kernel/time/tick-broadcast.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file contains functions which emulate a local clock-event
   4  * device via a broadcast event source.
   5  *
   6  * Copyright(C) 2005-2006, Thomas Gleixner <[email protected]>
   7  * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   8  * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
   9  */
  10 #include <linux/cpu.h>
  11 #include <linux/err.h>
  12 #include <linux/hrtimer.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/percpu.h>
  15 #include <linux/profile.h>
  16 #include <linux/sched.h>
  17 #include <linux/smp.h>
  18 #include <linux/module.h>
  19
  20 #include "tick-internal.h"
  21
  22 /*
  23  * Broadcast support for broken x86 hardware, where the local apic
  24  * timer stops in C3 state.
  25  */
  26
  27 static struct tick_device tick_broadcast_device;
  28 static cpumask_var_t tick_broadcast_mask __cpumask_var_read_mostly;
  29 static cpumask_var_t tick_broadcast_on __cpumask_var_read_mostly;
  30 static cpumask_var_t tmpmask __cpumask_var_read_mostly;
  31 static int tick_broadcast_forced;
  32
  33 static __cacheline_aligned_in_smp DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  34
  35 #ifdef CONFIG_TICK_ONESHOT
  36 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc);
  37 static void tick_broadcast_clear_oneshot(int cpu);
  38 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
  39 #else
  40 static inline void tick_broadcast_setup_oneshot(struct clock_event_device *bc) { BUG(); }
  41 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  42 static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
  43 #endif
  44
  45 /*
  46  * Debugging: see timer_list.c
  47  */
  48 struct tick_device *tick_get_broadcast_device(void)
  49 {
  50         return &tick_broadcast_device;
  51 }
  52
  53 struct cpumask *tick_get_broadcast_mask(void)
  54 {
  55         return tick_broadcast_mask;
  56 }
  57
  58 /*
  59  * Start the device in periodic mode
  60  */
  61 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  62 {
  63         if (bc)
  64                 tick_setup_periodic(bc, 1);
  65 }
  66
  67 /*
  68  * Check, if the device can be utilized as broadcast device:
  69  */
  70 static bool tick_check_broadcast_device(struct clock_event_device *curdev,
  71                                         struct clock_event_device *newdev)
  72 {
  73         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
  74             (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
  75             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
  76                 return false;
  77
  78         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
  79             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
  80                 return false;
  81
  82         return !curdev || newdev->rating > curdev->rating;
  83 }
  84
  85 /*
  86  * Conditionally install/replace broadcast device
  87  */
  88 void tick_install_broadcast_device(struct clock_event_device *dev)
  89 {
  90         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  91
  92         if (!tick_check_broadcast_device(cur, dev))
  93                 return;
  94
  95         if (!try_module_get(dev->owner))
  96                 return;
  97
  98         clockevents_exchange_device(cur, dev);
  99         if (cur)
 100                 cur->event_handler = clockevents_handle_noop;
 101         tick_broadcast_device.evtdev = dev;
 102         if (!cpumask_empty(tick_broadcast_mask))
 103                 tick_broadcast_start_periodic(dev);
 104         /*
 105          * Inform all cpus about this. We might be in a situation
 106          * where we did not switch to oneshot mode because the per cpu
 107          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
 108          * of a oneshot capable broadcast device. Without that
 109          * notification the systems stays stuck in periodic mode
 110          * forever.
 111          */
 112         if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
 113                 tick_clock_notify();
 114 }
 115
 116 /*
 117  * Check, if the device is the broadcast device
 118  */
 119 int tick_is_broadcast_device(struct clock_event_device *dev)
 120 {
 121         return (dev && tick_broadcast_device.evtdev == dev);
 122 }
 123
 124 int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
 125 {
 126         int ret = -ENODEV;
 127
 128         if (tick_is_broadcast_device(dev)) {
 129                 raw_spin_lock(&tick_broadcast_lock);
 130                 ret = __clockevents_update_freq(dev, freq);
 131                 raw_spin_unlock(&tick_broadcast_lock);
 132         }
 133         return ret;
 134 }
 135
 136
 137 static void err_broadcast(const struct cpumask *mask)
 138 {
 139         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 140 }
 141
 142 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 143 {
 144         if (!dev->broadcast)
 145                 dev->broadcast = tick_broadcast;
 146         if (!dev->broadcast) {
 147                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 148                              dev->name);
 149                 dev->broadcast = err_broadcast;
 150         }
 151 }
 152
 153 /*
 154  * Check, if the device is disfunctional and a place holder, which
 155  * needs to be handled by the broadcast device.
 156  */
 157 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 158 {
 159         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 160         unsigned long flags;
 161         int ret = 0;
 162
 163         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 164
 165         /*
 166          * Devices might be registered with both periodic and oneshot
 167          * mode disabled. This signals, that the device needs to be
 168          * operated from the broadcast device and is a placeholder for
 169          * the cpu local device.
 170          */
 171         if (!tick_device_is_functional(dev)) {
 172                 dev->event_handler = tick_handle_periodic;
 173                 tick_device_setup_broadcast_func(dev);
 174                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 175                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 176                         tick_broadcast_start_periodic(bc);
 177                 else
 178                         tick_broadcast_setup_oneshot(bc);
 179                 ret = 1;
 180         } else {
 181                 /*
 182                  * Clear the broadcast bit for this cpu if the
 183                  * device is not power state affected.
 184                  */
 185                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 186                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 187                 else
 188                         tick_device_setup_broadcast_func(dev);
 189
 190                 /*
 191                  * Clear the broadcast bit if the CPU is not in
 192                  * periodic broadcast on state.
 193                  */
 194                 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
 195                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 196
 197                 switch (tick_broadcast_device.mode) {
 198                 case TICKDEV_MODE_ONESHOT:
 199                         /*
 200                          * If the system is in oneshot mode we can
 201                          * unconditionally clear the oneshot mask bit,
 202                          * because the CPU is running and therefore
 203                          * not in an idle state which causes the power
 204                          * state affected device to stop. Let the
 205                          * caller initialize the device.
 206                          */
 207                         tick_broadcast_clear_oneshot(cpu);
 208                         ret = 0;
 209                         break;
 210
 211                 case TICKDEV_MODE_PERIODIC:
 212                         /*
 213                          * If the system is in periodic mode, check
 214                          * whether the broadcast device can be
 215                          * switched off now.
 216                          */
 217                         if (cpumask_empty(tick_broadcast_mask) && bc)
 218                                 clockevents_shutdown(bc);
 219                         /*
 220                          * If we kept the cpu in the broadcast mask,
 221                          * tell the caller to leave the per cpu device
 222                          * in shutdown state. The periodic interrupt
 223                          * is delivered by the broadcast device, if
 224                          * the broadcast device exists and is not
 225                          * hrtimer based.
 226                          */
 227                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 228                                 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 229                         break;
 230                 default:
 231                         break;
 232                 }
 233         }
 234         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 235         return ret;
 236 }
 237
 238 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 239 int tick_receive_broadcast(void)
 240 {
 241         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 242         struct clock_event_device *evt = td->evtdev;
 243
 244         if (!evt)
 245                 return -ENODEV;
 246
 247         if (!evt->event_handler)
 248                 return -EINVAL;
 249
 250         evt->event_handler(evt);
 251         return 0;
 252 }
 253 #endif
 254
 255 /*
 256  * Broadcast the event to the cpus, which are set in the mask (mangled).
 257  */
 258 static bool tick_do_broadcast(struct cpumask *mask)
 259 {
 260         int cpu = smp_processor_id();
 261         struct tick_device *td;
 262         bool local = false;
 263
 264         /*
 265          * Check, if the current cpu is in the mask
 266          */
 267         if (cpumask_test_cpu(cpu, mask)) {
 268                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 269
 270                 cpumask_clear_cpu(cpu, mask);
 271                 /*
 272                  * We only run the local handler, if the broadcast
 273                  * device is not hrtimer based. Otherwise we run into
 274                  * a hrtimer recursion.
 275                  *
 276                  * local timer_interrupt()
 277                  *   local_handler()
 278                  *     expire_hrtimers()
 279                  *       bc_handler()
 280                  *         local_handler()
 281                  *           expire_hrtimers()
 282                  */
 283                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 284         }
 285
 286         if (!cpumask_empty(mask)) {
 287                 /*
 288                  * It might be necessary to actually check whether the devices
 289                  * have different broadcast functions. For now, just use the
 290                  * one of the first device. This works as long as we have this
 291                  * misfeature only on x86 (lapic)
 292                  */
 293                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 294                 td->evtdev->broadcast(mask);
 295         }
 296         return local;
 297 }
 298
 299 /*
 300  * Periodic broadcast:
 301  * - invoke the broadcast handlers
 302  */
 303 static bool tick_do_periodic_broadcast(void)
 304 {
 305         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 306         return tick_do_broadcast(tmpmask);
 307 }
 308
 309 /*
 310  * Event handler for periodic broadcast ticks
 311  */
 312 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 313 {
 314         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 315         bool bc_local;
 316
 317         raw_spin_lock(&tick_broadcast_lock);
 318
 319         /* Handle spurious interrupts gracefully */
 320         if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
 321                 raw_spin_unlock(&tick_broadcast_lock);
 322                 return;
 323         }
 324
 325         bc_local = tick_do_periodic_broadcast();
 326
 327         if (clockevent_state_oneshot(dev)) {
 328                 ktime_t next = ktime_add(dev->next_event, tick_period);
 329
 330                 clockevents_program_event(dev, next, true);
 331         }
 332         raw_spin_unlock(&tick_broadcast_lock);
 333
 334         /*
 335          * We run the handler of the local cpu after dropping
 336          * tick_broadcast_lock because the handler might deadlock when
 337          * trying to switch to oneshot mode.
 338          */
 339         if (bc_local)
 340                 td->evtdev->event_handler(td->evtdev);
 341 }
 342
 343 /**
 344  * tick_broadcast_control - Enable/disable or force broadcast mode
 345  * @mode:       The selected broadcast mode
 346  *
 347  * Called when the system enters a state where affected tick devices
 348  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 349  */
 350 void tick_broadcast_control(enum tick_broadcast_mode mode)
 351 {
 352         struct clock_event_device *bc, *dev;
 353         struct tick_device *td;
 354         int cpu, bc_stopped;
 355         unsigned long flags;
 356
 357         /* Protects also the local clockevent device. */
 358         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 359         td = this_cpu_ptr(&tick_cpu_device);
 360         dev = td->evtdev;
 361
 362         /*
 363          * Is the device not affected by the powerstate ?
 364          */
 365         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 366                 goto out;
 367
 368         if (!tick_device_is_functional(dev))
 369                 goto out;
 370
 371         cpu = smp_processor_id();
 372         bc = tick_broadcast_device.evtdev;
 373         bc_stopped = cpumask_empty(tick_broadcast_mask);
 374
 375         switch (mode) {
 376         case TICK_BROADCAST_FORCE:
 377                 tick_broadcast_forced = 1;
 378         case TICK_BROADCAST_ON:
 379                 cpumask_set_cpu(cpu, tick_broadcast_on);
 380                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 381                         /*
 382                          * Only shutdown the cpu local device, if:
 383                          *
 384                          * - the broadcast device exists
 385                          * - the broadcast device is not a hrtimer based one
 386                          * - the broadcast device is in periodic mode to
 387                          *   avoid a hickup during switch to oneshot mode
 388                          */
 389                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 390                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 391                                 clockevents_shutdown(dev);
 392                 }
 393                 break;
 394
 395         case TICK_BROADCAST_OFF:
 396                 if (tick_broadcast_forced)
 397                         break;
 398                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 399                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 400                         if (tick_broadcast_device.mode ==
 401                             TICKDEV_MODE_PERIODIC)
 402                                 tick_setup_periodic(dev, 0);
 403                 }
 404                 break;
 405         }
 406
 407         if (bc) {
 408                 if (cpumask_empty(tick_broadcast_mask)) {
 409                         if (!bc_stopped)
 410                                 clockevents_shutdown(bc);
 411                 } else if (bc_stopped) {
 412                         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 413                                 tick_broadcast_start_periodic(bc);
 414                         else
 415                                 tick_broadcast_setup_oneshot(bc);
 416                 }
 417         }
 418 out:
 419         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 420 }
 421 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 422
 423 /*
 424  * Set the periodic handler depending on broadcast on/off
 425  */
 426 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 427 {
 428         if (!broadcast)
 429                 dev->event_handler = tick_handle_periodic;
 430         else
 431                 dev->event_handler = tick_handle_periodic_broadcast;
 432 }
 433
 434 #ifdef CONFIG_HOTPLUG_CPU
 435 /*
 436  * Remove a CPU from broadcasting
 437  */
 438 void tick_shutdown_broadcast(unsigned int cpu)
 439 {
 440         struct clock_event_device *bc;
 441         unsigned long flags;
 442
 443         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 444
 445         bc = tick_broadcast_device.evtdev;
 446         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 447         cpumask_clear_cpu(cpu, tick_broadcast_on);
 448
 449         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 450                 if (bc && cpumask_empty(tick_broadcast_mask))
 451                         clockevents_shutdown(bc);
 452         }
 453
 454         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 455 }
 456 #endif
 457
 458 void tick_suspend_broadcast(void)
 459 {
 460         struct clock_event_device *bc;
 461         unsigned long flags;
 462
 463         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 464
 465         bc = tick_broadcast_device.evtdev;
 466         if (bc)
 467                 clockevents_shutdown(bc);
 468
 469         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 470 }
 471
 472 /*
 473  * This is called from tick_resume_local() on a resuming CPU. That's
 474  * called from the core resume function, tick_unfreeze() and the magic XEN
 475  * resume hackery.
 476  *
 477  * In none of these cases the broadcast device mode can change and the
 478  * bit of the resuming CPU in the broadcast mask is safe as well.
 479  */
 480 bool tick_resume_check_broadcast(void)
 481 {
 482         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 483                 return false;
 484         else
 485                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 486 }
 487
 488 void tick_resume_broadcast(void)
 489 {
 490         struct clock_event_device *bc;
 491         unsigned long flags;
 492
 493         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 494
 495         bc = tick_broadcast_device.evtdev;
 496
 497         if (bc) {
 498                 clockevents_tick_resume(bc);
 499
 500                 switch (tick_broadcast_device.mode) {
 501                 case TICKDEV_MODE_PERIODIC:
 502                         if (!cpumask_empty(tick_broadcast_mask))
 503                                 tick_broadcast_start_periodic(bc);
 504                         break;
 505                 case TICKDEV_MODE_ONESHOT:
 506                         if (!cpumask_empty(tick_broadcast_mask))
 507                                 tick_resume_broadcast_oneshot(bc);
 508                         break;
 509                 }
 510         }
 511         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 512 }
 513
 514 #ifdef CONFIG_TICK_ONESHOT
 515
 516 static cpumask_var_t tick_broadcast_oneshot_mask __cpumask_var_read_mostly;
 517 static cpumask_var_t tick_broadcast_pending_mask __cpumask_var_read_mostly;
 518 static cpumask_var_t tick_broadcast_force_mask __cpumask_var_read_mostly;
 519
 520 /*
 521  * Exposed for debugging: see timer_list.c
 522  */
 523 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 524 {
 525         return tick_broadcast_oneshot_mask;
 526 }
 527
 528 /*
 529  * Called before going idle with interrupts disabled. Checks whether a
 530  * broadcast event from the other core is about to happen. We detected
 531  * that in tick_broadcast_oneshot_control(). The callsite can use this
 532  * to avoid a deep idle transition as we are about to get the
 533  * broadcast IPI right away.
 534  */
 535 int tick_check_broadcast_expired(void)
 536 {
 537         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 538 }
 539
 540 /*
 541  * Set broadcast interrupt affinity
 542  */
 543 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 544                                         const struct cpumask *cpumask)
 545 {
 546         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 547                 return;
 548
 549         if (cpumask_equal(bc->cpumask, cpumask))
 550                 return;
 551
 552         bc->cpumask = cpumask;
 553         irq_set_affinity(bc->irq, bc->cpumask);
 554 }
 555
 556 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 557                                      ktime_t expires)
 558 {
 559         if (!clockevent_state_oneshot(bc))
 560                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 561
 562         clockevents_program_event(bc, expires, 1);
 563         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 564 }
 565
 566 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 567 {
 568         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 569 }
 570
 571 /*
 572  * Called from irq_enter() when idle was interrupted to reenable the
 573  * per cpu device.
 574  */
 575 void tick_check_oneshot_broadcast_this_cpu(void)
 576 {
 577         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 578                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 579
 580                 /*
 581                  * We might be in the middle of switching over from
 582                  * periodic to oneshot. If the CPU has not yet
 583                  * switched over, leave the device alone.
 584                  */
 585                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 586                         clockevents_switch_state(td->evtdev,
 587                                               CLOCK_EVT_STATE_ONESHOT);
 588                 }
 589         }
 590 }
 591
 592 /*
 593  * Handle oneshot mode broadcasting
 594  */
 595 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 596 {
 597         struct tick_device *td;
 598         ktime_t now, next_event;
 599         int cpu, next_cpu = 0;
 600         bool bc_local;
 601
 602         raw_spin_lock(&tick_broadcast_lock);
 603         dev->next_event = KTIME_MAX;
 604         next_event = KTIME_MAX;
 605         cpumask_clear(tmpmask);
 606         now = ktime_get();
 607         /* Find all expired events */
 608         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 609                 /*
 610                  * Required for !SMP because for_each_cpu() reports
 611                  * unconditionally CPU0 as set on UP kernels.
 612                  */
 613                 if (!IS_ENABLED(CONFIG_SMP) &&
 614                     cpumask_empty(tick_broadcast_oneshot_mask))
 615                         break;
 616
 617                 td = &per_cpu(tick_cpu_device, cpu);
 618                 if (td->evtdev->next_event <= now) {
 619                         cpumask_set_cpu(cpu, tmpmask);
 620                         /*
 621                          * Mark the remote cpu in the pending mask, so
 622                          * it can avoid reprogramming the cpu local
 623                          * timer in tick_broadcast_oneshot_control().
 624                          */
 625                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 626                 } else if (td->evtdev->next_event < next_event) {
 627                         next_event = td->evtdev->next_event;
 628                         next_cpu = cpu;
 629                 }
 630         }
 631
 632         /*
 633          * Remove the current cpu from the pending mask. The event is
 634          * delivered immediately in tick_do_broadcast() !
 635          */
 636         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 637
 638         /* Take care of enforced broadcast requests */
 639         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 640         cpumask_clear(tick_broadcast_force_mask);
 641
 642         /*
 643          * Sanity check. Catch the case where we try to broadcast to
 644          * offline cpus.
 645          */
 646         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 647                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 648
 649         /*
 650          * Wakeup the cpus which have an expired event.
 651          */
 652         bc_local = tick_do_broadcast(tmpmask);
 653
 654         /*
 655          * Two reasons for reprogram:
 656          *
 657          * - The global event did not expire any CPU local
 658          * events. This happens in dyntick mode, as the maximum PIT
 659          * delta is quite small.
 660          *
 661          * - There are pending events on sleeping CPUs which were not
 662          * in the event mask
 663          */
 664         if (next_event != KTIME_MAX)
 665                 tick_broadcast_set_event(dev, next_cpu, next_event);
 666
 667         raw_spin_unlock(&tick_broadcast_lock);
 668
 669         if (bc_local) {
 670                 td = this_cpu_ptr(&tick_cpu_device);
 671                 td->evtdev->event_handler(td->evtdev);
 672         }
 673 }
 674
 675 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 676 {
 677         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 678                 return 0;
 679         if (bc->next_event == KTIME_MAX)
 680                 return 0;
 681         return bc->bound_on == cpu ? -EBUSY : 0;
 682 }
 683
 684 static void broadcast_shutdown_local(struct clock_event_device *bc,
 685                                      struct clock_event_device *dev)
 686 {
 687         /*
 688          * For hrtimer based broadcasting we cannot shutdown the cpu
 689          * local device if our own event is the first one to expire or
 690          * if we own the broadcast timer.
 691          */
 692         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 693                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 694                         return;
 695                 if (dev->next_event < bc->next_event)
 696                         return;
 697         }
 698         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 699 }
 700
 701 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 702 {
 703         struct clock_event_device *bc, *dev;
 704         int cpu, ret = 0;
 705         ktime_t now;
 706
 707         /*
 708          * If there is no broadcast device, tell the caller not to go
 709          * into deep idle.
 710          */
 711         if (!tick_broadcast_device.evtdev)
 712                 return -EBUSY;
 713
 714         dev = this_cpu_ptr(&tick_cpu_device)->evtdev;
 715
 716         raw_spin_lock(&tick_broadcast_lock);
 717         bc = tick_broadcast_device.evtdev;
 718         cpu = smp_processor_id();
 719
 720         if (state == TICK_BROADCAST_ENTER) {
 721                 /*
 722                  * If the current CPU owns the hrtimer broadcast
 723                  * mechanism, it cannot go deep idle and we do not add
 724                  * the CPU to the broadcast mask. We don't have to go
 725                  * through the EXIT path as the local timer is not
 726                  * shutdown.
 727                  */
 728                 ret = broadcast_needs_cpu(bc, cpu);
 729                 if (ret)
 730                         goto out;
 731
 732                 /*
 733                  * If the broadcast device is in periodic mode, we
 734                  * return.
 735                  */
 736                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 737                         /* If it is a hrtimer based broadcast, return busy */
 738                         if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
 739                                 ret = -EBUSY;
 740                         goto out;
 741                 }
 742
 743                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 744                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 745
 746                         /* Conditionally shut down the local timer. */
 747                         broadcast_shutdown_local(bc, dev);
 748
 749                         /*
 750                          * We only reprogram the broadcast timer if we
 751                          * did not mark ourself in the force mask and
 752                          * if the cpu local event is earlier than the
 753                          * broadcast event. If the current CPU is in
 754                          * the force mask, then we are going to be
 755                          * woken by the IPI right away; we return
 756                          * busy, so the CPU does not try to go deep
 757                          * idle.
 758                          */
 759                         if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
 760                                 ret = -EBUSY;
 761                         } else if (dev->next_event < bc->next_event) {
 762                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 763                                 /*
 764                                  * In case of hrtimer broadcasts the
 765                                  * programming might have moved the
 766                                  * timer to this cpu. If yes, remove
 767                                  * us from the broadcast mask and
 768                                  * return busy.
 769                                  */
 770                                 ret = broadcast_needs_cpu(bc, cpu);
 771                                 if (ret) {
 772                                         cpumask_clear_cpu(cpu,
 773                                                 tick_broadcast_oneshot_mask);
 774                                 }
 775                         }
 776                 }
 777         } else {
 778                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 779                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 780                         /*
 781                          * The cpu which was handling the broadcast
 782                          * timer marked this cpu in the broadcast
 783                          * pending mask and fired the broadcast
 784                          * IPI. So we are going to handle the expired
 785                          * event anyway via the broadcast IPI
 786                          * handler. No need to reprogram the timer
 787                          * with an already expired event.
 788                          */
 789                         if (cpumask_test_and_clear_cpu(cpu,
 790                                        tick_broadcast_pending_mask))
 791                                 goto out;
 792
 793                         /*
 794                          * Bail out if there is no next event.
 795                          */
 796                         if (dev->next_event == KTIME_MAX)
 797                                 goto out;
 798                         /*
 799                          * If the pending bit is not set, then we are
 800                          * either the CPU handling the broadcast
 801                          * interrupt or we got woken by something else.
 802                          *
 803                          * We are not longer in the broadcast mask, so
 804                          * if the cpu local expiry time is already
 805                          * reached, we would reprogram the cpu local
 806                          * timer with an already expired event.
 807                          *
 808                          * This can lead to a ping-pong when we return
 809                          * to idle and therefor rearm the broadcast
 810                          * timer before the cpu local timer was able
 811                          * to fire. This happens because the forced
 812                          * reprogramming makes sure that the event
 813                          * will happen in the future and depending on
 814                          * the min_delta setting this might be far
 815                          * enough out that the ping-pong starts.
 816                          *
 817                          * If the cpu local next_event has expired
 818                          * then we know that the broadcast timer
 819                          * next_event has expired as well and
 820                          * broadcast is about to be handled. So we
 821                          * avoid reprogramming and enforce that the
 822                          * broadcast handler, which did not run yet,
 823                          * will invoke the cpu local handler.
 824                          *
 825                          * We cannot call the handler directly from
 826                          * here, because we might be in a NOHZ phase
 827                          * and we did not go through the irq_enter()
 828                          * nohz fixups.
 829                          */
 830                         now = ktime_get();
 831                         if (dev->next_event <= now) {
 832                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 833                                 goto out;
 834                         }
 835                         /*
 836                          * We got woken by something else. Reprogram
 837                          * the cpu local timer device.
 838                          */
 839                         tick_program_event(dev->next_event, 1);
 840                 }
 841         }
 842 out:
 843         raw_spin_unlock(&tick_broadcast_lock);
 844         return ret;
 845 }
 846
 847 /*
 848  * Reset the one shot broadcast for a cpu
 849  *
 850  * Called with tick_broadcast_lock held
 851  */
 852 static void tick_broadcast_clear_oneshot(int cpu)
 853 {
 854         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 855         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 856 }
 857
 858 static void tick_broadcast_init_next_event(struct cpumask *mask,
 859                                            ktime_t expires)
 860 {
 861         struct tick_device *td;
 862         int cpu;
 863
 864         for_each_cpu(cpu, mask) {
 865                 td = &per_cpu(tick_cpu_device, cpu);
 866                 if (td->evtdev)
 867                         td->evtdev->next_event = expires;
 868         }
 869 }
 870
 871 /**
 872  * tick_broadcast_setup_oneshot - setup the broadcast device
 873  */
 874 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 875 {
 876         int cpu = smp_processor_id();
 877
 878         if (!bc)
 879                 return;
 880
 881         /* Set it up only once ! */
 882         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 883                 int was_periodic = clockevent_state_periodic(bc);
 884
 885                 bc->event_handler = tick_handle_oneshot_broadcast;
 886
 887                 /*
 888                  * We must be careful here. There might be other CPUs
 889                  * waiting for periodic broadcast. We need to set the
 890                  * oneshot_mask bits for those and program the
 891                  * broadcast device to fire.
 892                  */
 893                 cpumask_copy(tmpmask, tick_broadcast_mask);
 894                 cpumask_clear_cpu(cpu, tmpmask);
 895                 cpumask_or(tick_broadcast_oneshot_mask,
 896                            tick_broadcast_oneshot_mask, tmpmask);
 897
 898                 if (was_periodic && !cpumask_empty(tmpmask)) {
 899                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 900                         tick_broadcast_init_next_event(tmpmask,
 901                                                        tick_next_period);
 902                         tick_broadcast_set_event(bc, cpu, tick_next_period);
 903                 } else
 904                         bc->next_event = KTIME_MAX;
 905         } else {
 906                 /*
 907                  * The first cpu which switches to oneshot mode sets
 908                  * the bit for all other cpus which are in the general
 909                  * (periodic) broadcast mask. So the bit is set and
 910                  * would prevent the first broadcast enter after this
 911                  * to program the bc device.
 912                  */
 913                 tick_broadcast_clear_oneshot(cpu);
 914         }
 915 }
 916
 917 /*
 918  * Select oneshot operating mode for the broadcast device
 919  */
 920 void tick_broadcast_switch_to_oneshot(void)
 921 {
 922         struct clock_event_device *bc;
 923         unsigned long flags;
 924
 925         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 926
 927         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 928         bc = tick_broadcast_device.evtdev;
 929         if (bc)
 930                 tick_broadcast_setup_oneshot(bc);
 931
 932         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 933 }
 934
 935 #ifdef CONFIG_HOTPLUG_CPU
 936 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
 937 {
 938         struct clock_event_device *bc;
 939         unsigned long flags;
 940
 941         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 942         bc = tick_broadcast_device.evtdev;
 943
 944         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
 945                 /* This moves the broadcast assignment to this CPU: */
 946                 clockevents_program_event(bc, bc->next_event, 1);
 947         }
 948         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 949 }
 950
 951 /*
 952  * Remove a dead CPU from broadcasting
 953  */
 954 void tick_shutdown_broadcast_oneshot(unsigned int cpu)
 955 {
 956         unsigned long flags;
 957
 958         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 959
 960         /*
 961          * Clear the broadcast masks for the dead cpu, but do not stop
 962          * the broadcast device!
 963          */
 964         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 965         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 966         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
 967
 968         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 969 }
 970 #endif
 971
 972 /*
 973  * Check, whether the broadcast device is in one shot mode
 974  */
 975 int tick_broadcast_oneshot_active(void)
 976 {
 977         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 978 }
 979
 980 /*
 981  * Check whether the broadcast device supports oneshot.
 982  */
 983 bool tick_broadcast_oneshot_available(void)
 984 {
 985         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 986
 987         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 988 }
 989
 990 #else
 991 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 992 {
 993         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 994
 995         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
 996                 return -EBUSY;
 997
 998         return 0;
 999 }
1000 #endif
1001
1002 void __init tick_broadcast_init(void)
1003 {
1004         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
1005         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
1006         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1007 #ifdef CONFIG_TICK_ONESHOT
1008         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1009         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1010         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1011 #endif
1012 }