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 #include <trace/events/power.h>
  23
  24 #include <asm/irq_regs.h>
  25
  26 #include "tick-internal.h"
  27
  28 /*
  29  * Tick devices
  30  */
  31 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
  32 /*
  33  * Tick next event: keeps track of the tick time
  34  */
  35 ktime_t tick_next_period;
  36 ktime_t tick_period;
  37
  38 /*
  39  * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
  40  * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
  41  * variable has two functions:
  42  *
  43  * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
  44  *    timekeeping lock all at once. Only the CPU which is assigned to do the
  45  *    update is handling it.
  46  *
  47  * 2) Hand off the duty in the NOHZ idle case by setting the value to
  48  *    TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
  49  *    at it will take over and keep the time keeping alive.  The handover
  50  *    procedure also covers cpu hotplug.
  51  */
  52 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
  53
  54 /*
  55  * Debugging: see timer_list.c
  56  */
  57 struct tick_device *tick_get_device(int cpu)
  58 {
  59         return &per_cpu(tick_cpu_device, cpu);
  60 }
  61
  62 /**
  63  * tick_is_oneshot_available - check for a oneshot capable event device
  64  */
  65 int tick_is_oneshot_available(void)
  66 {
  67         struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
  68
  69         if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
  70                 return 0;
  71         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
  72                 return 1;
  73         return tick_broadcast_oneshot_available();
  74 }
  75
  76 /*
  77  * Periodic tick
  78  */
  79 static void tick_periodic(int cpu)
  80 {
  81         if (tick_do_timer_cpu == cpu) {
  82                 write_seqlock(&jiffies_lock);
  83
  84                 /* Keep track of the next tick event */
  85                 tick_next_period = ktime_add(tick_next_period, tick_period);
  86
  87                 do_timer(1);
  88                 write_sequnlock(&jiffies_lock);
  89                 update_wall_time();
  90         }
  91
  92         update_process_times(user_mode(get_irq_regs()));
  93         profile_tick(CPU_PROFILING);
  94 }
  95
  96 /*
  97  * Event handler for periodic ticks
  98  */
  99 void tick_handle_periodic(struct clock_event_device *dev)
 100 {
 101         int cpu = smp_processor_id();
 102         ktime_t next = dev->next_event;
 103
 104         tick_periodic(cpu);
 105
 106 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
 107         /*
 108          * The cpu might have transitioned to HIGHRES or NOHZ mode via
 109          * update_process_times() -> run_local_timers() ->
 110          * hrtimer_run_queues().
 111          */
 112         if (dev->event_handler != tick_handle_periodic)
 113                 return;
 114 #endif
 115
 116         if (!clockevent_state_oneshot(dev))
 117                 return;
 118         for (;;) {
 119                 /*
 120                  * Setup the next period for devices, which do not have
 121                  * periodic mode:
 122                  */
 123                 next = ktime_add(next, tick_period);
 124
 125                 if (!clockevents_program_event(dev, next, false))
 126                         return;
 127                 /*
 128                  * Have to be careful here. If we're in oneshot mode,
 129                  * before we call tick_periodic() in a loop, we need
 130                  * to be sure we're using a real hardware clocksource.
 131                  * Otherwise we could get trapped in an infinite
 132                  * loop, as the tick_periodic() increments jiffies,
 133                  * which then will increment time, possibly causing
 134                  * the loop to trigger again and again.
 135                  */
 136                 if (timekeeping_valid_for_hres())
 137                         tick_periodic(cpu);
 138         }
 139 }
 140
 141 /*
 142  * Setup the device for a periodic tick
 143  */
 144 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
 145 {
 146         tick_set_periodic_handler(dev, broadcast);
 147
 148         /* Broadcast setup ? */
 149         if (!tick_device_is_functional(dev))
 150                 return;
 151
 152         if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
 153             !tick_broadcast_oneshot_active()) {
 154                 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
 155         } else {
 156                 unsigned long seq;
 157                 ktime_t next;
 158
 159                 do {
 160                         seq = read_seqbegin(&jiffies_lock);
 161                         next = tick_next_period;
 162                 } while (read_seqretry(&jiffies_lock, seq));
 163
 164                 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 165
 166                 for (;;) {
 167                         if (!clockevents_program_event(dev, next, false))
 168                                 return;
 169                         next = ktime_add(next, tick_period);
 170                 }
 171         }
 172 }
 173
 174 /*
 175  * Setup the tick device
 176  */
 177 static void tick_setup_device(struct tick_device *td,
 178                               struct clock_event_device *newdev, int cpu,
 179                               const struct cpumask *cpumask)
 180 {
 181         void (*handler)(struct clock_event_device *) = NULL;
 182         ktime_t next_event = 0;
 183
 184         /*
 185          * First device setup ?
 186          */
 187         if (!td->evtdev) {
 188                 /*
 189                  * If no cpu took the do_timer update, assign it to
 190                  * this cpu:
 191                  */
 192                 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
 193                         if (!tick_nohz_full_cpu(cpu))
 194                                 tick_do_timer_cpu = cpu;
 195                         else
 196                                 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 197                         tick_next_period = ktime_get();
 198                         tick_period = NSEC_PER_SEC / HZ;
 199                 }
 200
 201                 /*
 202                  * Startup in periodic mode first.
 203                  */
 204                 td->mode = TICKDEV_MODE_PERIODIC;
 205         } else {
 206                 handler = td->evtdev->event_handler;
 207                 next_event = td->evtdev->next_event;
 208                 td->evtdev->event_handler = clockevents_handle_noop;
 209         }
 210
 211         td->evtdev = newdev;
 212
 213         /*
 214          * When the device is not per cpu, pin the interrupt to the
 215          * current cpu:
 216          */
 217         if (!cpumask_equal(newdev->cpumask, cpumask))
 218                 irq_set_affinity(newdev->irq, cpumask);
 219
 220         /*
 221          * When global broadcasting is active, check if the current
 222          * device is registered as a placeholder for broadcast mode.
 223          * This allows us to handle this x86 misfeature in a generic
 224          * way. This function also returns !=0 when we keep the
 225          * current active broadcast state for this CPU.
 226          */
 227         if (tick_device_uses_broadcast(newdev, cpu))
 228                 return;
 229
 230         if (td->mode == TICKDEV_MODE_PERIODIC)
 231                 tick_setup_periodic(newdev, 0);
 232         else
 233                 tick_setup_oneshot(newdev, handler, next_event);
 234 }
 235
 236 void tick_install_replacement(struct clock_event_device *newdev)
 237 {
 238         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 239         int cpu = smp_processor_id();
 240
 241         clockevents_exchange_device(td->evtdev, newdev);
 242         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 243         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 244                 tick_oneshot_notify();
 245 }
 246
 247 static bool tick_check_percpu(struct clock_event_device *curdev,
 248                               struct clock_event_device *newdev, int cpu)
 249 {
 250         if (!cpumask_test_cpu(cpu, newdev->cpumask))
 251                 return false;
 252         if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
 253                 return true;
 254         /* Check if irq affinity can be set */
 255         if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
 256                 return false;
 257         /* Prefer an existing cpu local device */
 258         if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
 259                 return false;
 260         return true;
 261 }
 262
 263 static bool tick_check_preferred(struct clock_event_device *curdev,
 264                                  struct clock_event_device *newdev)
 265 {
 266         /* Prefer oneshot capable device */
 267         if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
 268                 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
 269                         return false;
 270                 if (tick_oneshot_mode_active())
 271                         return false;
 272         }
 273
 274         /*
 275          * Use the higher rated one, but prefer a CPU local device with a lower
 276          * rating than a non-CPU local device
 277          */
 278         return !curdev ||
 279                 newdev->rating > curdev->rating ||
 280                !cpumask_equal(curdev->cpumask, newdev->cpumask);
 281 }
 282
 283 /*
 284  * Check whether the new device is a better fit than curdev. curdev
 285  * can be NULL !
 286  */
 287 bool tick_check_replacement(struct clock_event_device *curdev,
 288                             struct clock_event_device *newdev)
 289 {
 290         if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
 291                 return false;
 292
 293         return tick_check_preferred(curdev, newdev);
 294 }
 295
 296 /*
 297  * Check, if the new registered device should be used. Called with
 298  * clockevents_lock held and interrupts disabled.
 299  */
 300 void tick_check_new_device(struct clock_event_device *newdev)
 301 {
 302         struct clock_event_device *curdev;
 303         struct tick_device *td;
 304         int cpu;
 305
 306         cpu = smp_processor_id();
 307         td = &per_cpu(tick_cpu_device, cpu);
 308         curdev = td->evtdev;
 309
 310         /* cpu local device ? */
 311         if (!tick_check_percpu(curdev, newdev, cpu))
 312                 goto out_bc;
 313
 314         /* Preference decision */
 315         if (!tick_check_preferred(curdev, newdev))
 316                 goto out_bc;
 317
 318         if (!try_module_get(newdev->owner))
 319                 return;
 320
 321         /*
 322          * Replace the eventually existing device by the new
 323          * device. If the current device is the broadcast device, do
 324          * not give it back to the clockevents layer !
 325          */
 326         if (tick_is_broadcast_device(curdev)) {
 327                 clockevents_shutdown(curdev);
 328                 curdev = NULL;
 329         }
 330         clockevents_exchange_device(curdev, newdev);
 331         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 332         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 333                 tick_oneshot_notify();
 334         return;
 335
 336 out_bc:
 337         /*
 338          * Can the new device be used as a broadcast device ?
 339          */
 340         tick_install_broadcast_device(newdev);
 341 }
 342
 343 /**
 344  * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
 345  * @state:      The target state (enter/exit)
 346  *
 347  * The system enters/leaves a state, where affected devices might stop
 348  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
 349  *
 350  * Called with interrupts disabled, so clockevents_lock is not
 351  * required here because the local clock event device cannot go away
 352  * under us.
 353  */
 354 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 355 {
 356         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 357
 358         if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
 359                 return 0;
 360
 361         return __tick_broadcast_oneshot_control(state);
 362 }
 363 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
 364
 365 #ifdef CONFIG_HOTPLUG_CPU
 366 /*
 367  * Transfer the do_timer job away from a dying cpu.
 368  *
 369  * Called with interrupts disabled. Not locking required. If
 370  * tick_do_timer_cpu is owned by this cpu, nothing can change it.
 371  */
 372 void tick_handover_do_timer(void)
 373 {
 374         if (tick_do_timer_cpu == smp_processor_id()) {
 375                 int cpu = cpumask_first(cpu_online_mask);
 376
 377                 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
 378                         TICK_DO_TIMER_NONE;
 379         }
 380 }
 381
 382 /*
 383  * Shutdown an event device on a given cpu:
 384  *
 385  * This is called on a life CPU, when a CPU is dead. So we cannot
 386  * access the hardware device itself.
 387  * We just set the mode and remove it from the lists.
 388  */
 389 void tick_shutdown(unsigned int cpu)
 390 {
 391         struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 392         struct clock_event_device *dev = td->evtdev;
 393
 394         td->mode = TICKDEV_MODE_PERIODIC;
 395         if (dev) {
 396                 /*
 397                  * Prevent that the clock events layer tries to call
 398                  * the set mode function!
 399                  */
 400                 clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
 401                 clockevents_exchange_device(dev, NULL);
 402                 dev->event_handler = clockevents_handle_noop;
 403                 td->evtdev = NULL;
 404         }
 405 }
 406 #endif
 407
 408 /**
 409  * tick_suspend_local - Suspend the local tick device
 410  *
 411  * Called from the local cpu for freeze with interrupts disabled.
 412  *
 413  * No locks required. Nothing can change the per cpu device.
 414  */
 415 void tick_suspend_local(void)
 416 {
 417         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 418
 419         clockevents_shutdown(td->evtdev);
 420 }
 421
 422 /**
 423  * tick_resume_local - Resume the local tick device
 424  *
 425  * Called from the local CPU for unfreeze or XEN resume magic.
 426  *
 427  * No locks required. Nothing can change the per cpu device.
 428  */
 429 void tick_resume_local(void)
 430 {
 431         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 432         bool broadcast = tick_resume_check_broadcast();
 433
 434         clockevents_tick_resume(td->evtdev);
 435         if (!broadcast) {
 436                 if (td->mode == TICKDEV_MODE_PERIODIC)
 437                         tick_setup_periodic(td->evtdev, 0);
 438                 else
 439                         tick_resume_oneshot();
 440         }
 441 }
 442
 443 /**
 444  * tick_suspend - Suspend the tick and the broadcast device
 445  *
 446  * Called from syscore_suspend() via timekeeping_suspend with only one
 447  * CPU online and interrupts disabled or from tick_unfreeze() under
 448  * tick_freeze_lock.
 449  *
 450  * No locks required. Nothing can change the per cpu device.
 451  */
 452 void tick_suspend(void)
 453 {
 454         tick_suspend_local();
 455         tick_suspend_broadcast();
 456 }
 457
 458 /**
 459  * tick_resume - Resume the tick and the broadcast device
 460  *
 461  * Called from syscore_resume() via timekeeping_resume with only one
 462  * CPU online and interrupts disabled.
 463  *
 464  * No locks required. Nothing can change the per cpu device.
 465  */
 466 void tick_resume(void)
 467 {
 468         tick_resume_broadcast();
 469         tick_resume_local();
 470 }
 471
 472 #ifdef CONFIG_SUSPEND
 473 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
 474 static unsigned int tick_freeze_depth;
 475
 476 /**
 477  * tick_freeze - Suspend the local tick and (possibly) timekeeping.
 478  *
 479  * Check if this is the last online CPU executing the function and if so,
 480  * suspend timekeeping.  Otherwise suspend the local tick.
 481  *
 482  * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
 483  * Interrupts must not be enabled before the subsequent %tick_unfreeze().
 484  */
 485 void tick_freeze(void)
 486 {
 487         raw_spin_lock(&tick_freeze_lock);
 488
 489         tick_freeze_depth++;
 490         if (tick_freeze_depth == num_online_cpus()) {
 491                 trace_suspend_resume(TPS("timekeeping_freeze"),
 492                                      smp_processor_id(), true);
 493                 timekeeping_suspend();
 494         } else {
 495                 tick_suspend_local();
 496         }
 497
 498         raw_spin_unlock(&tick_freeze_lock);
 499 }
 500
 501 /**
 502  * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
 503  *
 504  * Check if this is the first CPU executing the function and if so, resume
 505  * timekeeping.  Otherwise resume the local tick.
 506  *
 507  * Call with interrupts disabled.  Must be balanced with %tick_freeze().
 508  * Interrupts must not be enabled after the preceding %tick_freeze().
 509  */
 510 void tick_unfreeze(void)
 511 {
 512         raw_spin_lock(&tick_freeze_lock);
 513
 514         if (tick_freeze_depth == num_online_cpus()) {
 515                 timekeeping_resume();
 516                 trace_suspend_resume(TPS("timekeeping_freeze"),
 517                                      smp_processor_id(), false);
 518         } else {
 519                 tick_resume_local();
 520         }
 521
 522         tick_freeze_depth--;
 523
 524         raw_spin_unlock(&tick_freeze_lock);
 525 }
 526 #endif /* CONFIG_SUSPEND */
 527
 528 /**
 529  * tick_init - initialize the tick control
 530  */
 531 void __init tick_init(void)
 532 {
 533         tick_broadcast_init();
 534         tick_nohz_init();
 535 }