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