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 __cpumask_var_read_mostly;
  33 static cpumask_var_t tick_broadcast_on __cpumask_var_read_mostly;
  34 static cpumask_var_t tmpmask __cpumask_var_read_mostly;
  35 static int tick_broadcast_forced;
  36
  37 static __cacheline_aligned_in_smp DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  38
  39 #ifdef CONFIG_TICK_ONESHOT
  40 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc);
  41 static void tick_broadcast_clear_oneshot(int cpu);
  42 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
  43 #else
  44 static inline void tick_broadcast_setup_oneshot(struct clock_event_device *bc) { BUG(); }
  45 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  46 static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
  47 #endif
  48
  49 /*
  50  * Debugging: see timer_list.c
  51  */
  52 struct tick_device *tick_get_broadcast_device(void)
  53 {
  54         return &tick_broadcast_device;
  55 }
  56
  57 struct cpumask *tick_get_broadcast_mask(void)
  58 {
  59         return tick_broadcast_mask;
  60 }
  61
  62 /*
  63  * Start the device in periodic mode
  64  */
  65 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  66 {
  67         if (bc)
  68                 tick_setup_periodic(bc, 1);
  69 }
  70
  71 /*
  72  * Check, if the device can be utilized as broadcast device:
  73  */
  74 static bool tick_check_broadcast_device(struct clock_event_device *curdev,
  75                                         struct clock_event_device *newdev)
  76 {
  77         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
  78             (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
  79             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
  80                 return false;
  81
  82         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
  83             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
  84                 return false;
  85
  86         return !curdev || newdev->rating > curdev->rating;
  87 }
  88
  89 /*
  90  * Conditionally install/replace broadcast device
  91  */
  92 void tick_install_broadcast_device(struct clock_event_device *dev)
  93 {
  94         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  95
  96         if (!tick_check_broadcast_device(cur, dev))
  97                 return;
  98
  99         if (!try_module_get(dev->owner))
 100                 return;
 101
 102         clockevents_exchange_device(cur, dev);
 103         if (cur)
 104                 cur->event_handler = clockevents_handle_noop;
 105         tick_broadcast_device.evtdev = dev;
 106         if (!cpumask_empty(tick_broadcast_mask))
 107                 tick_broadcast_start_periodic(dev);
 108         /*
 109          * Inform all cpus about this. We might be in a situation
 110          * where we did not switch to oneshot mode because the per cpu
 111          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
 112          * of a oneshot capable broadcast device. Without that
 113          * notification the systems stays stuck in periodic mode
 114          * forever.
 115          */
 116         if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
 117                 tick_clock_notify();
 118 }
 119
 120 /*
 121  * Check, if the device is the broadcast device
 122  */
 123 int tick_is_broadcast_device(struct clock_event_device *dev)
 124 {
 125         return (dev && tick_broadcast_device.evtdev == dev);
 126 }
 127
 128 int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
 129 {
 130         int ret = -ENODEV;
 131
 132         if (tick_is_broadcast_device(dev)) {
 133                 raw_spin_lock(&tick_broadcast_lock);
 134                 ret = __clockevents_update_freq(dev, freq);
 135                 raw_spin_unlock(&tick_broadcast_lock);
 136         }
 137         return ret;
 138 }
 139
 140
 141 static void err_broadcast(const struct cpumask *mask)
 142 {
 143         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 144 }
 145
 146 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 147 {
 148         if (!dev->broadcast)
 149                 dev->broadcast = tick_broadcast;
 150         if (!dev->broadcast) {
 151                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 152                              dev->name);
 153                 dev->broadcast = err_broadcast;
 154         }
 155 }
 156
 157 /*
 158  * Check, if the device is disfunctional and a place holder, which
 159  * needs to be handled by the broadcast device.
 160  */
 161 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 162 {
 163         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 164         unsigned long flags;
 165         int ret = 0;
 166
 167         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 168
 169         /*
 170          * Devices might be registered with both periodic and oneshot
 171          * mode disabled. This signals, that the device needs to be
 172          * operated from the broadcast device and is a placeholder for
 173          * the cpu local device.
 174          */
 175         if (!tick_device_is_functional(dev)) {
 176                 dev->event_handler = tick_handle_periodic;
 177                 tick_device_setup_broadcast_func(dev);
 178                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 179                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 180                         tick_broadcast_start_periodic(bc);
 181                 else
 182                         tick_broadcast_setup_oneshot(bc);
 183                 ret = 1;
 184         } else {
 185                 /*
 186                  * Clear the broadcast bit for this cpu if the
 187                  * device is not power state affected.
 188                  */
 189                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 190                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 191                 else
 192                         tick_device_setup_broadcast_func(dev);
 193
 194                 /*
 195                  * Clear the broadcast bit if the CPU is not in
 196                  * periodic broadcast on state.
 197                  */
 198                 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
 199                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 200
 201                 switch (tick_broadcast_device.mode) {
 202                 case TICKDEV_MODE_ONESHOT:
 203                         /*
 204                          * If the system is in oneshot mode we can
 205                          * unconditionally clear the oneshot mask bit,
 206                          * because the CPU is running and therefore
 207                          * not in an idle state which causes the power
 208                          * state affected device to stop. Let the
 209                          * caller initialize the device.
 210                          */
 211                         tick_broadcast_clear_oneshot(cpu);
 212                         ret = 0;
 213                         break;
 214
 215                 case TICKDEV_MODE_PERIODIC:
 216                         /*
 217                          * If the system is in periodic mode, check
 218                          * whether the broadcast device can be
 219                          * switched off now.
 220                          */
 221                         if (cpumask_empty(tick_broadcast_mask) && bc)
 222                                 clockevents_shutdown(bc);
 223                         /*
 224                          * If we kept the cpu in the broadcast mask,
 225                          * tell the caller to leave the per cpu device
 226                          * in shutdown state. The periodic interrupt
 227                          * is delivered by the broadcast device, if
 228                          * the broadcast device exists and is not
 229                          * hrtimer based.
 230                          */
 231                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 232                                 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 233                         break;
 234                 default:
 235                         break;
 236                 }
 237         }
 238         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 239         return ret;
 240 }
 241
 242 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 243 int tick_receive_broadcast(void)
 244 {
 245         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 246         struct clock_event_device *evt = td->evtdev;
 247
 248         if (!evt)
 249                 return -ENODEV;
 250
 251         if (!evt->event_handler)
 252                 return -EINVAL;
 253
 254         evt->event_handler(evt);
 255         return 0;
 256 }
 257 #endif
 258
 259 /*
 260  * Broadcast the event to the cpus, which are set in the mask (mangled).
 261  */
 262 static bool tick_do_broadcast(struct cpumask *mask)
 263 {
 264         int cpu = smp_processor_id();
 265         struct tick_device *td;
 266         bool local = false;
 267
 268         /*
 269          * Check, if the current cpu is in the mask
 270          */
 271         if (cpumask_test_cpu(cpu, mask)) {
 272                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 273
 274                 cpumask_clear_cpu(cpu, mask);
 275                 /*
 276                  * We only run the local handler, if the broadcast
 277                  * device is not hrtimer based. Otherwise we run into
 278                  * a hrtimer recursion.
 279                  *
 280                  * local timer_interrupt()
 281                  *   local_handler()
 282                  *     expire_hrtimers()
 283                  *       bc_handler()
 284                  *         local_handler()
 285                  *           expire_hrtimers()
 286                  */
 287                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 288         }
 289
 290         if (!cpumask_empty(mask)) {
 291                 /*
 292                  * It might be necessary to actually check whether the devices
 293                  * have different broadcast functions. For now, just use the
 294                  * one of the first device. This works as long as we have this
 295                  * misfeature only on x86 (lapic)
 296                  */
 297                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 298                 td->evtdev->broadcast(mask);
 299         }
 300         return local;
 301 }
 302
 303 /*
 304  * Periodic broadcast:
 305  * - invoke the broadcast handlers
 306  */
 307 static bool tick_do_periodic_broadcast(void)
 308 {
 309         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 310         return tick_do_broadcast(tmpmask);
 311 }
 312
 313 /*
 314  * Event handler for periodic broadcast ticks
 315  */
 316 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 317 {
 318         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 319         bool bc_local;
 320
 321         raw_spin_lock(&tick_broadcast_lock);
 322
 323         /* Handle spurious interrupts gracefully */
 324         if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
 325                 raw_spin_unlock(&tick_broadcast_lock);
 326                 return;
 327         }
 328
 329         bc_local = tick_do_periodic_broadcast();
 330
 331         if (clockevent_state_oneshot(dev)) {
 332                 ktime_t next = ktime_add(dev->next_event, tick_period);
 333
 334                 clockevents_program_event(dev, next, true);
 335         }
 336         raw_spin_unlock(&tick_broadcast_lock);
 337
 338         /*
 339          * We run the handler of the local cpu after dropping
 340          * tick_broadcast_lock because the handler might deadlock when
 341          * trying to switch to oneshot mode.
 342          */
 343         if (bc_local)
 344                 td->evtdev->event_handler(td->evtdev);
 345 }
 346
 347 /**
 348  * tick_broadcast_control - Enable/disable or force broadcast mode
 349  * @mode:       The selected broadcast mode
 350  *
 351  * Called when the system enters a state where affected tick devices
 352  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 353  */
 354 void tick_broadcast_control(enum tick_broadcast_mode mode)
 355 {
 356         struct clock_event_device *bc, *dev;
 357         struct tick_device *td;
 358         int cpu, bc_stopped;
 359         unsigned long flags;
 360
 361         /* Protects also the local clockevent device. */
 362         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 363         td = this_cpu_ptr(&tick_cpu_device);
 364         dev = td->evtdev;
 365
 366         /*
 367          * Is the device not affected by the powerstate ?
 368          */
 369         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 370                 goto out;
 371
 372         if (!tick_device_is_functional(dev))
 373                 goto out;
 374
 375         cpu = smp_processor_id();
 376         bc = tick_broadcast_device.evtdev;
 377         bc_stopped = cpumask_empty(tick_broadcast_mask);
 378
 379         switch (mode) {
 380         case TICK_BROADCAST_FORCE:
 381                 tick_broadcast_forced = 1;
 382         case TICK_BROADCAST_ON:
 383                 cpumask_set_cpu(cpu, tick_broadcast_on);
 384                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 385                         /*
 386                          * Only shutdown the cpu local device, if:
 387                          *
 388                          * - the broadcast device exists
 389                          * - the broadcast device is not a hrtimer based one
 390                          * - the broadcast device is in periodic mode to
 391                          *   avoid a hickup during switch to oneshot mode
 392                          */
 393                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 394                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 395                                 clockevents_shutdown(dev);
 396                 }
 397                 break;
 398
 399         case TICK_BROADCAST_OFF:
 400                 if (tick_broadcast_forced)
 401                         break;
 402                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 403                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 404                         if (tick_broadcast_device.mode ==
 405                             TICKDEV_MODE_PERIODIC)
 406                                 tick_setup_periodic(dev, 0);
 407                 }
 408                 break;
 409         }
 410
 411         if (bc) {
 412                 if (cpumask_empty(tick_broadcast_mask)) {
 413                         if (!bc_stopped)
 414                                 clockevents_shutdown(bc);
 415                 } else if (bc_stopped) {
 416                         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 417                                 tick_broadcast_start_periodic(bc);
 418                         else
 419                                 tick_broadcast_setup_oneshot(bc);
 420                 }
 421         }
 422 out:
 423         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 424 }
 425 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 426
 427 /*
 428  * Set the periodic handler depending on broadcast on/off
 429  */
 430 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 431 {
 432         if (!broadcast)
 433                 dev->event_handler = tick_handle_periodic;
 434         else
 435                 dev->event_handler = tick_handle_periodic_broadcast;
 436 }
 437
 438 #ifdef CONFIG_HOTPLUG_CPU
 439 /*
 440  * Remove a CPU from broadcasting
 441  */
 442 void tick_shutdown_broadcast(unsigned int cpu)
 443 {
 444         struct clock_event_device *bc;
 445         unsigned long flags;
 446
 447         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 448
 449         bc = tick_broadcast_device.evtdev;
 450         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 451         cpumask_clear_cpu(cpu, tick_broadcast_on);
 452
 453         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 454                 if (bc && cpumask_empty(tick_broadcast_mask))
 455                         clockevents_shutdown(bc);
 456         }
 457
 458         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 459 }
 460 #endif
 461
 462 void tick_suspend_broadcast(void)
 463 {
 464         struct clock_event_device *bc;
 465         unsigned long flags;
 466
 467         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 468
 469         bc = tick_broadcast_device.evtdev;
 470         if (bc)
 471                 clockevents_shutdown(bc);
 472
 473         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 474 }
 475
 476 /*
 477  * This is called from tick_resume_local() on a resuming CPU. That's
 478  * called from the core resume function, tick_unfreeze() and the magic XEN
 479  * resume hackery.
 480  *
 481  * In none of these cases the broadcast device mode can change and the
 482  * bit of the resuming CPU in the broadcast mask is safe as well.
 483  */
 484 bool tick_resume_check_broadcast(void)
 485 {
 486         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 487                 return false;
 488         else
 489                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 490 }
 491
 492 void tick_resume_broadcast(void)
 493 {
 494         struct clock_event_device *bc;
 495         unsigned long flags;
 496
 497         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 498
 499         bc = tick_broadcast_device.evtdev;
 500
 501         if (bc) {
 502                 clockevents_tick_resume(bc);
 503
 504                 switch (tick_broadcast_device.mode) {
 505                 case TICKDEV_MODE_PERIODIC:
 506                         if (!cpumask_empty(tick_broadcast_mask))
 507                                 tick_broadcast_start_periodic(bc);
 508                         break;
 509                 case TICKDEV_MODE_ONESHOT:
 510                         if (!cpumask_empty(tick_broadcast_mask))
 511                                 tick_resume_broadcast_oneshot(bc);
 512                         break;
 513                 }
 514         }
 515         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 516 }
 517
 518 #ifdef CONFIG_TICK_ONESHOT
 519
 520 static cpumask_var_t tick_broadcast_oneshot_mask __cpumask_var_read_mostly;
 521 static cpumask_var_t tick_broadcast_pending_mask __cpumask_var_read_mostly;
 522 static cpumask_var_t tick_broadcast_force_mask __cpumask_var_read_mostly;
 523
 524 /*
 525  * Exposed for debugging: see timer_list.c
 526  */
 527 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 528 {
 529         return tick_broadcast_oneshot_mask;
 530 }
 531
 532 /*
 533  * Called before going idle with interrupts disabled. Checks whether a
 534  * broadcast event from the other core is about to happen. We detected
 535  * that in tick_broadcast_oneshot_control(). The callsite can use this
 536  * to avoid a deep idle transition as we are about to get the
 537  * broadcast IPI right away.
 538  */
 539 int tick_check_broadcast_expired(void)
 540 {
 541         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 542 }
 543
 544 /*
 545  * Set broadcast interrupt affinity
 546  */
 547 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 548                                         const struct cpumask *cpumask)
 549 {
 550         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 551                 return;
 552
 553         if (cpumask_equal(bc->cpumask, cpumask))
 554                 return;
 555
 556         bc->cpumask = cpumask;
 557         irq_set_affinity(bc->irq, bc->cpumask);
 558 }
 559
 560 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 561                                      ktime_t expires)
 562 {
 563         if (!clockevent_state_oneshot(bc))
 564                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 565
 566         clockevents_program_event(bc, expires, 1);
 567         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 568 }
 569
 570 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 571 {
 572         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 573 }
 574
 575 /*
 576  * Called from irq_enter() when idle was interrupted to reenable the
 577  * per cpu device.
 578  */
 579 void tick_check_oneshot_broadcast_this_cpu(void)
 580 {
 581         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 582                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 583
 584                 /*
 585                  * We might be in the middle of switching over from
 586                  * periodic to oneshot. If the CPU has not yet
 587                  * switched over, leave the device alone.
 588                  */
 589                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 590                         clockevents_switch_state(td->evtdev,
 591                                               CLOCK_EVT_STATE_ONESHOT);
 592                 }
 593         }
 594 }
 595
 596 /*
 597  * Handle oneshot mode broadcasting
 598  */
 599 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 600 {
 601         struct tick_device *td;
 602         ktime_t now, next_event;
 603         int cpu, next_cpu = 0;
 604         bool bc_local;
 605
 606         raw_spin_lock(&tick_broadcast_lock);
 607         dev->next_event = KTIME_MAX;
 608         next_event = KTIME_MAX;
 609         cpumask_clear(tmpmask);
 610         now = ktime_get();
 611         /* Find all expired events */
 612         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 613                 /*
 614                  * Required for !SMP because for_each_cpu() reports
 615                  * unconditionally CPU0 as set on UP kernels.
 616                  */
 617                 if (!IS_ENABLED(CONFIG_SMP) &&
 618                     cpumask_empty(tick_broadcast_oneshot_mask))
 619                         break;
 620
 621                 td = &per_cpu(tick_cpu_device, cpu);
 622                 if (td->evtdev->next_event <= now) {
 623                         cpumask_set_cpu(cpu, tmpmask);
 624                         /*
 625                          * Mark the remote cpu in the pending mask, so
 626                          * it can avoid reprogramming the cpu local
 627                          * timer in tick_broadcast_oneshot_control().
 628                          */
 629                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 630                 } else if (td->evtdev->next_event < next_event) {
 631                         next_event = td->evtdev->next_event;
 632                         next_cpu = cpu;
 633                 }
 634         }
 635
 636         /*
 637          * Remove the current cpu from the pending mask. The event is
 638          * delivered immediately in tick_do_broadcast() !
 639          */
 640         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 641
 642         /* Take care of enforced broadcast requests */
 643         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 644         cpumask_clear(tick_broadcast_force_mask);
 645
 646         /*
 647          * Sanity check. Catch the case where we try to broadcast to
 648          * offline cpus.
 649          */
 650         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 651                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 652
 653         /*
 654          * Wakeup the cpus which have an expired event.
 655          */
 656         bc_local = tick_do_broadcast(tmpmask);
 657
 658         /*
 659          * Two reasons for reprogram:
 660          *
 661          * - The global event did not expire any CPU local
 662          * events. This happens in dyntick mode, as the maximum PIT
 663          * delta is quite small.
 664          *
 665          * - There are pending events on sleeping CPUs which were not
 666          * in the event mask
 667          */
 668         if (next_event != KTIME_MAX)
 669                 tick_broadcast_set_event(dev, next_cpu, next_event);
 670
 671         raw_spin_unlock(&tick_broadcast_lock);
 672
 673         if (bc_local) {
 674                 td = this_cpu_ptr(&tick_cpu_device);
 675                 td->evtdev->event_handler(td->evtdev);
 676         }
 677 }
 678
 679 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 680 {
 681         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 682                 return 0;
 683         if (bc->next_event == KTIME_MAX)
 684                 return 0;
 685         return bc->bound_on == cpu ? -EBUSY : 0;
 686 }
 687
 688 static void broadcast_shutdown_local(struct clock_event_device *bc,
 689                                      struct clock_event_device *dev)
 690 {
 691         /*
 692          * For hrtimer based broadcasting we cannot shutdown the cpu
 693          * local device if our own event is the first one to expire or
 694          * if we own the broadcast timer.
 695          */
 696         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 697                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 698                         return;
 699                 if (dev->next_event < bc->next_event)
 700                         return;
 701         }
 702         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 703 }
 704
 705 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 706 {
 707         struct clock_event_device *bc, *dev;
 708         int cpu, ret = 0;
 709         ktime_t now;
 710
 711         /*
 712          * If there is no broadcast device, tell the caller not to go
 713          * into deep idle.
 714          */
 715         if (!tick_broadcast_device.evtdev)
 716                 return -EBUSY;
 717
 718         dev = this_cpu_ptr(&tick_cpu_device)->evtdev;
 719
 720         raw_spin_lock(&tick_broadcast_lock);
 721         bc = tick_broadcast_device.evtdev;
 722         cpu = smp_processor_id();
 723
 724         if (state == TICK_BROADCAST_ENTER) {
 725                 /*
 726                  * If the current CPU owns the hrtimer broadcast
 727                  * mechanism, it cannot go deep idle and we do not add
 728                  * the CPU to the broadcast mask. We don't have to go
 729                  * through the EXIT path as the local timer is not
 730                  * shutdown.
 731                  */
 732                 ret = broadcast_needs_cpu(bc, cpu);
 733                 if (ret)
 734                         goto out;
 735
 736                 /*
 737                  * If the broadcast device is in periodic mode, we
 738                  * return.
 739                  */
 740                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 741                         /* If it is a hrtimer based broadcast, return busy */
 742                         if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
 743                                 ret = -EBUSY;
 744                         goto out;
 745                 }
 746
 747                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 748                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 749
 750                         /* Conditionally shut down the local timer. */
 751                         broadcast_shutdown_local(bc, dev);
 752
 753                         /*
 754                          * We only reprogram the broadcast timer if we
 755                          * did not mark ourself in the force mask and
 756                          * if the cpu local event is earlier than the
 757                          * broadcast event. If the current CPU is in
 758                          * the force mask, then we are going to be
 759                          * woken by the IPI right away; we return
 760                          * busy, so the CPU does not try to go deep
 761                          * idle.
 762                          */
 763                         if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
 764                                 ret = -EBUSY;
 765                         } else if (dev->next_event < bc->next_event) {
 766                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 767                                 /*
 768                                  * In case of hrtimer broadcasts the
 769                                  * programming might have moved the
 770                                  * timer to this cpu. If yes, remove
 771                                  * us from the broadcast mask and
 772                                  * return busy.
 773                                  */
 774                                 ret = broadcast_needs_cpu(bc, cpu);
 775                                 if (ret) {
 776                                         cpumask_clear_cpu(cpu,
 777                                                 tick_broadcast_oneshot_mask);
 778                                 }
 779                         }
 780                 }
 781         } else {
 782                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 783                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 784                         /*
 785                          * The cpu which was handling the broadcast
 786                          * timer marked this cpu in the broadcast
 787                          * pending mask and fired the broadcast
 788                          * IPI. So we are going to handle the expired
 789                          * event anyway via the broadcast IPI
 790                          * handler. No need to reprogram the timer
 791                          * with an already expired event.
 792                          */
 793                         if (cpumask_test_and_clear_cpu(cpu,
 794                                        tick_broadcast_pending_mask))
 795                                 goto out;
 796
 797                         /*
 798                          * Bail out if there is no next event.
 799                          */
 800                         if (dev->next_event == KTIME_MAX)
 801                                 goto out;
 802                         /*
 803                          * If the pending bit is not set, then we are
 804                          * either the CPU handling the broadcast
 805                          * interrupt or we got woken by something else.
 806                          *
 807                          * We are not longer in the broadcast mask, so
 808                          * if the cpu local expiry time is already
 809                          * reached, we would reprogram the cpu local
 810                          * timer with an already expired event.
 811                          *
 812                          * This can lead to a ping-pong when we return
 813                          * to idle and therefor rearm the broadcast
 814                          * timer before the cpu local timer was able
 815                          * to fire. This happens because the forced
 816                          * reprogramming makes sure that the event
 817                          * will happen in the future and depending on
 818                          * the min_delta setting this might be far
 819                          * enough out that the ping-pong starts.
 820                          *
 821                          * If the cpu local next_event has expired
 822                          * then we know that the broadcast timer
 823                          * next_event has expired as well and
 824                          * broadcast is about to be handled. So we
 825                          * avoid reprogramming and enforce that the
 826                          * broadcast handler, which did not run yet,
 827                          * will invoke the cpu local handler.
 828                          *
 829                          * We cannot call the handler directly from
 830                          * here, because we might be in a NOHZ phase
 831                          * and we did not go through the irq_enter()
 832                          * nohz fixups.
 833                          */
 834                         now = ktime_get();
 835                         if (dev->next_event <= now) {
 836                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 837                                 goto out;
 838                         }
 839                         /*
 840                          * We got woken by something else. Reprogram
 841                          * the cpu local timer device.
 842                          */
 843                         tick_program_event(dev->next_event, 1);
 844                 }
 845         }
 846 out:
 847         raw_spin_unlock(&tick_broadcast_lock);
 848         return ret;
 849 }
 850
 851 /*
 852  * Reset the one shot broadcast for a cpu
 853  *
 854  * Called with tick_broadcast_lock held
 855  */
 856 static void tick_broadcast_clear_oneshot(int cpu)
 857 {
 858         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 859         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 860 }
 861
 862 static void tick_broadcast_init_next_event(struct cpumask *mask,
 863                                            ktime_t expires)
 864 {
 865         struct tick_device *td;
 866         int cpu;
 867
 868         for_each_cpu(cpu, mask) {
 869                 td = &per_cpu(tick_cpu_device, cpu);
 870                 if (td->evtdev)
 871                         td->evtdev->next_event = expires;
 872         }
 873 }
 874
 875 /**
 876  * tick_broadcast_setup_oneshot - setup the broadcast device
 877  */
 878 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 879 {
 880         int cpu = smp_processor_id();
 881
 882         if (!bc)
 883                 return;
 884
 885         /* Set it up only once ! */
 886         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 887                 int was_periodic = clockevent_state_periodic(bc);
 888
 889                 bc->event_handler = tick_handle_oneshot_broadcast;
 890
 891                 /*
 892                  * We must be careful here. There might be other CPUs
 893                  * waiting for periodic broadcast. We need to set the
 894                  * oneshot_mask bits for those and program the
 895                  * broadcast device to fire.
 896                  */
 897                 cpumask_copy(tmpmask, tick_broadcast_mask);
 898                 cpumask_clear_cpu(cpu, tmpmask);
 899                 cpumask_or(tick_broadcast_oneshot_mask,
 900                            tick_broadcast_oneshot_mask, tmpmask);
 901
 902                 if (was_periodic && !cpumask_empty(tmpmask)) {
 903                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 904                         tick_broadcast_init_next_event(tmpmask,
 905                                                        tick_next_period);
 906                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 907                 } else
 908                         bc->next_event = KTIME_MAX;
 909         } else {
 910                 /*
 911                  * The first cpu which switches to oneshot mode sets
 912                  * the bit for all other cpus which are in the general
 913                  * (periodic) broadcast mask. So the bit is set and
 914                  * would prevent the first broadcast enter after this
 915                  * to program the bc device.
 916                  */
 917                 tick_broadcast_clear_oneshot(cpu);
 918         }
 919 }
 920
 921 /*
 922  * Select oneshot operating mode for the broadcast device
 923  */
 924 void tick_broadcast_switch_to_oneshot(void)
 925 {
 926         struct clock_event_device *bc;
 927         unsigned long flags;
 928
 929         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 930
 931         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 932         bc = tick_broadcast_device.evtdev;
 933         if (bc)
 934                 tick_broadcast_setup_oneshot(bc);
 935
 936         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 937 }
 938
 939 #ifdef CONFIG_HOTPLUG_CPU
 940 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 941 {
 942         struct clock_event_device *bc;
 943         unsigned long flags;
 944
 945         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 946         bc = tick_broadcast_device.evtdev;
 947
 948         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 949                 /* This moves the broadcast assignment to this CPU: */
 950                 clockevents_program_event(bc, bc->next_event, 1);
 951         }
 952         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 953 }
 954
 955 /*
 956  * Remove a dead CPU from broadcasting
 957  */
 958 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 959 {
 960         unsigned long flags;
 961
 962         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 963
 964         /*
 965          * Clear the broadcast masks for the dead cpu, but do not stop
 966          * the broadcast device!
 967          */
 968         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 969         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 970         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 971
 972         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 973 }
 974 #endif
 975
 976 /*
 977  * Check, whether the broadcast device is in one shot mode
 978  */
 979 int tick_broadcast_oneshot_active(void)
 980 {
 981         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 982 }
 983
 984 /*
 985  * Check whether the broadcast device supports oneshot.
 986  */
 987 bool tick_broadcast_oneshot_available(void)
 988 {
 989         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 990
 991         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 992 }
 993
 994 #else
 995 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 996 {
 997         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 998
 999         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
1000                 return -EBUSY;
1001
1002         return 0;
1003 }
1004 #endif
1005
1006 void __init tick_broadcast_init(void)
1007 {
1008         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
1009         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
1010         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1011 #ifdef CONFIG_TICK_ONESHOT
1012         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1013         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1014         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1015 #endif
1016 }