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