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/sched.h>
  28 #include <linux/percpu.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/ktime.h>
  31 #include <linux/module.h>
  32
  33 /*
  34  * Scheduler clock - returns current time in nanosec units.
  35  * This is default implementation.
  36  * Architectures and sub-architectures can override this.
  37  */
  38 unsigned long long __attribute__((weak)) sched_clock(void)
  39 {
  40         return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
  41 }
  42
  43 static __read_mostly int sched_clock_running;
  44
  45 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
  46
  47 struct sched_clock_data {
  48         /*
  49          * Raw spinlock - this is a special case: this might be called
  50          * from within instrumentation code so we dont want to do any
  51          * instrumentation ourselves.
  52          */
  53         raw_spinlock_t          lock;
  54
  55         u64                     tick_raw;
  56         u64                     tick_gtod;
  57         u64                     clock;
  58 };
  59
  60 static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
  61
  62 static inline struct sched_clock_data *this_scd(void)
  63 {
  64         return &__get_cpu_var(sched_clock_data);
  65 }
  66
  67 static inline struct sched_clock_data *cpu_sdc(int cpu)
  68 {
  69         return &per_cpu(sched_clock_data, cpu);
  70 }
  71
  72 void sched_clock_init(void)
  73 {
  74         u64 ktime_now = ktime_to_ns(ktime_get());
  75         int cpu;
  76
  77         for_each_possible_cpu(cpu) {
  78                 struct sched_clock_data *scd = cpu_sdc(cpu);
  79
  80                 scd->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
  81                 scd->tick_raw = 0;
  82                 scd->tick_gtod = ktime_now;
  83                 scd->clock = ktime_now;
  84         }
  85
  86         sched_clock_running = 1;
  87 }
  88
  89 /*
  90  * min,max except they take wrapping into account
  91  */
  92
  93 static inline u64 wrap_min(u64 x, u64 y)
  94 {
  95         return (s64)(x - y) < 0 ? x : y;
  96 }
  97
  98 static inline u64 wrap_max(u64 x, u64 y)
  99 {
 100         return (s64)(x - y) > 0 ? x : y;
 101 }
 102
 103 /*
 104  * update the percpu scd from the raw @now value
 105  *
 106  *  - filter out backward motion
 107  *  - use the GTOD tick value to create a window to filter crazy TSC values
 108  */
 109 static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
 110 {
 111         s64 delta = now - scd->tick_raw;
 112         u64 clock, min_clock, max_clock;
 113
 114         WARN_ON_ONCE(!irqs_disabled());
 115
 116         if (unlikely(delta < 0))
 117                 delta = 0;
 118
 119         /*
 120          * scd->clock = clamp(scd->tick_gtod + delta,
 121          *                    max(scd->tick_gtod, scd->clock),
 122          *                    scd->tick_gtod + TICK_NSEC);
 123          */
 124
 125         clock = scd->tick_gtod + delta;
 126         min_clock = wrap_max(scd->tick_gtod, scd->clock);
 127         max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC);
 128
 129         clock = wrap_max(clock, min_clock);
 130         clock = wrap_min(clock, max_clock);
 131
 132         scd->clock = clock;
 133
 134         return scd->clock;
 135 }
 136
 137 static void lock_double_clock(struct sched_clock_data *data1,
 138                                 struct sched_clock_data *data2)
 139 {
 140         if (data1 < data2) {
 141                 __raw_spin_lock(&data1->lock);
 142                 __raw_spin_lock(&data2->lock);
 143         } else {
 144                 __raw_spin_lock(&data2->lock);
 145                 __raw_spin_lock(&data1->lock);
 146         }
 147 }
 148
 149 u64 sched_clock_cpu(int cpu)
 150 {
 151         struct sched_clock_data *scd = cpu_sdc(cpu);
 152         u64 now, clock, this_clock, remote_clock;
 153
 154         if (unlikely(!sched_clock_running))
 155                 return 0ull;
 156
 157         WARN_ON_ONCE(!irqs_disabled());
 158         now = sched_clock();
 159
 160         if (cpu != raw_smp_processor_id()) {
 161                 struct sched_clock_data *my_scd = this_scd();
 162
 163                 lock_double_clock(scd, my_scd);
 164
 165                 this_clock = __update_sched_clock(my_scd, now);
 166                 remote_clock = scd->clock;
 167
 168                 /*
 169                  * Use the opportunity that we have both locks
 170                  * taken to couple the two clocks: we take the
 171                  * larger time as the latest time for both
 172                  * runqueues. (this creates monotonic movement)
 173                  */
 174                 if (likely((s64)(remote_clock - this_clock) < 0)) {
 175                         clock = this_clock;
 176                         scd->clock = clock;
 177                 } else {
 178                         /*
 179                          * Should be rare, but possible:
 180                          */
 181                         clock = remote_clock;
 182                         my_scd->clock = remote_clock;
 183                 }
 184
 185                 __raw_spin_unlock(&my_scd->lock);
 186         } else {
 187                 __raw_spin_lock(&scd->lock);
 188                 clock = __update_sched_clock(scd, now);
 189         }
 190
 191         __raw_spin_unlock(&scd->lock);
 192
 193         return clock;
 194 }
 195
 196 void sched_clock_tick(void)
 197 {
 198         struct sched_clock_data *scd = this_scd();
 199         u64 now, now_gtod;
 200
 201         if (unlikely(!sched_clock_running))
 202                 return;
 203
 204         WARN_ON_ONCE(!irqs_disabled());
 205
 206         now_gtod = ktime_to_ns(ktime_get());
 207         now = sched_clock();
 208
 209         __raw_spin_lock(&scd->lock);
 210         scd->tick_raw = now;
 211         scd->tick_gtod = now_gtod;
 212         __update_sched_clock(scd, now);
 213         __raw_spin_unlock(&scd->lock);
 214 }
 215
 216 /*
 217  * We are going deep-idle (irqs are disabled):
 218  */
 219 void sched_clock_idle_sleep_event(void)
 220 {
 221         sched_clock_cpu(smp_processor_id());
 222 }
 223 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 224
 225 /*
 226  * We just idled delta nanoseconds (called with irqs disabled):
 227  */
 228 void sched_clock_idle_wakeup_event(u64 delta_ns)
 229 {
 230         if (timekeeping_suspended)
 231                 return;
 232
 233         sched_clock_tick();
 234         touch_softlockup_watchdog();
 235 }
 236 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 237
 238 #else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 239
 240 void sched_clock_init(void)
 241 {
 242         sched_clock_running = 1;
 243 }
 244
 245 u64 sched_clock_cpu(int cpu)
 246 {
 247         if (unlikely(!sched_clock_running))
 248                 return 0;
 249
 250         return sched_clock();
 251 }
 252
 253 #endif
 254
 255 unsigned long long cpu_clock(int cpu)
 256 {
 257         unsigned long long clock;
 258         unsigned long flags;
 259
 260         local_irq_save(flags);
 261         clock = sched_clock_cpu(cpu);
 262         local_irq_restore(flags);
 263
 264         return clock;
 265 }
 266 EXPORT_SYMBOL_GPL(cpu_clock);