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  * Create a semi stable clock from a mixture of other events, including:
  14  *  - gtod
  15  *  - sched_clock()
  16  *  - explicit idle events
  17  *
  18  * We use gtod as base and the unstable clock deltas. The deltas are filtered,
  19  * making it monotonic and keeping it within an expected window.
  20  *
  21  * Furthermore, explicit sleep and wakeup hooks allow us to account for time
  22  * that is otherwise invisible (TSC gets stopped).
  23  *
  24  * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat
  25  * consistent between cpus (never more than 2 jiffies difference).
  26  */
  27 #include <linux/spinlock.h>
  28 #include <linux/hardirq.h>
  29 #include <linux/module.h>
  30 #include <linux/percpu.h>
  31 #include <linux/ktime.h>
  32 #include <linux/sched.h>
  33
  34 /*
  35  * Scheduler clock - returns current time in nanosec units.
  36  * This is default implementation.
  37  * Architectures and sub-architectures can override this.
  38  */
  39 unsigned long long __attribute__((weak)) sched_clock(void)
  40 {
  41         return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
  42 }
  43
  44 static __read_mostly int sched_clock_running;
  45
  46 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  47 __read_mostly int sched_clock_stable;
  48 #else
  49 static const int sched_clock_stable = 1;
  50 #endif
  51
  52 struct sched_clock_data {
  53         /*
  54          * Raw spinlock - this is a special case: this might be called
  55          * from within instrumentation code so we dont want to do any
  56          * instrumentation ourselves.
  57          */
  58         raw_spinlock_t          lock;
  59
  60         u64                     tick_raw;
  61         u64                     tick_gtod;
  62         u64                     clock;
  63 };
  64
  65 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  66
  67 static inline struct sched_clock_data *this_scd(void)
  68 {
  69         return &__get_cpu_var(sched_clock_data);
  70 }
  71
  72 static inline struct sched_clock_data *cpu_sdc(int cpu)
  73 {
  74         return &per_cpu(sched_clock_data, cpu);
  75 }
  76
  77 void sched_clock_init(void)
  78 {
  79         u64 ktime_now = ktime_to_ns(ktime_get());
  80         int cpu;
  81
  82         for_each_possible_cpu(cpu) {
  83                 struct sched_clock_data *scd = cpu_sdc(cpu);
  84
  85                 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  86                 scd->tick_raw = 0;
  87                 scd->tick_gtod = ktime_now;
  88                 scd->clock = ktime_now;
  89         }
  90
  91         sched_clock_running = 1;
  92 }
  93
  94 /*
  95  * min, max except they take wrapping into account
  96  */
  97
  98 static inline u64 wrap_min(u64 x, u64 y)
  99 {
 100         return (s64)(x - y) < 0 ? x : y;
 101 }
 102
 103 static inline u64 wrap_max(u64 x, u64 y)
 104 {
 105         return (s64)(x - y) > 0 ? x : y;
 106 }
 107
 108 /*
 109  * update the percpu scd from the raw @now value
 110  *
 111  *  - filter out backward motion
 112  *  - use the GTOD tick value to create a window to filter crazy TSC values
 113  */
 114 static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
 115 {
 116         s64 delta = now - scd->tick_raw;
 117         u64 clock, min_clock, max_clock;
 118
 119         WARN_ON_ONCE(!irqs_disabled());
 120
 121         if (unlikely(delta < 0))
 122                 delta = 0;
 123
 124         if (unlikely(!sched_clock_running))
 125                 return 0ull;
 126
 127         /*
 128          * scd->clock = clamp(scd->tick_gtod + delta,
 129          *                    max(scd->tick_gtod, scd->clock),
 130          *                    scd->tick_gtod + TICK_NSEC);
 131          */
 132
 133         clock = scd->tick_gtod + delta;
 134         min_clock = wrap_max(scd->tick_gtod, scd->clock);
 135         max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC);
 136
 137         clock = wrap_max(clock, min_clock);
 138         clock = wrap_min(clock, max_clock);
 139
 140         scd->clock = clock;
 141
 142         return scd->clock;
 143 }
 144
 145 static void lock_double_clock(struct sched_clock_data *data1,
 146                                 struct sched_clock_data *data2)
 147 {
 148         if (data1 < data2) {
 149                 __raw_spin_lock(&data1->lock);
 150                 __raw_spin_lock(&data2->lock);
 151         } else {
 152                 __raw_spin_lock(&data2->lock);
 153                 __raw_spin_lock(&data1->lock);
 154         }
 155 }
 156
 157 u64 sched_clock_cpu(int cpu)
 158 {
 159         u64 now, clock, this_clock, remote_clock;
 160         struct sched_clock_data *scd;
 161
 162         if (sched_clock_stable)
 163                 return sched_clock();
 164
 165         scd = cpu_sdc(cpu);
 166
 167         /*
 168          * Normally this is not called in NMI context - but if it is,
 169          * trying to do any locking here is totally lethal.
 170          */
 171         if (unlikely(in_nmi()))
 172                 return scd->clock;
 173
 174         if (unlikely(!sched_clock_running))
 175                 return 0ull;
 176
 177         WARN_ON_ONCE(!irqs_disabled());
 178         now = sched_clock();
 179
 180         if (cpu != raw_smp_processor_id()) {
 181                 struct sched_clock_data *my_scd = this_scd();
 182
 183                 lock_double_clock(scd, my_scd);
 184
 185                 this_clock = __update_sched_clock(my_scd, now);
 186                 remote_clock = scd->clock;
 187
 188                 /*
 189                  * Use the opportunity that we have both locks
 190                  * taken to couple the two clocks: we take the
 191                  * larger time as the latest time for both
 192                  * runqueues. (this creates monotonic movement)
 193                  */
 194                 if (likely((s64)(remote_clock - this_clock) < 0)) {
 195                         clock = this_clock;
 196                         scd->clock = clock;
 197                 } else {
 198                         /*
 199                          * Should be rare, but possible:
 200                          */
 201                         clock = remote_clock;
 202                         my_scd->clock = remote_clock;
 203                 }
 204
 205                 __raw_spin_unlock(&my_scd->lock);
 206         } else {
 207                 __raw_spin_lock(&scd->lock);
 208                 clock = __update_sched_clock(scd, now);
 209         }
 210
 211         __raw_spin_unlock(&scd->lock);
 212
 213         return clock;
 214 }
 215
 216 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
 217
 218 void sched_clock_tick(void)
 219 {
 220         struct sched_clock_data *scd = this_scd();
 221         u64 now, now_gtod;
 222
 223         if (unlikely(!sched_clock_running))
 224                 return;
 225
 226         WARN_ON_ONCE(!irqs_disabled());
 227
 228         now_gtod = ktime_to_ns(ktime_get());
 229         now = sched_clock();
 230
 231         __raw_spin_lock(&scd->lock);
 232         scd->tick_raw = now;
 233         scd->tick_gtod = now_gtod;
 234         __update_sched_clock(scd, now);
 235         __raw_spin_unlock(&scd->lock);
 236 }
 237
 238 /*
 239  * We are going deep-idle (irqs are disabled):
 240  */
 241 void sched_clock_idle_sleep_event(void)
 242 {
 243         sched_clock_cpu(smp_processor_id());
 244 }
 245 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 246
 247 /*
 248  * We just idled delta nanoseconds (called with irqs disabled):
 249  */
 250 void sched_clock_idle_wakeup_event(u64 delta_ns)
 251 {
 252         if (timekeeping_suspended)
 253                 return;
 254
 255         sched_clock_tick();
 256         touch_softlockup_watchdog();
 257 }
 258 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 259
 260 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 261
 262 unsigned long long cpu_clock(int cpu)
 263 {
 264         unsigned long long clock;
 265         unsigned long flags;
 266
 267         local_irq_save(flags);
 268         clock = sched_clock_cpu(cpu);
 269         local_irq_restore(flags);
 270
 271         return clock;
 272 }
 273 EXPORT_SYMBOL_GPL(cpu_clock);