }
 292 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 293
 294 /*
 295  * We just idled delta nanoseconds (called with irqs disabled):
 296  */
 297 void sched_clock_idle_wakeup_event(u64 delta_ns)
 298 {
 299         if (timekeeping_suspended)
 300                 return;
 301
 302         sched_clock_tick();
 303         touch_softlockup_watchdog();
 304 }
 305 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 306
 307 /*
 308  * As outlined at the top, provides a fast, high resolution, nanosecond
 309  * time source that is monotonic per cpu argument and has bounded drift
 310  * between cpus.
 311  *
 312  * ######################### BIG FAT WARNING ##########################
 313  * # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
 314  * # go backwards !!                                                  #
 315  * ####################################################################
 316  */
 317 u64 cpu_clock(int cpu)
 318 {
 319         u64 clock;
 320         unsigned long flags;
 321
 322         local_irq_save(flags);
 323         clock = sched_clock_cpu(cpu);
 324         local_irq_restore(flags);
 325
 326         return clock;
 327 }
 328
 329 /*
 330  * Similar to cpu_clock() for the current cpu. Time will only be observed
 331  * to be monotonic if care is taken to only compare timestampt taken on the
 332  * same CPU.
 333  *
 334  * See cpu_clock().
 335  */
 336 u64 local_clock(void)
 337 {
 338         u64 clock;
 339         unsigned long flags;
 340
 341         local_irq_save(flags);
 342         clock = sched_clock_cpu(smp_processor_id());
 343         local_irq_restore(flags);
 344
 345         return clock;
 346 }
 347
 348 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 349
 350 void sched_clock_init(void)
 351 {
 352         sched_clock_running = 1;
 353 }
 354
 355 u64 sched_clock_cpu(int cpu)
 356 {
 357         if (unlikely(!sched_clock_running))
 358                 return 0;
 359
 360         return sched_clock();
 361 }
 362
 363 u64 cpu_clock(int cpu)
 364 {
 365         return sched_clock_cpu(cpu);
 366 }
 367
 368 u64 local_clock(void)
 369 {
 370         return sched_clock_cpu(0);
 371 }
 372
 373 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 374
 375 EXPORT_SYMBOL_GPL(cpu_clock);
 376 EXPORT_SYMBOL_GPL(local_clock);