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
 320         /* Handle spurious interrupts gracefully */
 321         if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
 322                 raw_spin_unlock(&tick_broadcast_lock);
 323                 return;
 324         }
 325
 326         bc_local = tick_do_periodic_broadcast();
 327
 328         if (clockevent_state_oneshot(dev)) {
 329                 ktime_t next = ktime_add(dev->next_event, tick_period);
 330
 331                 clockevents_program_event(dev, next, true);
 332         }
 333         raw_spin_unlock(&tick_broadcast_lock);
 334
 335         /*
 336          * We run the handler of the local cpu after dropping
 337          * tick_broadcast_lock because the handler might deadlock when
 338          * trying to switch to oneshot mode.
 339          */
 340         if (bc_local)
 341                 td->evtdev->event_handler(td->evtdev);
 342 }
 343
 344 /**
 345  * tick_broadcast_control - Enable/disable or force broadcast mode
 346  * @mode:       The selected broadcast mode
 347  *
 348  * Called when the system enters a state where affected tick devices
 349  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 350  *
 351  * Called with interrupts disabled, so clockevents_lock is not
 352  * required here because the local clock event device cannot go away
 353  * under us.
 354  */
 355 void tick_broadcast_control(enum tick_broadcast_mode mode)
 356 {
 357         struct clock_event_device *bc, *dev;
 358         struct tick_device *td;
 359         int cpu, bc_stopped;
 360
 361         td = this_cpu_ptr(&tick_cpu_device);
 362         dev = td->evtdev;
 363
 364         /*
 365          * Is the device not affected by the powerstate ?
 366          */
 367         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 368                 return;
 369
 370         if (!tick_device_is_functional(dev))
 371                 return;
 372
 373         raw_spin_lock(&tick_broadcast_lock);
 374         cpu = smp_processor_id();
 375         bc = tick_broadcast_device.evtdev;
 376         bc_stopped = cpumask_empty(tick_broadcast_mask);
 377
 378         switch (mode) {
 379         case TICK_BROADCAST_FORCE:
 380                 tick_broadcast_forced = 1;
 381         case TICK_BROADCAST_ON:
 382                 cpumask_set_cpu(cpu, tick_broadcast_on);
 383                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 384                         /*
 385                          * Only shutdown the cpu local device, if:
 386                          *
 387                          * - the broadcast device exists
 388                          * - the broadcast device is not a hrtimer based one
 389                          * - the broadcast device is in periodic mode to
 390                          *   avoid a hickup during switch to oneshot mode
 391                          */
 392                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 393                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 394                                 clockevents_shutdown(dev);
 395                 }
 396                 break;
 397
 398         case TICK_BROADCAST_OFF:
 399                 if (tick_broadcast_forced)
 400                         break;
 401                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 402                 if (!tick_device_is_functional(dev))
 403                         break;
 404                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 405                         if (tick_broadcast_device.mode ==
 406                             TICKDEV_MODE_PERIODIC)
 407                                 tick_setup_periodic(dev, 0);
 408                 }
 409                 break;
 410         }
 411
 412         if (bc) {
 413                 if (cpumask_empty(tick_broadcast_mask)) {
 414                         if (!bc_stopped)
 415                                 clockevents_shutdown(bc);
 416                 } else if (bc_stopped) {
 417                         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 418                                 tick_broadcast_start_periodic(bc);
 419                         else
 420                                 tick_broadcast_setup_oneshot(bc);
 421                 }
 422         }
 423         raw_spin_unlock(&tick_broadcast_lock);
 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;
 521 static cpumask_var_t tick_broadcast_pending_mask;
 522 static cpumask_var_t tick_broadcast_force_mask;
 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.tv64 = KTIME_MAX;
 608         next_event.tv64 = 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                 td = &per_cpu(tick_cpu_device, cpu);
 614                 if (td->evtdev->next_event.tv64 <= now.tv64) {
 615                         cpumask_set_cpu(cpu, tmpmask);
 616                         /*
 617                          * Mark the remote cpu in the pending mask, so
 618                          * it can avoid reprogramming the cpu local
 619                          * timer in tick_broadcast_oneshot_control().
 620                          */
 621                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 622                 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
 623                         next_event.tv64 = td->evtdev->next_event.tv64;
 624                         next_cpu = cpu;
 625                 }
 626         }
 627
 628         /*
 629          * Remove the current cpu from the pending mask. The event is
 630          * delivered immediately in tick_do_broadcast() !
 631          */
 632         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 633
 634         /* Take care of enforced broadcast requests */
 635         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 636         cpumask_clear(tick_broadcast_force_mask);
 637
 638         /*
 639          * Sanity check. Catch the case where we try to broadcast to
 640          * offline cpus.
 641          */
 642         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 643                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 644
 645         /*
 646          * Wakeup the cpus which have an expired event.
 647          */
 648         bc_local = tick_do_broadcast(tmpmask);
 649
 650         /*
 651          * Two reasons for reprogram:
 652          *
 653          * - The global event did not expire any CPU local
 654          * events. This happens in dyntick mode, as the maximum PIT
 655          * delta is quite small.
 656          *
 657          * - There are pending events on sleeping CPUs which were not
 658          * in the event mask
 659          */
 660         if (next_event.tv64 != KTIME_MAX)
 661                 tick_broadcast_set_event(dev, next_cpu, next_event);
 662
 663         raw_spin_unlock(&tick_broadcast_lock);
 664
 665         if (bc_local) {
 666                 td = this_cpu_ptr(&tick_cpu_device);
 667                 td->evtdev->event_handler(td->evtdev);
 668         }
 669 }
 670
 671 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 672 {
 673         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 674                 return 0;
 675         if (bc->next_event.tv64 == KTIME_MAX)
 676                 return 0;
 677         return bc->bound_on == cpu ? -EBUSY : 0;
 678 }
 679
 680 static void broadcast_shutdown_local(struct clock_event_device *bc,
 681                                      struct clock_event_device *dev)
 682 {
 683         /*
 684          * For hrtimer based broadcasting we cannot shutdown the cpu
 685          * local device if our own event is the first one to expire or
 686          * if we own the broadcast timer.
 687          */
 688         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 689                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 690                         return;
 691                 if (dev->next_event.tv64 < bc->next_event.tv64)
 692                         return;
 693         }
 694         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 695 }
 696
 697 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 698 {
 699         struct clock_event_device *bc, *dev;
 700         int cpu, ret = 0;
 701         ktime_t now;
 702
 703         /*
 704          * If there is no broadcast device, tell the caller not to go
 705          * into deep idle.
 706          */
 707         if (!tick_broadcast_device.evtdev)
 708                 return -EBUSY;
 709
 710         dev = this_cpu_ptr(&tick_cpu_device)->evtdev;
 711
 712         raw_spin_lock(&tick_broadcast_lock);
 713         bc = tick_broadcast_device.evtdev;
 714         cpu = smp_processor_id();
 715
 716         if (state == TICK_BROADCAST_ENTER) {
 717                 /*
 718                  * If the current CPU owns the hrtimer broadcast
 719                  * mechanism, it cannot go deep idle and we do not add
 720                  * the CPU to the broadcast mask. We don't have to go
 721                  * through the EXIT path as the local timer is not
 722                  * shutdown.
 723                  */
 724                 ret = broadcast_needs_cpu(bc, cpu);
 725                 if (ret)
 726                         goto out;
 727
 728                 /*
 729                  * If the broadcast device is in periodic mode, we
 730                  * return.
 731                  */
 732                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 733                         /* If it is a hrtimer based broadcast, return busy */
 734                         if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
 735                                 ret = -EBUSY;
 736                         goto out;
 737                 }
 738
 739                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 740                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 741
 742                         /* Conditionally shut down the local timer. */
 743                         broadcast_shutdown_local(bc, dev);
 744
 745                         /*
 746                          * We only reprogram the broadcast timer if we
 747                          * did not mark ourself in the force mask and
 748                          * if the cpu local event is earlier than the
 749                          * broadcast event. If the current CPU is in
 750                          * the force mask, then we are going to be
 751                          * woken by the IPI right away; we return
 752                          * busy, so the CPU does not try to go deep
 753                          * idle.
 754                          */
 755                         if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
 756                                 ret = -EBUSY;
 757                         } else if (dev->next_event.tv64 < bc->next_event.tv64) {
 758                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 759                                 /*
 760                                  * In case of hrtimer broadcasts the
 761                                  * programming might have moved the
 762                                  * timer to this cpu. If yes, remove
 763                                  * us from the broadcast mask and
 764                                  * return busy.
 765                                  */
 766                                 ret = broadcast_needs_cpu(bc, cpu);
 767                                 if (ret) {
 768                                         cpumask_clear_cpu(cpu,
 769                                                 tick_broadcast_oneshot_mask);
 770                                 }
 771                         }
 772                 }
 773         } else {
 774                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 775                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 776                         /*
 777                          * The cpu which was handling the broadcast
 778                          * timer marked this cpu in the broadcast
 779                          * pending mask and fired the broadcast
 780                          * IPI. So we are going to handle the expired
 781                          * event anyway via the broadcast IPI
 782                          * handler. No need to reprogram the timer
 783                          * with an already expired event.
 784                          */
 785                         if (cpumask_test_and_clear_cpu(cpu,
 786                                        tick_broadcast_pending_mask))
 787                                 goto out;
 788
 789                         /*
 790                          * Bail out if there is no next event.
 791                          */
 792                         if (dev->next_event.tv64 == KTIME_MAX)
 793                                 goto out;
 794                         /*
 795                          * If the pending bit is not set, then we are
 796                          * either the CPU handling the broadcast
 797                          * interrupt or we got woken by something else.
 798                          *
 799                          * We are not longer in the broadcast mask, so
 800                          * if the cpu local expiry time is already
 801                          * reached, we would reprogram the cpu local
 802                          * timer with an already expired event.
 803                          *
 804                          * This can lead to a ping-pong when we return
 805                          * to idle and therefor rearm the broadcast
 806                          * timer before the cpu local timer was able
 807                          * to fire. This happens because the forced
 808                          * reprogramming makes sure that the event
 809                          * will happen in the future and depending on
 810                          * the min_delta setting this might be far
 811                          * enough out that the ping-pong starts.
 812                          *
 813                          * If the cpu local next_event has expired
 814                          * then we know that the broadcast timer
 815                          * next_event has expired as well and
 816                          * broadcast is about to be handled. So we
 817                          * avoid reprogramming and enforce that the
 818                          * broadcast handler, which did not run yet,
 819                          * will invoke the cpu local handler.
 820                          *
 821                          * We cannot call the handler directly from
 822                          * here, because we might be in a NOHZ phase
 823                          * and we did not go through the irq_enter()
 824                          * nohz fixups.
 825                          */
 826                         now = ktime_get();
 827                         if (dev->next_event.tv64 <= now.tv64) {
 828                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 829                                 goto out;
 830                         }
 831                         /*
 832                          * We got woken by something else. Reprogram
 833                          * the cpu local timer device.
 834                          */
 835                         tick_program_event(dev->next_event, 1);
 836                 }
 837         }
 838 out:
 839         raw_spin_unlock(&tick_broadcast_lock);
 840         return ret;
 841 }
 842
 843 /*
 844  * Reset the one shot broadcast for a cpu
 845  *
 846  * Called with tick_broadcast_lock held
 847  */
 848 static void tick_broadcast_clear_oneshot(int cpu)
 849 {
 850         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 851         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 852 }
 853
 854 static void tick_broadcast_init_next_event(struct cpumask *mask,
 855                                            ktime_t expires)
 856 {
 857         struct tick_device *td;
 858         int cpu;
 859
 860         for_each_cpu(cpu, mask) {
 861                 td = &per_cpu(tick_cpu_device, cpu);
 862                 if (td->evtdev)
 863                         td->evtdev->next_event = expires;
 864         }
 865 }
 866
 867 /**
 868  * tick_broadcast_setup_oneshot - setup the broadcast device
 869  */
 870 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 871 {
 872         int cpu = smp_processor_id();
 873
 874         /* Set it up only once ! */
 875         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 876                 int was_periodic = clockevent_state_periodic(bc);
 877
 878                 bc->event_handler = tick_handle_oneshot_broadcast;
 879
 880                 /*
 881                  * We must be careful here. There might be other CPUs
 882                  * waiting for periodic broadcast. We need to set the
 883                  * oneshot_mask bits for those and program the
 884                  * broadcast device to fire.
 885                  */
 886                 cpumask_copy(tmpmask, tick_broadcast_mask);
 887                 cpumask_clear_cpu(cpu, tmpmask);
 888                 cpumask_or(tick_broadcast_oneshot_mask,
 889                            tick_broadcast_oneshot_mask, tmpmask);
 890
 891                 if (was_periodic && !cpumask_empty(tmpmask)) {
 892                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 893                         tick_broadcast_init_next_event(tmpmask,
 894                                                        tick_next_period);
 895                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 896                 } else
 897                         bc->next_event.tv64 = KTIME_MAX;
 898         } else {
 899                 /*
 900                  * The first cpu which switches to oneshot mode sets
 901                  * the bit for all other cpus which are in the general
 902                  * (periodic) broadcast mask. So the bit is set and
 903                  * would prevent the first broadcast enter after this
 904                  * to program the bc device.
 905                  */
 906                 tick_broadcast_clear_oneshot(cpu);
 907         }
 908 }
 909
 910 /*
 911  * Select oneshot operating mode for the broadcast device
 912  */
 913 void tick_broadcast_switch_to_oneshot(void)
 914 {
 915         struct clock_event_device *bc;
 916         unsigned long flags;
 917
 918         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 919
 920         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 921         bc = tick_broadcast_device.evtdev;
 922         if (bc)
 923                 tick_broadcast_setup_oneshot(bc);
 924
 925         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 926 }
 927
 928 #ifdef CONFIG_HOTPLUG_CPU
 929 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 930 {
 931         struct clock_event_device *bc;
 932         unsigned long flags;
 933
 934         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 935         bc = tick_broadcast_device.evtdev;
 936
 937         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 938                 /* This moves the broadcast assignment to this CPU: */
 939                 clockevents_program_event(bc, bc->next_event, 1);
 940         }
 941         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 942 }
 943
 944 /*
 945  * Remove a dead CPU from broadcasting
 946  */
 947 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 948 {
 949         unsigned long flags;
 950
 951         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 952
 953         /*
 954          * Clear the broadcast masks for the dead cpu, but do not stop
 955          * the broadcast device!
 956          */
 957         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 958         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 959         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 960
 961         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 962 }
 963 #endif
 964
 965 /*
 966  * Check, whether the broadcast device is in one shot mode
 967  */
 968 int tick_broadcast_oneshot_active(void)
 969 {
 970         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 971 }
 972
 973 /*
 974  * Check whether the broadcast device supports oneshot.
 975  */
 976 bool tick_broadcast_oneshot_available(void)
 977 {
 978         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 979
 980         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 981 }
 982
 983 #else
 984 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 985 {
 986         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 987
 988         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
 989                 return -EBUSY;
 990
 991         return 0;
 992 }
 993 #endif
 994
 995 void __init tick_broadcast_init(void)
 996 {
 997         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 998         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
 999         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1000 #ifdef CONFIG_TICK_ONESHOT
1001         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1002         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1003         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1004 #endif
1005 }