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
  45 static __read_mostly int sched_clock_running;
  46
  47 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  48 __read_mostly int sched_clock_stable;
  49
  50 struct sched_clock_data {
  51         /*
  52          * Raw spinlock - this is a special case: this might be called
  53          * from within instrumentation code so we dont want to do any
  54          * instrumentation ourselves.
  55          */
  56         raw_spinlock_t          lock;
  57
  58         u64                     tick_raw;
  59         u64                     tick_gtod;
  60         u64                     clock;
  61 };
  62
  63 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  64
  65 static inline struct sched_clock_data *this_scd(void)
  66 {
  67         return &__get_cpu_var(sched_clock_data);
  68 }
  69
  70 static inline struct sched_clock_data *cpu_sdc(int cpu)
  71 {
  72         return &per_cpu(sched_clock_data, cpu);
  73 }
  74
  75 void sched_clock_init(void)
  76 {
  77         u64 ktime_now = ktime_to_ns(ktime_get());
  78         int cpu;
  79
  80         for_each_possible_cpu(cpu) {
  81                 struct sched_clock_data *scd = cpu_sdc(cpu);
  82
  83                 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  84                 scd->tick_raw = 0;
  85                 scd->tick_gtod = ktime_now;
  86                 scd->clock = ktime_now;
  87         }
  88
  89         sched_clock_running = 1;
  90 }
  91
  92 /*
  93  * min, max except they take wrapping into account
  94  */
  95
  96 static inline u64 wrap_min(u64 x, u64 y)
  97 {
  98         return (s64)(x - y) < 0 ? x : y;
  99 }
 100
 101 static inline u64 wrap_max(u64 x, u64 y)
 102 {
 103         return (s64)(x - y) > 0 ? x : y;
 104 }
 105
 106 /*
 107  * update the percpu scd from the raw @now value
 108  *
 109  *  - filter out backward motion
 110  *  - use the GTOD tick value to create a window to filter crazy TSC values
 111  */
 112 static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
 113 {
 114         s64 delta = now - scd->tick_raw;
 115         u64 clock, min_clock, max_clock;
 116
 117         if (unlikely(delta < 0))
 118                 delta = 0;
 119
 120         /*
 121          * scd->clock = clamp(scd->tick_gtod + delta,
 122          *                    max(scd->tick_gtod, scd->clock),
 123          *                    scd->tick_gtod + TICK_NSEC);
 124          */
 125
 126         clock = scd->tick_gtod + delta;
 127         min_clock = wrap_max(scd->tick_gtod, scd->clock);
 128         max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC);
 129
 130         clock = wrap_max(clock, min_clock);
 131         clock = wrap_min(clock, max_clock);
 132
 133         scd->clock = clock;
 134
 135         return scd->clock;
 136 }
 137
 138 static void lock_double_clock(struct sched_clock_data *data1,
 139                                 struct sched_clock_data *data2)
 140 {
 141         if (data1 < data2) {
 142                 __raw_spin_lock(&data1->lock);
 143                 __raw_spin_lock(&data2->lock);
 144         } else {
 145                 __raw_spin_lock(&data2->lock);
 146                 __raw_spin_lock(&data1->lock);
 147         }
 148 }
 149
 150 u64 sched_clock_cpu(int cpu)
 151 {
 152         u64 now, clock, this_clock, remote_clock;
 153         struct sched_clock_data *scd;
 154
 155         if (sched_clock_stable)
 156                 return sched_clock();
 157
 158         scd = cpu_sdc(cpu);
 159
 160         /*
 161          * Normally this is not called in NMI context - but if it is,
 162          * trying to do any locking here is totally lethal.
 163          */
 164         if (unlikely(in_nmi()))
 165                 return scd->clock;
 166
 167         if (unlikely(!sched_clock_running))
 168                 return 0ull;
 169
 170         WARN_ON_ONCE(!irqs_disabled());
 171         now = sched_clock();
 172
 173         if (cpu != raw_smp_processor_id()) {
 174                 struct sched_clock_data *my_scd = this_scd();
 175
 176                 lock_double_clock(scd, my_scd);
 177
 178                 this_clock = __update_sched_clock(my_scd, now);
 179                 remote_clock = scd->clock;
 180
 181                 /*
 182                  * Use the opportunity that we have both locks
 183                  * taken to couple the two clocks: we take the
 184                  * larger time as the latest time for both
 185                  * runqueues. (this creates monotonic movement)
 186                  */
 187                 if (likely((s64)(remote_clock - this_clock) < 0)) {
 188                         clock = this_clock;
 189                         scd->clock = clock;
 190                 } else {
 191                         /*
 192                          * Should be rare, but possible:
 193                          */
 194                         clock = remote_clock;
 195                         my_scd->clock = remote_clock;
 196                 }
 197
 198                 __raw_spin_unlock(&my_scd->lock);
 199         } else {
 200                 __raw_spin_lock(&scd->lock);
 201                 clock = __update_sched_clock(scd, now);
 202         }
 203
 204         __raw_spin_unlock(&scd->lock);
 205
 206         return clock;
 207 }
 208
 209 void sched_clock_tick(void)
 210 {
 211         struct sched_clock_data *scd;
 212         u64 now, now_gtod;
 213
 214         if (sched_clock_stable)
 215                 return;
 216
 217         if (unlikely(!sched_clock_running))
 218                 return;
 219
 220         WARN_ON_ONCE(!irqs_disabled());
 221
 222         scd = this_scd();
 223         now_gtod = ktime_to_ns(ktime_get());
 224         now = sched_clock();
 225
 226         __raw_spin_lock(&scd->lock);
 227         scd->tick_raw = now;
 228         scd->tick_gtod = now_gtod;
 229         __update_sched_clock(scd, now);
 230         __raw_spin_unlock(&scd->lock);
 231 }
 232
 233 /*
 234  * We are going deep-idle (irqs are disabled):
 235  */
 236 void sched_clock_idle_sleep_event(void)
 237 {
 238         sched_clock_cpu(smp_processor_id());
 239 }
 240 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 241
 242 /*
 243  * We just idled delta nanoseconds (called with irqs disabled):
 244  */
 245 void sched_clock_idle_wakeup_event(u64 delta_ns)
 246 {
 247         if (timekeeping_suspended)
 248                 return;
 249
 250         sched_clock_tick();
 251         touch_softlockup_watchdog();
 252 }
 253 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 254
 255 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 256
 257 void sched_clock_init(void)
 258 {
 259         sched_clock_running = 1;
 260 }
 261
 262 u64 sched_clock_cpu(int cpu)
 263 {
 264         if (unlikely(!sched_clock_running))
 265                 return 0;
 266
 267         return sched_clock();
 268 }
 269
 270 #endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 271
 272 unsigned long long cpu_clock(int cpu)
 273 {
 274         unsigned long long clock;
 275         unsigned long flags;
 276
 277         local_irq_save(flags);
 278         clock = sched_clock_cpu(cpu);
 279         local_irq_restore(flags);
 280
 281         return clock;
 282 }
 283 EXPORT_SYMBOL_GPL(cpu_clock);