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 - INITIAL_JIFFIES)
  42                                         * (NSEC_PER_SEC / HZ);
  43 }
  44 EXPORT_SYMBOL_GPL(sched_clock);
  45
  46 static __read_mostly int sched_clock_running;
  47
  48 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  49 __read_mostly int sched_clock_stable;
  50
  51 struct sched_clock_data {
  52         u64                     tick_raw;
  53         u64                     tick_gtod;
  54         u64                     clock;
  55 };
  56
  57 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  58
  59 static inline struct sched_clock_data *this_scd(void)
  60 {
  61         return &__get_cpu_var(sched_clock_data);
  62 }
  63
  64 static inline struct sched_clock_data *cpu_sdc(int cpu)
  65 {
  66         return &per_cpu(sched_clock_data, cpu);
  67 }
  68
  69 void sched_clock_init(void)
  70 {
  71         u64 ktime_now = ktime_to_ns(ktime_get());
  72         int cpu;
  73
  74         for_each_possible_cpu(cpu) {
  75                 struct sched_clock_data *scd = cpu_sdc(cpu);
  76
  77                 scd->tick_raw = 0;
  78                 scd->tick_gtod = ktime_now;
  79                 scd->clock = ktime_now;
  80         }
  81
  82         sched_clock_running = 1;
  83 }
  84
  85 /*
  86  * min, max except they take wrapping into account
  87  */
  88
  89 static inline u64 wrap_min(u64 x, u64 y)
  90 {
  91         return (s64)(x - y) < 0 ? x : y;
  92 }
  93
  94 static inline u64 wrap_max(u64 x, u64 y)
  95 {
  96         return (s64)(x - y) > 0 ? x : y;
  97 }
  98
  99 /*
 100  * update the percpu scd from the raw @now value
 101  *
 102  *  - filter out backward motion
 103  *  - use the GTOD tick value to create a window to filter crazy TSC values
 104  */
 105 static u64 sched_clock_local(struct sched_clock_data *scd)
 106 {
 107         u64 now, clock, old_clock, min_clock, max_clock;
 108         s64 delta;
 109
 110 again:
 111         now = sched_clock();
 112         delta = now - scd->tick_raw;
 113         if (unlikely(delta < 0))
 114                 delta = 0;
 115
 116         old_clock = scd->clock;
 117
 118         /*
 119          * scd->clock = clamp(scd->tick_gtod + delta,
 120          *                    max(scd->tick_gtod, scd->clock),
 121          *                    scd->tick_gtod + TICK_NSEC);
 122          */
 123
 124         clock = scd->tick_gtod + delta;
 125         min_clock = wrap_max(scd->tick_gtod, old_clock);
 126         max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
 127
 128         clock = wrap_max(clock, min_clock);
 129         clock = wrap_min(clock, max_clock);
 130
 131         if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
 132                 goto again;
 133
 134         return clock;
 135 }
 136
 137 static u64 sched_clock_remote(struct sched_clock_data *scd)
 138 {
 139         struct sched_clock_data *my_scd = this_scd();
 140         u64 this_clock, remote_clock;
 141         u64 *ptr, old_val, val;
 142
 143         sched_clock_local(my_scd);
 144 again:
 145         this_clock = my_scd->clock;
 146         remote_clock = scd->clock;
 147
 148         /*
 149          * Use the opportunity that we have both locks
 150          * taken to couple the two clocks: we take the
 151          * larger time as the latest time for both
 152          * runqueues. (this creates monotonic movement)
 153          */
 154         if (likely((s64)(remote_clock - this_clock) < 0)) {
 155                 ptr = &scd->clock;
 156                 old_val = remote_clock;
 157                 val = this_clock;
 158         } else {
 159                 /*
 160                  * Should be rare, but possible:
 161                  */
 162                 ptr = &my_scd->clock;
 163                 old_val = this_clock;
 164                 val = remote_clock;
 165         }
 166
 167         if (cmpxchg64(ptr, old_val, val) != old_val)
 168                 goto again;
 169
 170         return val;
 171 }
 172
 173 u64 sched_clock_cpu(int cpu)
 174 {
 175         struct sched_clock_data *scd;
 176         u64 clock;
 177
 178         WARN_ON_ONCE(!irqs_disabled());
 179
 180         if (sched_clock_stable)
 181                 return sched_clock();
 182
 183         if (unlikely(!sched_clock_running))
 184                 return 0ull;
 185
 186         scd = cpu_sdc(cpu);
 187
 188         if (cpu != smp_processor_id())
 189                 clock = sched_clock_remote(scd);
 190         else
 191                 clock = sched_clock_local(scd);
 192
 193         return clock;
 194 }
 195
 196 void sched_clock_tick(void)
 197 {
 198         struct sched_clock_data *scd;
 199         u64 now, now_gtod;
 200
 201         if (sched_clock_stable)
 202                 return;
 203
 204         if (unlikely(!sched_clock_running))
 205                 return;
 206
 207         WARN_ON_ONCE(!irqs_disabled());
 208
 209         scd = this_scd();
 210         now_gtod = ktime_to_ns(ktime_get());
 211         now = sched_clock();
 212
 213         scd->tick_raw = now;
 214         scd->tick_gtod = now_gtod;
 215         sched_clock_local(scd);
 216 }
 217
 218 /*
 219  * We are going deep-idle (irqs are disabled):
 220  */
 221 void sched_clock_idle_sleep_event(void)
 222 {
 223         sched_clock_cpu(smp_processor_id());
 224 }
 225 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 226
 227 /*
 228  * We just idled delta nanoseconds (called with irqs disabled):
 229  */
 230 void sched_clock_idle_wakeup_event(u64 delta_ns)
 231 {
 232         if (timekeeping_suspended)
 233                 return;
 234
 235         sched_clock_tick();
 236         touch_softlockup_watchdog();
 237 }
 238 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 239
 240 unsigned long long cpu_clock(int cpu)
 241 {
 242         unsigned long long clock;
 243         unsigned long flags;
 244
 245         local_irq_save(flags);
 246         clock = sched_clock_cpu(cpu);
 247         local_irq_restore(flags);
 248
 249         return clock;
 250 }
 251
 252 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 253
 254 void sched_clock_init(void)
 255 {
 256         sched_clock_running = 1;
 257 }
 258
 259 u64 sched_clock_cpu(int cpu)
 260 {
 261         if (unlikely(!sched_clock_running))
 262                 return 0;
 263
 264         return sched_clock();
 265 }
 266
 267
 268 unsigned long long cpu_clock(int cpu)
 269 {
 270         return sched_clock_cpu(cpu);
 271 }
 272
 273 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 274
 275 EXPORT_SYMBOL_GPL(cpu_clock);