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
  31  */
  32 ktime_t tick_next_period;
  33 ktime_t tick_period;
  34
  35 /*
  36  * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
  37  * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
  38  * variable has two functions:
  39  *
  40  * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
  41  *    timekeeping lock all at once. Only the CPU which is assigned to do the
  42  *    update is handling it.
  43  *
  44  * 2) Hand off the duty in the NOHZ idle case by setting the value to
  45  *    TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
  46  *    at it will take over and keep the time keeping alive.  The handover
  47  *    procedure also covers cpu hotplug.
  48  */
  49 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
  50 #ifdef CONFIG_NO_HZ_FULL
  51 /*
  52  * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
  53  * tick_do_timer_cpu and it should be taken over by an eligible secondary
  54  * when one comes online.
  55  */
  56 static int tick_do_timer_boot_cpu __read_mostly = -1;
  57 #endif
  58
  59 /*
  60  * Debugging: see timer_list.c
  61  */
  62 struct tick_device *tick_get_device(int cpu)
  63 {
  64         return &per_cpu(tick_cpu_device, cpu);
  65 }
  66
  67 /**
  68  * tick_is_oneshot_available - check for a oneshot capable event device
  69  */
  70 int tick_is_oneshot_available(void)
  71 {
  72         struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
  73
  74         if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
  75                 return 0;
  76         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
  77                 return 1;
  78         return tick_broadcast_oneshot_available();
  79 }
  80
  81 /*
  82  * Periodic tick
  83  */
  84 static void tick_periodic(int cpu)
  85 {
  86         if (tick_do_timer_cpu == cpu) {
  87                 write_seqlock(&jiffies_lock);
  88
  89                 /* Keep track of the next tick event */
  90                 tick_next_period = ktime_add(tick_next_period, tick_period);
  91
  92                 do_timer(1);
  93                 write_sequnlock(&jiffies_lock);
  94                 update_wall_time();
  95         }
  96
  97         update_process_times(user_mode(get_irq_regs()));
  98         profile_tick(CPU_PROFILING);
  99 }
 100
 101 /*
 102  * Event handler for periodic ticks
 103  */
 104 void tick_handle_periodic(struct clock_event_device *dev)
 105 {
 106         int cpu = smp_processor_id();
 107         ktime_t next = dev->next_event;
 108
 109         tick_periodic(cpu);
 110
 111 #if defined(CONFIG_HIGH_RES_TIMERS) || defined(CONFIG_NO_HZ_COMMON)
 112         /*
 113          * The cpu might have transitioned to HIGHRES or NOHZ mode via
 114          * update_process_times() -> run_local_timers() ->
 115          * hrtimer_run_queues().
 116          */
 117         if (dev->event_handler != tick_handle_periodic)
 118                 return;
 119 #endif
 120
 121         if (!clockevent_state_oneshot(dev))
 122                 return;
 123         for (;;) {
 124                 /*
 125                  * Setup the next period for devices, which do not have
 126                  * periodic mode:
 127                  */
 128                 next = ktime_add(next, tick_period);
 129
 130                 if (!clockevents_program_event(dev, next, false))
 131                         return;
 132                 /*
 133                  * Have to be careful here. If we're in oneshot mode,
 134                  * before we call tick_periodic() in a loop, we need
 135                  * to be sure we're using a real hardware clocksource.
 136                  * Otherwise we could get trapped in an infinite
 137                  * loop, as the tick_periodic() increments jiffies,
 138                  * which then will increment time, possibly causing
 139                  * the loop to trigger again and again.
 140                  */
 141                 if (timekeeping_valid_for_hres())
 142                         tick_periodic(cpu);
 143         }
 144 }
 145
 146 /*
 147  * Setup the device for a periodic tick
 148  */
 149 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
 150 {
 151         tick_set_periodic_handler(dev, broadcast);
 152
 153         /* Broadcast setup ? */
 154         if (!tick_device_is_functional(dev))
 155                 return;
 156
 157         if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
 158             !tick_broadcast_oneshot_active()) {
 159                 clockevents_switch_state(dev, CLOCK_EVT_STATE_PERIODIC);
 160         } else {
 161                 unsigned int seq;
 162                 ktime_t next;
 163
 164                 do {
 165                         seq = read_seqbegin(&jiffies_lock);
 166                         next = tick_next_period;
 167                 } while (read_seqretry(&jiffies_lock, seq));
 168
 169                 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 170
 171                 for (;;) {
 172                         if (!clockevents_program_event(dev, next, false))
 173                                 return;
 174                         next = ktime_add(next, tick_period);
 175                 }
 176         }
 177 }
 178
 179 #ifdef CONFIG_NO_HZ_FULL
 180 static void giveup_do_timer(void *info)
 181 {
 182         int cpu = *(unsigned int *)info;
 183
 184         WARN_ON(tick_do_timer_cpu != smp_processor_id());
 185
 186         tick_do_timer_cpu = cpu;
 187 }
 188
 189 static void tick_take_do_timer_from_boot(void)
 190 {
 191         int cpu = smp_processor_id();
 192         int from = tick_do_timer_boot_cpu;
 193
 194         if (from >= 0 && from != cpu)
 195                 smp_call_function_single(from, giveup_do_timer, &cpu, 1);
 196 }
 197 #endif
 198
 199 /*
 200  * Setup the tick device
 201  */
 202 static void tick_setup_device(struct tick_device *td,
 203                               struct clock_event_device *newdev, int cpu,
 204                               const struct cpumask *cpumask)
 205 {
 206         void (*handler)(struct clock_event_device *) = NULL;
 207         ktime_t next_event = 0;
 208
 209         /*
 210          * First device setup ?
 211          */
 212         if (!td->evtdev) {
 213                 /*
 214                  * If no cpu took the do_timer update, assign it to
 215                  * this cpu:
 216                  */
 217                 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
 218                         tick_do_timer_cpu = cpu;
 219
 220                         tick_next_period = ktime_get();
 221                         tick_period = NSEC_PER_SEC / HZ;
 222 #ifdef CONFIG_NO_HZ_FULL
 223                         /*
 224                          * The boot CPU may be nohz_full, in which case set
 225                          * tick_do_timer_boot_cpu so the first housekeeping
 226                          * secondary that comes up will take do_timer from
 227                          * us.
 228                          */
 229                         if (tick_nohz_full_cpu(cpu))
 230                                 tick_do_timer_boot_cpu = cpu;
 231
 232                 } else if (tick_do_timer_boot_cpu != -1 &&
 233                                                 !tick_nohz_full_cpu(cpu)) {
 234                         tick_take_do_timer_from_boot();
 235                         tick_do_timer_boot_cpu = -1;
 236                         WARN_ON(tick_do_timer_cpu != cpu);
 237 #endif
 238                 }
 239
 240                 /*
 241                  * Startup in periodic mode first.
 242                  */
 243                 td->mode = TICKDEV_MODE_PERIODIC;
 244         } else {
 245                 handler = td->evtdev->event_handler;
 246                 next_event = td->evtdev->next_event;
 247                 td->evtdev->event_handler = clockevents_handle_noop;
 248         }
 249
 250         td->evtdev = newdev;
 251
 252         /*
 253          * When the device is not per cpu, pin the interrupt to the
 254          * current cpu:
 255          */
 256         if (!cpumask_equal(newdev->cpumask, cpumask))
 257                 irq_set_affinity(newdev->irq, cpumask);
 258
 259         /*
 260          * When global broadcasting is active, check if the current
 261          * device is registered as a placeholder for broadcast mode.
 262          * This allows us to handle this x86 misfeature in a generic
 263          * way. This function also returns !=0 when we keep the
 264          * current active broadcast state for this CPU.
 265          */
 266         if (tick_device_uses_broadcast(newdev, cpu))
 267                 return;
 268
 269         if (td->mode == TICKDEV_MODE_PERIODIC)
 270                 tick_setup_periodic(newdev, 0);
 271         else
 272                 tick_setup_oneshot(newdev, handler, next_event);
 273 }
 274
 275 void tick_install_replacement(struct clock_event_device *newdev)
 276 {
 277         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 278         int cpu = smp_processor_id();
 279
 280         clockevents_exchange_device(td->evtdev, newdev);
 281         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 282         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 283                 tick_oneshot_notify();
 284 }
 285
 286 static bool tick_check_percpu(struct clock_event_device *curdev,
 287                               struct clock_event_device *newdev, int cpu)
 288 {
 289         if (!cpumask_test_cpu(cpu, newdev->cpumask))
 290                 return false;
 291         if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
 292                 return true;
 293         /* Check if irq affinity can be set */
 294         if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
 295                 return false;
 296         /* Prefer an existing cpu local device */
 297         if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
 298                 return false;
 299         return true;
 300 }
 301
 302 static bool tick_check_preferred(struct clock_event_device *curdev,
 303                                  struct clock_event_device *newdev)
 304 {
 305         /* Prefer oneshot capable device */
 306         if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
 307                 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
 308                         return false;
 309                 if (tick_oneshot_mode_active())
 310                         return false;
 311         }
 312
 313         /*
 314          * Use the higher rated one, but prefer a CPU local device with a lower
 315          * rating than a non-CPU local device
 316          */
 317         return !curdev ||
 318                 newdev->rating > curdev->rating ||
 319                !cpumask_equal(curdev->cpumask, newdev->cpumask);
 320 }
 321
 322 /*
 323  * Check whether the new device is a better fit than curdev. curdev
 324  * can be NULL !
 325  */
 326 bool tick_check_replacement(struct clock_event_device *curdev,
 327                             struct clock_event_device *newdev)
 328 {
 329         if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
 330                 return false;
 331
 332         return tick_check_preferred(curdev, newdev);
 333 }
 334
 335 /*
 336  * Check, if the new registered device should be used. Called with
 337  * clockevents_lock held and interrupts disabled.
 338  */
 339 void tick_check_new_device(struct clock_event_device *newdev)
 340 {
 341         struct clock_event_device *curdev;
 342         struct tick_device *td;
 343         int cpu;
 344
 345         cpu = smp_processor_id();
 346         td = &per_cpu(tick_cpu_device, cpu);
 347         curdev = td->evtdev;
 348
 349         /* cpu local device ? */
 350         if (!tick_check_percpu(curdev, newdev, cpu))
 351                 goto out_bc;
 352
 353         /* Preference decision */
 354         if (!tick_check_preferred(curdev, newdev))
 355                 goto out_bc;
 356
 357         if (!try_module_get(newdev->owner))
 358                 return;
 359
 360         /*
 361          * Replace the eventually existing device by the new
 362          * device. If the current device is the broadcast device, do
 363          * not give it back to the clockevents layer !
 364          */
 365         if (tick_is_broadcast_device(curdev)) {
 366                 clockevents_shutdown(curdev);
 367                 curdev = NULL;
 368         }
 369         clockevents_exchange_device(curdev, newdev);
 370         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 371         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 372                 tick_oneshot_notify();
 373         return;
 374
 375 out_bc:
 376         /*
 377          * Can the new device be used as a broadcast device ?
 378          */
 379         tick_install_broadcast_device(newdev);
 380 }
 381
 382 /**
 383  * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
 384  * @state:      The target state (enter/exit)
 385  *
 386  * The system enters/leaves a state, where affected devices might stop
 387  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
 388  *
 389  * Called with interrupts disabled, so clockevents_lock is not
 390  * required here because the local clock event device cannot go away
 391  * under us.
 392  */
 393 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 394 {
 395         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 396
 397         if (!(td->evtdev->features & CLOCK_EVT_FEAT_C3STOP))
 398                 return 0;
 399
 400         return __tick_broadcast_oneshot_control(state);
 401 }
 402 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
 403
 404 #ifdef CONFIG_HOTPLUG_CPU
 405 /*
 406  * Transfer the do_timer job away from a dying cpu.
 407  *
 408  * Called with interrupts disabled. Not locking required. If
 409  * tick_do_timer_cpu is owned by this cpu, nothing can change it.
 410  */
 411 void tick_handover_do_timer(void)
 412 {
 413         if (tick_do_timer_cpu == smp_processor_id()) {
 414                 int cpu = cpumask_first(cpu_online_mask);
 415
 416                 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
 417                         TICK_DO_TIMER_NONE;
 418         }
 419 }
 420
 421 /*
 422  * Shutdown an event device on a given cpu:
 423  *
 424  * This is called on a life CPU, when a CPU is dead. So we cannot
 425  * access the hardware device itself.
 426  * We just set the mode and remove it from the lists.
 427  */
 428 void tick_shutdown(unsigned int cpu)
 429 {
 430         struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 431         struct clock_event_device *dev = td->evtdev;
 432
 433         td->mode = TICKDEV_MODE_PERIODIC;
 434         if (dev) {
 435                 /*
 436                  * Prevent that the clock events layer tries to call
 437                  * the set mode function!
 438                  */
 439                 clockevent_set_state(dev, CLOCK_EVT_STATE_DETACHED);
 440                 clockevents_exchange_device(dev, NULL);
 441                 dev->event_handler = clockevents_handle_noop;
 442                 td->evtdev = NULL;
 443         }
 444 }
 445 #endif
 446
 447 /**
 448  * tick_suspend_local - Suspend the local tick device
 449  *
 450  * Called from the local cpu for freeze with interrupts disabled.
 451  *
 452  * No locks required. Nothing can change the per cpu device.
 453  */
 454 void tick_suspend_local(void)
 455 {
 456         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 457
 458         clockevents_shutdown(td->evtdev);
 459 }
 460
 461 /**
 462  * tick_resume_local - Resume the local tick device
 463  *
 464  * Called from the local CPU for unfreeze or XEN resume magic.
 465  *
 466  * No locks required. Nothing can change the per cpu device.
 467  */
 468 void tick_resume_local(void)
 469 {
 470         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 471         bool broadcast = tick_resume_check_broadcast();
 472
 473         clockevents_tick_resume(td->evtdev);
 474         if (!broadcast) {
 475                 if (td->mode == TICKDEV_MODE_PERIODIC)
 476                         tick_setup_periodic(td->evtdev, 0);
 477                 else
 478                         tick_resume_oneshot();
 479         }
 480 }
 481
 482 /**
 483  * tick_suspend - Suspend the tick and the broadcast device
 484  *
 485  * Called from syscore_suspend() via timekeeping_suspend with only one
 486  * CPU online and interrupts disabled or from tick_unfreeze() under
 487  * tick_freeze_lock.
 488  *
 489  * No locks required. Nothing can change the per cpu device.
 490  */
 491 void tick_suspend(void)
 492 {
 493         tick_suspend_local();
 494         tick_suspend_broadcast();
 495 }
 496
 497 /**
 498  * tick_resume - Resume the tick and the broadcast device
 499  *
 500  * Called from syscore_resume() via timekeeping_resume with only one
 501  * CPU online and interrupts disabled.
 502  *
 503  * No locks required. Nothing can change the per cpu device.
 504  */
 505 void tick_resume(void)
 506 {
 507         tick_resume_broadcast();
 508         tick_resume_local();
 509 }
 510
 511 #ifdef CONFIG_SUSPEND
 512 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
 513 static unsigned int tick_freeze_depth;
 514
 515 /**
 516  * tick_freeze - Suspend the local tick and (possibly) timekeeping.
 517  *
 518  * Check if this is the last online CPU executing the function and if so,
 519  * suspend timekeeping.  Otherwise suspend the local tick.
 520  *
 521  * Call with interrupts disabled.  Must be balanced with %tick_unfreeze().
 522  * Interrupts must not be enabled before the subsequent %tick_unfreeze().
 523  */
 524 void tick_freeze(void)
 525 {
 526         raw_spin_lock(&tick_freeze_lock);
 527
 528         tick_freeze_depth++;
 529         if (tick_freeze_depth == num_online_cpus()) {
 530                 trace_suspend_resume(TPS("timekeeping_freeze"),
 531                                      smp_processor_id(), true);
 532                 system_state = SYSTEM_SUSPEND;
 533                 sched_clock_suspend();
 534                 timekeeping_suspend();
 535         } else {
 536                 tick_suspend_local();
 537         }
 538
 539         raw_spin_unlock(&tick_freeze_lock);
 540 }
 541
 542 /**
 543  * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
 544  *
 545  * Check if this is the first CPU executing the function and if so, resume
 546  * timekeeping.  Otherwise resume the local tick.
 547  *
 548  * Call with interrupts disabled.  Must be balanced with %tick_freeze().
 549  * Interrupts must not be enabled after the preceding %tick_freeze().
 550  */
 551 void tick_unfreeze(void)
 552 {
 553         raw_spin_lock(&tick_freeze_lock);
 554
 555         if (tick_freeze_depth == num_online_cpus()) {
 556                 timekeeping_resume();
 557                 sched_clock_resume();
 558                 system_state = SYSTEM_RUNNING;
 559                 trace_suspend_resume(TPS("timekeeping_freeze"),
 560                                      smp_processor_id(), false);
 561         } else {
 562                 touch_softlockup_watchdog();
 563                 tick_resume_local();
 564         }
 565
 566         tick_freeze_depth--;
 567
 568         raw_spin_unlock(&tick_freeze_lock);
 569 }
 570 #endif /* CONFIG_SUSPEND */
 571
 572 /**
 573  * tick_init - initialize the tick control
 574  */
 575 void __init tick_init(void)
 576 {
 577         tick_broadcast_init();
 578         tick_nohz_init();
 579 }