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_clear_oneshot(int cpu);
  41 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
  42 #else
  43 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  44 static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
  45 #endif
  46
  47 /*
  48  * Debugging: see timer_list.c
  49  */
  50 struct tick_device *tick_get_broadcast_device(void)
  51 {
  52         return &tick_broadcast_device;
  53 }
  54
  55 struct cpumask *tick_get_broadcast_mask(void)
  56 {
  57         return tick_broadcast_mask;
  58 }
  59
  60 /*
  61  * Start the device in periodic mode
  62  */
  63 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  64 {
  65         if (bc)
  66                 tick_setup_periodic(bc, 1);
  67 }
  68
  69 /*
  70  * Check, if the device can be utilized as broadcast device:
  71  */
  72 static bool tick_check_broadcast_device(struct clock_event_device *curdev,
  73                                         struct clock_event_device *newdev)
  74 {
  75         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
  76             (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
  77             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
  78                 return false;
  79
  80         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
  81             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
  82                 return false;
  83
  84         return !curdev || newdev->rating > curdev->rating;
  85 }
  86
  87 /*
  88  * Conditionally install/replace broadcast device
  89  */
  90 void tick_install_broadcast_device(struct clock_event_device *dev)
  91 {
  92         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  93
  94         if (!tick_check_broadcast_device(cur, dev))
  95                 return;
  96
  97         if (!try_module_get(dev->owner))
  98                 return;
  99
 100         clockevents_exchange_device(cur, dev);
 101         if (cur)
 102                 cur->event_handler = clockevents_handle_noop;
 103         tick_broadcast_device.evtdev = dev;
 104         if (!cpumask_empty(tick_broadcast_mask))
 105                 tick_broadcast_start_periodic(dev);
 106         /*
 107          * Inform all cpus about this. We might be in a situation
 108          * where we did not switch to oneshot mode because the per cpu
 109          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
 110          * of a oneshot capable broadcast device. Without that
 111          * notification the systems stays stuck in periodic mode
 112          * forever.
 113          */
 114         if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
 115                 tick_clock_notify();
 116 }
 117
 118 /*
 119  * Check, if the device is the broadcast device
 120  */
 121 int tick_is_broadcast_device(struct clock_event_device *dev)
 122 {
 123         return (dev && tick_broadcast_device.evtdev == dev);
 124 }
 125
 126 int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
 127 {
 128         int ret = -ENODEV;
 129
 130         if (tick_is_broadcast_device(dev)) {
 131                 raw_spin_lock(&tick_broadcast_lock);
 132                 ret = __clockevents_update_freq(dev, freq);
 133                 raw_spin_unlock(&tick_broadcast_lock);
 134         }
 135         return ret;
 136 }
 137
 138
 139 static void err_broadcast(const struct cpumask *mask)
 140 {
 141         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 142 }
 143
 144 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 145 {
 146         if (!dev->broadcast)
 147                 dev->broadcast = tick_broadcast;
 148         if (!dev->broadcast) {
 149                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 150                              dev->name);
 151                 dev->broadcast = err_broadcast;
 152         }
 153 }
 154
 155 /*
 156  * Check, if the device is disfunctional and a place holder, which
 157  * needs to be handled by the broadcast device.
 158  */
 159 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 160 {
 161         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 162         unsigned long flags;
 163         int ret = 0;
 164
 165         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 166
 167         /*
 168          * Devices might be registered with both periodic and oneshot
 169          * mode disabled. This signals, that the device needs to be
 170          * operated from the broadcast device and is a placeholder for
 171          * the cpu local device.
 172          */
 173         if (!tick_device_is_functional(dev)) {
 174                 dev->event_handler = tick_handle_periodic;
 175                 tick_device_setup_broadcast_func(dev);
 176                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 177                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 178                         tick_broadcast_start_periodic(bc);
 179                 else
 180                         tick_broadcast_setup_oneshot(bc);
 181                 ret = 1;
 182         } else {
 183                 /*
 184                  * Clear the broadcast bit for this cpu if the
 185                  * device is not power state affected.
 186                  */
 187                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 188                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 189                 else
 190                         tick_device_setup_broadcast_func(dev);
 191
 192                 /*
 193                  * Clear the broadcast bit if the CPU is not in
 194                  * periodic broadcast on state.
 195                  */
 196                 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
 197                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 198
 199                 switch (tick_broadcast_device.mode) {
 200                 case TICKDEV_MODE_ONESHOT:
 201                         /*
 202                          * If the system is in oneshot mode we can
 203                          * unconditionally clear the oneshot mask bit,
 204                          * because the CPU is running and therefore
 205                          * not in an idle state which causes the power
 206                          * state affected device to stop. Let the
 207                          * caller initialize the device.
 208                          */
 209                         tick_broadcast_clear_oneshot(cpu);
 210                         ret = 0;
 211                         break;
 212
 213                 case TICKDEV_MODE_PERIODIC:
 214                         /*
 215                          * If the system is in periodic mode, check
 216                          * whether the broadcast device can be
 217                          * switched off now.
 218                          */
 219                         if (cpumask_empty(tick_broadcast_mask) && bc)
 220                                 clockevents_shutdown(bc);
 221                         /*
 222                          * If we kept the cpu in the broadcast mask,
 223                          * tell the caller to leave the per cpu device
 224                          * in shutdown state. The periodic interrupt
 225                          * is delivered by the broadcast device, if
 226                          * the broadcast device exists and is not
 227                          * hrtimer based.
 228                          */
 229                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 230                                 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 231                         break;
 232                 default:
 233                         break;
 234                 }
 235         }
 236         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 237         return ret;
 238 }
 239
 240 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 241 int tick_receive_broadcast(void)
 242 {
 243         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 244         struct clock_event_device *evt = td->evtdev;
 245
 246         if (!evt)
 247                 return -ENODEV;
 248
 249         if (!evt->event_handler)
 250                 return -EINVAL;
 251
 252         evt->event_handler(evt);
 253         return 0;
 254 }
 255 #endif
 256
 257 /*
 258  * Broadcast the event to the cpus, which are set in the mask (mangled).
 259  */
 260 static bool tick_do_broadcast(struct cpumask *mask)
 261 {
 262         int cpu = smp_processor_id();
 263         struct tick_device *td;
 264         bool local = false;
 265
 266         /*
 267          * Check, if the current cpu is in the mask
 268          */
 269         if (cpumask_test_cpu(cpu, mask)) {
 270                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 271
 272                 cpumask_clear_cpu(cpu, mask);
 273                 /*
 274                  * We only run the local handler, if the broadcast
 275                  * device is not hrtimer based. Otherwise we run into
 276                  * a hrtimer recursion.
 277                  *
 278                  * local timer_interrupt()
 279                  *   local_handler()
 280                  *     expire_hrtimers()
 281                  *       bc_handler()
 282                  *         local_handler()
 283                  *           expire_hrtimers()
 284                  */
 285                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 286         }
 287
 288         if (!cpumask_empty(mask)) {
 289                 /*
 290                  * It might be necessary to actually check whether the devices
 291                  * have different broadcast functions. For now, just use the
 292                  * one of the first device. This works as long as we have this
 293                  * misfeature only on x86 (lapic)
 294                  */
 295                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 296                 td->evtdev->broadcast(mask);
 297         }
 298         return local;
 299 }
 300
 301 /*
 302  * Periodic broadcast:
 303  * - invoke the broadcast handlers
 304  */
 305 static bool tick_do_periodic_broadcast(void)
 306 {
 307         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 308         return tick_do_broadcast(tmpmask);
 309 }
 310
 311 /*
 312  * Event handler for periodic broadcast ticks
 313  */
 314 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 315 {
 316         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 317         bool bc_local;
 318
 319         raw_spin_lock(&tick_broadcast_lock);
 320
 321         /* Handle spurious interrupts gracefully */
 322         if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
 323                 raw_spin_unlock(&tick_broadcast_lock);
 324                 return;
 325         }
 326
 327         bc_local = tick_do_periodic_broadcast();
 328
 329         if (clockevent_state_oneshot(dev)) {
 330                 ktime_t next = ktime_add(dev->next_event, tick_period);
 331
 332                 clockevents_program_event(dev, next, true);
 333         }
 334         raw_spin_unlock(&tick_broadcast_lock);
 335
 336         /*
 337          * We run the handler of the local cpu after dropping
 338          * tick_broadcast_lock because the handler might deadlock when
 339          * trying to switch to oneshot mode.
 340          */
 341         if (bc_local)
 342                 td->evtdev->event_handler(td->evtdev);
 343 }
 344
 345 /**
 346  * tick_broadcast_control - Enable/disable or force broadcast mode
 347  * @mode:       The selected broadcast mode
 348  *
 349  * Called when the system enters a state where affected tick devices
 350  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 351  */
 352 void tick_broadcast_control(enum tick_broadcast_mode mode)
 353 {
 354         struct clock_event_device *bc, *dev;
 355         struct tick_device *td;
 356         int cpu, bc_stopped;
 357         unsigned long flags;
 358
 359         /* Protects also the local clockevent device. */
 360         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 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                 goto out;
 369
 370         if (!tick_device_is_functional(dev))
 371                 goto out;
 372
 373         cpu = smp_processor_id();
 374         bc = tick_broadcast_device.evtdev;
 375         bc_stopped = cpumask_empty(tick_broadcast_mask);
 376
 377         switch (mode) {
 378         case TICK_BROADCAST_FORCE:
 379                 tick_broadcast_forced = 1;
 380         case TICK_BROADCAST_ON:
 381                 cpumask_set_cpu(cpu, tick_broadcast_on);
 382                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 383                         /*
 384                          * Only shutdown the cpu local device, if:
 385                          *
 386                          * - the broadcast device exists
 387                          * - the broadcast device is not a hrtimer based one
 388                          * - the broadcast device is in periodic mode to
 389                          *   avoid a hickup during switch to oneshot mode
 390                          */
 391                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 392                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 393                                 clockevents_shutdown(dev);
 394                 }
 395                 break;
 396
 397         case TICK_BROADCAST_OFF:
 398                 if (tick_broadcast_forced)
 399                         break;
 400                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 401                 if (!tick_device_is_functional(dev))
 402                         break;
 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                 td = &per_cpu(tick_cpu_device, cpu);
 614                 if (td->evtdev->next_event <= now) {
 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 < next_event) {
 623                         next_event = td->evtdev->next_event;
 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 != 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 == 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 < bc->next_event)
 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 < bc->next_event) {
 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 == 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 <= now) {
 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         if (!bc)
 875                 return;
 876
 877         /* Set it up only once ! */
 878         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 879                 int was_periodic = clockevent_state_periodic(bc);
 880
 881                 bc->event_handler = tick_handle_oneshot_broadcast;
 882
 883                 /*
 884                  * We must be careful here. There might be other CPUs
 885                  * waiting for periodic broadcast. We need to set the
 886                  * oneshot_mask bits for those and program the
 887                  * broadcast device to fire.
 888                  */
 889                 cpumask_copy(tmpmask, tick_broadcast_mask);
 890                 cpumask_clear_cpu(cpu, tmpmask);
 891                 cpumask_or(tick_broadcast_oneshot_mask,
 892                            tick_broadcast_oneshot_mask, tmpmask);
 893
 894                 if (was_periodic && !cpumask_empty(tmpmask)) {
 895                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 896                         tick_broadcast_init_next_event(tmpmask,
 897                                                        tick_next_period);
 898                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 899                 } else
 900                         bc->next_event = KTIME_MAX;
 901         } else {
 902                 /*
 903                  * The first cpu which switches to oneshot mode sets
 904                  * the bit for all other cpus which are in the general
 905                  * (periodic) broadcast mask. So the bit is set and
 906                  * would prevent the first broadcast enter after this
 907                  * to program the bc device.
 908                  */
 909                 tick_broadcast_clear_oneshot(cpu);
 910         }
 911 }
 912
 913 /*
 914  * Select oneshot operating mode for the broadcast device
 915  */
 916 void tick_broadcast_switch_to_oneshot(void)
 917 {
 918         struct clock_event_device *bc;
 919         unsigned long flags;
 920
 921         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 922
 923         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 924         bc = tick_broadcast_device.evtdev;
 925         if (bc)
 926                 tick_broadcast_setup_oneshot(bc);
 927
 928         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 929 }
 930
 931 #ifdef CONFIG_HOTPLUG_CPU
 932 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 933 {
 934         struct clock_event_device *bc;
 935         unsigned long flags;
 936
 937         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 938         bc = tick_broadcast_device.evtdev;
 939
 940         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 941                 /* This moves the broadcast assignment to this CPU: */
 942                 clockevents_program_event(bc, bc->next_event, 1);
 943         }
 944         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 945 }
 946
 947 /*
 948  * Remove a dead CPU from broadcasting
 949  */
 950 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 951 {
 952         unsigned long flags;
 953
 954         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 955
 956         /*
 957          * Clear the broadcast masks for the dead cpu, but do not stop
 958          * the broadcast device!
 959          */
 960         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 961         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 962         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 963
 964         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 965 }
 966 #endif
 967
 968 /*
 969  * Check, whether the broadcast device is in one shot mode
 970  */
 971 int tick_broadcast_oneshot_active(void)
 972 {
 973         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 974 }
 975
 976 /*
 977  * Check whether the broadcast device supports oneshot.
 978  */
 979 bool tick_broadcast_oneshot_available(void)
 980 {
 981         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 982
 983         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 984 }
 985
 986 #else
 987 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 988 {
 989         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 990
 991         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
 992                 return -EBUSY;
 993
 994         return 0;
 995 }
 996 #endif
 997
 998 void __init tick_broadcast_init(void)
 999 {
1000         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
1001         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
1002         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1003 #ifdef CONFIG_TICK_ONESHOT
1004         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1005         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1006         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1007 #endif
1008 }