kernel/time/tick-common.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file contains the base functions to manage periodic tick
   4  * related events.
   5  *
   6  * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
   7  * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   8  * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
   9  */
  10 #include <linux/cpu.h>
  11 #include <linux/err.h>
  12 #include <linux/hrtimer.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/nmi.h>
  15 #include <linux/percpu.h>
  16 #include <linux/profile.h>
  17 #include <linux/sched.h>
  18 #include <linux/module.h>
  19 #include <trace/events/power.h>
  20
  21 #include <asm/irq_regs.h>
  22
  23 #include "tick-internal.h"
  24
  25 /*
  26  * Tick devices
  27  */
  28 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
  29 /*
  30  * Tick next event: keeps track of the tick time. It's updated by the
  31  * CPU which handles the tick and protected by jiffies_lock. There is
  32  * no requirement to write hold the jiffies seqcount for it.
  33  */
  34 ktime_t tick_next_period;
  35
  36 /*
  37  * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
  38  * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
  39  * variable has two functions:
  40  *
  41  * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
  42  *    timekeeping lock all at once. Only the CPU which is assigned to do the
  43  *    update is handling it.
  44  *
  45  * 2) Hand off the duty in the NOHZ idle case by setting the value to
  46  *    TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
  47  *    at it will take over and keep the time keeping alive.  The handover
  48  *    procedure also covers cpu hotplug.
  49  */
  50 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
  51 #ifdef CONFIG_NO_HZ_FULL
  52 /*
  53  * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
  54  * tick_do_timer_cpu and it should be taken over by an eligible secondary
  55  * when one comes online.
  56  */
  57 static int tick_do_timer_boot_cpu __read_mostly = -1;
  58 #endif
  59
  60 /*
  61  * Debugging: see timer_list.c
  62  */
  63 struct tick_device *tick_get_device(int cpu)
  64 {
  65         return &per_cpu(tick_cpu_device, cpu);
  66 }
  67
  68 /**
  69  * tick_is_oneshot_available - check for a oneshot capable event device
  70  */
  71 int tick_is_oneshot_available(void)
  72 {
  73         struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
  74
  75         if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
  76                 return 0;
  77         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
  78                 return 1;
  79         return tick_broadcast_oneshot_available();
  80 }
  81
  82 /*
  83  * Periodic tick
  84  */
  85 static void tick_periodic(int cpu)
  86 {
  87         if (tick_do_timer_cpu == cpu) {
  88                 raw_spin_lock(&jiffies_lock);
  89                 write_seqcount_begin(&jiffies_seq);
  90
  91                 /* Keep track of the next tick event */
  92                 tick_next_period = ktime_add_ns(tick_next_period, TICK_NSEC);
  93
  94                 do_timer(1);
  95                 write_seqcount_end(&jiffies_seq);
  96                 raw_spin_unlock(&jiffies_lock);
  97                 update_wall_time();
  98         }
  99
 100         update_process_times(user_mode(get_irq_regs()));
 101         profile_tick(CPU_PROFILING);
 102 }
 103
 104 /*
 105  * Event handler for periodic ticks
 106  */
 107 void tick_handle_periodic(struct clock_event_device *dev)
 108 {
 109         int cpu = smp_processor_id();
 110         ktime_t next = dev->next_event;
 111
 112         tick_periodic(cpu);
 113
 114 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
 115         /*
 116          * The cpu might have transitioned to HIGHRES or NOHZ mode via
 117          * update_process_times() -> run_local_timers() ->
 118          * hrtimer_run_queues().
 119          */
 120         if (dev->event_handler != tick_handle_periodic)
 121                 return;
 122 #endif
 123
 124         if (!clockevent_state_oneshot(dev))
 125                 return;
 126         for (;;) {
 127                 /*
 128                  * Setup the next period for devices, which do not have
 129                  * periodic mode:
 130                  */
 131                 next = ktime_add_ns(next, TICK_NSEC);
 132
 133                 if (!clockevents_program_event(dev, next, false))
 134                         return;
 135                 /*
 136                  * Have to be careful here. If we're in oneshot mode,
 137                  * before we call tick_periodic() in a loop, we need
 138                  * to be sure we're using a real hardware clocksource.
 139                  * Otherwise we could get trapped in an infinite
 140                  * loop, as the tick_periodic() increments jiffies,
 141                  * which then will increment time, possibly causing
 142                  * the loop to trigger again and again.
 143                  */
 144                 if (timekeeping_valid_for_hres())
 145                         tick_periodic(cpu);
 146         }
 147 }
 148
 149 /*
 150  * Setup the device for a periodic tick
 151  */
 152 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
 153 {
 154         tick_set_periodic_handler(dev, broadcast);
 155
 156         /* Broadcast setup ? */
 157         if (!tick_device_is_functional(dev))
 158                 return;
 159
 160         if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
 161             !tick_broadcast_oneshot_active()) {
 162                 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
 163         } else {
 164                 unsigned int seq;
 165                 ktime_t next;
 166
 167                 do {
 168                         seq = read_seqcount_begin(&jiffies_seq);
 169                         next = tick_next_period;
 170                 } while (read_seqcount_retry(&jiffies_seq, seq));
 171
 172                 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 173
 174                 for (;;) {
 175                         if (!clockevents_program_event(dev, next, false))
 176                                 return;
 177                         next = ktime_add_ns(next, TICK_NSEC);
 178                 }
 179         }
 180 }
 181
 182 #ifdef CONFIG_NO_HZ_FULL
 183 static void giveup_do_timer(void *info)
 184 {
 185         int cpu = *(unsigned int *)info;
 186
 187         WARN_ON(tick_do_timer_cpu != smp_processor_id());
 188
 189         tick_do_timer_cpu = cpu;
 190 }
 191
 192 static void tick_take_do_timer_from_boot(void)
 193 {
 194         int cpu = smp_processor_id();
 195         int from = tick_do_timer_boot_cpu;
 196
 197         if (from >= 0 && from != cpu)
 198                 smp_call_function_single(from, giveup_do_timer, &cpu, 1);
 199 }
 200 #endif
 201
 202 /*
 203  * Setup the tick device
 204  */
 205 static void tick_setup_device(struct tick_device *td,
 206                               struct clock_event_device *newdev, int cpu,
 207                               const struct cpumask *cpumask)
 208 {
 209         void (*handler)(struct clock_event_device *) = NULL;
 210         ktime_t next_event = 0;
 211
 212         /*
 213          * First device setup ?
 214          */
 215         if (!td->evtdev) {
 216                 /*
 217                  * If no cpu took the do_timer update, assign it to
 218                  * this cpu:
 219                  */
 220                 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
 221                         tick_do_timer_cpu = cpu;
 222
 223                         tick_next_period = ktime_get();
 224 #ifdef CONFIG_NO_HZ_FULL
 225                         /*
 226                          * The boot CPU may be nohz_full, in which case set
 227                          * tick_do_timer_boot_cpu so the first housekeeping
 228                          * secondary that comes up will take do_timer from
 229                          * us.
 230                          */
 231                         if (tick_nohz_full_cpu(cpu))
 232                                 tick_do_timer_boot_cpu = cpu;
 233
 234                 } else if (tick_do_timer_boot_cpu != -1 &&
 235                                                 !tick_nohz_full_cpu(cpu)) {
 236                         tick_take_do_timer_from_boot();
 237                         tick_do_timer_boot_cpu = -1;
 238                         WARN_ON(tick_do_timer_cpu != cpu);
 239 #endif
 240                 }
 241
 242                 /*
 243                  * Startup in periodic mode first.
 244                  */
 245                 td->mode = TICKDEV_MODE_PERIODIC;
 246         } else {
 247                 handler = td->evtdev->event_handler;
 248                 next_event = td->evtdev->next_event;
 249                 td->evtdev->event_handler = clockevents_handle_noop;
 250         }
 251
 252         td->evtdev = newdev;
 253
 254         /*
 255          * When the device is not per cpu, pin the interrupt to the
 256          * current cpu:
 257          */
 258         if (!cpumask_equal(newdev->cpumask, cpumask))
 259                 irq_set_affinity(newdev->irq, cpumask);
 260
 261         /*
 262          * When global broadcasting is active, check if the current
 263          * device is registered as a placeholder for broadcast mode.
 264          * This allows us to handle this x86 misfeature in a generic
 265          * way. This function also returns !=0 when we keep the
 266          * current active broadcast state for this CPU.
 267          */
 268         if (tick_device_uses_broadcast(newdev, cpu))
 269                 return;
 270
 271         if (td->mode == TICKDEV_MODE_PERIODIC)
 272                 tick_setup_periodic(newdev, 0);
 273         else
 274                 tick_setup_oneshot(newdev, handler, next_event);
 275 }
 276
 277 void tick_install_replacement(struct clock_event_device *newdev)
 278 {
 279         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 280         int cpu = smp_processor_id();
 281
 282         clockevents_exchange_device(td->evtdev, newdev);
 283         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 284         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 285                 tick_oneshot_notify();
 286 }
 287
 288 static bool tick_check_percpu(struct clock_event_device *curdev,
 289                               struct clock_event_device *newdev, int cpu)
 290 {
 291         if (!cpumask_test_cpu(cpu, newdev->cpumask))
 292                 return false;
 293         if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
 294                 return true;
 295         /* Check if irq affinity can be set */
 296         if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
 297                 return false;
 298         /* Prefer an existing cpu local device */
 299         if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
 300                 return false;
 301         return true;
 302 }
 303
 304 static bool tick_check_preferred(struct clock_event_device *curdev,
 305                                  struct clock_event_device *newdev)
 306 {
 307         /* Prefer oneshot capable device */
 308         if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
 309                 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
 310                         return false;
 311                 if (tick_oneshot_mode_active())
 312                         return false;
 313         }
 314
 315         /*
 316          * Use the higher rated one, but prefer a CPU local device with a lower
 317          * rating than a non-CPU local device
 318          */
 319         return !curdev ||
 320                 newdev->rating > curdev->rating ||
 321                !cpumask_equal(curdev->cpumask, newdev->cpumask);
 322 }
 323
 324 /*
 325  * Check whether the new device is a better fit than curdev. curdev
 326  * can be NULL !
 327  */
 328 bool tick_check_replacement(struct clock_event_device *curdev,
 329                             struct clock_event_device *newdev)
 330 {
 331         if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
 332                 return false;
 333
 334         return tick_check_preferred(curdev, newdev);
 335 }
 336
 337 /*
 338  * Check, if the new registered device should be used. Called with
 339  * clockevents_lock held and interrupts disabled.
 340  */
 341 void tick_check_new_device(struct clock_event_device *newdev)
 342 {
 343         struct clock_event_device *curdev;
 344         struct tick_device *td;
 345         int cpu;
 346
 347         cpu = smp_processor_id();
 348         td = &per_cpu(tick_cpu_device, cpu);
 349         curdev = td->evtdev;
 350
 351         /* cpu local device ? */
 352         if (!tick_check_percpu(curdev, newdev, cpu))
 353                 goto out_bc;
 354
 355         /* Preference decision */
 356         if (!tick_check_preferred(curdev, newdev))
 357                 goto out_bc;
 358
 359         if (!try_module_get(newdev->owner))
 360                 return;
 361
 362         /*
 363          * Replace the eventually existing device by the new
 364          * device. If the current device is the broadcast device, do
 365          * not give it back to the clockevents layer !
 366          */
 367         if (tick_is_broadcast_device(curdev)) {
 368                 clockevents_shutdown(curdev);
 369                 curdev = NULL;
 370         }
 371         clockevents_exchange_device(curdev, newdev);
 372         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 373         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 374                 tick_oneshot_notify();
 375         return;
 376
 377 out_bc:
 378         /*
 379          * Can the new device be used as a broadcast device ?
 380          */
 381         tick_install_broadcast_device(newdev);
 382 }
 383
 384 /**
 385  * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
 386  * @state:      The target state (enter/exit)
 387  *
 388  * The system enters/leaves a state, where affected devices might stop
 389  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
 390  *
 391  * Called with interrupts disabled, so clockevents_lock is not
 392  * required here because the local clock event device cannot go away
 393  * under us.
 394  */
 395 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 396 {
 397         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 398
 399         if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
 400                 return 0;
 401
 402         return __tick_broadcast_oneshot_control(state);
 403 }
 404 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
 405
 406 #ifdef CONFIG_HOTPLUG_CPU
 407 /*
 408  * Transfer the do_timer job away from a dying cpu.
 409  *
 410  * Called with interrupts disabled. No locking required. If
 411  * tick_do_timer_cpu is owned by this cpu, nothing can change it.
 412  */
 413 void tick_handover_do_timer(void)
 414 {
 415         if (tick_do_timer_cpu == smp_processor_id())
 416                 tick_do_timer_cpu = cpumask_first(cpu_online_mask);
 417 }
 418
 419 /*
 420  * Shutdown an event device on a given cpu:
 421  *
 422  * This is called on a life CPU, when a CPU is dead. So we cannot
 423  * access the hardware device itself.
 424  * We just set the mode and remove it from the lists.
 425  */
 426 void tick_shutdown(unsigned int cpu)
 427 {
 428         struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 429         struct clock_event_device *dev = td->evtdev;
 430
 431         td->mode = TICKDEV_MODE_PERIODIC;
 432         if (dev) {
 433                 /*
 434                  * Prevent that the clock events layer tries to call
 435                  * the set mode function!
 436                  */
 437                 clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
 438                 clockevents_exchange_device(dev, NULL);
 439                 dev->event_handler = clockevents_handle_noop;
 440                 td->evtdev = NULL;
 441         }
 442 }
 443 #endif
 444
 445 /**
 446  * tick_suspend_local - Suspend the local tick device
 447  *
 448  * Called from the local cpu for freeze with interrupts disabled.
 449  *
 450  * No locks required. Nothing can change the per cpu device.
 451  */
 452 void tick_suspend_local(void)
 453 {
 454         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 455
 456         clockevents_shutdown(td->evtdev);
 457 }
 458
 459 /**
 460  * tick_resume_local - Resume the local tick device
 461  *
 462  * Called from the local CPU for unfreeze or XEN resume magic.
 463  *
 464  * No locks required. Nothing can change the per cpu device.
 465  */
 466 void tick_resume_local(void)
 467 {
 468         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 469         bool broadcast = tick_resume_check_broadcast();
 470
 471         clockevents_tick_resume(td->evtdev);
 472         if (!broadcast) {
 473                 if (td->mode == TICKDEV_MODE_PERIODIC)
 474                         tick_setup_periodic(td->evtdev, 0);
 475                 else
 476                         tick_resume_oneshot();
 477         }
 478 }
 479
 480 /**
 481  * tick_suspend - Suspend the tick and the broadcast device
 482  *
 483  * Called from syscore_suspend() via timekeeping_suspend with only one
 484  * CPU online and interrupts disabled or from tick_unfreeze() under
 485  * tick_freeze_lock.
 486  *
 487  * No locks required. Nothing can change the per cpu device.
 488  */
 489 void tick_suspend(void)
 490 {
 491         tick_suspend_local();
 492         tick_suspend_broadcast();
 493 }
 494
 495 /**
 496  * tick_resume - Resume the tick and the broadcast device
 497  *
 498  * Called from syscore_resume() via timekeeping_resume with only one
 499  * CPU online and interrupts disabled.
 500  *
 501  * No locks required. Nothing can change the per cpu device.
 502  */
 503 void tick_resume(void)
 504 {
 505         tick_resume_broadcast();
 506         tick_resume_local();
 507 }
 508
 509 #ifdef CONFIG_SUSPEND
 510 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
 511 static unsigned int tick_freeze_depth;
 512
 513 /**
 514  * tick_freeze - Suspend the local tick and (possibly) timekeeping.
 515  *
 516  * Check if this is the last online CPU executing the function and if so,
 517  * suspend timekeeping.  Otherwise suspend the local tick.
 518  *
 519  * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
 520  * Interrupts must not be enabled before the subsequent %tick_unfreeze().
 521  */
 522 void tick_freeze(void)
 523 {
 524         raw_spin_lock(&tick_freeze_lock);
 525
 526         tick_freeze_depth++;
 527         if (tick_freeze_depth == num_online_cpus()) {
 528                 trace_suspend_resume(TPS("timekeeping_freeze"),
 529                                      smp_processor_id(), true);
 530                 system_state = SYSTEM_SUSPEND;
 531                 sched_clock_suspend();
 532                 timekeeping_suspend();
 533         } else {
 534                 tick_suspend_local();
 535         }
 536
 537         raw_spin_unlock(&tick_freeze_lock);
 538 }
 539
 540 /**
 541  * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
 542  *
 543  * Check if this is the first CPU executing the function and if so, resume
 544  * timekeeping.  Otherwise resume the local tick.
 545  *
 546  * Call with interrupts disabled.  Must be balanced with %tick_freeze().
 547  * Interrupts must not be enabled after the preceding %tick_freeze().
 548  */
 549 void tick_unfreeze(void)
 550 {
 551         raw_spin_lock(&tick_freeze_lock);
 552
 553         if (tick_freeze_depth == num_online_cpus()) {
 554                 timekeeping_resume();
 555                 sched_clock_resume();
 556                 system_state = SYSTEM_RUNNING;
 557                 trace_suspend_resume(TPS("timekeeping_freeze"),
 558                                      smp_processor_id(), false);
 559         } else {
 560                 touch_softlockup_watchdog();
 561                 tick_resume_local();
 562         }
 563
 564         tick_freeze_depth--;
 565
 566         raw_spin_unlock(&tick_freeze_lock);
 567 }
 568 #endif /* CONFIG_SUSPEND */
 569
 570 /**
 571  * tick_init - initialize the tick control
 572  */
 573 void __init tick_init(void)
 574 {
 575         tick_broadcast_init();
 576         tick_nohz_init();
 577 }