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
  23 #include "tick-internal.h"
  24
  25 /*
  26  * Broadcast support for broken x86 hardware, where the local apic
  27  * timer stops in C3 state.
  28  */
  29
  30 static struct tick_device tick_broadcast_device;
  31 static cpumask_var_t tick_broadcast_mask;
  32 static cpumask_var_t tmpmask;
  33 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  34 static int tick_broadcast_force;
  35
  36 #ifdef CONFIG_TICK_ONESHOT
  37 static void tick_broadcast_clear_oneshot(int cpu);
  38 #else
  39 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  40 #endif
  41
  42 /*
  43  * Debugging: see timer_list.c
  44  */
  45 struct tick_device *tick_get_broadcast_device(void)
  46 {
  47         return &tick_broadcast_device;
  48 }
  49
  50 struct cpumask *tick_get_broadcast_mask(void)
  51 {
  52         return tick_broadcast_mask;
  53 }
  54
  55 /*
  56  * Start the device in periodic mode
  57  */
  58 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  59 {
  60         if (bc)
  61                 tick_setup_periodic(bc, 1);
  62 }
  63
  64 /*
  65  * Check, if the device can be utilized as broadcast device:
  66  */
  67 void tick_install_broadcast_device(struct clock_event_device *dev)
  68 {
  69         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  70
  71         if ((dev->features & CLOCK_EVT_FEAT_DUMMY) ||
  72             (tick_broadcast_device.evtdev &&
  73              tick_broadcast_device.evtdev->rating >= dev->rating) ||
  74              (dev->features & CLOCK_EVT_FEAT_C3STOP))
  75                 return;
  76
  77         clockevents_exchange_device(tick_broadcast_device.evtdev, dev);
  78         if (cur)
  79                 cur->event_handler = clockevents_handle_noop;
  80         tick_broadcast_device.evtdev = dev;
  81         if (!cpumask_empty(tick_broadcast_mask))
  82                 tick_broadcast_start_periodic(dev);
  83         /*
  84          * Inform all cpus about this. We might be in a situation
  85          * where we did not switch to oneshot mode because the per cpu
  86          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
  87          * of a oneshot capable broadcast device. Without that
  88          * notification the systems stays stuck in periodic mode
  89          * forever.
  90          */
  91         if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
  92                 tick_clock_notify();
  93 }
  94
  95 /*
  96  * Check, if the device is the broadcast device
  97  */
  98 int tick_is_broadcast_device(struct clock_event_device *dev)
  99 {
 100         return (dev && tick_broadcast_device.evtdev == dev);
 101 }
 102
 103 static void err_broadcast(const struct cpumask *mask)
 104 {
 105         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 106 }
 107
 108 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 109 {
 110         if (!dev->broadcast)
 111                 dev->broadcast = tick_broadcast;
 112         if (!dev->broadcast) {
 113                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 114                              dev->name);
 115                 dev->broadcast = err_broadcast;
 116         }
 117 }
 118
 119 /*
 120  * Check, if the device is disfunctional and a place holder, which
 121  * needs to be handled by the broadcast device.
 122  */
 123 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 124 {
 125         unsigned long flags;
 126         int ret = 0;
 127
 128         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 129
 130         /*
 131          * Devices might be registered with both periodic and oneshot
 132          * mode disabled. This signals, that the device needs to be
 133          * operated from the broadcast device and is a placeholder for
 134          * the cpu local device.
 135          */
 136         if (!tick_device_is_functional(dev)) {
 137                 dev->event_handler = tick_handle_periodic;
 138                 tick_device_setup_broadcast_func(dev);
 139                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 140                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
 141                 ret = 1;
 142         } else {
 143                 /*
 144                  * When the new device is not affected by the stop
 145                  * feature and the cpu is marked in the broadcast mask
 146                  * then clear the broadcast bit.
 147                  */
 148                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
 149                         int cpu = smp_processor_id();
 150                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 151                         tick_broadcast_clear_oneshot(cpu);
 152                 } else {
 153                         tick_device_setup_broadcast_func(dev);
 154                 }
 155         }
 156         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 157         return ret;
 158 }
 159
 160 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 161 int tick_receive_broadcast(void)
 162 {
 163         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 164         struct clock_event_device *evt = td->evtdev;
 165
 166         if (!evt)
 167                 return -ENODEV;
 168
 169         if (!evt->event_handler)
 170                 return -EINVAL;
 171
 172         evt->event_handler(evt);
 173         return 0;
 174 }
 175 #endif
 176
 177 /*
 178  * Broadcast the event to the cpus, which are set in the mask (mangled).
 179  */
 180 static void tick_do_broadcast(struct cpumask *mask)
 181 {
 182         int cpu = smp_processor_id();
 183         struct tick_device *td;
 184
 185         /*
 186          * Check, if the current cpu is in the mask
 187          */
 188         if (cpumask_test_cpu(cpu, mask)) {
 189                 cpumask_clear_cpu(cpu, mask);
 190                 td = &per_cpu(tick_cpu_device, cpu);
 191                 td->evtdev->event_handler(td->evtdev);
 192         }
 193
 194         if (!cpumask_empty(mask)) {
 195                 /*
 196                  * It might be necessary to actually check whether the devices
 197                  * have different broadcast functions. For now, just use the
 198                  * one of the first device. This works as long as we have this
 199                  * misfeature only on x86 (lapic)
 200                  */
 201                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 202                 td->evtdev->broadcast(mask);
 203         }
 204 }
 205
 206 /*
 207  * Periodic broadcast:
 208  * - invoke the broadcast handlers
 209  */
 210 static void tick_do_periodic_broadcast(void)
 211 {
 212         raw_spin_lock(&tick_broadcast_lock);
 213
 214         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 215         tick_do_broadcast(tmpmask);
 216
 217         raw_spin_unlock(&tick_broadcast_lock);
 218 }
 219
 220 /*
 221  * Event handler for periodic broadcast ticks
 222  */
 223 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 224 {
 225         ktime_t next;
 226
 227         tick_do_periodic_broadcast();
 228
 229         /*
 230          * The device is in periodic mode. No reprogramming necessary:
 231          */
 232         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
 233                 return;
 234
 235         /*
 236          * Setup the next period for devices, which do not have
 237          * periodic mode. We read dev->next_event first and add to it
 238          * when the event already expired. clockevents_program_event()
 239          * sets dev->next_event only when the event is really
 240          * programmed to the device.
 241          */
 242         for (next = dev->next_event; ;) {
 243                 next = ktime_add(next, tick_period);
 244
 245                 if (!clockevents_program_event(dev, next, false))
 246                         return;
 247                 tick_do_periodic_broadcast();
 248         }
 249 }
 250
 251 /*
 252  * Powerstate information: The system enters/leaves a state, where
 253  * affected devices might stop
 254  */
 255 static void tick_do_broadcast_on_off(unsigned long *reason)
 256 {
 257         struct clock_event_device *bc, *dev;
 258         struct tick_device *td;
 259         unsigned long flags;
 260         int cpu, bc_stopped;
 261
 262         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 263
 264         cpu = smp_processor_id();
 265         td = &per_cpu(tick_cpu_device, cpu);
 266         dev = td->evtdev;
 267         bc = tick_broadcast_device.evtdev;
 268
 269         /*
 270          * Is the device not affected by the powerstate ?
 271          */
 272         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 273                 goto out;
 274
 275         if (!tick_device_is_functional(dev))
 276                 goto out;
 277
 278         bc_stopped = cpumask_empty(tick_broadcast_mask);
 279
 280         switch (*reason) {
 281         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
 282         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
 283                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 284                         if (tick_broadcast_device.mode ==
 285                             TICKDEV_MODE_PERIODIC)
 286                                 clockevents_shutdown(dev);
 287                 }
 288                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
 289                         tick_broadcast_force = 1;
 290                 break;
 291         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
 292                 if (!tick_broadcast_force &&
 293                     cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 294                         if (tick_broadcast_device.mode ==
 295                             TICKDEV_MODE_PERIODIC)
 296                                 tick_setup_periodic(dev, 0);
 297                 }
 298                 break;
 299         }
 300
 301         if (cpumask_empty(tick_broadcast_mask)) {
 302                 if (!bc_stopped)
 303                         clockevents_shutdown(bc);
 304         } else if (bc_stopped) {
 305                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 306                         tick_broadcast_start_periodic(bc);
 307                 else
 308                         tick_broadcast_setup_oneshot(bc);
 309         }
 310 out:
 311         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 312 }
 313
 314 /*
 315  * Powerstate information: The system enters/leaves a state, where
 316  * affected devices might stop.
 317  */
 318 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
 319 {
 320         if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
 321                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
 322                        "offline CPU #%d\n", *oncpu);
 323         else
 324                 tick_do_broadcast_on_off(&reason);
 325 }
 326
 327 /*
 328  * Set the periodic handler depending on broadcast on/off
 329  */
 330 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 331 {
 332         if (!broadcast)
 333                 dev->event_handler = tick_handle_periodic;
 334         else
 335                 dev->event_handler = tick_handle_periodic_broadcast;
 336 }
 337
 338 /*
 339  * Remove a CPU from broadcasting
 340  */
 341 void tick_shutdown_broadcast(unsigned int *cpup)
 342 {
 343         struct clock_event_device *bc;
 344         unsigned long flags;
 345         unsigned int cpu = *cpup;
 346
 347         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 348
 349         bc = tick_broadcast_device.evtdev;
 350         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 351
 352         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 353                 if (bc && cpumask_empty(tick_broadcast_mask))
 354                         clockevents_shutdown(bc);
 355         }
 356
 357         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 358 }
 359
 360 void tick_suspend_broadcast(void)
 361 {
 362         struct clock_event_device *bc;
 363         unsigned long flags;
 364
 365         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 366
 367         bc = tick_broadcast_device.evtdev;
 368         if (bc)
 369                 clockevents_shutdown(bc);
 370
 371         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 372 }
 373
 374 int tick_resume_broadcast(void)
 375 {
 376         struct clock_event_device *bc;
 377         unsigned long flags;
 378         int broadcast = 0;
 379
 380         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 381
 382         bc = tick_broadcast_device.evtdev;
 383
 384         if (bc) {
 385                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
 386
 387                 switch (tick_broadcast_device.mode) {
 388                 case TICKDEV_MODE_PERIODIC:
 389                         if (!cpumask_empty(tick_broadcast_mask))
 390                                 tick_broadcast_start_periodic(bc);
 391                         broadcast = cpumask_test_cpu(smp_processor_id(),
 392                                                      tick_broadcast_mask);
 393                         break;
 394                 case TICKDEV_MODE_ONESHOT:
 395                         if (!cpumask_empty(tick_broadcast_mask))
 396                                 broadcast = tick_resume_broadcast_oneshot(bc);
 397                         break;
 398                 }
 399         }
 400         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 401
 402         return broadcast;
 403 }
 404
 405
 406 #ifdef CONFIG_TICK_ONESHOT
 407
 408 static cpumask_var_t tick_broadcast_oneshot_mask;
 409 static cpumask_var_t tick_broadcast_pending_mask;
 410 static cpumask_var_t tick_broadcast_force_mask;
 411
 412 /*
 413  * Exposed for debugging: see timer_list.c
 414  */
 415 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 416 {
 417         return tick_broadcast_oneshot_mask;
 418 }
 419
 420 /*
 421  * Called before going idle with interrupts disabled. Checks whether a
 422  * broadcast event from the other core is about to happen. We detected
 423  * that in tick_broadcast_oneshot_control(). The callsite can use this
 424  * to avoid a deep idle transition as we are about to get the
 425  * broadcast IPI right away.
 426  */
 427 int tick_check_broadcast_expired(void)
 428 {
 429         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 430 }
 431
 432 /*
 433  * Set broadcast interrupt affinity
 434  */
 435 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 436                                         const struct cpumask *cpumask)
 437 {
 438         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 439                 return;
 440
 441         if (cpumask_equal(bc->cpumask, cpumask))
 442                 return;
 443
 444         bc->cpumask = cpumask;
 445         irq_set_affinity(bc->irq, bc->cpumask);
 446 }
 447
 448 static int tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 449                                     ktime_t expires, int force)
 450 {
 451         int ret;
 452
 453         if (bc->mode != CLOCK_EVT_MODE_ONESHOT)
 454                 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 455
 456         ret = clockevents_program_event(bc, expires, force);
 457         if (!ret)
 458                 tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 459         return ret;
 460 }
 461
 462 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 463 {
 464         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 465         return 0;
 466 }
 467
 468 /*
 469  * Called from irq_enter() when idle was interrupted to reenable the
 470  * per cpu device.
 471  */
 472 void tick_check_oneshot_broadcast(int cpu)
 473 {
 474         if (cpumask_test_cpu(cpu, tick_broadcast_oneshot_mask)) {
 475                 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 476
 477                 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
 478         }
 479 }
 480
 481 /*
 482  * Handle oneshot mode broadcasting
 483  */
 484 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 485 {
 486         struct tick_device *td;
 487         ktime_t now, next_event;
 488         int cpu, next_cpu = 0;
 489
 490         raw_spin_lock(&tick_broadcast_lock);
 491 again:
 492         dev->next_event.tv64 = KTIME_MAX;
 493         next_event.tv64 = KTIME_MAX;
 494         cpumask_clear(tmpmask);
 495         now = ktime_get();
 496         /* Find all expired events */
 497         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 498                 td = &per_cpu(tick_cpu_device, cpu);
 499                 if (td->evtdev->next_event.tv64 <= now.tv64) {
 500                         cpumask_set_cpu(cpu, tmpmask);
 501                         /*
 502                          * Mark the remote cpu in the pending mask, so
 503                          * it can avoid reprogramming the cpu local
 504                          * timer in tick_broadcast_oneshot_control().
 505                          */
 506                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 507                 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
 508                         next_event.tv64 = td->evtdev->next_event.tv64;
 509                         next_cpu = cpu;
 510                 }
 511         }
 512
 513         /*
 514          * Remove the current cpu from the pending mask. The event is
 515          * delivered immediately in tick_do_broadcast() !
 516          */
 517         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 518
 519         /* Take care of enforced broadcast requests */
 520         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 521         cpumask_clear(tick_broadcast_force_mask);
 522
 523         /*
 524          * Wakeup the cpus which have an expired event.
 525          */
 526         tick_do_broadcast(tmpmask);
 527
 528         /*
 529          * Two reasons for reprogram:
 530          *
 531          * - The global event did not expire any CPU local
 532          * events. This happens in dyntick mode, as the maximum PIT
 533          * delta is quite small.
 534          *
 535          * - There are pending events on sleeping CPUs which were not
 536          * in the event mask
 537          */
 538         if (next_event.tv64 != KTIME_MAX) {
 539                 /*
 540                  * Rearm the broadcast device. If event expired,
 541                  * repeat the above
 542                  */
 543                 if (tick_broadcast_set_event(dev, next_cpu, next_event, 0))
 544                         goto again;
 545         }
 546         raw_spin_unlock(&tick_broadcast_lock);
 547 }
 548
 549 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 550 {
 551         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 552                 return 0;
 553         if (bc->next_event.tv64 == KTIME_MAX)
 554                 return 0;
 555         return bc->bound_on == cpu ? -EBUSY : 0;
 556 }
 557
 558 static void broadcast_shutdown_local(struct clock_event_device *bc,
 559                                      struct clock_event_device *dev)
 560 {
 561         /*
 562          * For hrtimer based broadcasting we cannot shutdown the cpu
 563          * local device if our own event is the first one to expire or
 564          * if we own the broadcast timer.
 565          */
 566         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 567                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 568                         return;
 569                 if (dev->next_event.tv64 < bc->next_event.tv64)
 570                         return;
 571         }
 572         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
 573 }
 574
 575 static void broadcast_move_bc(int deadcpu)
 576 {
 577         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 578
 579         if (!bc || !broadcast_needs_cpu(bc, deadcpu))
 580                 return;
 581         /* This moves the broadcast assignment to this cpu */
 582         clockevents_program_event(bc, bc->next_event, 1);
 583 }
 584
 585 /*
 586  * Powerstate information: The system enters/leaves a state, where
 587  * affected devices might stop
 588  * Returns 0 on success, -EBUSY if the cpu is used to broadcast wakeups.
 589  */
 590 int tick_broadcast_oneshot_control(unsigned long reason)
 591 {
 592         struct clock_event_device *bc, *dev;
 593         struct tick_device *td;
 594         unsigned long flags;
 595         ktime_t now;
 596         int cpu, ret = 0;
 597
 598         /*
 599          * Periodic mode does not care about the enter/exit of power
 600          * states
 601          */
 602         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 603                 return 0;
 604
 605         /*
 606          * We are called with preemtion disabled from the depth of the
 607          * idle code, so we can't be moved away.
 608          */
 609         cpu = smp_processor_id();
 610         td = &per_cpu(tick_cpu_device, cpu);
 611         dev = td->evtdev;
 612
 613         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 614                 return 0;
 615
 616         bc = tick_broadcast_device.evtdev;
 617
 618         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 619         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
 620                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 621                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 622                         broadcast_shutdown_local(bc, dev);
 623                         /*
 624                          * We only reprogram the broadcast timer if we
 625                          * did not mark ourself in the force mask and
 626                          * if the cpu local event is earlier than the
 627                          * broadcast event. If the current CPU is in
 628                          * the force mask, then we are going to be
 629                          * woken by the IPI right away.
 630                          */
 631                         if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
 632                             dev->next_event.tv64 < bc->next_event.tv64)
 633                                 tick_broadcast_set_event(bc, cpu, dev->next_event, 1);
 634                 }
 635                 /*
 636                  * If the current CPU owns the hrtimer broadcast
 637                  * mechanism, it cannot go deep idle and we remove the
 638                  * CPU from the broadcast mask. We don't have to go
 639                  * through the EXIT path as the local timer is not
 640                  * shutdown.
 641                  */
 642                 ret = broadcast_needs_cpu(bc, cpu);
 643                 if (ret)
 644                         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 645         } else {
 646                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 647                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 648                         /*
 649                          * The cpu which was handling the broadcast
 650                          * timer marked this cpu in the broadcast
 651                          * pending mask and fired the broadcast
 652                          * IPI. So we are going to handle the expired
 653                          * event anyway via the broadcast IPI
 654                          * handler. No need to reprogram the timer
 655                          * with an already expired event.
 656                          */
 657                         if (cpumask_test_and_clear_cpu(cpu,
 658                                        tick_broadcast_pending_mask))
 659                                 goto out;
 660
 661                         /*
 662                          * Bail out if there is no next event.
 663                          */
 664                         if (dev->next_event.tv64 == KTIME_MAX)
 665                                 goto out;
 666                         /*
 667                          * If the pending bit is not set, then we are
 668                          * either the CPU handling the broadcast
 669                          * interrupt or we got woken by something else.
 670                          *
 671                          * We are not longer in the broadcast mask, so
 672                          * if the cpu local expiry time is already
 673                          * reached, we would reprogram the cpu local
 674                          * timer with an already expired event.
 675                          *
 676                          * This can lead to a ping-pong when we return
 677                          * to idle and therefor rearm the broadcast
 678                          * timer before the cpu local timer was able
 679                          * to fire. This happens because the forced
 680                          * reprogramming makes sure that the event
 681                          * will happen in the future and depending on
 682                          * the min_delta setting this might be far
 683                          * enough out that the ping-pong starts.
 684                          *
 685                          * If the cpu local next_event has expired
 686                          * then we know that the broadcast timer
 687                          * next_event has expired as well and
 688                          * broadcast is about to be handled. So we
 689                          * avoid reprogramming and enforce that the
 690                          * broadcast handler, which did not run yet,
 691                          * will invoke the cpu local handler.
 692                          *
 693                          * We cannot call the handler directly from
 694                          * here, because we might be in a NOHZ phase
 695                          * and we did not go through the irq_enter()
 696                          * nohz fixups.
 697                          */
 698                         now = ktime_get();
 699                         if (dev->next_event.tv64 <= now.tv64) {
 700                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 701                                 goto out;
 702                         }
 703                         /*
 704                          * We got woken by something else. Reprogram
 705                          * the cpu local timer device.
 706                          */
 707                         tick_program_event(dev->next_event, 1);
 708                 }
 709         }
 710 out:
 711         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 712         return ret;
 713 }
 714
 715 /*
 716  * Reset the one shot broadcast for a cpu
 717  *
 718  * Called with tick_broadcast_lock held
 719  */
 720 static void tick_broadcast_clear_oneshot(int cpu)
 721 {
 722         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 723 }
 724
 725 static void tick_broadcast_init_next_event(struct cpumask *mask,
 726                                            ktime_t expires)
 727 {
 728         struct tick_device *td;
 729         int cpu;
 730
 731         for_each_cpu(cpu, mask) {
 732                 td = &per_cpu(tick_cpu_device, cpu);
 733                 if (td->evtdev)
 734                         td->evtdev->next_event = expires;
 735         }
 736 }
 737
 738 /**
 739  * tick_broadcast_setup_oneshot - setup the broadcast device
 740  */
 741 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 742 {
 743         int cpu = smp_processor_id();
 744
 745         /* Set it up only once ! */
 746         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 747                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
 748
 749                 bc->event_handler = tick_handle_oneshot_broadcast;
 750
 751                 /*
 752                  * We must be careful here. There might be other CPUs
 753                  * waiting for periodic broadcast. We need to set the
 754                  * oneshot_mask bits for those and program the
 755                  * broadcast device to fire.
 756                  */
 757                 cpumask_copy(tmpmask, tick_broadcast_mask);
 758                 cpumask_clear_cpu(cpu, tmpmask);
 759                 cpumask_or(tick_broadcast_oneshot_mask,
 760                            tick_broadcast_oneshot_mask, tmpmask);
 761
 762                 if (was_periodic && !cpumask_empty(tmpmask)) {
 763                         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 764                         tick_broadcast_init_next_event(tmpmask,
 765                                                        tick_next_period);
 766                         tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
 767                 } else
 768                         bc->next_event.tv64 = KTIME_MAX;
 769         } else {
 770                 /*
 771                  * The first cpu which switches to oneshot mode sets
 772                  * the bit for all other cpus which are in the general
 773                  * (periodic) broadcast mask. So the bit is set and
 774                  * would prevent the first broadcast enter after this
 775                  * to program the bc device.
 776                  */
 777                 tick_broadcast_clear_oneshot(cpu);
 778         }
 779 }
 780
 781 /*
 782  * Select oneshot operating mode for the broadcast device
 783  */
 784 void tick_broadcast_switch_to_oneshot(void)
 785 {
 786         struct clock_event_device *bc;
 787         unsigned long flags;
 788
 789         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 790
 791         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 792         bc = tick_broadcast_device.evtdev;
 793         if (bc)
 794                 tick_broadcast_setup_oneshot(bc);
 795
 796         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 797 }
 798
 799
 800 /*
 801  * Remove a dead CPU from broadcasting
 802  */
 803 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
 804 {
 805         unsigned long flags;
 806         unsigned int cpu = *cpup;
 807
 808         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 809
 810         /*
 811          * Clear the broadcast mask flag for the dead cpu, but do not
 812          * stop the broadcast device!
 813          */
 814         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 815
 816         broadcast_move_bc(cpu);
 817
 818         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 819 }
 820
 821 /*
 822  * Check, whether the broadcast device is in one shot mode
 823  */
 824 int tick_broadcast_oneshot_active(void)
 825 {
 826         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 827 }
 828
 829 /*
 830  * Check whether the broadcast device supports oneshot.
 831  */
 832 bool tick_broadcast_oneshot_available(void)
 833 {
 834         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 835
 836         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 837 }
 838
 839 #endif
 840
 841 void __init tick_broadcast_init(void)
 842 {
 843         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 844         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
 845 #ifdef CONFIG_TICK_ONESHOT
 846         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
 847         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
 848         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
 849 #endif
 850 }