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 bool tick_do_broadcast(struct cpumask *mask)
 259 {
 260         int cpu = smp_processor_id();
 261         struct tick_device *td;
 262         bool local = false;
 263
 264         /*
 265          * Check, if the current cpu is in the mask
 266          */
 267         if (cpumask_test_cpu(cpu, mask)) {
 268                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 269
 270                 cpumask_clear_cpu(cpu, mask);
 271                 /*
 272                  * We only run the local handler, if the broadcast
 273                  * device is not hrtimer based. Otherwise we run into
 274                  * a hrtimer recursion.
 275                  *
 276                  * local timer_interrupt()
 277                  *   local_handler()
 278                  *     expire_hrtimers()
 279                  *       bc_handler()
 280                  *         local_handler()
 281                  *           expire_hrtimers()
 282                  */
 283                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 284         }
 285
 286         if (!cpumask_empty(mask)) {
 287                 /*
 288                  * It might be necessary to actually check whether the devices
 289                  * have different broadcast functions. For now, just use the
 290                  * one of the first device. This works as long as we have this
 291                  * misfeature only on x86 (lapic)
 292                  */
 293                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 294                 td->evtdev->broadcast(mask);
 295         }
 296         return local;
 297 }
 298
 299 /*
 300  * Periodic broadcast:
 301  * - invoke the broadcast handlers
 302  */
 303 static bool tick_do_periodic_broadcast(void)
 304 {
 305         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 306         return tick_do_broadcast(tmpmask);
 307 }
 308
 309 /*
 310  * Event handler for periodic broadcast ticks
 311  */
 312 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 313 {
 314         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 315         bool bc_local;
 316
 317         raw_spin_lock(&tick_broadcast_lock);
 318         bc_local = tick_do_periodic_broadcast();
 319
 320         if (clockevent_state_oneshot(dev)) {
 321                 ktime_t next = ktime_add(dev->next_event, tick_period);
 322
 323                 clockevents_program_event(dev, next, true);
 324         }
 325         raw_spin_unlock(&tick_broadcast_lock);
 326
 327         /*
 328          * We run the handler of the local cpu after dropping
 329          * tick_broadcast_lock because the handler might deadlock when
 330          * trying to switch to oneshot mode.
 331          */
 332         if (bc_local)
 333                 td->evtdev->event_handler(td->evtdev);
 334 }
 335
 336 /**
 337  * tick_broadcast_control - Enable/disable or force broadcast mode
 338  * @mode:       The selected broadcast mode
 339  *
 340  * Called when the system enters a state where affected tick devices
 341  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 342  *
 343  * Called with interrupts disabled, so clockevents_lock is not
 344  * required here because the local clock event device cannot go away
 345  * under us.
 346  */
 347 void tick_broadcast_control(enum tick_broadcast_mode mode)
 348 {
 349         struct clock_event_device *bc, *dev;
 350         struct tick_device *td;
 351         int cpu, bc_stopped;
 352
 353         td = this_cpu_ptr(&tick_cpu_device);
 354         dev = td->evtdev;
 355
 356         /*
 357          * Is the device not affected by the powerstate ?
 358          */
 359         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 360                 return;
 361
 362         if (!tick_device_is_functional(dev))
 363                 return;
 364
 365         raw_spin_lock(&tick_broadcast_lock);
 366         cpu = smp_processor_id();
 367         bc = tick_broadcast_device.evtdev;
 368         bc_stopped = cpumask_empty(tick_broadcast_mask);
 369
 370         switch (mode) {
 371         case TICK_BROADCAST_FORCE:
 372                 tick_broadcast_forced = 1;
 373         case TICK_BROADCAST_ON:
 374                 cpumask_set_cpu(cpu, tick_broadcast_on);
 375                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 376                         if (tick_broadcast_device.mode ==
 377                             TICKDEV_MODE_PERIODIC)
 378                                 clockevents_shutdown(dev);
 379                 }
 380                 break;
 381
 382         case TICK_BROADCAST_OFF:
 383                 if (tick_broadcast_forced)
 384                         break;
 385                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 386                 if (!tick_device_is_functional(dev))
 387                         break;
 388                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 389                         if (tick_broadcast_device.mode ==
 390                             TICKDEV_MODE_PERIODIC)
 391                                 tick_setup_periodic(dev, 0);
 392                 }
 393                 break;
 394         }
 395
 396         if (cpumask_empty(tick_broadcast_mask)) {
 397                 if (!bc_stopped)
 398                         clockevents_shutdown(bc);
 399         } else if (bc_stopped) {
 400                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 401                         tick_broadcast_start_periodic(bc);
 402                 else
 403                         tick_broadcast_setup_oneshot(bc);
 404         }
 405         raw_spin_unlock(&tick_broadcast_lock);
 406 }
 407 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 408
 409 /*
 410  * Set the periodic handler depending on broadcast on/off
 411  */
 412 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 413 {
 414         if (!broadcast)
 415                 dev->event_handler = tick_handle_periodic;
 416         else
 417                 dev->event_handler = tick_handle_periodic_broadcast;
 418 }
 419
 420 #ifdef CONFIG_HOTPLUG_CPU
 421 /*
 422  * Remove a CPU from broadcasting
 423  */
 424 void tick_shutdown_broadcast(unsigned int cpu)
 425 {
 426         struct clock_event_device *bc;
 427         unsigned long flags;
 428
 429         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 430
 431         bc = tick_broadcast_device.evtdev;
 432         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 433         cpumask_clear_cpu(cpu, tick_broadcast_on);
 434
 435         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 436                 if (bc && cpumask_empty(tick_broadcast_mask))
 437                         clockevents_shutdown(bc);
 438         }
 439
 440         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 441 }
 442 #endif
 443
 444 void tick_suspend_broadcast(void)
 445 {
 446         struct clock_event_device *bc;
 447         unsigned long flags;
 448
 449         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 450
 451         bc = tick_broadcast_device.evtdev;
 452         if (bc)
 453                 clockevents_shutdown(bc);
 454
 455         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 456 }
 457
 458 /*
 459  * This is called from tick_resume_local() on a resuming CPU. That's
 460  * called from the core resume function, tick_unfreeze() and the magic XEN
 461  * resume hackery.
 462  *
 463  * In none of these cases the broadcast device mode can change and the
 464  * bit of the resuming CPU in the broadcast mask is safe as well.
 465  */
 466 bool tick_resume_check_broadcast(void)
 467 {
 468         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 469                 return false;
 470         else
 471                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 472 }
 473
 474 void tick_resume_broadcast(void)
 475 {
 476         struct clock_event_device *bc;
 477         unsigned long flags;
 478
 479         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 480
 481         bc = tick_broadcast_device.evtdev;
 482
 483         if (bc) {
 484                 clockevents_tick_resume(bc);
 485
 486                 switch (tick_broadcast_device.mode) {
 487                 case TICKDEV_MODE_PERIODIC:
 488                         if (!cpumask_empty(tick_broadcast_mask))
 489                                 tick_broadcast_start_periodic(bc);
 490                         break;
 491                 case TICKDEV_MODE_ONESHOT:
 492                         if (!cpumask_empty(tick_broadcast_mask))
 493                                 tick_resume_broadcast_oneshot(bc);
 494                         break;
 495                 }
 496         }
 497         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 498 }
 499
 500 #ifdef CONFIG_TICK_ONESHOT
 501
 502 static cpumask_var_t tick_broadcast_oneshot_mask;
 503 static cpumask_var_t tick_broadcast_pending_mask;
 504 static cpumask_var_t tick_broadcast_force_mask;
 505
 506 /*
 507  * Exposed for debugging: see timer_list.c
 508  */
 509 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 510 {
 511         return tick_broadcast_oneshot_mask;
 512 }
 513
 514 /*
 515  * Called before going idle with interrupts disabled. Checks whether a
 516  * broadcast event from the other core is about to happen. We detected
 517  * that in tick_broadcast_oneshot_control(). The callsite can use this
 518  * to avoid a deep idle transition as we are about to get the
 519  * broadcast IPI right away.
 520  */
 521 int tick_check_broadcast_expired(void)
 522 {
 523         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 524 }
 525
 526 /*
 527  * Set broadcast interrupt affinity
 528  */
 529 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 530                                         const struct cpumask *cpumask)
 531 {
 532         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 533                 return;
 534
 535         if (cpumask_equal(bc->cpumask, cpumask))
 536                 return;
 537
 538         bc->cpumask = cpumask;
 539         irq_set_affinity(bc->irq, bc->cpumask);
 540 }
 541
 542 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 543                                      ktime_t expires)
 544 {
 545         if (!clockevent_state_oneshot(bc))
 546                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 547
 548         clockevents_program_event(bc, expires, 1);
 549         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 550 }
 551
 552 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 553 {
 554         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 555 }
 556
 557 /*
 558  * Called from irq_enter() when idle was interrupted to reenable the
 559  * per cpu device.
 560  */
 561 void tick_check_oneshot_broadcast_this_cpu(void)
 562 {
 563         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 564                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 565
 566                 /*
 567                  * We might be in the middle of switching over from
 568                  * periodic to oneshot. If the CPU has not yet
 569                  * switched over, leave the device alone.
 570                  */
 571                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 572                         clockevents_switch_state(td->evtdev,
 573                                               CLOCK_EVT_STATE_ONESHOT);
 574                 }
 575         }
 576 }
 577
 578 /*
 579  * Handle oneshot mode broadcasting
 580  */
 581 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 582 {
 583         struct tick_device *td;
 584         ktime_t now, next_event;
 585         int cpu, next_cpu = 0;
 586         bool bc_local;
 587
 588         raw_spin_lock(&tick_broadcast_lock);
 589         dev->next_event.tv64 = KTIME_MAX;
 590         next_event.tv64 = KTIME_MAX;
 591         cpumask_clear(tmpmask);
 592         now = ktime_get();
 593         /* Find all expired events */
 594         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 595                 td = &per_cpu(tick_cpu_device, cpu);
 596                 if (td->evtdev->next_event.tv64 <= now.tv64) {
 597                         cpumask_set_cpu(cpu, tmpmask);
 598                         /*
 599                          * Mark the remote cpu in the pending mask, so
 600                          * it can avoid reprogramming the cpu local
 601                          * timer in tick_broadcast_oneshot_control().
 602                          */
 603                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 604                 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
 605                         next_event.tv64 = td->evtdev->next_event.tv64;
 606                         next_cpu = cpu;
 607                 }
 608         }
 609
 610         /*
 611          * Remove the current cpu from the pending mask. The event is
 612          * delivered immediately in tick_do_broadcast() !
 613          */
 614         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 615
 616         /* Take care of enforced broadcast requests */
 617         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 618         cpumask_clear(tick_broadcast_force_mask);
 619
 620         /*
 621          * Sanity check. Catch the case where we try to broadcast to
 622          * offline cpus.
 623          */
 624         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 625                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 626
 627         /*
 628          * Wakeup the cpus which have an expired event.
 629          */
 630         bc_local = tick_do_broadcast(tmpmask);
 631
 632         /*
 633          * Two reasons for reprogram:
 634          *
 635          * - The global event did not expire any CPU local
 636          * events. This happens in dyntick mode, as the maximum PIT
 637          * delta is quite small.
 638          *
 639          * - There are pending events on sleeping CPUs which were not
 640          * in the event mask
 641          */
 642         if (next_event.tv64 != KTIME_MAX)
 643                 tick_broadcast_set_event(dev, next_cpu, next_event);
 644
 645         raw_spin_unlock(&tick_broadcast_lock);
 646
 647         if (bc_local) {
 648                 td = this_cpu_ptr(&tick_cpu_device);
 649                 td->evtdev->event_handler(td->evtdev);
 650         }
 651 }
 652
 653 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 654 {
 655         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 656                 return 0;
 657         if (bc->next_event.tv64 == KTIME_MAX)
 658                 return 0;
 659         return bc->bound_on == cpu ? -EBUSY : 0;
 660 }
 661
 662 static void broadcast_shutdown_local(struct clock_event_device *bc,
 663                                      struct clock_event_device *dev)
 664 {
 665         /*
 666          * For hrtimer based broadcasting we cannot shutdown the cpu
 667          * local device if our own event is the first one to expire or
 668          * if we own the broadcast timer.
 669          */
 670         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 671                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 672                         return;
 673                 if (dev->next_event.tv64 < bc->next_event.tv64)
 674                         return;
 675         }
 676         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 677 }
 678
 679 /**
 680  * tick_broadcast_oneshot_control - Enter/exit broadcast oneshot mode
 681  * @state:      The target state (enter/exit)
 682  *
 683  * The system enters/leaves a state, where affected devices might stop
 684  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
 685  *
 686  * Called with interrupts disabled, so clockevents_lock is not
 687  * required here because the local clock event device cannot go away
 688  * under us.
 689  */
 690 int tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 691 {
 692         struct clock_event_device *bc, *dev;
 693         struct tick_device *td;
 694         int cpu, ret = 0;
 695         ktime_t now;
 696
 697         /*
 698          * Periodic mode does not care about the enter/exit of power
 699          * states
 700          */
 701         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 702                 return 0;
 703
 704         /*
 705          * We are called with preemtion disabled from the depth of the
 706          * idle code, so we can't be moved away.
 707          */
 708         td = this_cpu_ptr(&tick_cpu_device);
 709         dev = td->evtdev;
 710
 711         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 712                 return 0;
 713
 714         raw_spin_lock(&tick_broadcast_lock);
 715         bc = tick_broadcast_device.evtdev;
 716         cpu = smp_processor_id();
 717
 718         if (state == TICK_BROADCAST_ENTER) {
 719                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 720                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 721                         broadcast_shutdown_local(bc, dev);
 722                         /*
 723                          * We only reprogram the broadcast timer if we
 724                          * did not mark ourself in the force mask and
 725                          * if the cpu local event is earlier than the
 726                          * broadcast event. If the current CPU is in
 727                          * the force mask, then we are going to be
 728                          * woken by the IPI right away.
 729                          */
 730                         if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
 731                             dev->next_event.tv64 < bc->next_event.tv64)
 732                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 733                 }
 734                 /*
 735                  * If the current CPU owns the hrtimer broadcast
 736                  * mechanism, it cannot go deep idle and we remove the
 737                  * CPU from the broadcast mask. We don't have to go
 738                  * through the EXIT path as the local timer is not
 739                  * shutdown.
 740                  */
 741                 ret = broadcast_needs_cpu(bc, cpu);
 742                 if (ret)
 743                         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 744         } else {
 745                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 746                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 747                         /*
 748                          * The cpu which was handling the broadcast
 749                          * timer marked this cpu in the broadcast
 750                          * pending mask and fired the broadcast
 751                          * IPI. So we are going to handle the expired
 752                          * event anyway via the broadcast IPI
 753                          * handler. No need to reprogram the timer
 754                          * with an already expired event.
 755                          */
 756                         if (cpumask_test_and_clear_cpu(cpu,
 757                                        tick_broadcast_pending_mask))
 758                                 goto out;
 759
 760                         /*
 761                          * Bail out if there is no next event.
 762                          */
 763                         if (dev->next_event.tv64 == KTIME_MAX)
 764                                 goto out;
 765                         /*
 766                          * If the pending bit is not set, then we are
 767                          * either the CPU handling the broadcast
 768                          * interrupt or we got woken by something else.
 769                          *
 770                          * We are not longer in the broadcast mask, so
 771                          * if the cpu local expiry time is already
 772                          * reached, we would reprogram the cpu local
 773                          * timer with an already expired event.
 774                          *
 775                          * This can lead to a ping-pong when we return
 776                          * to idle and therefor rearm the broadcast
 777                          * timer before the cpu local timer was able
 778                          * to fire. This happens because the forced
 779                          * reprogramming makes sure that the event
 780                          * will happen in the future and depending on
 781                          * the min_delta setting this might be far
 782                          * enough out that the ping-pong starts.
 783                          *
 784                          * If the cpu local next_event has expired
 785                          * then we know that the broadcast timer
 786                          * next_event has expired as well and
 787                          * broadcast is about to be handled. So we
 788                          * avoid reprogramming and enforce that the
 789                          * broadcast handler, which did not run yet,
 790                          * will invoke the cpu local handler.
 791                          *
 792                          * We cannot call the handler directly from
 793                          * here, because we might be in a NOHZ phase
 794                          * and we did not go through the irq_enter()
 795                          * nohz fixups.
 796                          */
 797                         now = ktime_get();
 798                         if (dev->next_event.tv64 <= now.tv64) {
 799                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 800                                 goto out;
 801                         }
 802                         /*
 803                          * We got woken by something else. Reprogram
 804                          * the cpu local timer device.
 805                          */
 806                         tick_program_event(dev->next_event, 1);
 807                 }
 808         }
 809 out:
 810         raw_spin_unlock(&tick_broadcast_lock);
 811         return ret;
 812 }
 813 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
 814
 815 /*
 816  * Reset the one shot broadcast for a cpu
 817  *
 818  * Called with tick_broadcast_lock held
 819  */
 820 static void tick_broadcast_clear_oneshot(int cpu)
 821 {
 822         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 823         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 824 }
 825
 826 static void tick_broadcast_init_next_event(struct cpumask *mask,
 827                                            ktime_t expires)
 828 {
 829         struct tick_device *td;
 830         int cpu;
 831
 832         for_each_cpu(cpu, mask) {
 833                 td = &per_cpu(tick_cpu_device, cpu);
 834                 if (td->evtdev)
 835                         td->evtdev->next_event = expires;
 836         }
 837 }
 838
 839 /**
 840  * tick_broadcast_setup_oneshot - setup the broadcast device
 841  */
 842 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 843 {
 844         int cpu = smp_processor_id();
 845
 846         /* Set it up only once ! */
 847         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 848                 int was_periodic = clockevent_state_periodic(bc);
 849
 850                 bc->event_handler = tick_handle_oneshot_broadcast;
 851
 852                 /*
 853                  * We must be careful here. There might be other CPUs
 854                  * waiting for periodic broadcast. We need to set the
 855                  * oneshot_mask bits for those and program the
 856                  * broadcast device to fire.
 857                  */
 858                 cpumask_copy(tmpmask, tick_broadcast_mask);
 859                 cpumask_clear_cpu(cpu, tmpmask);
 860                 cpumask_or(tick_broadcast_oneshot_mask,
 861                            tick_broadcast_oneshot_mask, tmpmask);
 862
 863                 if (was_periodic && !cpumask_empty(tmpmask)) {
 864                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 865                         tick_broadcast_init_next_event(tmpmask,
 866                                                        tick_next_period);
 867                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 868                 } else
 869                         bc->next_event.tv64 = KTIME_MAX;
 870         } else {
 871                 /*
 872                  * The first cpu which switches to oneshot mode sets
 873                  * the bit for all other cpus which are in the general
 874                  * (periodic) broadcast mask. So the bit is set and
 875                  * would prevent the first broadcast enter after this
 876                  * to program the bc device.
 877                  */
 878                 tick_broadcast_clear_oneshot(cpu);
 879         }
 880 }
 881
 882 /*
 883  * Select oneshot operating mode for the broadcast device
 884  */
 885 void tick_broadcast_switch_to_oneshot(void)
 886 {
 887         struct clock_event_device *bc;
 888         unsigned long flags;
 889
 890         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 891
 892         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 893         bc = tick_broadcast_device.evtdev;
 894         if (bc)
 895                 tick_broadcast_setup_oneshot(bc);
 896
 897         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 898 }
 899
 900 #ifdef CONFIG_HOTPLUG_CPU
 901 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 902 {
 903         struct clock_event_device *bc;
 904         unsigned long flags;
 905
 906         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 907         bc = tick_broadcast_device.evtdev;
 908
 909         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 910                 /* This moves the broadcast assignment to this CPU: */
 911                 clockevents_program_event(bc, bc->next_event, 1);
 912         }
 913         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 914 }
 915
 916 /*
 917  * Remove a dead CPU from broadcasting
 918  */
 919 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 920 {
 921         unsigned long flags;
 922
 923         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 924
 925         /*
 926          * Clear the broadcast masks for the dead cpu, but do not stop
 927          * the broadcast device!
 928          */
 929         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 930         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 931         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 932
 933         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 934 }
 935 #endif
 936
 937 /*
 938  * Check, whether the broadcast device is in one shot mode
 939  */
 940 int tick_broadcast_oneshot_active(void)
 941 {
 942         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 943 }
 944
 945 /*
 946  * Check whether the broadcast device supports oneshot.
 947  */
 948 bool tick_broadcast_oneshot_available(void)
 949 {
 950         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 951
 952         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 953 }
 954
 955 #endif
 956
 957 void __init tick_broadcast_init(void)
 958 {
 959         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 960         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
 961         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
 962 #ifdef CONFIG_TICK_ONESHOT
 963         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
 964         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
 965         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
 966 #endif
 967 }