kernel/time/tick-broadcast.c

   1 /*
   2  * linux/kernel/time/tick-broadcast.c
   3  *
   4  * This file contains functions which emulate a local clock-event
   5  * device via a broadcast event source.
   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/smp.h>
  22 #include <linux/module.h>
  23
  24 #include "tick-internal.h"
  25
  26 /*
  27  * Broadcast support for broken x86 hardware, where the local apic
  28  * timer stops in C3 state.
  29  */
  30
  31 static struct tick_device tick_broadcast_device;
  32 static cpumask_var_t tick_broadcast_mask;
  33 static cpumask_var_t tick_broadcast_on;
  34 static cpumask_var_t tmpmask;
  35 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  36 static int tick_broadcast_forced;
  37
  38 #ifdef CONFIG_TICK_ONESHOT
  39 static void tick_broadcast_clear_oneshot(int cpu);
  40 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
  41 #else
  42 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  43 static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
  44 #endif
  45
  46 /*
  47  * Debugging: see timer_list.c
  48  */
  49 struct tick_device *tick_get_broadcast_device(void)
  50 {
  51         return &tick_broadcast_device;
  52 }
  53
  54 struct cpumask *tick_get_broadcast_mask(void)
  55 {
  56         return tick_broadcast_mask;
  57 }
  58
  59 /*
  60  * Start the device in periodic mode
  61  */
  62 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  63 {
  64         if (bc)
  65                 tick_setup_periodic(bc, 1);
  66 }
  67
  68 /*
  69  * Check, if the device can be utilized as broadcast device:
  70  */
  71 static bool tick_check_broadcast_device(struct clock_event_device *curdev,
  72                                         struct clock_event_device *newdev)
  73 {
  74         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
  75             (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
  76             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
  77                 return false;
  78
  79         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
  80             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
  81                 return false;
  82
  83         return !curdev || newdev->rating > curdev->rating;
  84 }
  85
  86 /*
  87  * Conditionally install/replace broadcast device
  88  */
  89 void tick_install_broadcast_device(struct clock_event_device *dev)
  90 {
  91         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  92
  93         if (!tick_check_broadcast_device(cur, dev))
  94                 return;
  95
  96         if (!try_module_get(dev->owner))
  97                 return;
  98
  99         clockevents_exchange_device(cur, dev);
 100         if (cur)
 101                 cur->event_handler = clockevents_handle_noop;
 102         tick_broadcast_device.evtdev = dev;
 103         if (!cpumask_empty(tick_broadcast_mask))
 104                 tick_broadcast_start_periodic(dev);
 105         /*
 106          * Inform all cpus about this. We might be in a situation
 107          * where we did not switch to oneshot mode because the per cpu
 108          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
 109          * of a oneshot capable broadcast device. Without that
 110          * notification the systems stays stuck in periodic mode
 111          * forever.
 112          */
 113         if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
 114                 tick_clock_notify();
 115 }
 116
 117 /*
 118  * Check, if the device is the broadcast device
 119  */
 120 int tick_is_broadcast_device(struct clock_event_device *dev)
 121 {
 122         return (dev && tick_broadcast_device.evtdev == dev);
 123 }
 124
 125 int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
 126 {
 127         int ret = -ENODEV;
 128
 129         if (tick_is_broadcast_device(dev)) {
 130                 raw_spin_lock(&tick_broadcast_lock);
 131                 ret = __clockevents_update_freq(dev, freq);
 132                 raw_spin_unlock(&tick_broadcast_lock);
 133         }
 134         return ret;
 135 }
 136
 137
 138 static void err_broadcast(const struct cpumask *mask)
 139 {
 140         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 141 }
 142
 143 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 144 {
 145         if (!dev->broadcast)
 146                 dev->broadcast = tick_broadcast;
 147         if (!dev->broadcast) {
 148                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 149                              dev->name);
 150                 dev->broadcast = err_broadcast;
 151         }
 152 }
 153
 154 /*
 155  * Check, if the device is disfunctional and a place holder, which
 156  * needs to be handled by the broadcast device.
 157  */
 158 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 159 {
 160         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 161         unsigned long flags;
 162         int ret;
 163
 164         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 165
 166         /*
 167          * Devices might be registered with both periodic and oneshot
 168          * mode disabled. This signals, that the device needs to be
 169          * operated from the broadcast device and is a placeholder for
 170          * the cpu local device.
 171          */
 172         if (!tick_device_is_functional(dev)) {
 173                 dev->event_handler = tick_handle_periodic;
 174                 tick_device_setup_broadcast_func(dev);
 175                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 176                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 177                         tick_broadcast_start_periodic(bc);
 178                 else
 179                         tick_broadcast_setup_oneshot(bc);
 180                 ret = 1;
 181         } else {
 182                 /*
 183                  * Clear the broadcast bit for this cpu if the
 184                  * device is not power state affected.
 185                  */
 186                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 187                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 188                 else
 189                         tick_device_setup_broadcast_func(dev);
 190
 191                 /*
 192                  * Clear the broadcast bit if the CPU is not in
 193                  * periodic broadcast on state.
 194                  */
 195                 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
 196                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 197
 198                 switch (tick_broadcast_device.mode) {
 199                 case TICKDEV_MODE_ONESHOT:
 200                         /*
 201                          * If the system is in oneshot mode we can
 202                          * unconditionally clear the oneshot mask bit,
 203                          * because the CPU is running and therefore
 204                          * not in an idle state which causes the power
 205                          * state affected device to stop. Let the
 206                          * caller initialize the device.
 207                          */
 208                         tick_broadcast_clear_oneshot(cpu);
 209                         ret = 0;
 210                         break;
 211
 212                 case TICKDEV_MODE_PERIODIC:
 213                         /*
 214                          * If the system is in periodic mode, check
 215                          * whether the broadcast device can be
 216                          * switched off now.
 217                          */
 218                         if (cpumask_empty(tick_broadcast_mask) && bc)
 219                                 clockevents_shutdown(bc);
 220                         /*
 221                          * If we kept the cpu in the broadcast mask,
 222                          * tell the caller to leave the per cpu device
 223                          * in shutdown state. The periodic interrupt
 224                          * is delivered by the broadcast device.
 225                          */
 226                         ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 227                         break;
 228                 default:
 229                         /* Nothing to do */
 230                         ret = 0;
 231                         break;
 232                 }
 233         }
 234         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 235         return ret;
 236 }
 237
 238 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 239 int tick_receive_broadcast(void)
 240 {
 241         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 242         struct clock_event_device *evt = td->evtdev;
 243
 244         if (!evt)
 245                 return -ENODEV;
 246
 247         if (!evt->event_handler)
 248                 return -EINVAL;
 249
 250         evt->event_handler(evt);
 251         return 0;
 252 }
 253 #endif
 254
 255 /*
 256  * Broadcast the event to the cpus, which are set in the mask (mangled).
 257  */
 258 static void tick_do_broadcast(struct cpumask *mask)
 259 {
 260         int cpu = smp_processor_id();
 261         struct tick_device *td;
 262
 263         /*
 264          * Check, if the current cpu is in the mask
 265          */
 266         if (cpumask_test_cpu(cpu, mask)) {
 267                 cpumask_clear_cpu(cpu, mask);
 268                 td = &per_cpu(tick_cpu_device, cpu);
 269                 td->evtdev->event_handler(td->evtdev);
 270         }
 271
 272         if (!cpumask_empty(mask)) {
 273                 /*
 274                  * It might be necessary to actually check whether the devices
 275                  * have different broadcast functions. For now, just use the
 276                  * one of the first device. This works as long as we have this
 277                  * misfeature only on x86 (lapic)
 278                  */
 279                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 280                 td->evtdev->broadcast(mask);
 281         }
 282 }
 283
 284 /*
 285  * Periodic broadcast:
 286  * - invoke the broadcast handlers
 287  */
 288 static void tick_do_periodic_broadcast(void)
 289 {
 290         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 291         tick_do_broadcast(tmpmask);
 292 }
 293
 294 /*
 295  * Event handler for periodic broadcast ticks
 296  */
 297 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 298 {
 299         ktime_t next;
 300
 301         raw_spin_lock(&tick_broadcast_lock);
 302
 303         tick_do_periodic_broadcast();
 304
 305         /*
 306          * The device is in periodic mode. No reprogramming necessary:
 307          */
 308         if (dev->state == CLOCK_EVT_STATE_PERIODIC)
 309                 goto unlock;
 310
 311         /*
 312          * Setup the next period for devices, which do not have
 313          * periodic mode. We read dev->next_event first and add to it
 314          * when the event already expired. clockevents_program_event()
 315          * sets dev->next_event only when the event is really
 316          * programmed to the device.
 317          */
 318         for (next = dev->next_event; ;) {
 319                 next = ktime_add(next, tick_period);
 320
 321                 if (!clockevents_program_event(dev, next, false))
 322                         goto unlock;
 323                 tick_do_periodic_broadcast();
 324         }
 325 unlock:
 326         raw_spin_unlock(&tick_broadcast_lock);
 327 }
 328
 329 /**
 330  * tick_broadcast_control - Enable/disable or force broadcast mode
 331  * @mode:       The selected broadcast mode
 332  *
 333  * Called when the system enters a state where affected tick devices
 334  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 335  *
 336  * Called with interrupts disabled, so clockevents_lock is not
 337  * required here because the local clock event device cannot go away
 338  * under us.
 339  */
 340 void tick_broadcast_control(enum tick_broadcast_mode mode)
 341 {
 342         struct clock_event_device *bc, *dev;
 343         struct tick_device *td;
 344         int cpu, bc_stopped;
 345
 346         td = this_cpu_ptr(&tick_cpu_device);
 347         dev = td->evtdev;
 348
 349         /*
 350          * Is the device not affected by the powerstate ?
 351          */
 352         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 353                 return;
 354
 355         if (!tick_device_is_functional(dev))
 356                 return;
 357
 358         raw_spin_lock(&tick_broadcast_lock);
 359         cpu = smp_processor_id();
 360         bc = tick_broadcast_device.evtdev;
 361         bc_stopped = cpumask_empty(tick_broadcast_mask);
 362
 363         switch (mode) {
 364         case TICK_BROADCAST_FORCE:
 365                 tick_broadcast_forced = 1;
 366         case TICK_BROADCAST_ON:
 367                 cpumask_set_cpu(cpu, tick_broadcast_on);
 368                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 369                         if (tick_broadcast_device.mode ==
 370                             TICKDEV_MODE_PERIODIC)
 371                                 clockevents_shutdown(dev);
 372                 }
 373                 break;
 374
 375         case TICK_BROADCAST_OFF:
 376                 if (tick_broadcast_forced)
 377                         break;
 378                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 379                 if (!tick_device_is_functional(dev))
 380                         break;
 381                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 382                         if (tick_broadcast_device.mode ==
 383                             TICKDEV_MODE_PERIODIC)
 384                                 tick_setup_periodic(dev, 0);
 385                 }
 386                 break;
 387         }
 388
 389         if (cpumask_empty(tick_broadcast_mask)) {
 390                 if (!bc_stopped)
 391                         clockevents_shutdown(bc);
 392         } else if (bc_stopped) {
 393                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 394                         tick_broadcast_start_periodic(bc);
 395                 else
 396                         tick_broadcast_setup_oneshot(bc);
 397         }
 398         raw_spin_unlock(&tick_broadcast_lock);
 399 }
 400 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 401
 402 /*
 403  * Set the periodic handler depending on broadcast on/off
 404  */
 405 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 406 {
 407         if (!broadcast)
 408                 dev->event_handler = tick_handle_periodic;
 409         else
 410                 dev->event_handler = tick_handle_periodic_broadcast;
 411 }
 412
 413 /*
 414  * Remove a CPU from broadcasting
 415  */
 416 void tick_shutdown_broadcast(unsigned int *cpup)
 417 {
 418         struct clock_event_device *bc;
 419         unsigned long flags;
 420         unsigned int cpu = *cpup;
 421
 422         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 423
 424         bc = tick_broadcast_device.evtdev;
 425         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 426         cpumask_clear_cpu(cpu, tick_broadcast_on);
 427
 428         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 429                 if (bc && cpumask_empty(tick_broadcast_mask))
 430                         clockevents_shutdown(bc);
 431         }
 432
 433         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 434 }
 435
 436 void tick_suspend_broadcast(void)
 437 {
 438         struct clock_event_device *bc;
 439         unsigned long flags;
 440
 441         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 442
 443         bc = tick_broadcast_device.evtdev;
 444         if (bc)
 445                 clockevents_shutdown(bc);
 446
 447         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 448 }
 449
 450 /*
 451  * This is called from tick_resume_local() on a resuming CPU. That's
 452  * called from the core resume function, tick_unfreeze() and the magic XEN
 453  * resume hackery.
 454  *
 455  * In none of these cases the broadcast device mode can change and the
 456  * bit of the resuming CPU in the broadcast mask is safe as well.
 457  */
 458 bool tick_resume_check_broadcast(void)
 459 {
 460         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 461                 return false;
 462         else
 463                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 464 }
 465
 466 void tick_resume_broadcast(void)
 467 {
 468         struct clock_event_device *bc;
 469         unsigned long flags;
 470
 471         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 472
 473         bc = tick_broadcast_device.evtdev;
 474
 475         if (bc) {
 476                 clockevents_tick_resume(bc);
 477
 478                 switch (tick_broadcast_device.mode) {
 479                 case TICKDEV_MODE_PERIODIC:
 480                         if (!cpumask_empty(tick_broadcast_mask))
 481                                 tick_broadcast_start_periodic(bc);
 482                         break;
 483                 case TICKDEV_MODE_ONESHOT:
 484                         if (!cpumask_empty(tick_broadcast_mask))
 485                                 tick_resume_broadcast_oneshot(bc);
 486                         break;
 487                 }
 488         }
 489         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 490 }
 491
 492 #ifdef CONFIG_TICK_ONESHOT
 493
 494 static cpumask_var_t tick_broadcast_oneshot_mask;
 495 static cpumask_var_t tick_broadcast_pending_mask;
 496 static cpumask_var_t tick_broadcast_force_mask;
 497
 498 /*
 499  * Exposed for debugging: see timer_list.c
 500  */
 501 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 502 {
 503         return tick_broadcast_oneshot_mask;
 504 }
 505
 506 /*
 507  * Called before going idle with interrupts disabled. Checks whether a
 508  * broadcast event from the other core is about to happen. We detected
 509  * that in tick_broadcast_oneshot_control(). The callsite can use this
 510  * to avoid a deep idle transition as we are about to get the
 511  * broadcast IPI right away.
 512  */
 513 int tick_check_broadcast_expired(void)
 514 {
 515         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 516 }
 517
 518 /*
 519  * Set broadcast interrupt affinity
 520  */
 521 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 522                                         const struct cpumask *cpumask)
 523 {
 524         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 525                 return;
 526
 527         if (cpumask_equal(bc->cpumask, cpumask))
 528                 return;
 529
 530         bc->cpumask = cpumask;
 531         irq_set_affinity(bc->irq, bc->cpumask);
 532 }
 533
 534 static int tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 535                                     ktime_t expires, int force)
 536 {
 537         int ret;
 538
 539         if (bc->state != CLOCK_EVT_STATE_ONESHOT)
 540                 clockevents_set_state(bc, CLOCK_EVT_STATE_ONESHOT);
 541
 542         ret = clockevents_program_event(bc, expires, force);
 543         if (!ret)
 544                 tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 545         return ret;
 546 }
 547
 548 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 549 {
 550         clockevents_set_state(bc, CLOCK_EVT_STATE_ONESHOT);
 551 }
 552
 553 /*
 554  * Called from irq_enter() when idle was interrupted to reenable the
 555  * per cpu device.
 556  */
 557 void tick_check_oneshot_broadcast_this_cpu(void)
 558 {
 559         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 560                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 561
 562                 /*
 563                  * We might be in the middle of switching over from
 564                  * periodic to oneshot. If the CPU has not yet
 565                  * switched over, leave the device alone.
 566                  */
 567                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 568                         clockevents_set_state(td->evtdev,
 569                                               CLOCK_EVT_STATE_ONESHOT);
 570                 }
 571         }
 572 }
 573
 574 /*
 575  * Handle oneshot mode broadcasting
 576  */
 577 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 578 {
 579         struct tick_device *td;
 580         ktime_t now, next_event;
 581         int cpu, next_cpu = 0;
 582
 583         raw_spin_lock(&tick_broadcast_lock);
 584 again:
 585         dev->next_event.tv64 = KTIME_MAX;
 586         next_event.tv64 = KTIME_MAX;
 587         cpumask_clear(tmpmask);
 588         now = ktime_get();
 589         /* Find all expired events */
 590         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 591                 td = &per_cpu(tick_cpu_device, cpu);
 592                 if (td->evtdev->next_event.tv64 <= now.tv64) {
 593                         cpumask_set_cpu(cpu, tmpmask);
 594                         /*
 595                          * Mark the remote cpu in the pending mask, so
 596                          * it can avoid reprogramming the cpu local
 597                          * timer in tick_broadcast_oneshot_control().
 598                          */
 599                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 600                 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
 601                         next_event.tv64 = td->evtdev->next_event.tv64;
 602                         next_cpu = cpu;
 603                 }
 604         }
 605
 606         /*
 607          * Remove the current cpu from the pending mask. The event is
 608          * delivered immediately in tick_do_broadcast() !
 609          */
 610         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 611
 612         /* Take care of enforced broadcast requests */
 613         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 614         cpumask_clear(tick_broadcast_force_mask);
 615
 616         /*
 617          * Sanity check. Catch the case where we try to broadcast to
 618          * offline cpus.
 619          */
 620         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 621                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 622
 623         /*
 624          * Wakeup the cpus which have an expired event.
 625          */
 626         tick_do_broadcast(tmpmask);
 627
 628         /*
 629          * Two reasons for reprogram:
 630          *
 631          * - The global event did not expire any CPU local
 632          * events. This happens in dyntick mode, as the maximum PIT
 633          * delta is quite small.
 634          *
 635          * - There are pending events on sleeping CPUs which were not
 636          * in the event mask
 637          */
 638         if (next_event.tv64 != KTIME_MAX) {
 639                 /*
 640                  * Rearm the broadcast device. If event expired,
 641                  * repeat the above
 642                  */
 643                 if (tick_broadcast_set_event(dev, next_cpu, next_event, 0))
 644                         goto again;
 645         }
 646         raw_spin_unlock(&tick_broadcast_lock);
 647 }
 648
 649 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 650 {
 651         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 652                 return 0;
 653         if (bc->next_event.tv64 == KTIME_MAX)
 654                 return 0;
 655         return bc->bound_on == cpu ? -EBUSY : 0;
 656 }
 657
 658 static void broadcast_shutdown_local(struct clock_event_device *bc,
 659                                      struct clock_event_device *dev)
 660 {
 661         /*
 662          * For hrtimer based broadcasting we cannot shutdown the cpu
 663          * local device if our own event is the first one to expire or
 664          * if we own the broadcast timer.
 665          */
 666         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 667                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 668                         return;
 669                 if (dev->next_event.tv64 < bc->next_event.tv64)
 670                         return;
 671         }
 672         clockevents_set_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 673 }
 674
 675 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 676 {
 677         struct clock_event_device *bc;
 678         unsigned long flags;
 679
 680         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 681         bc = tick_broadcast_device.evtdev;
 682
 683         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 684                 /* This moves the broadcast assignment to this CPU: */
 685                 clockevents_program_event(bc, bc->next_event, 1);
 686         }
 687         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 688 }
 689
 690 /**
 691  * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
 692  * @state:      The target state (enter/exit)
 693  *
 694  * The system enters/leaves a state, where affected devices might stop
 695  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
 696  *
 697  * Called with interrupts disabled, so clockevents_lock is not
 698  * required here because the local clock event device cannot go away
 699  * under us.
 700  */
 701 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 702 {
 703         struct clock_event_device *bc, *dev;
 704         struct tick_device *td;
 705         int cpu, ret = 0;
 706         ktime_t now;
 707
 708         /*
 709          * Periodic mode does not care about the enter/exit of power
 710          * states
 711          */
 712         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 713                 return 0;
 714
 715         /*
 716          * We are called with preemtion disabled from the depth of the
 717          * idle code, so we can't be moved away.
 718          */
 719         td = this_cpu_ptr(&tick_cpu_device);
 720         dev = td->evtdev;
 721
 722         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 723                 return 0;
 724
 725         raw_spin_lock(&tick_broadcast_lock);
 726         bc = tick_broadcast_device.evtdev;
 727         cpu = smp_processor_id();
 728
 729         if (state == TICK_BROADCAST_ENTER) {
 730                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 731                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 732                         broadcast_shutdown_local(bc, dev);
 733                         /*
 734                          * We only reprogram the broadcast timer if we
 735                          * did not mark ourself in the force mask and
 736                          * if the cpu local event is earlier than the
 737                          * broadcast event. If the current CPU is in
 738                          * the force mask, then we are going to be
 739                          * woken by the IPI right away.
 740                          */
 741                         if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
 742                             dev->next_event.tv64 < bc->next_event.tv64)
 743                                 tick_broadcast_set_event(bc, cpu, dev->next_event, 1);
 744                 }
 745                 /*
 746                  * If the current CPU owns the hrtimer broadcast
 747                  * mechanism, it cannot go deep idle and we remove the
 748                  * CPU from the broadcast mask. We don't have to go
 749                  * through the EXIT path as the local timer is not
 750                  * shutdown.
 751                  */
 752                 ret = broadcast_needs_cpu(bc, cpu);
 753                 if (ret)
 754                         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 755         } else {
 756                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 757                         clockevents_set_state(dev, CLOCK_EVT_STATE_ONESHOT);
 758                         /*
 759                          * The cpu which was handling the broadcast
 760                          * timer marked this cpu in the broadcast
 761                          * pending mask and fired the broadcast
 762                          * IPI. So we are going to handle the expired
 763                          * event anyway via the broadcast IPI
 764                          * handler. No need to reprogram the timer
 765                          * with an already expired event.
 766                          */
 767                         if (cpumask_test_and_clear_cpu(cpu,
 768                                        tick_broadcast_pending_mask))
 769                                 goto out;
 770
 771                         /*
 772                          * Bail out if there is no next event.
 773                          */
 774                         if (dev->next_event.tv64 == KTIME_MAX)
 775                                 goto out;
 776                         /*
 777                          * If the pending bit is not set, then we are
 778                          * either the CPU handling the broadcast
 779                          * interrupt or we got woken by something else.
 780                          *
 781                          * We are not longer in the broadcast mask, so
 782                          * if the cpu local expiry time is already
 783                          * reached, we would reprogram the cpu local
 784                          * timer with an already expired event.
 785                          *
 786                          * This can lead to a ping-pong when we return
 787                          * to idle and therefor rearm the broadcast
 788                          * timer before the cpu local timer was able
 789                          * to fire. This happens because the forced
 790                          * reprogramming makes sure that the event
 791                          * will happen in the future and depending on
 792                          * the min_delta setting this might be far
 793                          * enough out that the ping-pong starts.
 794                          *
 795                          * If the cpu local next_event has expired
 796                          * then we know that the broadcast timer
 797                          * next_event has expired as well and
 798                          * broadcast is about to be handled. So we
 799                          * avoid reprogramming and enforce that the
 800                          * broadcast handler, which did not run yet,
 801                          * will invoke the cpu local handler.
 802                          *
 803                          * We cannot call the handler directly from
 804                          * here, because we might be in a NOHZ phase
 805                          * and we did not go through the irq_enter()
 806                          * nohz fixups.
 807                          */
 808                         now = ktime_get();
 809                         if (dev->next_event.tv64 <= now.tv64) {
 810                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 811                                 goto out;
 812                         }
 813                         /*
 814                          * We got woken by something else. Reprogram
 815                          * the cpu local timer device.
 816                          */
 817                         tick_program_event(dev->next_event, 1);
 818                 }
 819         }
 820 out:
 821         raw_spin_unlock(&tick_broadcast_lock);
 822         return ret;
 823 }
 824 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
 825
 826 /*
 827  * Reset the one shot broadcast for a cpu
 828  *
 829  * Called with tick_broadcast_lock held
 830  */
 831 static void tick_broadcast_clear_oneshot(int cpu)
 832 {
 833         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 834         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 835 }
 836
 837 static void tick_broadcast_init_next_event(struct cpumask *mask,
 838                                            ktime_t expires)
 839 {
 840         struct tick_device *td;
 841         int cpu;
 842
 843         for_each_cpu(cpu, mask) {
 844                 td = &per_cpu(tick_cpu_device, cpu);
 845                 if (td->evtdev)
 846                         td->evtdev->next_event = expires;
 847         }
 848 }
 849
 850 /**
 851  * tick_broadcast_setup_oneshot - setup the broadcast device
 852  */
 853 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 854 {
 855         int cpu = smp_processor_id();
 856
 857         /* Set it up only once ! */
 858         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 859                 int was_periodic = bc->state == CLOCK_EVT_STATE_PERIODIC;
 860
 861                 bc->event_handler = tick_handle_oneshot_broadcast;
 862
 863                 /*
 864                  * We must be careful here. There might be other CPUs
 865                  * waiting for periodic broadcast. We need to set the
 866                  * oneshot_mask bits for those and program the
 867                  * broadcast device to fire.
 868                  */
 869                 cpumask_copy(tmpmask, tick_broadcast_mask);
 870                 cpumask_clear_cpu(cpu, tmpmask);
 871                 cpumask_or(tick_broadcast_oneshot_mask,
 872                            tick_broadcast_oneshot_mask, tmpmask);
 873
 874                 if (was_periodic && !cpumask_empty(tmpmask)) {
 875                         clockevents_set_state(bc, CLOCK_EVT_STATE_ONESHOT);
 876                         tick_broadcast_init_next_event(tmpmask,
 877                                                        tick_next_period);
 878                         tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
 879                 } else
 880                         bc->next_event.tv64 = KTIME_MAX;
 881         } else {
 882                 /*
 883                  * The first cpu which switches to oneshot mode sets
 884                  * the bit for all other cpus which are in the general
 885                  * (periodic) broadcast mask. So the bit is set and
 886                  * would prevent the first broadcast enter after this
 887                  * to program the bc device.
 888                  */
 889                 tick_broadcast_clear_oneshot(cpu);
 890         }
 891 }
 892
 893 /*
 894  * Select oneshot operating mode for the broadcast device
 895  */
 896 void tick_broadcast_switch_to_oneshot(void)
 897 {
 898         struct clock_event_device *bc;
 899         unsigned long flags;
 900
 901         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 902
 903         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 904         bc = tick_broadcast_device.evtdev;
 905         if (bc)
 906                 tick_broadcast_setup_oneshot(bc);
 907
 908         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 909 }
 910
 911
 912 /*
 913  * Remove a dead CPU from broadcasting
 914  */
 915 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
 916 {
 917         unsigned long flags;
 918         unsigned int cpu = *cpup;
 919
 920         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 921
 922         /*
 923          * Clear the broadcast masks for the dead cpu, but do not stop
 924          * the broadcast device!
 925          */
 926         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 927         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 928         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 929
 930         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 931 }
 932
 933 /*
 934  * Check, whether the broadcast device is in one shot mode
 935  */
 936 int tick_broadcast_oneshot_active(void)
 937 {
 938         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 939 }
 940
 941 /*
 942  * Check whether the broadcast device supports oneshot.
 943  */
 944 bool tick_broadcast_oneshot_available(void)
 945 {
 946         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 947
 948         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 949 }
 950
 951 #endif
 952
 953 void __init tick_broadcast_init(void)
 954 {
 955         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 956         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
 957         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
 958 #ifdef CONFIG_TICK_ONESHOT
 959         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
 960         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
 961         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
 962 #endif
 963 }