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  * Called with interrupts disabled, so clockevents_lock is not
 353  * required here because the local clock event device cannot go away
 354  * under us.
 355  */
 356 void tick_broadcast_control(enum tick_broadcast_mode mode)
 357 {
 358         struct clock_event_device *bc, *dev;
 359         struct tick_device *td;
 360         int cpu, bc_stopped;
 361
 362         td = this_cpu_ptr(&tick_cpu_device);
 363         dev = td->evtdev;
 364
 365         /*
 366          * Is the device not affected by the powerstate ?
 367          */
 368         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 369                 return;
 370
 371         if (!tick_device_is_functional(dev))
 372                 return;
 373
 374         raw_spin_lock(&tick_broadcast_lock);
 375         cpu = smp_processor_id();
 376         bc = tick_broadcast_device.evtdev;
 377         bc_stopped = cpumask_empty(tick_broadcast_mask);
 378
 379         switch (mode) {
 380         case TICK_BROADCAST_FORCE:
 381                 tick_broadcast_forced = 1;
 382         case TICK_BROADCAST_ON:
 383                 cpumask_set_cpu(cpu, tick_broadcast_on);
 384                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 385                         /*
 386                          * Only shutdown the cpu local device, if:
 387                          *
 388                          * - the broadcast device exists
 389                          * - the broadcast device is not a hrtimer based one
 390                          * - the broadcast device is in periodic mode to
 391                          *   avoid a hickup during switch to oneshot mode
 392                          */
 393                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 394                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 395                                 clockevents_shutdown(dev);
 396                 }
 397                 break;
 398
 399         case TICK_BROADCAST_OFF:
 400                 if (tick_broadcast_forced)
 401                         break;
 402                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 403                 if (!tick_device_is_functional(dev))
 404                         break;
 405                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 406                         if (tick_broadcast_device.mode ==
 407                             TICKDEV_MODE_PERIODIC)
 408                                 tick_setup_periodic(dev, 0);
 409                 }
 410                 break;
 411         }
 412
 413         if (bc) {
 414                 if (cpumask_empty(tick_broadcast_mask)) {
 415                         if (!bc_stopped)
 416                                 clockevents_shutdown(bc);
 417                 } else if (bc_stopped) {
 418                         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 419                                 tick_broadcast_start_periodic(bc);
 420                         else
 421                                 tick_broadcast_setup_oneshot(bc);
 422                 }
 423         }
 424         raw_spin_unlock(&tick_broadcast_lock);
 425 }
 426 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 427
 428 /*
 429  * Set the periodic handler depending on broadcast on/off
 430  */
 431 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 432 {
 433         if (!broadcast)
 434                 dev->event_handler = tick_handle_periodic;
 435         else
 436                 dev->event_handler = tick_handle_periodic_broadcast;
 437 }
 438
 439 #ifdef CONFIG_HOTPLUG_CPU
 440 /*
 441  * Remove a CPU from broadcasting
 442  */
 443 void tick_shutdown_broadcast(unsigned int cpu)
 444 {
 445         struct clock_event_device *bc;
 446         unsigned long flags;
 447
 448         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 449
 450         bc = tick_broadcast_device.evtdev;
 451         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 452         cpumask_clear_cpu(cpu, tick_broadcast_on);
 453
 454         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 455                 if (bc && cpumask_empty(tick_broadcast_mask))
 456                         clockevents_shutdown(bc);
 457         }
 458
 459         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 460 }
 461 #endif
 462
 463 void tick_suspend_broadcast(void)
 464 {
 465         struct clock_event_device *bc;
 466         unsigned long flags;
 467
 468         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 469
 470         bc = tick_broadcast_device.evtdev;
 471         if (bc)
 472                 clockevents_shutdown(bc);
 473
 474         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 475 }
 476
 477 /*
 478  * This is called from tick_resume_local() on a resuming CPU. That's
 479  * called from the core resume function, tick_unfreeze() and the magic XEN
 480  * resume hackery.
 481  *
 482  * In none of these cases the broadcast device mode can change and the
 483  * bit of the resuming CPU in the broadcast mask is safe as well.
 484  */
 485 bool tick_resume_check_broadcast(void)
 486 {
 487         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 488                 return false;
 489         else
 490                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 491 }
 492
 493 void tick_resume_broadcast(void)
 494 {
 495         struct clock_event_device *bc;
 496         unsigned long flags;
 497
 498         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 499
 500         bc = tick_broadcast_device.evtdev;
 501
 502         if (bc) {
 503                 clockevents_tick_resume(bc);
 504
 505                 switch (tick_broadcast_device.mode) {
 506                 case TICKDEV_MODE_PERIODIC:
 507                         if (!cpumask_empty(tick_broadcast_mask))
 508                                 tick_broadcast_start_periodic(bc);
 509                         break;
 510                 case TICKDEV_MODE_ONESHOT:
 511                         if (!cpumask_empty(tick_broadcast_mask))
 512                                 tick_resume_broadcast_oneshot(bc);
 513                         break;
 514                 }
 515         }
 516         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 517 }
 518
 519 #ifdef CONFIG_TICK_ONESHOT
 520
 521 static cpumask_var_t tick_broadcast_oneshot_mask __cpumask_var_read_mostly;
 522 static cpumask_var_t tick_broadcast_pending_mask __cpumask_var_read_mostly;
 523 static cpumask_var_t tick_broadcast_force_mask __cpumask_var_read_mostly;
 524
 525 /*
 526  * Exposed for debugging: see timer_list.c
 527  */
 528 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 529 {
 530         return tick_broadcast_oneshot_mask;
 531 }
 532
 533 /*
 534  * Called before going idle with interrupts disabled. Checks whether a
 535  * broadcast event from the other core is about to happen. We detected
 536  * that in tick_broadcast_oneshot_control(). The callsite can use this
 537  * to avoid a deep idle transition as we are about to get the
 538  * broadcast IPI right away.
 539  */
 540 int tick_check_broadcast_expired(void)
 541 {
 542         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 543 }
 544
 545 /*
 546  * Set broadcast interrupt affinity
 547  */
 548 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 549                                         const struct cpumask *cpumask)
 550 {
 551         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 552                 return;
 553
 554         if (cpumask_equal(bc->cpumask, cpumask))
 555                 return;
 556
 557         bc->cpumask = cpumask;
 558         irq_set_affinity(bc->irq, bc->cpumask);
 559 }
 560
 561 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 562                                      ktime_t expires)
 563 {
 564         if (!clockevent_state_oneshot(bc))
 565                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 566
 567         clockevents_program_event(bc, expires, 1);
 568         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 569 }
 570
 571 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 572 {
 573         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 574 }
 575
 576 /*
 577  * Called from irq_enter() when idle was interrupted to reenable the
 578  * per cpu device.
 579  */
 580 void tick_check_oneshot_broadcast_this_cpu(void)
 581 {
 582         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 583                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 584
 585                 /*
 586                  * We might be in the middle of switching over from
 587                  * periodic to oneshot. If the CPU has not yet
 588                  * switched over, leave the device alone.
 589                  */
 590                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 591                         clockevents_switch_state(td->evtdev,
 592                                               CLOCK_EVT_STATE_ONESHOT);
 593                 }
 594         }
 595 }
 596
 597 /*
 598  * Handle oneshot mode broadcasting
 599  */
 600 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 601 {
 602         struct tick_device *td;
 603         ktime_t now, next_event;
 604         int cpu, next_cpu = 0;
 605         bool bc_local;
 606
 607         raw_spin_lock(&tick_broadcast_lock);
 608         dev->next_event = KTIME_MAX;
 609         next_event = KTIME_MAX;
 610         cpumask_clear(tmpmask);
 611         now = ktime_get();
 612         /* Find all expired events */
 613         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 614                 td = &per_cpu(tick_cpu_device, cpu);
 615                 if (td->evtdev->next_event <= now) {
 616                         cpumask_set_cpu(cpu, tmpmask);
 617                         /*
 618                          * Mark the remote cpu in the pending mask, so
 619                          * it can avoid reprogramming the cpu local
 620                          * timer in tick_broadcast_oneshot_control().
 621                          */
 622                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 623                 } else if (td->evtdev->next_event < next_event) {
 624                         next_event = td->evtdev->next_event;
 625                         next_cpu = cpu;
 626                 }
 627         }
 628
 629         /*
 630          * Remove the current cpu from the pending mask. The event is
 631          * delivered immediately in tick_do_broadcast() !
 632          */
 633         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 634
 635         /* Take care of enforced broadcast requests */
 636         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 637         cpumask_clear(tick_broadcast_force_mask);
 638
 639         /*
 640          * Sanity check. Catch the case where we try to broadcast to
 641          * offline cpus.
 642          */
 643         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 644                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 645
 646         /*
 647          * Wakeup the cpus which have an expired event.
 648          */
 649         bc_local = tick_do_broadcast(tmpmask);
 650
 651         /*
 652          * Two reasons for reprogram:
 653          *
 654          * - The global event did not expire any CPU local
 655          * events. This happens in dyntick mode, as the maximum PIT
 656          * delta is quite small.
 657          *
 658          * - There are pending events on sleeping CPUs which were not
 659          * in the event mask
 660          */
 661         if (next_event != KTIME_MAX)
 662                 tick_broadcast_set_event(dev, next_cpu, next_event);
 663
 664         raw_spin_unlock(&tick_broadcast_lock);
 665
 666         if (bc_local) {
 667                 td = this_cpu_ptr(&tick_cpu_device);
 668                 td->evtdev->event_handler(td->evtdev);
 669         }
 670 }
 671
 672 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 673 {
 674         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 675                 return 0;
 676         if (bc->next_event == KTIME_MAX)
 677                 return 0;
 678         return bc->bound_on == cpu ? -EBUSY : 0;
 679 }
 680
 681 static void broadcast_shutdown_local(struct clock_event_device *bc,
 682                                      struct clock_event_device *dev)
 683 {
 684         /*
 685          * For hrtimer based broadcasting we cannot shutdown the cpu
 686          * local device if our own event is the first one to expire or
 687          * if we own the broadcast timer.
 688          */
 689         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 690                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 691                         return;
 692                 if (dev->next_event < bc->next_event)
 693                         return;
 694         }
 695         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 696 }
 697
 698 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 699 {
 700         struct clock_event_device *bc, *dev;
 701         int cpu, ret = 0;
 702         ktime_t now;
 703
 704         /*
 705          * If there is no broadcast device, tell the caller not to go
 706          * into deep idle.
 707          */
 708         if (!tick_broadcast_device.evtdev)
 709                 return -EBUSY;
 710
 711         dev = this_cpu_ptr(&tick_cpu_device)->evtdev;
 712
 713         raw_spin_lock(&tick_broadcast_lock);
 714         bc = tick_broadcast_device.evtdev;
 715         cpu = smp_processor_id();
 716
 717         if (state == TICK_BROADCAST_ENTER) {
 718                 /*
 719                  * If the current CPU owns the hrtimer broadcast
 720                  * mechanism, it cannot go deep idle and we do not add
 721                  * the CPU to the broadcast mask. We don't have to go
 722                  * through the EXIT path as the local timer is not
 723                  * shutdown.
 724                  */
 725                 ret = broadcast_needs_cpu(bc, cpu);
 726                 if (ret)
 727                         goto out;
 728
 729                 /*
 730                  * If the broadcast device is in periodic mode, we
 731                  * return.
 732                  */
 733                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 734                         /* If it is a hrtimer based broadcast, return busy */
 735                         if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
 736                                 ret = -EBUSY;
 737                         goto out;
 738                 }
 739
 740                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 741                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 742
 743                         /* Conditionally shut down the local timer. */
 744                         broadcast_shutdown_local(bc, dev);
 745
 746                         /*
 747                          * We only reprogram the broadcast timer if we
 748                          * did not mark ourself in the force mask and
 749                          * if the cpu local event is earlier than the
 750                          * broadcast event. If the current CPU is in
 751                          * the force mask, then we are going to be
 752                          * woken by the IPI right away; we return
 753                          * busy, so the CPU does not try to go deep
 754                          * idle.
 755                          */
 756                         if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
 757                                 ret = -EBUSY;
 758                         } else if (dev->next_event < bc->next_event) {
 759                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 760                                 /*
 761                                  * In case of hrtimer broadcasts the
 762                                  * programming might have moved the
 763                                  * timer to this cpu. If yes, remove
 764                                  * us from the broadcast mask and
 765                                  * return busy.
 766                                  */
 767                                 ret = broadcast_needs_cpu(bc, cpu);
 768                                 if (ret) {
 769                                         cpumask_clear_cpu(cpu,
 770                                                 tick_broadcast_oneshot_mask);
 771                                 }
 772                         }
 773                 }
 774         } else {
 775                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 776                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 777                         /*
 778                          * The cpu which was handling the broadcast
 779                          * timer marked this cpu in the broadcast
 780                          * pending mask and fired the broadcast
 781                          * IPI. So we are going to handle the expired
 782                          * event anyway via the broadcast IPI
 783                          * handler. No need to reprogram the timer
 784                          * with an already expired event.
 785                          */
 786                         if (cpumask_test_and_clear_cpu(cpu,
 787                                        tick_broadcast_pending_mask))
 788                                 goto out;
 789
 790                         /*
 791                          * Bail out if there is no next event.
 792                          */
 793                         if (dev->next_event == KTIME_MAX)
 794                                 goto out;
 795                         /*
 796                          * If the pending bit is not set, then we are
 797                          * either the CPU handling the broadcast
 798                          * interrupt or we got woken by something else.
 799                          *
 800                          * We are not longer in the broadcast mask, so
 801                          * if the cpu local expiry time is already
 802                          * reached, we would reprogram the cpu local
 803                          * timer with an already expired event.
 804                          *
 805                          * This can lead to a ping-pong when we return
 806                          * to idle and therefor rearm the broadcast
 807                          * timer before the cpu local timer was able
 808                          * to fire. This happens because the forced
 809                          * reprogramming makes sure that the event
 810                          * will happen in the future and depending on
 811                          * the min_delta setting this might be far
 812                          * enough out that the ping-pong starts.
 813                          *
 814                          * If the cpu local next_event has expired
 815                          * then we know that the broadcast timer
 816                          * next_event has expired as well and
 817                          * broadcast is about to be handled. So we
 818                          * avoid reprogramming and enforce that the
 819                          * broadcast handler, which did not run yet,
 820                          * will invoke the cpu local handler.
 821                          *
 822                          * We cannot call the handler directly from
 823                          * here, because we might be in a NOHZ phase
 824                          * and we did not go through the irq_enter()
 825                          * nohz fixups.
 826                          */
 827                         now = ktime_get();
 828                         if (dev->next_event <= now) {
 829                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 830                                 goto out;
 831                         }
 832                         /*
 833                          * We got woken by something else. Reprogram
 834                          * the cpu local timer device.
 835                          */
 836                         tick_program_event(dev->next_event, 1);
 837                 }
 838         }
 839 out:
 840         raw_spin_unlock(&tick_broadcast_lock);
 841         return ret;
 842 }
 843
 844 /*
 845  * Reset the one shot broadcast for a cpu
 846  *
 847  * Called with tick_broadcast_lock held
 848  */
 849 static void tick_broadcast_clear_oneshot(int cpu)
 850 {
 851         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 852         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 853 }
 854
 855 static void tick_broadcast_init_next_event(struct cpumask *mask,
 856                                            ktime_t expires)
 857 {
 858         struct tick_device *td;
 859         int cpu;
 860
 861         for_each_cpu(cpu, mask) {
 862                 td = &per_cpu(tick_cpu_device, cpu);
 863                 if (td->evtdev)
 864                         td->evtdev->next_event = expires;
 865         }
 866 }
 867
 868 /**
 869  * tick_broadcast_setup_oneshot - setup the broadcast device
 870  */
 871 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 872 {
 873         int cpu = smp_processor_id();
 874
 875         if (!bc)
 876                 return;
 877
 878         /* Set it up only once ! */
 879         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 880                 int was_periodic = clockevent_state_periodic(bc);
 881
 882                 bc->event_handler = tick_handle_oneshot_broadcast;
 883
 884                 /*
 885                  * We must be careful here. There might be other CPUs
 886                  * waiting for periodic broadcast. We need to set the
 887                  * oneshot_mask bits for those and program the
 888                  * broadcast device to fire.
 889                  */
 890                 cpumask_copy(tmpmask, tick_broadcast_mask);
 891                 cpumask_clear_cpu(cpu, tmpmask);
 892                 cpumask_or(tick_broadcast_oneshot_mask,
 893                            tick_broadcast_oneshot_mask, tmpmask);
 894
 895                 if (was_periodic && !cpumask_empty(tmpmask)) {
 896                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 897                         tick_broadcast_init_next_event(tmpmask,
 898                                                        tick_next_period);
 899                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 900                 } else
 901                         bc->next_event = KTIME_MAX;
 902         } else {
 903                 /*
 904                  * The first cpu which switches to oneshot mode sets
 905                  * the bit for all other cpus which are in the general
 906                  * (periodic) broadcast mask. So the bit is set and
 907                  * would prevent the first broadcast enter after this
 908                  * to program the bc device.
 909                  */
 910                 tick_broadcast_clear_oneshot(cpu);
 911         }
 912 }
 913
 914 /*
 915  * Select oneshot operating mode for the broadcast device
 916  */
 917 void tick_broadcast_switch_to_oneshot(void)
 918 {
 919         struct clock_event_device *bc;
 920         unsigned long flags;
 921
 922         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 923
 924         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 925         bc = tick_broadcast_device.evtdev;
 926         if (bc)
 927                 tick_broadcast_setup_oneshot(bc);
 928
 929         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 930 }
 931
 932 #ifdef CONFIG_HOTPLUG_CPU
 933 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 934 {
 935         struct clock_event_device *bc;
 936         unsigned long flags;
 937
 938         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 939         bc = tick_broadcast_device.evtdev;
 940
 941         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 942                 /* This moves the broadcast assignment to this CPU: */
 943                 clockevents_program_event(bc, bc->next_event, 1);
 944         }
 945         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 946 }
 947
 948 /*
 949  * Remove a dead CPU from broadcasting
 950  */
 951 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 952 {
 953         unsigned long flags;
 954
 955         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 956
 957         /*
 958          * Clear the broadcast masks for the dead cpu, but do not stop
 959          * the broadcast device!
 960          */
 961         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 962         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 963         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 964
 965         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 966 }
 967 #endif
 968
 969 /*
 970  * Check, whether the broadcast device is in one shot mode
 971  */
 972 int tick_broadcast_oneshot_active(void)
 973 {
 974         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 975 }
 976
 977 /*
 978  * Check whether the broadcast device supports oneshot.
 979  */
 980 bool tick_broadcast_oneshot_available(void)
 981 {
 982         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 983
 984         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 985 }
 986
 987 #else
 988 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 989 {
 990         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 991
 992         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
 993                 return -EBUSY;
 994
 995         return 0;
 996 }
 997 #endif
 998
 999 void __init tick_broadcast_init(void)
1000 {
1001         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
1002         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
1003         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1004 #ifdef CONFIG_TICK_ONESHOT
1005         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1006         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1007         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1008 #endif
1009 }