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 = 0;
 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, if
 225                          * the broadcast device exists and is not
 226                          * hrtimer based.
 227                          */
 228                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 229                                 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 230                         break;
 231                 default:
 232                         break;
 233                 }
 234         }
 235         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 236         return ret;
 237 }
 238
 239 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 240 int tick_receive_broadcast(void)
 241 {
 242         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 243         struct clock_event_device *evt = td->evtdev;
 244
 245         if (!evt)
 246                 return -ENODEV;
 247
 248         if (!evt->event_handler)
 249                 return -EINVAL;
 250
 251         evt->event_handler(evt);
 252         return 0;
 253 }
 254 #endif
 255
 256 /*
 257  * Broadcast the event to the cpus, which are set in the mask (mangled).
 258  */
 259 static bool tick_do_broadcast(struct cpumask *mask)
 260 {
 261         int cpu = smp_processor_id();
 262         struct tick_device *td;
 263         bool local = false;
 264
 265         /*
 266          * Check, if the current cpu is in the mask
 267          */
 268         if (cpumask_test_cpu(cpu, mask)) {
 269                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 270
 271                 cpumask_clear_cpu(cpu, mask);
 272                 /*
 273                  * We only run the local handler, if the broadcast
 274                  * device is not hrtimer based. Otherwise we run into
 275                  * a hrtimer recursion.
 276                  *
 277                  * local timer_interrupt()
 278                  *   local_handler()
 279                  *     expire_hrtimers()
 280                  *       bc_handler()
 281                  *         local_handler()
 282                  *           expire_hrtimers()
 283                  */
 284                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 285         }
 286
 287         if (!cpumask_empty(mask)) {
 288                 /*
 289                  * It might be necessary to actually check whether the devices
 290                  * have different broadcast functions. For now, just use the
 291                  * one of the first device. This works as long as we have this
 292                  * misfeature only on x86 (lapic)
 293                  */
 294                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 295                 td->evtdev->broadcast(mask);
 296         }
 297         return local;
 298 }
 299
 300 /*
 301  * Periodic broadcast:
 302  * - invoke the broadcast handlers
 303  */
 304 static bool tick_do_periodic_broadcast(void)
 305 {
 306         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 307         return tick_do_broadcast(tmpmask);
 308 }
 309
 310 /*
 311  * Event handler for periodic broadcast ticks
 312  */
 313 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 314 {
 315         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 316         bool bc_local;
 317
 318         raw_spin_lock(&tick_broadcast_lock);
 319         bc_local = tick_do_periodic_broadcast();
 320
 321         if (clockevent_state_oneshot(dev)) {
 322                 ktime_t next = ktime_add(dev->next_event, tick_period);
 323
 324                 clockevents_program_event(dev, next, true);
 325         }
 326         raw_spin_unlock(&tick_broadcast_lock);
 327
 328         /*
 329          * We run the handler of the local cpu after dropping
 330          * tick_broadcast_lock because the handler might deadlock when
 331          * trying to switch to oneshot mode.
 332          */
 333         if (bc_local)
 334                 td->evtdev->event_handler(td->evtdev);
 335 }
 336
 337 /**
 338  * tick_broadcast_control - Enable/disable or force broadcast mode
 339  * @mode:       The selected broadcast mode
 340  *
 341  * Called when the system enters a state where affected tick devices
 342  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 343  *
 344  * Called with interrupts disabled, so clockevents_lock is not
 345  * required here because the local clock event device cannot go away
 346  * under us.
 347  */
 348 void tick_broadcast_control(enum tick_broadcast_mode mode)
 349 {
 350         struct clock_event_device *bc, *dev;
 351         struct tick_device *td;
 352         int cpu, bc_stopped;
 353
 354         td = this_cpu_ptr(&tick_cpu_device);
 355         dev = td->evtdev;
 356
 357         /*
 358          * Is the device not affected by the powerstate ?
 359          */
 360         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 361                 return;
 362
 363         if (!tick_device_is_functional(dev))
 364                 return;
 365
 366         raw_spin_lock(&tick_broadcast_lock);
 367         cpu = smp_processor_id();
 368         bc = tick_broadcast_device.evtdev;
 369         bc_stopped = cpumask_empty(tick_broadcast_mask);
 370
 371         switch (mode) {
 372         case TICK_BROADCAST_FORCE:
 373                 tick_broadcast_forced = 1;
 374         case TICK_BROADCAST_ON:
 375                 cpumask_set_cpu(cpu, tick_broadcast_on);
 376                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 377                         /*
 378                          * Only shutdown the cpu local device, if:
 379                          *
 380                          * - the broadcast device exists
 381                          * - the broadcast device is not a hrtimer based one
 382                          * - the broadcast device is in periodic mode to
 383                          *   avoid a hickup during switch to oneshot mode
 384                          */
 385                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 386                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 387                                 clockevents_shutdown(dev);
 388                 }
 389                 break;
 390
 391         case TICK_BROADCAST_OFF:
 392                 if (tick_broadcast_forced)
 393                         break;
 394                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 395                 if (!tick_device_is_functional(dev))
 396                         break;
 397                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 398                         if (tick_broadcast_device.mode ==
 399                             TICKDEV_MODE_PERIODIC)
 400                                 tick_setup_periodic(dev, 0);
 401                 }
 402                 break;
 403         }
 404
 405         if (cpumask_empty(tick_broadcast_mask)) {
 406                 if (!bc_stopped)
 407                         clockevents_shutdown(bc);
 408         } else if (bc_stopped) {
 409                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 410                         tick_broadcast_start_periodic(bc);
 411                 else
 412                         tick_broadcast_setup_oneshot(bc);
 413         }
 414         raw_spin_unlock(&tick_broadcast_lock);
 415 }
 416 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 417
 418 /*
 419  * Set the periodic handler depending on broadcast on/off
 420  */
 421 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 422 {
 423         if (!broadcast)
 424                 dev->event_handler = tick_handle_periodic;
 425         else
 426                 dev->event_handler = tick_handle_periodic_broadcast;
 427 }
 428
 429 #ifdef CONFIG_HOTPLUG_CPU
 430 /*
 431  * Remove a CPU from broadcasting
 432  */
 433 void tick_shutdown_broadcast(unsigned int cpu)
 434 {
 435         struct clock_event_device *bc;
 436         unsigned long flags;
 437
 438         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 439
 440         bc = tick_broadcast_device.evtdev;
 441         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 442         cpumask_clear_cpu(cpu, tick_broadcast_on);
 443
 444         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 445                 if (bc && cpumask_empty(tick_broadcast_mask))
 446                         clockevents_shutdown(bc);
 447         }
 448
 449         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 450 }
 451 #endif
 452
 453 void tick_suspend_broadcast(void)
 454 {
 455         struct clock_event_device *bc;
 456         unsigned long flags;
 457
 458         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 459
 460         bc = tick_broadcast_device.evtdev;
 461         if (bc)
 462                 clockevents_shutdown(bc);
 463
 464         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 465 }
 466
 467 /*
 468  * This is called from tick_resume_local() on a resuming CPU. That's
 469  * called from the core resume function, tick_unfreeze() and the magic XEN
 470  * resume hackery.
 471  *
 472  * In none of these cases the broadcast device mode can change and the
 473  * bit of the resuming CPU in the broadcast mask is safe as well.
 474  */
 475 bool tick_resume_check_broadcast(void)
 476 {
 477         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 478                 return false;
 479         else
 480                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 481 }
 482
 483 void tick_resume_broadcast(void)
 484 {
 485         struct clock_event_device *bc;
 486         unsigned long flags;
 487
 488         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 489
 490         bc = tick_broadcast_device.evtdev;
 491
 492         if (bc) {
 493                 clockevents_tick_resume(bc);
 494
 495                 switch (tick_broadcast_device.mode) {
 496                 case TICKDEV_MODE_PERIODIC:
 497                         if (!cpumask_empty(tick_broadcast_mask))
 498                                 tick_broadcast_start_periodic(bc);
 499                         break;
 500                 case TICKDEV_MODE_ONESHOT:
 501                         if (!cpumask_empty(tick_broadcast_mask))
 502                                 tick_resume_broadcast_oneshot(bc);
 503                         break;
 504                 }
 505         }
 506         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 507 }
 508
 509 #ifdef CONFIG_TICK_ONESHOT
 510
 511 static cpumask_var_t tick_broadcast_oneshot_mask;
 512 static cpumask_var_t tick_broadcast_pending_mask;
 513 static cpumask_var_t tick_broadcast_force_mask;
 514
 515 /*
 516  * Exposed for debugging: see timer_list.c
 517  */
 518 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 519 {
 520         return tick_broadcast_oneshot_mask;
 521 }
 522
 523 /*
 524  * Called before going idle with interrupts disabled. Checks whether a
 525  * broadcast event from the other core is about to happen. We detected
 526  * that in tick_broadcast_oneshot_control(). The callsite can use this
 527  * to avoid a deep idle transition as we are about to get the
 528  * broadcast IPI right away.
 529  */
 530 int tick_check_broadcast_expired(void)
 531 {
 532         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 533 }
 534
 535 /*
 536  * Set broadcast interrupt affinity
 537  */
 538 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 539                                         const struct cpumask *cpumask)
 540 {
 541         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 542                 return;
 543
 544         if (cpumask_equal(bc->cpumask, cpumask))
 545                 return;
 546
 547         bc->cpumask = cpumask;
 548         irq_set_affinity(bc->irq, bc->cpumask);
 549 }
 550
 551 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 552                                      ktime_t expires)
 553 {
 554         if (!clockevent_state_oneshot(bc))
 555                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 556
 557         clockevents_program_event(bc, expires, 1);
 558         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 559 }
 560
 561 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 562 {
 563         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 564 }
 565
 566 /*
 567  * Called from irq_enter() when idle was interrupted to reenable the
 568  * per cpu device.
 569  */
 570 void tick_check_oneshot_broadcast_this_cpu(void)
 571 {
 572         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 573                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 574
 575                 /*
 576                  * We might be in the middle of switching over from
 577                  * periodic to oneshot. If the CPU has not yet
 578                  * switched over, leave the device alone.
 579                  */
 580                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 581                         clockevents_switch_state(td->evtdev,
 582                                               CLOCK_EVT_STATE_ONESHOT);
 583                 }
 584         }
 585 }
 586
 587 /*
 588  * Handle oneshot mode broadcasting
 589  */
 590 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 591 {
 592         struct tick_device *td;
 593         ktime_t now, next_event;
 594         int cpu, next_cpu = 0;
 595         bool bc_local;
 596
 597         raw_spin_lock(&tick_broadcast_lock);
 598         dev->next_event.tv64 = KTIME_MAX;
 599         next_event.tv64 = KTIME_MAX;
 600         cpumask_clear(tmpmask);
 601         now = ktime_get();
 602         /* Find all expired events */
 603         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 604                 td = &per_cpu(tick_cpu_device, cpu);
 605                 if (td->evtdev->next_event.tv64 <= now.tv64) {
 606                         cpumask_set_cpu(cpu, tmpmask);
 607                         /*
 608                          * Mark the remote cpu in the pending mask, so
 609                          * it can avoid reprogramming the cpu local
 610                          * timer in tick_broadcast_oneshot_control().
 611                          */
 612                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 613                 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
 614                         next_event.tv64 = td->evtdev->next_event.tv64;
 615                         next_cpu = cpu;
 616                 }
 617         }
 618
 619         /*
 620          * Remove the current cpu from the pending mask. The event is
 621          * delivered immediately in tick_do_broadcast() !
 622          */
 623         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 624
 625         /* Take care of enforced broadcast requests */
 626         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 627         cpumask_clear(tick_broadcast_force_mask);
 628
 629         /*
 630          * Sanity check. Catch the case where we try to broadcast to
 631          * offline cpus.
 632          */
 633         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 634                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 635
 636         /*
 637          * Wakeup the cpus which have an expired event.
 638          */
 639         bc_local = tick_do_broadcast(tmpmask);
 640
 641         /*
 642          * Two reasons for reprogram:
 643          *
 644          * - The global event did not expire any CPU local
 645          * events. This happens in dyntick mode, as the maximum PIT
 646          * delta is quite small.
 647          *
 648          * - There are pending events on sleeping CPUs which were not
 649          * in the event mask
 650          */
 651         if (next_event.tv64 != KTIME_MAX)
 652                 tick_broadcast_set_event(dev, next_cpu, next_event);
 653
 654         raw_spin_unlock(&tick_broadcast_lock);
 655
 656         if (bc_local) {
 657                 td = this_cpu_ptr(&tick_cpu_device);
 658                 td->evtdev->event_handler(td->evtdev);
 659         }
 660 }
 661
 662 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 663 {
 664         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 665                 return 0;
 666         if (bc->next_event.tv64 == KTIME_MAX)
 667                 return 0;
 668         return bc->bound_on == cpu ? -EBUSY : 0;
 669 }
 670
 671 static void broadcast_shutdown_local(struct clock_event_device *bc,
 672                                      struct clock_event_device *dev)
 673 {
 674         /*
 675          * For hrtimer based broadcasting we cannot shutdown the cpu
 676          * local device if our own event is the first one to expire or
 677          * if we own the broadcast timer.
 678          */
 679         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 680                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 681                         return;
 682                 if (dev->next_event.tv64 < bc->next_event.tv64)
 683                         return;
 684         }
 685         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 686 }
 687
 688 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 689 {
 690         struct clock_event_device *bc, *dev;
 691         int cpu, ret = 0;
 692         ktime_t now;
 693
 694         /*
 695          * If there is no broadcast device, tell the caller not to go
 696          * into deep idle.
 697          */
 698         if (!tick_broadcast_device.evtdev)
 699                 return -EBUSY;
 700
 701         dev = this_cpu_ptr(&tick_cpu_device)->evtdev;
 702
 703         raw_spin_lock(&tick_broadcast_lock);
 704         bc = tick_broadcast_device.evtdev;
 705         cpu = smp_processor_id();
 706
 707         if (state == TICK_BROADCAST_ENTER) {
 708                 /*
 709                  * If the broadcast device is in periodic mode, we
 710                  * return.
 711                  */
 712                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 713                         goto out;
 714
 715                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 716                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 717                         broadcast_shutdown_local(bc, dev);
 718                         /*
 719                          * We only reprogram the broadcast timer if we
 720                          * did not mark ourself in the force mask and
 721                          * if the cpu local event is earlier than the
 722                          * broadcast event. If the current CPU is in
 723                          * the force mask, then we are going to be
 724                          * woken by the IPI right away.
 725                          */
 726                         if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
 727                             dev->next_event.tv64 < bc->next_event.tv64)
 728                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 729                 }
 730                 /*
 731                  * If the current CPU owns the hrtimer broadcast
 732                  * mechanism, it cannot go deep idle and we remove the
 733                  * CPU from the broadcast mask. We don't have to go
 734                  * through the EXIT path as the local timer is not
 735                  * shutdown.
 736                  */
 737                 ret = broadcast_needs_cpu(bc, cpu);
 738                 if (ret)
 739                         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 740         } else {
 741                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 742                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 743                         /*
 744                          * The cpu which was handling the broadcast
 745                          * timer marked this cpu in the broadcast
 746                          * pending mask and fired the broadcast
 747                          * IPI. So we are going to handle the expired
 748                          * event anyway via the broadcast IPI
 749                          * handler. No need to reprogram the timer
 750                          * with an already expired event.
 751                          */
 752                         if (cpumask_test_and_clear_cpu(cpu,
 753                                        tick_broadcast_pending_mask))
 754                                 goto out;
 755
 756                         /*
 757                          * Bail out if there is no next event.
 758                          */
 759                         if (dev->next_event.tv64 == KTIME_MAX)
 760                                 goto out;
 761                         /*
 762                          * If the pending bit is not set, then we are
 763                          * either the CPU handling the broadcast
 764                          * interrupt or we got woken by something else.
 765                          *
 766                          * We are not longer in the broadcast mask, so
 767                          * if the cpu local expiry time is already
 768                          * reached, we would reprogram the cpu local
 769                          * timer with an already expired event.
 770                          *
 771                          * This can lead to a ping-pong when we return
 772                          * to idle and therefor rearm the broadcast
 773                          * timer before the cpu local timer was able
 774                          * to fire. This happens because the forced
 775                          * reprogramming makes sure that the event
 776                          * will happen in the future and depending on
 777                          * the min_delta setting this might be far
 778                          * enough out that the ping-pong starts.
 779                          *
 780                          * If the cpu local next_event has expired
 781                          * then we know that the broadcast timer
 782                          * next_event has expired as well and
 783                          * broadcast is about to be handled. So we
 784                          * avoid reprogramming and enforce that the
 785                          * broadcast handler, which did not run yet,
 786                          * will invoke the cpu local handler.
 787                          *
 788                          * We cannot call the handler directly from
 789                          * here, because we might be in a NOHZ phase
 790                          * and we did not go through the irq_enter()
 791                          * nohz fixups.
 792                          */
 793                         now = ktime_get();
 794                         if (dev->next_event.tv64 <= now.tv64) {
 795                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 796                                 goto out;
 797                         }
 798                         /*
 799                          * We got woken by something else. Reprogram
 800                          * the cpu local timer device.
 801                          */
 802                         tick_program_event(dev->next_event, 1);
 803                 }
 804         }
 805 out:
 806         raw_spin_unlock(&tick_broadcast_lock);
 807         return ret;
 808 }
 809 EXPORT_SYMBOL_GPL(tick_broadcast_oneshot_control);
 810
 811 /*
 812  * Reset the one shot broadcast for a cpu
 813  *
 814  * Called with tick_broadcast_lock held
 815  */
 816 static void tick_broadcast_clear_oneshot(int cpu)
 817 {
 818         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 819         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 820 }
 821
 822 static void tick_broadcast_init_next_event(struct cpumask *mask,
 823                                            ktime_t expires)
 824 {
 825         struct tick_device *td;
 826         int cpu;
 827
 828         for_each_cpu(cpu, mask) {
 829                 td = &per_cpu(tick_cpu_device, cpu);
 830                 if (td->evtdev)
 831                         td->evtdev->next_event = expires;
 832         }
 833 }
 834
 835 /**
 836  * tick_broadcast_setup_oneshot - setup the broadcast device
 837  */
 838 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 839 {
 840         int cpu = smp_processor_id();
 841
 842         /* Set it up only once ! */
 843         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 844                 int was_periodic = clockevent_state_periodic(bc);
 845
 846                 bc->event_handler = tick_handle_oneshot_broadcast;
 847
 848                 /*
 849                  * We must be careful here. There might be other CPUs
 850                  * waiting for periodic broadcast. We need to set the
 851                  * oneshot_mask bits for those and program the
 852                  * broadcast device to fire.
 853                  */
 854                 cpumask_copy(tmpmask, tick_broadcast_mask);
 855                 cpumask_clear_cpu(cpu, tmpmask);
 856                 cpumask_or(tick_broadcast_oneshot_mask,
 857                            tick_broadcast_oneshot_mask, tmpmask);
 858
 859                 if (was_periodic && !cpumask_empty(tmpmask)) {
 860                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 861                         tick_broadcast_init_next_event(tmpmask,
 862                                                        tick_next_period);
 863                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 864                 } else
 865                         bc->next_event.tv64 = KTIME_MAX;
 866         } else {
 867                 /*
 868                  * The first cpu which switches to oneshot mode sets
 869                  * the bit for all other cpus which are in the general
 870                  * (periodic) broadcast mask. So the bit is set and
 871                  * would prevent the first broadcast enter after this
 872                  * to program the bc device.
 873                  */
 874                 tick_broadcast_clear_oneshot(cpu);
 875         }
 876 }
 877
 878 /*
 879  * Select oneshot operating mode for the broadcast device
 880  */
 881 void tick_broadcast_switch_to_oneshot(void)
 882 {
 883         struct clock_event_device *bc;
 884         unsigned long flags;
 885
 886         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 887
 888         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 889         bc = tick_broadcast_device.evtdev;
 890         if (bc)
 891                 tick_broadcast_setup_oneshot(bc);
 892
 893         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 894 }
 895
 896 #ifdef CONFIG_HOTPLUG_CPU
 897 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 898 {
 899         struct clock_event_device *bc;
 900         unsigned long flags;
 901
 902         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 903         bc = tick_broadcast_device.evtdev;
 904
 905         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 906                 /* This moves the broadcast assignment to this CPU: */
 907                 clockevents_program_event(bc, bc->next_event, 1);
 908         }
 909         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 910 }
 911
 912 /*
 913  * Remove a dead CPU from broadcasting
 914  */
 915 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 916 {
 917         unsigned long flags;
 918
 919         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 920
 921         /*
 922          * Clear the broadcast masks for the dead cpu, but do not stop
 923          * the broadcast device!
 924          */
 925         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 926         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 927         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 928
 929         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 930 }
 931 #endif
 932
 933 /*
 934  * Check, whether the broadcast device is in one shot mode
 935  */
 936 int tick_broadcast_oneshot_active(void)
 937 {
 938         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 939 }
 940
 941 /*
 942  * Check whether the broadcast device supports oneshot.
 943  */
 944 bool tick_broadcast_oneshot_available(void)
 945 {
 946         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 947
 948         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 949 }
 950
 951 #else
 952 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 953 {
 954         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 955
 956         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
 957                 return -EBUSY;
 958
 959         return 0;
 960 }
 961 #endif
 962
 963 void __init tick_broadcast_init(void)
 964 {
 965         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 966         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
 967         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
 968 #ifdef CONFIG_TICK_ONESHOT
 969         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
 970         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
 971         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
 972 #endif
 973 }