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/nmi.h>
  62 #include <linux/sched/clock.h>
  63 #include <linux/static_key.h>
  64 #include <linux/workqueue.h>
  65 #include <linux/compiler.h>
  66 #include <linux/tick.h>
  67 #include <linux/init.h>
  68
  69 /*
  70  * Scheduler clock - returns current time in nanosec units.
  71  * This is default implementation.
  72  * Architectures and sub-architectures can override this.
  73  */
  74 unsigned long long __weak sched_clock(void)
  75 {
  76         return (unsigned long long)(jiffies - INITIAL_JIFFIES)
  77                                         * (NSEC_PER_SEC / HZ);
  78 }
  79 EXPORT_SYMBOL_GPL(sched_clock);
  80
  81 __read_mostly int sched_clock_running;
  82
  83 void sched_clock_init(void)
  84 {
  85         sched_clock_running = 1;
  86 }
  87
  88 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  89 /*
  90  * We must start with !__sched_clock_stable because the unstable -> stable
  91  * transition is accurate, while the stable -> unstable transition is not.
  92  *
  93  * Similarly we start with __sched_clock_stable_early, thereby assuming we
  94  * will become stable, such that there's only a single 1 -> 0 transition.
  95  */
  96 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
  97 static int __sched_clock_stable_early = 1;
  98
  99 /*
 100  * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
 101  */
 102 __read_mostly u64 __sched_clock_offset;
 103 static __read_mostly u64 __gtod_offset;
 104
 105 struct sched_clock_data {
 106         u64                     tick_raw;
 107         u64                     tick_gtod;
 108         u64                     clock;
 109 };
 110
 111 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
 112
 113 static inline struct sched_clock_data *this_scd(void)
 114 {
 115         return this_cpu_ptr(&sched_clock_data);
 116 }
 117
 118 static inline struct sched_clock_data *cpu_sdc(int cpu)
 119 {
 120         return &per_cpu(sched_clock_data, cpu);
 121 }
 122
 123 int sched_clock_stable(void)
 124 {
 125         return static_branch_likely(&__sched_clock_stable);
 126 }
 127
 128 static void __scd_stamp(struct sched_clock_data *scd)
 129 {
 130         scd->tick_gtod = ktime_get_ns();
 131         scd->tick_raw = sched_clock();
 132 }
 133
 134 static void __set_sched_clock_stable(void)
 135 {
 136         struct sched_clock_data *scd;
 137
 138         /*
 139          * Since we're still unstable and the tick is already running, we have
 140          * to disable IRQs in order to get a consistent scd->tick* reading.
 141          */
 142         local_irq_disable();
 143         scd = this_scd();
 144         /*
 145          * Attempt to make the (initial) unstable->stable transition continuous.
 146          */
 147         __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
 148         local_irq_enable();
 149
 150         printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
 151                         scd->tick_gtod, __gtod_offset,
 152                         scd->tick_raw,  __sched_clock_offset);
 153
 154         static_branch_enable(&__sched_clock_stable);
 155         tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
 156 }
 157
 158 /*
 159  * If we ever get here, we're screwed, because we found out -- typically after
 160  * the fact -- that TSC wasn't good. This means all our clocksources (including
 161  * ktime) could have reported wrong values.
 162  *
 163  * What we do here is an attempt to fix up and continue sort of where we left
 164  * off in a coherent manner.
 165  *
 166  * The only way to fully avoid random clock jumps is to boot with:
 167  * "tsc=unstable".
 168  */
 169 static void __sched_clock_work(struct work_struct *work)
 170 {
 171         struct sched_clock_data *scd;
 172         int cpu;
 173
 174         /* take a current timestamp and set 'now' */
 175         preempt_disable();
 176         scd = this_scd();
 177         __scd_stamp(scd);
 178         scd->clock = scd->tick_gtod + __gtod_offset;
 179         preempt_enable();
 180
 181         /* clone to all CPUs */
 182         for_each_possible_cpu(cpu)
 183                 per_cpu(sched_clock_data, cpu) = *scd;
 184
 185         printk(KERN_WARNING "TSC found unstable after boot, most likely due to broken BIOS. Use 'tsc=unstable'.\n");
 186         printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
 187                         scd->tick_gtod, __gtod_offset,
 188                         scd->tick_raw,  __sched_clock_offset);
 189
 190         static_branch_disable(&__sched_clock_stable);
 191 }
 192
 193 static DECLARE_WORK(sched_clock_work, __sched_clock_work);
 194
 195 static void __clear_sched_clock_stable(void)
 196 {
 197         if (!sched_clock_stable())
 198                 return;
 199
 200         tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
 201         schedule_work(&sched_clock_work);
 202 }
 203
 204 void clear_sched_clock_stable(void)
 205 {
 206         __sched_clock_stable_early = 0;
 207
 208         smp_mb(); /* matches sched_clock_init_late() */
 209
 210         if (sched_clock_running == 2)
 211                 __clear_sched_clock_stable();
 212 }
 213
 214 /*
 215  * We run this as late_initcall() such that it runs after all built-in drivers,
 216  * notably: acpi_processor and intel_idle, which can mark the TSC as unstable.
 217  */
 218 static int __init sched_clock_init_late(void)
 219 {
 220         sched_clock_running = 2;
 221         /*
 222          * Ensure that it is impossible to not do a static_key update.
 223          *
 224          * Either {set,clear}_sched_clock_stable() must see sched_clock_running
 225          * and do the update, or we must see their __sched_clock_stable_early
 226          * and do the update, or both.
 227          */
 228         smp_mb(); /* matches {set,clear}_sched_clock_stable() */
 229
 230         if (__sched_clock_stable_early)
 231                 __set_sched_clock_stable();
 232
 233         return 0;
 234 }
 235 late_initcall(sched_clock_init_late);
 236
 237 /*
 238  * min, max except they take wrapping into account
 239  */
 240
 241 static inline u64 wrap_min(u64 x, u64 y)
 242 {
 243         return (s64)(x - y) < 0 ? x : y;
 244 }
 245
 246 static inline u64 wrap_max(u64 x, u64 y)
 247 {
 248         return (s64)(x - y) > 0 ? x : y;
 249 }
 250
 251 /*
 252  * update the percpu scd from the raw @now value
 253  *
 254  *  - filter out backward motion
 255  *  - use the GTOD tick value to create a window to filter crazy TSC values
 256  */
 257 static u64 sched_clock_local(struct sched_clock_data *scd)
 258 {
 259         u64 now, clock, old_clock, min_clock, max_clock, gtod;
 260         s64 delta;
 261
 262 again:
 263         now = sched_clock();
 264         delta = now - scd->tick_raw;
 265         if (unlikely(delta < 0))
 266                 delta = 0;
 267
 268         old_clock = scd->clock;
 269
 270         /*
 271          * scd->clock = clamp(scd->tick_gtod + delta,
 272          *                    max(scd->tick_gtod, scd->clock),
 273          *                    scd->tick_gtod + TICK_NSEC);
 274          */
 275
 276         gtod = scd->tick_gtod + __gtod_offset;
 277         clock = gtod + delta;
 278         min_clock = wrap_max(gtod, old_clock);
 279         max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
 280
 281         clock = wrap_max(clock, min_clock);
 282         clock = wrap_min(clock, max_clock);
 283
 284         if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
 285                 goto again;
 286
 287         return clock;
 288 }
 289
 290 static u64 sched_clock_remote(struct sched_clock_data *scd)
 291 {
 292         struct sched_clock_data *my_scd = this_scd();
 293         u64 this_clock, remote_clock;
 294         u64 *ptr, old_val, val;
 295
 296 #if BITS_PER_LONG != 64
 297 again:
 298         /*
 299          * Careful here: The local and the remote clock values need to
 300          * be read out atomic as we need to compare the values and
 301          * then update either the local or the remote side. So the
 302          * cmpxchg64 below only protects one readout.
 303          *
 304          * We must reread via sched_clock_local() in the retry case on
 305          * 32bit as an NMI could use sched_clock_local() via the
 306          * tracer and hit between the readout of
 307          * the low32bit and the high 32bit portion.
 308          */
 309         this_clock = sched_clock_local(my_scd);
 310         /*
 311          * We must enforce atomic readout on 32bit, otherwise the
 312          * update on the remote cpu can hit inbetween the readout of
 313          * the low32bit and the high 32bit portion.
 314          */
 315         remote_clock = cmpxchg64(&scd->clock, 0, 0);
 316 #else
 317         /*
 318          * On 64bit the read of [my]scd->clock is atomic versus the
 319          * update, so we can avoid the above 32bit dance.
 320          */
 321         sched_clock_local(my_scd);
 322 again:
 323         this_clock = my_scd->clock;
 324         remote_clock = scd->clock;
 325 #endif
 326
 327         /*
 328          * Use the opportunity that we have both locks
 329          * taken to couple the two clocks: we take the
 330          * larger time as the latest time for both
 331          * runqueues. (this creates monotonic movement)
 332          */
 333         if (likely((s64)(remote_clock - this_clock) < 0)) {
 334                 ptr = &scd->clock;
 335                 old_val = remote_clock;
 336                 val = this_clock;
 337         } else {
 338                 /*
 339                  * Should be rare, but possible:
 340                  */
 341                 ptr = &my_scd->clock;
 342                 old_val = this_clock;
 343                 val = remote_clock;
 344         }
 345
 346         if (cmpxchg64(ptr, old_val, val) != old_val)
 347                 goto again;
 348
 349         return val;
 350 }
 351
 352 /*
 353  * Similar to cpu_clock(), but requires local IRQs to be disabled.
 354  *
 355  * See cpu_clock().
 356  */
 357 u64 sched_clock_cpu(int cpu)
 358 {
 359         struct sched_clock_data *scd;
 360         u64 clock;
 361
 362         if (sched_clock_stable())
 363                 return sched_clock() + __sched_clock_offset;
 364
 365         if (unlikely(!sched_clock_running))
 366                 return 0ull;
 367
 368         preempt_disable_notrace();
 369         scd = cpu_sdc(cpu);
 370
 371         if (cpu != smp_processor_id())
 372                 clock = sched_clock_remote(scd);
 373         else
 374                 clock = sched_clock_local(scd);
 375         preempt_enable_notrace();
 376
 377         return clock;
 378 }
 379 EXPORT_SYMBOL_GPL(sched_clock_cpu);
 380
 381 void sched_clock_tick(void)
 382 {
 383         struct sched_clock_data *scd;
 384
 385         if (sched_clock_stable())
 386                 return;
 387
 388         if (unlikely(!sched_clock_running))
 389                 return;
 390
 391         WARN_ON_ONCE(!irqs_disabled());
 392
 393         scd = this_scd();
 394         __scd_stamp(scd);
 395         sched_clock_local(scd);
 396 }
 397
 398 void sched_clock_tick_stable(void)
 399 {
 400         u64 gtod, clock;
 401
 402         if (!sched_clock_stable())
 403                 return;
 404
 405         /*
 406          * Called under watchdog_lock.
 407          *
 408          * The watchdog just found this TSC to (still) be stable, so now is a
 409          * good moment to update our __gtod_offset. Because once we find the
 410          * TSC to be unstable, any computation will be computing crap.
 411          */
 412         local_irq_disable();
 413         gtod = ktime_get_ns();
 414         clock = sched_clock();
 415         __gtod_offset = (clock + __sched_clock_offset) - gtod;
 416         local_irq_enable();
 417 }
 418
 419 /*
 420  * We are going deep-idle (irqs are disabled):
 421  */
 422 void sched_clock_idle_sleep_event(void)
 423 {
 424         sched_clock_cpu(smp_processor_id());
 425 }
 426 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 427
 428 /*
 429  * We just idled; resync with ktime.
 430  */
 431 void sched_clock_idle_wakeup_event(void)
 432 {
 433         unsigned long flags;
 434
 435         if (sched_clock_stable())
 436                 return;
 437
 438         if (unlikely(timekeeping_suspended))
 439                 return;
 440
 441         local_irq_save(flags);
 442         sched_clock_tick();
 443         local_irq_restore(flags);
 444 }
 445 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 446
 447 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 448
 449 u64 sched_clock_cpu(int cpu)
 450 {
 451         if (unlikely(!sched_clock_running))
 452                 return 0;
 453
 454         return sched_clock();
 455 }
 456
 457 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 458
 459 /*
 460  * Running clock - returns the time that has elapsed while a guest has been
 461  * running.
 462  * On a guest this value should be local_clock minus the time the guest was
 463  * suspended by the hypervisor (for any reason).
 464  * On bare metal this function should return the same as local_clock.
 465  * Architectures and sub-architectures can override this.
 466  */
 467 u64 __weak running_clock(void)
 468 {
 469         return local_clock();
 470 }