kernel/time/tick-broadcast.c

   1 /*
   2  * linux/kernel/time/tick-broadcast.c
   3  *
   4  * This file contains functions which emulate a local clock-event
   5  * device via a broadcast event source.
   6  *
   7  * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
   8  * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   9  * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
  10  *
  11  * This code is licenced under the GPL version 2. For details see
  12  * kernel-base/COPYING.
  13  */
  14 #include <linux/cpu.h>
  15 #include <linux/err.h>
  16 #include <linux/hrtimer.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/percpu.h>
  19 #include <linux/profile.h>
  20 #include <linux/sched.h>
  21 #include <linux/smp.h>
  22 #include <linux/module.h>
  23
  24 #include "tick-internal.h"
  25
  26 /*
  27  * Broadcast support for broken x86 hardware, where the local apic
  28  * timer stops in C3 state.
  29  */
  30
  31 static struct tick_device tick_broadcast_device;
  32 static cpumask_var_t tick_broadcast_mask;
  33 static cpumask_var_t tmpmask;
  34 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  35 static int tick_broadcast_force;
  36
  37 #ifdef CONFIG_TICK_ONESHOT
  38 static void tick_broadcast_clear_oneshot(int cpu);
  39 #else
  40 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  41 #endif
  42
  43 /*
  44  * Debugging: see timer_list.c
  45  */
  46 struct tick_device *tick_get_broadcast_device(void)
  47 {
  48         return &tick_broadcast_device;
  49 }
  50
  51 struct cpumask *tick_get_broadcast_mask(void)
  52 {
  53         return tick_broadcast_mask;
  54 }
  55
  56 /*
  57  * Start the device in periodic mode
  58  */
  59 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  60 {
  61         if (bc)
  62                 tick_setup_periodic(bc, 1);
  63 }
  64
  65 /*
  66  * Check, if the device can be utilized as broadcast device:
  67  */
  68 void tick_install_broadcast_device(struct clock_event_device *dev)
  69 {
  70         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  71
  72         if ((dev->features & CLOCK_EVT_FEAT_DUMMY) ||
  73             (tick_broadcast_device.evtdev &&
  74              tick_broadcast_device.evtdev->rating >= dev->rating) ||
  75              (dev->features & CLOCK_EVT_FEAT_C3STOP))
  76                 return;
  77         if (!try_module_get(dev->owner))
  78                 return;
  79
  80         clockevents_exchange_device(tick_broadcast_device.evtdev, dev);
  81         if (cur)
  82                 cur->event_handler = clockevents_handle_noop;
  83         tick_broadcast_device.evtdev = dev;
  84         if (!cpumask_empty(tick_broadcast_mask))
  85                 tick_broadcast_start_periodic(dev);
  86         /*
  87          * Inform all cpus about this. We might be in a situation
  88          * where we did not switch to oneshot mode because the per cpu
  89          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
  90          * of a oneshot capable broadcast device. Without that
  91          * notification the systems stays stuck in periodic mode
  92          * forever.
  93          */
  94         if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
  95                 tick_clock_notify();
  96 }
  97
  98 /*
  99  * Check, if the device is the broadcast device
 100  */
 101 int tick_is_broadcast_device(struct clock_event_device *dev)
 102 {
 103         return (dev && tick_broadcast_device.evtdev == dev);
 104 }
 105
 106 static void err_broadcast(const struct cpumask *mask)
 107 {
 108         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 109 }
 110
 111 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 112 {
 113         if (!dev->broadcast)
 114                 dev->broadcast = tick_broadcast;
 115         if (!dev->broadcast) {
 116                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 117                              dev->name);
 118                 dev->broadcast = err_broadcast;
 119         }
 120 }
 121
 122 /*
 123  * Check, if the device is disfunctional and a place holder, which
 124  * needs to be handled by the broadcast device.
 125  */
 126 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 127 {
 128         unsigned long flags;
 129         int ret = 0;
 130
 131         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 132
 133         /*
 134          * Devices might be registered with both periodic and oneshot
 135          * mode disabled. This signals, that the device needs to be
 136          * operated from the broadcast device and is a placeholder for
 137          * the cpu local device.
 138          */
 139         if (!tick_device_is_functional(dev)) {
 140                 dev->event_handler = tick_handle_periodic;
 141                 tick_device_setup_broadcast_func(dev);
 142                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 143                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
 144                 ret = 1;
 145         } else {
 146                 /*
 147                  * When the new device is not affected by the stop
 148                  * feature and the cpu is marked in the broadcast mask
 149                  * then clear the broadcast bit.
 150                  */
 151                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
 152                         int cpu = smp_processor_id();
 153                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 154                         tick_broadcast_clear_oneshot(cpu);
 155                 } else {
 156                         tick_device_setup_broadcast_func(dev);
 157                 }
 158         }
 159         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 160         return ret;
 161 }
 162
 163 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 164 int tick_receive_broadcast(void)
 165 {
 166         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 167         struct clock_event_device *evt = td->evtdev;
 168
 169         if (!evt)
 170                 return -ENODEV;
 171
 172         if (!evt->event_handler)
 173                 return -EINVAL;
 174
 175         evt->event_handler(evt);
 176         return 0;
 177 }
 178 #endif
 179
 180 /*
 181  * Broadcast the event to the cpus, which are set in the mask (mangled).
 182  */
 183 static void tick_do_broadcast(struct cpumask *mask)
 184 {
 185         int cpu = smp_processor_id();
 186         struct tick_device *td;
 187
 188         /*
 189          * Check, if the current cpu is in the mask
 190          */
 191         if (cpumask_test_cpu(cpu, mask)) {
 192                 cpumask_clear_cpu(cpu, mask);
 193                 td = &per_cpu(tick_cpu_device, cpu);
 194                 td->evtdev->event_handler(td->evtdev);
 195         }
 196
 197         if (!cpumask_empty(mask)) {
 198                 /*
 199                  * It might be necessary to actually check whether the devices
 200                  * have different broadcast functions. For now, just use the
 201                  * one of the first device. This works as long as we have this
 202                  * misfeature only on x86 (lapic)
 203                  */
 204                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 205                 td->evtdev->broadcast(mask);
 206         }
 207 }
 208
 209 /*
 210  * Periodic broadcast:
 211  * - invoke the broadcast handlers
 212  */
 213 static void tick_do_periodic_broadcast(void)
 214 {
 215         raw_spin_lock(&tick_broadcast_lock);
 216
 217         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 218         tick_do_broadcast(tmpmask);
 219
 220         raw_spin_unlock(&tick_broadcast_lock);
 221 }
 222
 223 /*
 224  * Event handler for periodic broadcast ticks
 225  */
 226 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 227 {
 228         ktime_t next;
 229
 230         tick_do_periodic_broadcast();
 231
 232         /*
 233          * The device is in periodic mode. No reprogramming necessary:
 234          */
 235         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
 236                 return;
 237
 238         /*
 239          * Setup the next period for devices, which do not have
 240          * periodic mode. We read dev->next_event first and add to it
 241          * when the event already expired. clockevents_program_event()
 242          * sets dev->next_event only when the event is really
 243          * programmed to the device.
 244          */
 245         for (next = dev->next_event; ;) {
 246                 next = ktime_add(next, tick_period);
 247
 248                 if (!clockevents_program_event(dev, next, false))
 249                         return;
 250                 tick_do_periodic_broadcast();
 251         }
 252 }
 253
 254 /*
 255  * Powerstate information: The system enters/leaves a state, where
 256  * affected devices might stop
 257  */
 258 static void tick_do_broadcast_on_off(unsigned long *reason)
 259 {
 260         struct clock_event_device *bc, *dev;
 261         struct tick_device *td;
 262         unsigned long flags;
 263         int cpu, bc_stopped;
 264
 265         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 266
 267         cpu = smp_processor_id();
 268         td = &per_cpu(tick_cpu_device, cpu);
 269         dev = td->evtdev;
 270         bc = tick_broadcast_device.evtdev;
 271
 272         /*
 273          * Is the device not affected by the powerstate ?
 274          */
 275         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 276                 goto out;
 277
 278         if (!tick_device_is_functional(dev))
 279                 goto out;
 280
 281         bc_stopped = cpumask_empty(tick_broadcast_mask);
 282
 283         switch (*reason) {
 284         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
 285         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
 286                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 287                         if (tick_broadcast_device.mode ==
 288                             TICKDEV_MODE_PERIODIC)
 289                                 clockevents_shutdown(dev);
 290                 }
 291                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
 292                         tick_broadcast_force = 1;
 293                 break;
 294         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
 295                 if (!tick_broadcast_force &&
 296                     cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 297                         if (tick_broadcast_device.mode ==
 298                             TICKDEV_MODE_PERIODIC)
 299                                 tick_setup_periodic(dev, 0);
 300                 }
 301                 break;
 302         }
 303
 304         if (cpumask_empty(tick_broadcast_mask)) {
 305                 if (!bc_stopped)
 306                         clockevents_shutdown(bc);
 307         } else if (bc_stopped) {
 308                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 309                         tick_broadcast_start_periodic(bc);
 310                 else
 311                         tick_broadcast_setup_oneshot(bc);
 312         }
 313 out:
 314         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 315 }
 316
 317 /*
 318  * Powerstate information: The system enters/leaves a state, where
 319  * affected devices might stop.
 320  */
 321 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
 322 {
 323         if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
 324                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
 325                        "offline CPU #%d\n", *oncpu);
 326         else
 327                 tick_do_broadcast_on_off(&reason);
 328 }
 329
 330 /*
 331  * Set the periodic handler depending on broadcast on/off
 332  */
 333 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 334 {
 335         if (!broadcast)
 336                 dev->event_handler = tick_handle_periodic;
 337         else
 338                 dev->event_handler = tick_handle_periodic_broadcast;
 339 }
 340
 341 /*
 342  * Remove a CPU from broadcasting
 343  */
 344 void tick_shutdown_broadcast(unsigned int *cpup)
 345 {
 346         struct clock_event_device *bc;
 347         unsigned long flags;
 348         unsigned int cpu = *cpup;
 349
 350         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 351
 352         bc = tick_broadcast_device.evtdev;
 353         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 354
 355         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 356                 if (bc && cpumask_empty(tick_broadcast_mask))
 357                         clockevents_shutdown(bc);
 358         }
 359
 360         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 361 }
 362
 363 void tick_suspend_broadcast(void)
 364 {
 365         struct clock_event_device *bc;
 366         unsigned long flags;
 367
 368         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 369
 370         bc = tick_broadcast_device.evtdev;
 371         if (bc)
 372                 clockevents_shutdown(bc);
 373
 374         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 375 }
 376
 377 int tick_resume_broadcast(void)
 378 {
 379         struct clock_event_device *bc;
 380         unsigned long flags;
 381         int broadcast = 0;
 382
 383         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 384
 385         bc = tick_broadcast_device.evtdev;
 386
 387         if (bc) {
 388                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
 389
 390                 switch (tick_broadcast_device.mode) {
 391                 case TICKDEV_MODE_PERIODIC:
 392                         if (!cpumask_empty(tick_broadcast_mask))
 393                                 tick_broadcast_start_periodic(bc);
 394                         broadcast = cpumask_test_cpu(smp_processor_id(),
 395                                                      tick_broadcast_mask);
 396                         break;
 397                 case TICKDEV_MODE_ONESHOT:
 398                         if (!cpumask_empty(tick_broadcast_mask))
 399                                 broadcast = tick_resume_broadcast_oneshot(bc);
 400                         break;
 401                 }
 402         }
 403         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 404
 405         return broadcast;
 406 }
 407
 408
 409 #ifdef CONFIG_TICK_ONESHOT
 410
 411 static cpumask_var_t tick_broadcast_oneshot_mask;
 412 static cpumask_var_t tick_broadcast_pending_mask;
 413 static cpumask_var_t tick_broadcast_force_mask;
 414
 415 /*
 416  * Exposed for debugging: see timer_list.c
 417  */
 418 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 419 {
 420         return tick_broadcast_oneshot_mask;
 421 }
 422
 423 /*
 424  * Called before going idle with interrupts disabled. Checks whether a
 425  * broadcast event from the other core is about to happen. We detected
 426  * that in tick_broadcast_oneshot_control(). The callsite can use this
 427  * to avoid a deep idle transition as we are about to get the
 428  * broadcast IPI right away.
 429  */
 430 int tick_check_broadcast_expired(void)
 431 {
 432         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 433 }
 434
 435 /*
 436  * Set broadcast interrupt affinity
 437  */
 438 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 439                                         const struct cpumask *cpumask)
 440 {
 441         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 442                 return;
 443
 444         if (cpumask_equal(bc->cpumask, cpumask))
 445                 return;
 446
 447         bc->cpumask = cpumask;
 448         irq_set_affinity(bc->irq, bc->cpumask);
 449 }
 450
 451 static int tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 452                                     ktime_t expires, int force)
 453 {
 454         int ret;
 455
 456         if (bc->mode != CLOCK_EVT_MODE_ONESHOT)
 457                 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 458
 459         ret = clockevents_program_event(bc, expires, force);
 460         if (!ret)
 461                 tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 462         return ret;
 463 }
 464
 465 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 466 {
 467         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 468         return 0;
 469 }
 470
 471 /*
 472  * Called from irq_enter() when idle was interrupted to reenable the
 473  * per cpu device.
 474  */
 475 void tick_check_oneshot_broadcast(int cpu)
 476 {
 477         if (cpumask_test_cpu(cpu, tick_broadcast_oneshot_mask)) {
 478                 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 479
 480                 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
 481         }
 482 }
 483
 484 /*
 485  * Handle oneshot mode broadcasting
 486  */
 487 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 488 {
 489         struct tick_device *td;
 490         ktime_t now, next_event;
 491         int cpu, next_cpu = 0;
 492
 493         raw_spin_lock(&tick_broadcast_lock);
 494 again:
 495         dev->next_event.tv64 = KTIME_MAX;
 496         next_event.tv64 = KTIME_MAX;
 497         cpumask_clear(tmpmask);
 498         now = ktime_get();
 499         /* Find all expired events */
 500         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 501                 td = &per_cpu(tick_cpu_device, cpu);
 502                 if (td->evtdev->next_event.tv64 <= now.tv64) {
 503                         cpumask_set_cpu(cpu, tmpmask);
 504                         /*
 505                          * Mark the remote cpu in the pending mask, so
 506                          * it can avoid reprogramming the cpu local
 507                          * timer in tick_broadcast_oneshot_control().
 508                          */
 509                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 510                 } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
 511                         next_event.tv64 = td->evtdev->next_event.tv64;
 512                         next_cpu = cpu;
 513                 }
 514         }
 515
 516         /* Take care of enforced broadcast requests */
 517         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 518         cpumask_clear(tick_broadcast_force_mask);
 519
 520         /*
 521          * Wakeup the cpus which have an expired event.
 522          */
 523         tick_do_broadcast(tmpmask);
 524
 525         /*
 526          * Two reasons for reprogram:
 527          *
 528          * - The global event did not expire any CPU local
 529          * events. This happens in dyntick mode, as the maximum PIT
 530          * delta is quite small.
 531          *
 532          * - There are pending events on sleeping CPUs which were not
 533          * in the event mask
 534          */
 535         if (next_event.tv64 != KTIME_MAX) {
 536                 /*
 537                  * Rearm the broadcast device. If event expired,
 538                  * repeat the above
 539                  */
 540                 if (tick_broadcast_set_event(dev, next_cpu, next_event, 0))
 541                         goto again;
 542         }
 543         raw_spin_unlock(&tick_broadcast_lock);
 544 }
 545
 546 /*
 547  * Powerstate information: The system enters/leaves a state, where
 548  * affected devices might stop
 549  */
 550 void tick_broadcast_oneshot_control(unsigned long reason)
 551 {
 552         struct clock_event_device *bc, *dev;
 553         struct tick_device *td;
 554         unsigned long flags;
 555         ktime_t now;
 556         int cpu;
 557
 558         /*
 559          * Periodic mode does not care about the enter/exit of power
 560          * states
 561          */
 562         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 563                 return;
 564
 565         /*
 566          * We are called with preemtion disabled from the depth of the
 567          * idle code, so we can't be moved away.
 568          */
 569         cpu = smp_processor_id();
 570         td = &per_cpu(tick_cpu_device, cpu);
 571         dev = td->evtdev;
 572
 573         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 574                 return;
 575
 576         bc = tick_broadcast_device.evtdev;
 577
 578         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 579         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
 580                 WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 581                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 582                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
 583                         /*
 584                          * We only reprogram the broadcast timer if we
 585                          * did not mark ourself in the force mask and
 586                          * if the cpu local event is earlier than the
 587                          * broadcast event. If the current CPU is in
 588                          * the force mask, then we are going to be
 589                          * woken by the IPI right away.
 590                          */
 591                         if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
 592                             dev->next_event.tv64 < bc->next_event.tv64)
 593                                 tick_broadcast_set_event(bc, cpu, dev->next_event, 1);
 594                 }
 595         } else {
 596                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 597                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 598                         if (dev->next_event.tv64 == KTIME_MAX)
 599                                 goto out;
 600                         /*
 601                          * The cpu which was handling the broadcast
 602                          * timer marked this cpu in the broadcast
 603                          * pending mask and fired the broadcast
 604                          * IPI. So we are going to handle the expired
 605                          * event anyway via the broadcast IPI
 606                          * handler. No need to reprogram the timer
 607                          * with an already expired event.
 608                          */
 609                         if (cpumask_test_and_clear_cpu(cpu,
 610                                        tick_broadcast_pending_mask))
 611                                 goto out;
 612
 613                         /*
 614                          * If the pending bit is not set, then we are
 615                          * either the CPU handling the broadcast
 616                          * interrupt or we got woken by something else.
 617                          *
 618                          * We are not longer in the broadcast mask, so
 619                          * if the cpu local expiry time is already
 620                          * reached, we would reprogram the cpu local
 621                          * timer with an already expired event.
 622                          *
 623                          * This can lead to a ping-pong when we return
 624                          * to idle and therefor rearm the broadcast
 625                          * timer before the cpu local timer was able
 626                          * to fire. This happens because the forced
 627                          * reprogramming makes sure that the event
 628                          * will happen in the future and depending on
 629                          * the min_delta setting this might be far
 630                          * enough out that the ping-pong starts.
 631                          *
 632                          * If the cpu local next_event has expired
 633                          * then we know that the broadcast timer
 634                          * next_event has expired as well and
 635                          * broadcast is about to be handled. So we
 636                          * avoid reprogramming and enforce that the
 637                          * broadcast handler, which did not run yet,
 638                          * will invoke the cpu local handler.
 639                          *
 640                          * We cannot call the handler directly from
 641                          * here, because we might be in a NOHZ phase
 642                          * and we did not go through the irq_enter()
 643                          * nohz fixups.
 644                          */
 645                         now = ktime_get();
 646                         if (dev->next_event.tv64 <= now.tv64) {
 647                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 648                                 goto out;
 649                         }
 650                         /*
 651                          * We got woken by something else. Reprogram
 652                          * the cpu local timer device.
 653                          */
 654                         tick_program_event(dev->next_event, 1);
 655                 }
 656         }
 657 out:
 658         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 659 }
 660
 661 /*
 662  * Reset the one shot broadcast for a cpu
 663  *
 664  * Called with tick_broadcast_lock held
 665  */
 666 static void tick_broadcast_clear_oneshot(int cpu)
 667 {
 668         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 669 }
 670
 671 static void tick_broadcast_init_next_event(struct cpumask *mask,
 672                                            ktime_t expires)
 673 {
 674         struct tick_device *td;
 675         int cpu;
 676
 677         for_each_cpu(cpu, mask) {
 678                 td = &per_cpu(tick_cpu_device, cpu);
 679                 if (td->evtdev)
 680                         td->evtdev->next_event = expires;
 681         }
 682 }
 683
 684 /**
 685  * tick_broadcast_setup_oneshot - setup the broadcast device
 686  */
 687 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 688 {
 689         int cpu = smp_processor_id();
 690
 691         /* Set it up only once ! */
 692         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 693                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
 694
 695                 bc->event_handler = tick_handle_oneshot_broadcast;
 696
 697                 /* Take the do_timer update */
 698                 if (!tick_nohz_full_cpu(cpu))
 699                         tick_do_timer_cpu = cpu;
 700
 701                 /*
 702                  * We must be careful here. There might be other CPUs
 703                  * waiting for periodic broadcast. We need to set the
 704                  * oneshot_mask bits for those and program the
 705                  * broadcast device to fire.
 706                  */
 707                 cpumask_copy(tmpmask, tick_broadcast_mask);
 708                 cpumask_clear_cpu(cpu, tmpmask);
 709                 cpumask_or(tick_broadcast_oneshot_mask,
 710                            tick_broadcast_oneshot_mask, tmpmask);
 711
 712                 if (was_periodic && !cpumask_empty(tmpmask)) {
 713                         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 714                         tick_broadcast_init_next_event(tmpmask,
 715                                                        tick_next_period);
 716                         tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
 717                 } else
 718                         bc->next_event.tv64 = KTIME_MAX;
 719         } else {
 720                 /*
 721                  * The first cpu which switches to oneshot mode sets
 722                  * the bit for all other cpus which are in the general
 723                  * (periodic) broadcast mask. So the bit is set and
 724                  * would prevent the first broadcast enter after this
 725                  * to program the bc device.
 726                  */
 727                 tick_broadcast_clear_oneshot(cpu);
 728         }
 729 }
 730
 731 /*
 732  * Select oneshot operating mode for the broadcast device
 733  */
 734 void tick_broadcast_switch_to_oneshot(void)
 735 {
 736         struct clock_event_device *bc;
 737         unsigned long flags;
 738
 739         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 740
 741         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 742         bc = tick_broadcast_device.evtdev;
 743         if (bc)
 744                 tick_broadcast_setup_oneshot(bc);
 745
 746         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 747 }
 748
 749
 750 /*
 751  * Remove a dead CPU from broadcasting
 752  */
 753 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
 754 {
 755         unsigned long flags;
 756         unsigned int cpu = *cpup;
 757
 758         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 759
 760         /*
 761          * Clear the broadcast mask flag for the dead cpu, but do not
 762          * stop the broadcast device!
 763          */
 764         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 765
 766         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 767 }
 768
 769 /*
 770  * Check, whether the broadcast device is in one shot mode
 771  */
 772 int tick_broadcast_oneshot_active(void)
 773 {
 774         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 775 }
 776
 777 /*
 778  * Check whether the broadcast device supports oneshot.
 779  */
 780 bool tick_broadcast_oneshot_available(void)
 781 {
 782         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 783
 784         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 785 }
 786
 787 #endif
 788
 789 void __init tick_broadcast_init(void)
 790 {
 791         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 792         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
 793 #ifdef CONFIG_TICK_ONESHOT
 794         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
 795         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
 796         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
 797 #endif
 798 }