kernel/time/tick-common.c

   1 /*
   2  * linux/kernel/time/tick-common.c
   3  *
   4  * This file contains the base functions to manage periodic tick
   5  * related events.
   6  *
   7  * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
   8  * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   9  * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
  10  *
  11  * This code is licenced under the GPL version 2. For details see
  12  * kernel-base/COPYING.
  13  */
  14 #include <linux/cpu.h>
  15 #include <linux/err.h>
  16 #include <linux/hrtimer.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/percpu.h>
  19 #include <linux/profile.h>
  20 #include <linux/sched.h>
  21 #include <linux/module.h>
  22
  23 #include <asm/irq_regs.h>
  24
  25 #include "tick-internal.h"
  26
  27 /*
  28  * Tick devices
  29  */
  30 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
  31 /*
  32  * Tick next event: keeps track of the tick time
  33  */
  34 ktime_t tick_next_period;
  35 ktime_t tick_period;
  36
  37 /*
  38  * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
  39  * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
  40  * variable has two functions:
  41  *
  42  * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
  43  *    timekeeping lock all at once. Only the CPU which is assigned to do the
  44  *    update is handling it.
  45  *
  46  * 2) Hand off the duty in the NOHZ idle case by setting the value to
  47  *    TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
  48  *    at it will take over and keep the time keeping alive.  The handover
  49  *    procedure also covers cpu hotplug.
  50  */
  51 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
  52
  53 /*
  54  * Debugging: see timer_list.c
  55  */
  56 struct tick_device *tick_get_device(int cpu)
  57 {
  58         return &per_cpu(tick_cpu_device, cpu);
  59 }
  60
  61 /**
  62  * tick_is_oneshot_available - check for a oneshot capable event device
  63  */
  64 int tick_is_oneshot_available(void)
  65 {
  66         struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
  67
  68         if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
  69                 return 0;
  70         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
  71                 return 1;
  72         return tick_broadcast_oneshot_available();
  73 }
  74
  75 /*
  76  * Periodic tick
  77  */
  78 static void tick_periodic(int cpu)
  79 {
  80         if (tick_do_timer_cpu == cpu) {
  81                 write_seqlock(&jiffies_lock);
  82
  83                 /* Keep track of the next tick event */
  84                 tick_next_period = ktime_add(tick_next_period, tick_period);
  85
  86                 do_timer(1);
  87                 write_sequnlock(&jiffies_lock);
  88                 update_wall_time();
  89         }
  90
  91         update_process_times(user_mode(get_irq_regs()));
  92         profile_tick(CPU_PROFILING);
  93 }
  94
  95 /*
  96  * Event handler for periodic ticks
  97  */
  98 void tick_handle_periodic(struct clock_event_device *dev)
  99 {
 100         int cpu = smp_processor_id();
 101         ktime_t next = dev->next_event;
 102
 103         tick_periodic(cpu);
 104
 105         if (dev->mode != CLOCK_EVT_MODE_ONESHOT)
 106                 return;
 107         for (;;) {
 108                 /*
 109                  * Setup the next period for devices, which do not have
 110                  * periodic mode:
 111                  */
 112                 next = ktime_add(next, tick_period);
 113
 114                 if (!clockevents_program_event(dev, next, false))
 115                         return;
 116                 /*
 117                  * Have to be careful here. If we're in oneshot mode,
 118                  * before we call tick_periodic() in a loop, we need
 119                  * to be sure we're using a real hardware clocksource.
 120                  * Otherwise we could get trapped in an infinite
 121                  * loop, as the tick_periodic() increments jiffies,
 122                  * which then will increment time, possibly causing
 123                  * the loop to trigger again and again.
 124                  */
 125                 if (timekeeping_valid_for_hres())
 126                         tick_periodic(cpu);
 127         }
 128 }
 129
 130 /*
 131  * Setup the device for a periodic tick
 132  */
 133 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
 134 {
 135         tick_set_periodic_handler(dev, broadcast);
 136
 137         /* Broadcast setup ? */
 138         if (!tick_device_is_functional(dev))
 139                 return;
 140
 141         if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
 142             !tick_broadcast_oneshot_active()) {
 143                 clockevents_set_mode(dev, CLOCK_EVT_MODE_PERIODIC);
 144         } else {
 145                 unsigned long seq;
 146                 ktime_t next;
 147
 148                 do {
 149                         seq = read_seqbegin(&jiffies_lock);
 150                         next = tick_next_period;
 151                 } while (read_seqretry(&jiffies_lock, seq));
 152
 153                 clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 154
 155                 for (;;) {
 156                         if (!clockevents_program_event(dev, next, false))
 157                                 return;
 158                         next = ktime_add(next, tick_period);
 159                 }
 160         }
 161 }
 162
 163 /*
 164  * Setup the tick device
 165  */
 166 static void tick_setup_device(struct tick_device *td,
 167                               struct clock_event_device *newdev, int cpu,
 168                               const struct cpumask *cpumask)
 169 {
 170         ktime_t next_event;
 171         void (*handler)(struct clock_event_device *) = NULL;
 172
 173         /*
 174          * First device setup ?
 175          */
 176         if (!td->evtdev) {
 177                 /*
 178                  * If no cpu took the do_timer update, assign it to
 179                  * this cpu:
 180                  */
 181                 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
 182                         if (!tick_nohz_full_cpu(cpu))
 183                                 tick_do_timer_cpu = cpu;
 184                         else
 185                                 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 186                         tick_next_period = ktime_get();
 187                         tick_period = ktime_set(0, NSEC_PER_SEC / HZ);
 188                 }
 189
 190                 /*
 191                  * Startup in periodic mode first.
 192                  */
 193                 td->mode = TICKDEV_MODE_PERIODIC;
 194         } else {
 195                 handler = td->evtdev->event_handler;
 196                 next_event = td->evtdev->next_event;
 197                 td->evtdev->event_handler = clockevents_handle_noop;
 198         }
 199
 200         td->evtdev = newdev;
 201
 202         /*
 203          * When the device is not per cpu, pin the interrupt to the
 204          * current cpu:
 205          */
 206         if (!cpumask_equal(newdev->cpumask, cpumask))
 207                 irq_set_affinity(newdev->irq, cpumask);
 208
 209         /*
 210          * When global broadcasting is active, check if the current
 211          * device is registered as a placeholder for broadcast mode.
 212          * This allows us to handle this x86 misfeature in a generic
 213          * way. This function also returns !=0 when we keep the
 214          * current active broadcast state for this CPU.
 215          */
 216         if (tick_device_uses_broadcast(newdev, cpu))
 217                 return;
 218
 219         if (td->mode == TICKDEV_MODE_PERIODIC)
 220                 tick_setup_periodic(newdev, 0);
 221         else
 222                 tick_setup_oneshot(newdev, handler, next_event);
 223 }
 224
 225 void tick_install_replacement(struct clock_event_device *newdev)
 226 {
 227         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 228         int cpu = smp_processor_id();
 229
 230         clockevents_exchange_device(td->evtdev, newdev);
 231         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 232         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 233                 tick_oneshot_notify();
 234 }
 235
 236 static bool tick_check_percpu(struct clock_event_device *curdev,
 237                               struct clock_event_device *newdev, int cpu)
 238 {
 239         if (!cpumask_test_cpu(cpu, newdev->cpumask))
 240                 return false;
 241         if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
 242                 return true;
 243         /* Check if irq affinity can be set */
 244         if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
 245                 return false;
 246         /* Prefer an existing cpu local device */
 247         if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
 248                 return false;
 249         return true;
 250 }
 251
 252 static bool tick_check_preferred(struct clock_event_device *curdev,
 253                                  struct clock_event_device *newdev)
 254 {
 255         /* Prefer oneshot capable device */
 256         if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
 257                 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
 258                         return false;
 259                 if (tick_oneshot_mode_active())
 260                         return false;
 261         }
 262
 263         /*
 264          * Use the higher rated one, but prefer a CPU local device with a lower
 265          * rating than a non-CPU local device
 266          */
 267         return !curdev ||
 268                 newdev->rating > curdev->rating ||
 269                !cpumask_equal(curdev->cpumask, newdev->cpumask);
 270 }
 271
 272 /*
 273  * Check whether the new device is a better fit than curdev. curdev
 274  * can be NULL !
 275  */
 276 bool tick_check_replacement(struct clock_event_device *curdev,
 277                             struct clock_event_device *newdev)
 278 {
 279         if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
 280                 return false;
 281
 282         return tick_check_preferred(curdev, newdev);
 283 }
 284
 285 /*
 286  * Check, if the new registered device should be used. Called with
 287  * clockevents_lock held and interrupts disabled.
 288  */
 289 void tick_check_new_device(struct clock_event_device *newdev)
 290 {
 291         struct clock_event_device *curdev;
 292         struct tick_device *td;
 293         int cpu;
 294
 295         cpu = smp_processor_id();
 296         if (!cpumask_test_cpu(cpu, newdev->cpumask))
 297                 goto out_bc;
 298
 299         td = &per_cpu(tick_cpu_device, cpu);
 300         curdev = td->evtdev;
 301
 302         /* cpu local device ? */
 303         if (!tick_check_percpu(curdev, newdev, cpu))
 304                 goto out_bc;
 305
 306         /* Preference decision */
 307         if (!tick_check_preferred(curdev, newdev))
 308                 goto out_bc;
 309
 310         if (!try_module_get(newdev->owner))
 311                 return;
 312
 313         /*
 314          * Replace the eventually existing device by the new
 315          * device. If the current device is the broadcast device, do
 316          * not give it back to the clockevents layer !
 317          */
 318         if (tick_is_broadcast_device(curdev)) {
 319                 clockevents_shutdown(curdev);
 320                 curdev = NULL;
 321         }
 322         clockevents_exchange_device(curdev, newdev);
 323         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 324         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 325                 tick_oneshot_notify();
 326         return;
 327
 328 out_bc:
 329         /*
 330          * Can the new device be used as a broadcast device ?
 331          */
 332         tick_install_broadcast_device(newdev);
 333 }
 334
 335 /*
 336  * Transfer the do_timer job away from a dying cpu.
 337  *
 338  * Called with interrupts disabled.
 339  */
 340 void tick_handover_do_timer(int *cpup)
 341 {
 342         if (*cpup == tick_do_timer_cpu) {
 343                 int cpu = cpumask_first(cpu_online_mask);
 344
 345                 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
 346                         TICK_DO_TIMER_NONE;
 347         }
 348 }
 349
 350 /*
 351  * Shutdown an event device on a given cpu:
 352  *
 353  * This is called on a life CPU, when a CPU is dead. So we cannot
 354  * access the hardware device itself.
 355  * We just set the mode and remove it from the lists.
 356  */
 357 void tick_shutdown(unsigned int *cpup)
 358 {
 359         struct tick_device *td = &per_cpu(tick_cpu_device, *cpup);
 360         struct clock_event_device *dev = td->evtdev;
 361
 362         td->mode = TICKDEV_MODE_PERIODIC;
 363         if (dev) {
 364                 /*
 365                  * Prevent that the clock events layer tries to call
 366                  * the set mode function!
 367                  */
 368                 dev->mode = CLOCK_EVT_MODE_UNUSED;
 369                 clockevents_exchange_device(dev, NULL);
 370                 dev->event_handler = clockevents_handle_noop;
 371                 td->evtdev = NULL;
 372         }
 373 }
 374
 375 void tick_suspend(void)
 376 {
 377         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 378
 379         clockevents_shutdown(td->evtdev);
 380 }
 381
 382 void tick_resume(void)
 383 {
 384         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 385         int broadcast = tick_resume_broadcast();
 386
 387         clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_RESUME);
 388
 389         if (!broadcast) {
 390                 if (td->mode == TICKDEV_MODE_PERIODIC)
 391                         tick_setup_periodic(td->evtdev, 0);
 392                 else
 393                         tick_resume_oneshot();
 394         }
 395 }
 396
 397 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
 398 static unsigned int tick_freeze_depth;
 399
 400 /**
 401  * tick_freeze - Suspend the local tick and (possibly) timekeeping.
 402  *
 403  * Check if this is the last online CPU executing the function and if so,
 404  * suspend timekeeping.  Otherwise suspend the local tick.
 405  *
 406  * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
 407  * Interrupts must not be enabled before the subsequent %tick_unfreeze().
 408  */
 409 void tick_freeze(void)
 410 {
 411         raw_spin_lock(&tick_freeze_lock);
 412
 413         tick_freeze_depth++;
 414         if (tick_freeze_depth == num_online_cpus()) {
 415                 timekeeping_suspend();
 416         } else {
 417                 tick_suspend();
 418                 tick_suspend_broadcast();
 419         }
 420
 421         raw_spin_unlock(&tick_freeze_lock);
 422 }
 423
 424 /**
 425  * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
 426  *
 427  * Check if this is the first CPU executing the function and if so, resume
 428  * timekeeping.  Otherwise resume the local tick.
 429  *
 430  * Call with interrupts disabled.  Must be balanced with %tick_freeze().
 431  * Interrupts must not be enabled after the preceding %tick_freeze().
 432  */
 433 void tick_unfreeze(void)
 434 {
 435         raw_spin_lock(&tick_freeze_lock);
 436
 437         if (tick_freeze_depth == num_online_cpus())
 438                 timekeeping_resume();
 439         else
 440                 tick_resume();
 441
 442         tick_freeze_depth--;
 443
 444         raw_spin_unlock(&tick_freeze_lock);
 445 }
 446
 447 /**
 448  * tick_init - initialize the tick control
 449  */
 450 void __init tick_init(void)
 451 {
 452         tick_broadcast_init();
 453         tick_nohz_init();
 454 }