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
  68 /*
  69  * Scheduler clock - returns current time in nanosec units.
  70  * This is default implementation.
  71  * Architectures and sub-architectures can override this.
  72  */
  73 unsigned long long __weak sched_clock(void)
  74 {
  75         return (unsigned long long)(jiffies - INITIAL_JIFFIES)
  76                                         * (NSEC_PER_SEC / HZ);
  77 }
  78 EXPORT_SYMBOL_GPL(sched_clock);
  79
  80 __read_mostly int sched_clock_running;
  81
  82 void sched_clock_init(void)
  83 {
  84         sched_clock_running = 1;
  85 }
  86
  87 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  88 /*
  89  * We must start with !__sched_clock_stable because the unstable -> stable
  90  * transition is accurate, while the stable -> unstable transition is not.
  91  *
  92  * Similarly we start with __sched_clock_stable_early, thereby assuming we
  93  * will become stable, such that there's only a single 1 -> 0 transition.
  94  */
  95 static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
  96 static int __sched_clock_stable_early = 1;
  97
  98 /*
  99  * We want: ktime_get_ns() + __gtod_offset == sched_clock() + __sched_clock_offset
 100  */
 101 __read_mostly u64 __sched_clock_offset;
 102 static __read_mostly u64 __gtod_offset;
 103
 104 struct sched_clock_data {
 105         u64                     tick_raw;
 106         u64                     tick_gtod;
 107         u64                     clock;
 108 };
 109
 110 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
 111
 112 static inline struct sched_clock_data *this_scd(void)
 113 {
 114         return this_cpu_ptr(&sched_clock_data);
 115 }
 116
 117 static inline struct sched_clock_data *cpu_sdc(int cpu)
 118 {
 119         return &per_cpu(sched_clock_data, cpu);
 120 }
 121
 122 int sched_clock_stable(void)
 123 {
 124         return static_branch_likely(&__sched_clock_stable);
 125 }
 126
 127 static void __set_sched_clock_stable(void)
 128 {
 129         struct sched_clock_data *scd = this_scd();
 130
 131         /*
 132          * Attempt to make the (initial) unstable->stable transition continuous.
 133          */
 134         __sched_clock_offset = (scd->tick_gtod + __gtod_offset) - (scd->tick_raw);
 135
 136         printk(KERN_INFO "sched_clock: Marking stable (%lld, %lld)->(%lld, %lld)\n",
 137                         scd->tick_gtod, __gtod_offset,
 138                         scd->tick_raw,  __sched_clock_offset);
 139
 140         static_branch_enable(&__sched_clock_stable);
 141         tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
 142 }
 143
 144 static void __sched_clock_work(struct work_struct *work)
 145 {
 146         static_branch_disable(&__sched_clock_stable);
 147 }
 148
 149 static DECLARE_WORK(sched_clock_work, __sched_clock_work);
 150
 151 static void __clear_sched_clock_stable(void)
 152 {
 153         struct sched_clock_data *scd = this_scd();
 154
 155         /*
 156          * Attempt to make the stable->unstable transition continuous.
 157          *
 158          * Trouble is, this is typically called from the TSC watchdog
 159          * timer, which is late per definition. This means the tick
 160          * values can already be screwy.
 161          *
 162          * Still do what we can.
 163          */
 164         __gtod_offset = (scd->tick_raw + __sched_clock_offset) - (scd->tick_gtod);
 165
 166         printk(KERN_INFO "sched_clock: Marking unstable (%lld, %lld)<-(%lld, %lld)\n",
 167                         scd->tick_gtod, __gtod_offset,
 168                         scd->tick_raw,  __sched_clock_offset);
 169
 170         tick_dep_set(TICK_DEP_BIT_CLOCK_UNSTABLE);
 171
 172         if (sched_clock_stable())
 173                 schedule_work(&sched_clock_work);
 174 }
 175
 176 void clear_sched_clock_stable(void)
 177 {
 178         __sched_clock_stable_early = 0;
 179
 180         smp_mb(); /* matches sched_clock_init_late() */
 181
 182         if (sched_clock_running == 2)
 183                 __clear_sched_clock_stable();
 184 }
 185
 186 void sched_clock_init_late(void)
 187 {
 188         sched_clock_running = 2;
 189         /*
 190          * Ensure that it is impossible to not do a static_key update.
 191          *
 192          * Either {set,clear}_sched_clock_stable() must see sched_clock_running
 193          * and do the update, or we must see their __sched_clock_stable_early
 194          * and do the update, or both.
 195          */
 196         smp_mb(); /* matches {set,clear}_sched_clock_stable() */
 197
 198         if (__sched_clock_stable_early)
 199                 __set_sched_clock_stable();
 200 }
 201
 202 /*
 203  * min, max except they take wrapping into account
 204  */
 205
 206 static inline u64 wrap_min(u64 x, u64 y)
 207 {
 208         return (s64)(x - y) < 0 ? x : y;
 209 }
 210
 211 static inline u64 wrap_max(u64 x, u64 y)
 212 {
 213         return (s64)(x - y) > 0 ? x : y;
 214 }
 215
 216 /*
 217  * update the percpu scd from the raw @now value
 218  *
 219  *  - filter out backward motion
 220  *  - use the GTOD tick value to create a window to filter crazy TSC values
 221  */
 222 static u64 sched_clock_local(struct sched_clock_data *scd)
 223 {
 224         u64 now, clock, old_clock, min_clock, max_clock, gtod;
 225         s64 delta;
 226
 227 again:
 228         now = sched_clock();
 229         delta = now - scd->tick_raw;
 230         if (unlikely(delta < 0))
 231                 delta = 0;
 232
 233         old_clock = scd->clock;
 234
 235         /*
 236          * scd->clock = clamp(scd->tick_gtod + delta,
 237          *                    max(scd->tick_gtod, scd->clock),
 238          *                    scd->tick_gtod + TICK_NSEC);
 239          */
 240
 241         gtod = scd->tick_gtod + __gtod_offset;
 242         clock = gtod + delta;
 243         min_clock = wrap_max(gtod, old_clock);
 244         max_clock = wrap_max(old_clock, gtod + TICK_NSEC);
 245
 246         clock = wrap_max(clock, min_clock);
 247         clock = wrap_min(clock, max_clock);
 248
 249         if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
 250                 goto again;
 251
 252         return clock;
 253 }
 254
 255 static u64 sched_clock_remote(struct sched_clock_data *scd)
 256 {
 257         struct sched_clock_data *my_scd = this_scd();
 258         u64 this_clock, remote_clock;
 259         u64 *ptr, old_val, val;
 260
 261 #if BITS_PER_LONG != 64
 262 again:
 263         /*
 264          * Careful here: The local and the remote clock values need to
 265          * be read out atomic as we need to compare the values and
 266          * then update either the local or the remote side. So the
 267          * cmpxchg64 below only protects one readout.
 268          *
 269          * We must reread via sched_clock_local() in the retry case on
 270          * 32bit as an NMI could use sched_clock_local() via the
 271          * tracer and hit between the readout of
 272          * the low32bit and the high 32bit portion.
 273          */
 274         this_clock = sched_clock_local(my_scd);
 275         /*
 276          * We must enforce atomic readout on 32bit, otherwise the
 277          * update on the remote cpu can hit inbetween the readout of
 278          * the low32bit and the high 32bit portion.
 279          */
 280         remote_clock = cmpxchg64(&scd->clock, 0, 0);
 281 #else
 282         /*
 283          * On 64bit the read of [my]scd->clock is atomic versus the
 284          * update, so we can avoid the above 32bit dance.
 285          */
 286         sched_clock_local(my_scd);
 287 again:
 288         this_clock = my_scd->clock;
 289         remote_clock = scd->clock;
 290 #endif
 291
 292         /*
 293          * Use the opportunity that we have both locks
 294          * taken to couple the two clocks: we take the
 295          * larger time as the latest time for both
 296          * runqueues. (this creates monotonic movement)
 297          */
 298         if (likely((s64)(remote_clock - this_clock) < 0)) {
 299                 ptr = &scd->clock;
 300                 old_val = remote_clock;
 301                 val = this_clock;
 302         } else {
 303                 /*
 304                  * Should be rare, but possible:
 305                  */
 306                 ptr = &my_scd->clock;
 307                 old_val = this_clock;
 308                 val = remote_clock;
 309         }
 310
 311         if (cmpxchg64(ptr, old_val, val) != old_val)
 312                 goto again;
 313
 314         return val;
 315 }
 316
 317 /*
 318  * Similar to cpu_clock(), but requires local IRQs to be disabled.
 319  *
 320  * See cpu_clock().
 321  */
 322 u64 sched_clock_cpu(int cpu)
 323 {
 324         struct sched_clock_data *scd;
 325         u64 clock;
 326
 327         if (sched_clock_stable())
 328                 return sched_clock() + __sched_clock_offset;
 329
 330         if (unlikely(!sched_clock_running))
 331                 return 0ull;
 332
 333         preempt_disable_notrace();
 334         scd = cpu_sdc(cpu);
 335
 336         if (cpu != smp_processor_id())
 337                 clock = sched_clock_remote(scd);
 338         else
 339                 clock = sched_clock_local(scd);
 340         preempt_enable_notrace();
 341
 342         return clock;
 343 }
 344 EXPORT_SYMBOL_GPL(sched_clock_cpu);
 345
 346 void sched_clock_tick(void)
 347 {
 348         struct sched_clock_data *scd;
 349
 350         WARN_ON_ONCE(!irqs_disabled());
 351
 352         /*
 353          * Update these values even if sched_clock_stable(), because it can
 354          * become unstable at any point in time at which point we need some
 355          * values to fall back on.
 356          *
 357          * XXX arguably we can skip this if we expose tsc_clocksource_reliable
 358          */
 359         scd = this_scd();
 360         scd->tick_raw  = sched_clock();
 361         scd->tick_gtod = ktime_get_ns();
 362
 363         if (!sched_clock_stable() && likely(sched_clock_running))
 364                 sched_clock_local(scd);
 365 }
 366
 367 /*
 368  * We are going deep-idle (irqs are disabled):
 369  */
 370 void sched_clock_idle_sleep_event(void)
 371 {
 372         sched_clock_cpu(smp_processor_id());
 373 }
 374 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 375
 376 /*
 377  * We just idled delta nanoseconds (called with irqs disabled):
 378  */
 379 void sched_clock_idle_wakeup_event(u64 delta_ns)
 380 {
 381         if (timekeeping_suspended)
 382                 return;
 383
 384         sched_clock_tick();
 385         touch_softlockup_watchdog_sched();
 386 }
 387 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 388
 389 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 390
 391 u64 sched_clock_cpu(int cpu)
 392 {
 393         if (unlikely(!sched_clock_running))
 394                 return 0;
 395
 396         return sched_clock();
 397 }
 398
 399 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 400
 401 /*
 402  * Running clock - returns the time that has elapsed while a guest has been
 403  * running.
 404  * On a guest this value should be local_clock minus the time the guest was
 405  * suspended by the hypervisor (for any reason).
 406  * On bare metal this function should return the same as local_clock.
 407  * Architectures and sub-architectures can override this.
 408  */
 409 u64 __weak running_clock(void)
 410 {
 411         return local_clock();
 412 }