kernel/time/tick-broadcast.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file contains functions which emulate a local clock-event
   4  * device via a broadcast event source.
   5  *
   6  * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
   7  * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   8  * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
   9  */
  10 #include <linux/cpu.h>
  11 #include <linux/err.h>
  12 #include <linux/hrtimer.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/percpu.h>
  15 #include <linux/profile.h>
  16 #include <linux/sched.h>
  17 #include <linux/smp.h>
  18 #include <linux/module.h>
  19
  20 #include "tick-internal.h"
  21
  22 /*
  23  * Broadcast support for broken x86 hardware, where the local apic
  24  * timer stops in C3 state.
  25  */
  26
  27 static struct tick_device tick_broadcast_device;
  28 static cpumask_var_t tick_broadcast_mask __cpumask_var_read_mostly;
  29 static cpumask_var_t tick_broadcast_on __cpumask_var_read_mostly;
  30 static cpumask_var_t tmpmask __cpumask_var_read_mostly;
  31 static int tick_broadcast_forced;
  32
  33 static __cacheline_aligned_in_smp DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  34
  35 #ifdef CONFIG_TICK_ONESHOT
  36 static DEFINE_PER_CPU(struct clock_event_device *, tick_oneshot_wakeup_device);
  37
  38 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc);
  39 static void tick_broadcast_clear_oneshot(int cpu);
  40 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc);
  41 # ifdef CONFIG_HOTPLUG_CPU
  42 static void tick_broadcast_oneshot_offline(unsigned int cpu);
  43 # endif
  44 #else
  45 static inline void tick_broadcast_setup_oneshot(struct clock_event_device *bc) { BUG(); }
  46 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  47 static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
  48 # ifdef CONFIG_HOTPLUG_CPU
  49 static inline void tick_broadcast_oneshot_offline(unsigned int cpu) { }
  50 # endif
  51 #endif
  52
  53 /*
  54  * Debugging: see timer_list.c
  55  */
  56 struct tick_device *tick_get_broadcast_device(void)
  57 {
  58         return &tick_broadcast_device;
  59 }
  60
  61 struct cpumask *tick_get_broadcast_mask(void)
  62 {
  63         return tick_broadcast_mask;
  64 }
  65
  66 static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu);
  67
  68 const struct clock_event_device *tick_get_wakeup_device(int cpu)
  69 {
  70         return tick_get_oneshot_wakeup_device(cpu);
  71 }
  72
  73 /*
  74  * Start the device in periodic mode
  75  */
  76 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  77 {
  78         if (bc)
  79                 tick_setup_periodic(bc, 1);
  80 }
  81
  82 /*
  83  * Check, if the device can be utilized as broadcast device:
  84  */
  85 static bool tick_check_broadcast_device(struct clock_event_device *curdev,
  86                                         struct clock_event_device *newdev)
  87 {
  88         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
  89             (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
  90             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
  91                 return false;
  92
  93         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
  94             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
  95                 return false;
  96
  97         return !curdev || newdev->rating > curdev->rating;
  98 }
  99
 100 #ifdef CONFIG_TICK_ONESHOT
 101 static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu)
 102 {
 103         return per_cpu(tick_oneshot_wakeup_device, cpu);
 104 }
 105
 106 static void tick_oneshot_wakeup_handler(struct clock_event_device *wd)
 107 {
 108         /*
 109          * If we woke up early and the tick was reprogrammed in the
 110          * meantime then this may be spurious but harmless.
 111          */
 112         tick_receive_broadcast();
 113 }
 114
 115 static bool tick_set_oneshot_wakeup_device(struct clock_event_device *newdev,
 116                                            int cpu)
 117 {
 118         struct clock_event_device *curdev = tick_get_oneshot_wakeup_device(cpu);
 119
 120         if (!newdev)
 121                 goto set_device;
 122
 123         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
 124             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
 125                  return false;
 126
 127         if (!(newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
 128             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
 129                 return false;
 130
 131         if (!cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
 132                 return false;
 133
 134         if (curdev && newdev->rating <= curdev->rating)
 135                 return false;
 136
 137         if (!try_module_get(newdev->owner))
 138                 return false;
 139
 140         newdev->event_handler = tick_oneshot_wakeup_handler;
 141 set_device:
 142         clockevents_exchange_device(curdev, newdev);
 143         per_cpu(tick_oneshot_wakeup_device, cpu) = newdev;
 144         return true;
 145 }
 146 #else
 147 static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu)
 148 {
 149         return NULL;
 150 }
 151
 152 static bool tick_set_oneshot_wakeup_device(struct clock_event_device *newdev,
 153                                            int cpu)
 154 {
 155         return false;
 156 }
 157 #endif
 158
 159 /*
 160  * Conditionally install/replace broadcast device
 161  */
 162 void tick_install_broadcast_device(struct clock_event_device *dev, int cpu)
 163 {
 164         struct clock_event_device *cur = tick_broadcast_device.evtdev;
 165
 166         if (tick_set_oneshot_wakeup_device(dev, cpu))
 167                 return;
 168
 169         if (!tick_check_broadcast_device(cur, dev))
 170                 return;
 171
 172         if (!try_module_get(dev->owner))
 173                 return;
 174
 175         clockevents_exchange_device(cur, dev);
 176         if (cur)
 177                 cur->event_handler = clockevents_handle_noop;
 178         tick_broadcast_device.evtdev = dev;
 179         if (!cpumask_empty(tick_broadcast_mask))
 180                 tick_broadcast_start_periodic(dev);
 181
 182         if (!(dev->features & CLOCK_EVT_FEAT_ONESHOT))
 183                 return;
 184
 185         /*
 186          * If the system already runs in oneshot mode, switch the newly
 187          * registered broadcast device to oneshot mode explicitly.
 188          */
 189         if (tick_broadcast_oneshot_active()) {
 190                 tick_broadcast_switch_to_oneshot();
 191                 return;
 192         }
 193
 194         /*
 195          * Inform all cpus about this. We might be in a situation
 196          * where we did not switch to oneshot mode because the per cpu
 197          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
 198          * of a oneshot capable broadcast device. Without that
 199          * notification the systems stays stuck in periodic mode
 200          * forever.
 201          */
 202         tick_clock_notify();
 203 }
 204
 205 /*
 206  * Check, if the device is the broadcast device
 207  */
 208 int tick_is_broadcast_device(struct clock_event_device *dev)
 209 {
 210         return (dev && tick_broadcast_device.evtdev == dev);
 211 }
 212
 213 int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
 214 {
 215         int ret = -ENODEV;
 216
 217         if (tick_is_broadcast_device(dev)) {
 218                 raw_spin_lock(&tick_broadcast_lock);
 219                 ret = __clockevents_update_freq(dev, freq);
 220                 raw_spin_unlock(&tick_broadcast_lock);
 221         }
 222         return ret;
 223 }
 224
 225
 226 static void err_broadcast(const struct cpumask *mask)
 227 {
 228         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 229 }
 230
 231 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 232 {
 233         if (!dev->broadcast)
 234                 dev->broadcast = tick_broadcast;
 235         if (!dev->broadcast) {
 236                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 237                              dev->name);
 238                 dev->broadcast = err_broadcast;
 239         }
 240 }
 241
 242 /*
 243  * Check, if the device is dysfunctional and a placeholder, which
 244  * needs to be handled by the broadcast device.
 245  */
 246 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 247 {
 248         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 249         unsigned long flags;
 250         int ret = 0;
 251
 252         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 253
 254         /*
 255          * Devices might be registered with both periodic and oneshot
 256          * mode disabled. This signals, that the device needs to be
 257          * operated from the broadcast device and is a placeholder for
 258          * the cpu local device.
 259          */
 260         if (!tick_device_is_functional(dev)) {
 261                 dev->event_handler = tick_handle_periodic;
 262                 tick_device_setup_broadcast_func(dev);
 263                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 264                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 265                         tick_broadcast_start_periodic(bc);
 266                 else
 267                         tick_broadcast_setup_oneshot(bc);
 268                 ret = 1;
 269         } else {
 270                 /*
 271                  * Clear the broadcast bit for this cpu if the
 272                  * device is not power state affected.
 273                  */
 274                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 275                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 276                 else
 277                         tick_device_setup_broadcast_func(dev);
 278
 279                 /*
 280                  * Clear the broadcast bit if the CPU is not in
 281                  * periodic broadcast on state.
 282                  */
 283                 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
 284                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 285
 286                 switch (tick_broadcast_device.mode) {
 287                 case TICKDEV_MODE_ONESHOT:
 288                         /*
 289                          * If the system is in oneshot mode we can
 290                          * unconditionally clear the oneshot mask bit,
 291                          * because the CPU is running and therefore
 292                          * not in an idle state which causes the power
 293                          * state affected device to stop. Let the
 294                          * caller initialize the device.
 295                          */
 296                         tick_broadcast_clear_oneshot(cpu);
 297                         ret = 0;
 298                         break;
 299
 300                 case TICKDEV_MODE_PERIODIC:
 301                         /*
 302                          * If the system is in periodic mode, check
 303                          * whether the broadcast device can be
 304                          * switched off now.
 305                          */
 306                         if (cpumask_empty(tick_broadcast_mask) && bc)
 307                                 clockevents_shutdown(bc);
 308                         /*
 309                          * If we kept the cpu in the broadcast mask,
 310                          * tell the caller to leave the per cpu device
 311                          * in shutdown state. The periodic interrupt
 312                          * is delivered by the broadcast device, if
 313                          * the broadcast device exists and is not
 314                          * hrtimer based.
 315                          */
 316                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 317                                 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 318                         break;
 319                 default:
 320                         break;
 321                 }
 322         }
 323         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 324         return ret;
 325 }
 326
 327 int tick_receive_broadcast(void)
 328 {
 329         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 330         struct clock_event_device *evt = td->evtdev;
 331
 332         if (!evt)
 333                 return -ENODEV;
 334
 335         if (!evt->event_handler)
 336                 return -EINVAL;
 337
 338         evt->event_handler(evt);
 339         return 0;
 340 }
 341
 342 /*
 343  * Broadcast the event to the cpus, which are set in the mask (mangled).
 344  */
 345 static bool tick_do_broadcast(struct cpumask *mask)
 346 {
 347         int cpu = smp_processor_id();
 348         struct tick_device *td;
 349         bool local = false;
 350
 351         /*
 352          * Check, if the current cpu is in the mask
 353          */
 354         if (cpumask_test_cpu(cpu, mask)) {
 355                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 356
 357                 cpumask_clear_cpu(cpu, mask);
 358                 /*
 359                  * We only run the local handler, if the broadcast
 360                  * device is not hrtimer based. Otherwise we run into
 361                  * a hrtimer recursion.
 362                  *
 363                  * local timer_interrupt()
 364                  *   local_handler()
 365                  *     expire_hrtimers()
 366                  *       bc_handler()
 367                  *         local_handler()
 368                  *           expire_hrtimers()
 369                  */
 370                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 371         }
 372
 373         if (!cpumask_empty(mask)) {
 374                 /*
 375                  * It might be necessary to actually check whether the devices
 376                  * have different broadcast functions. For now, just use the
 377                  * one of the first device. This works as long as we have this
 378                  * misfeature only on x86 (lapic)
 379                  */
 380                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 381                 td->evtdev->broadcast(mask);
 382         }
 383         return local;
 384 }
 385
 386 /*
 387  * Periodic broadcast:
 388  * - invoke the broadcast handlers
 389  */
 390 static bool tick_do_periodic_broadcast(void)
 391 {
 392         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 393         return tick_do_broadcast(tmpmask);
 394 }
 395
 396 /*
 397  * Event handler for periodic broadcast ticks
 398  */
 399 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 400 {
 401         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 402         bool bc_local;
 403
 404         raw_spin_lock(&tick_broadcast_lock);
 405
 406         /* Handle spurious interrupts gracefully */
 407         if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
 408                 raw_spin_unlock(&tick_broadcast_lock);
 409                 return;
 410         }
 411
 412         bc_local = tick_do_periodic_broadcast();
 413
 414         if (clockevent_state_oneshot(dev)) {
 415                 ktime_t next = ktime_add_ns(dev->next_event, TICK_NSEC);
 416
 417                 clockevents_program_event(dev, next, true);
 418         }
 419         raw_spin_unlock(&tick_broadcast_lock);
 420
 421         /*
 422          * We run the handler of the local cpu after dropping
 423          * tick_broadcast_lock because the handler might deadlock when
 424          * trying to switch to oneshot mode.
 425          */
 426         if (bc_local)
 427                 td->evtdev->event_handler(td->evtdev);
 428 }
 429
 430 /**
 431  * tick_broadcast_control - Enable/disable or force broadcast mode
 432  * @mode:       The selected broadcast mode
 433  *
 434  * Called when the system enters a state where affected tick devices
 435  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 436  */
 437 void tick_broadcast_control(enum tick_broadcast_mode mode)
 438 {
 439         struct clock_event_device *bc, *dev;
 440         struct tick_device *td;
 441         int cpu, bc_stopped;
 442         unsigned long flags;
 443
 444         /* Protects also the local clockevent device. */
 445         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 446         td = this_cpu_ptr(&tick_cpu_device);
 447         dev = td->evtdev;
 448
 449         /*
 450          * Is the device not affected by the powerstate ?
 451          */
 452         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 453                 goto out;
 454
 455         if (!tick_device_is_functional(dev))
 456                 goto out;
 457
 458         cpu = smp_processor_id();
 459         bc = tick_broadcast_device.evtdev;
 460         bc_stopped = cpumask_empty(tick_broadcast_mask);
 461
 462         switch (mode) {
 463         case TICK_BROADCAST_FORCE:
 464                 tick_broadcast_forced = 1;
 465                 fallthrough;
 466         case TICK_BROADCAST_ON:
 467                 cpumask_set_cpu(cpu, tick_broadcast_on);
 468                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 469                         /*
 470                          * Only shutdown the cpu local device, if:
 471                          *
 472                          * - the broadcast device exists
 473                          * - the broadcast device is not a hrtimer based one
 474                          * - the broadcast device is in periodic mode to
 475                          *   avoid a hiccup during switch to oneshot mode
 476                          */
 477                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 478                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 479                                 clockevents_shutdown(dev);
 480                 }
 481                 break;
 482
 483         case TICK_BROADCAST_OFF:
 484                 if (tick_broadcast_forced)
 485                         break;
 486                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 487                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 488                         if (tick_broadcast_device.mode ==
 489                             TICKDEV_MODE_PERIODIC)
 490                                 tick_setup_periodic(dev, 0);
 491                 }
 492                 break;
 493         }
 494
 495         if (bc) {
 496                 if (cpumask_empty(tick_broadcast_mask)) {
 497                         if (!bc_stopped)
 498                                 clockevents_shutdown(bc);
 499                 } else if (bc_stopped) {
 500                         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 501                                 tick_broadcast_start_periodic(bc);
 502                         else
 503                                 tick_broadcast_setup_oneshot(bc);
 504                 }
 505         }
 506 out:
 507         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 508 }
 509 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 510
 511 /*
 512  * Set the periodic handler depending on broadcast on/off
 513  */
 514 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 515 {
 516         if (!broadcast)
 517                 dev->event_handler = tick_handle_periodic;
 518         else
 519                 dev->event_handler = tick_handle_periodic_broadcast;
 520 }
 521
 522 #ifdef CONFIG_HOTPLUG_CPU
 523 static void tick_shutdown_broadcast(void)
 524 {
 525         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 526
 527         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 528                 if (bc && cpumask_empty(tick_broadcast_mask))
 529                         clockevents_shutdown(bc);
 530         }
 531 }
 532
 533 /*
 534  * Remove a CPU from broadcasting
 535  */
 536 void tick_broadcast_offline(unsigned int cpu)
 537 {
 538         raw_spin_lock(&tick_broadcast_lock);
 539         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 540         cpumask_clear_cpu(cpu, tick_broadcast_on);
 541         tick_broadcast_oneshot_offline(cpu);
 542         tick_shutdown_broadcast();
 543         raw_spin_unlock(&tick_broadcast_lock);
 544 }
 545
 546 #endif
 547
 548 void tick_suspend_broadcast(void)
 549 {
 550         struct clock_event_device *bc;
 551         unsigned long flags;
 552
 553         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 554
 555         bc = tick_broadcast_device.evtdev;
 556         if (bc)
 557                 clockevents_shutdown(bc);
 558
 559         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 560 }
 561
 562 /*
 563  * This is called from tick_resume_local() on a resuming CPU. That's
 564  * called from the core resume function, tick_unfreeze() and the magic XEN
 565  * resume hackery.
 566  *
 567  * In none of these cases the broadcast device mode can change and the
 568  * bit of the resuming CPU in the broadcast mask is safe as well.
 569  */
 570 bool tick_resume_check_broadcast(void)
 571 {
 572         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 573                 return false;
 574         else
 575                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 576 }
 577
 578 void tick_resume_broadcast(void)
 579 {
 580         struct clock_event_device *bc;
 581         unsigned long flags;
 582
 583         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 584
 585         bc = tick_broadcast_device.evtdev;
 586
 587         if (bc) {
 588                 clockevents_tick_resume(bc);
 589
 590                 switch (tick_broadcast_device.mode) {
 591                 case TICKDEV_MODE_PERIODIC:
 592                         if (!cpumask_empty(tick_broadcast_mask))
 593                                 tick_broadcast_start_periodic(bc);
 594                         break;
 595                 case TICKDEV_MODE_ONESHOT:
 596                         if (!cpumask_empty(tick_broadcast_mask))
 597                                 tick_resume_broadcast_oneshot(bc);
 598                         break;
 599                 }
 600         }
 601         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 602 }
 603
 604 #ifdef CONFIG_TICK_ONESHOT
 605
 606 static cpumask_var_t tick_broadcast_oneshot_mask __cpumask_var_read_mostly;
 607 static cpumask_var_t tick_broadcast_pending_mask __cpumask_var_read_mostly;
 608 static cpumask_var_t tick_broadcast_force_mask __cpumask_var_read_mostly;
 609
 610 /*
 611  * Exposed for debugging: see timer_list.c
 612  */
 613 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 614 {
 615         return tick_broadcast_oneshot_mask;
 616 }
 617
 618 /*
 619  * Called before going idle with interrupts disabled. Checks whether a
 620  * broadcast event from the other core is about to happen. We detected
 621  * that in tick_broadcast_oneshot_control(). The callsite can use this
 622  * to avoid a deep idle transition as we are about to get the
 623  * broadcast IPI right away.
 624  */
 625 int tick_check_broadcast_expired(void)
 626 {
 627         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 628 }
 629
 630 /*
 631  * Set broadcast interrupt affinity
 632  */
 633 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 634                                         const struct cpumask *cpumask)
 635 {
 636         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 637                 return;
 638
 639         if (cpumask_equal(bc->cpumask, cpumask))
 640                 return;
 641
 642         bc->cpumask = cpumask;
 643         irq_set_affinity(bc->irq, bc->cpumask);
 644 }
 645
 646 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 647                                      ktime_t expires)
 648 {
 649         if (!clockevent_state_oneshot(bc))
 650                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 651
 652         clockevents_program_event(bc, expires, 1);
 653         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 654 }
 655
 656 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 657 {
 658         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 659 }
 660
 661 /*
 662  * Called from irq_enter() when idle was interrupted to reenable the
 663  * per cpu device.
 664  */
 665 void tick_check_oneshot_broadcast_this_cpu(void)
 666 {
 667         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 668                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 669
 670                 /*
 671                  * We might be in the middle of switching over from
 672                  * periodic to oneshot. If the CPU has not yet
 673                  * switched over, leave the device alone.
 674                  */
 675                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 676                         clockevents_switch_state(td->evtdev,
 677                                               CLOCK_EVT_STATE_ONESHOT);
 678                 }
 679         }
 680 }
 681
 682 /*
 683  * Handle oneshot mode broadcasting
 684  */
 685 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 686 {
 687         struct tick_device *td;
 688         ktime_t now, next_event;
 689         int cpu, next_cpu = 0;
 690         bool bc_local;
 691
 692         raw_spin_lock(&tick_broadcast_lock);
 693         dev->next_event = KTIME_MAX;
 694         next_event = KTIME_MAX;
 695         cpumask_clear(tmpmask);
 696         now = ktime_get();
 697         /* Find all expired events */
 698         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 699                 /*
 700                  * Required for !SMP because for_each_cpu() reports
 701                  * unconditionally CPU0 as set on UP kernels.
 702                  */
 703                 if (!IS_ENABLED(CONFIG_SMP) &&
 704                     cpumask_empty(tick_broadcast_oneshot_mask))
 705                         break;
 706
 707                 td = &per_cpu(tick_cpu_device, cpu);
 708                 if (td->evtdev->next_event <= now) {
 709                         cpumask_set_cpu(cpu, tmpmask);
 710                         /*
 711                          * Mark the remote cpu in the pending mask, so
 712                          * it can avoid reprogramming the cpu local
 713                          * timer in tick_broadcast_oneshot_control().
 714                          */
 715                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 716                 } else if (td->evtdev->next_event < next_event) {
 717                         next_event = td->evtdev->next_event;
 718                         next_cpu = cpu;
 719                 }
 720         }
 721
 722         /*
 723          * Remove the current cpu from the pending mask. The event is
 724          * delivered immediately in tick_do_broadcast() !
 725          */
 726         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 727
 728         /* Take care of enforced broadcast requests */
 729         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 730         cpumask_clear(tick_broadcast_force_mask);
 731
 732         /*
 733          * Sanity check. Catch the case where we try to broadcast to
 734          * offline cpus.
 735          */
 736         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 737                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 738
 739         /*
 740          * Wakeup the cpus which have an expired event.
 741          */
 742         bc_local = tick_do_broadcast(tmpmask);
 743
 744         /*
 745          * Two reasons for reprogram:
 746          *
 747          * - The global event did not expire any CPU local
 748          * events. This happens in dyntick mode, as the maximum PIT
 749          * delta is quite small.
 750          *
 751          * - There are pending events on sleeping CPUs which were not
 752          * in the event mask
 753          */
 754         if (next_event != KTIME_MAX)
 755                 tick_broadcast_set_event(dev, next_cpu, next_event);
 756
 757         raw_spin_unlock(&tick_broadcast_lock);
 758
 759         if (bc_local) {
 760                 td = this_cpu_ptr(&tick_cpu_device);
 761                 td->evtdev->event_handler(td->evtdev);
 762         }
 763 }
 764
 765 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 766 {
 767         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 768                 return 0;
 769         if (bc->next_event == KTIME_MAX)
 770                 return 0;
 771         return bc->bound_on == cpu ? -EBUSY : 0;
 772 }
 773
 774 static void broadcast_shutdown_local(struct clock_event_device *bc,
 775                                      struct clock_event_device *dev)
 776 {
 777         /*
 778          * For hrtimer based broadcasting we cannot shutdown the cpu
 779          * local device if our own event is the first one to expire or
 780          * if we own the broadcast timer.
 781          */
 782         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 783                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 784                         return;
 785                 if (dev->next_event < bc->next_event)
 786                         return;
 787         }
 788         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 789 }
 790
 791 static int ___tick_broadcast_oneshot_control(enum tick_broadcast_state state,
 792                                              struct tick_device *td,
 793                                              int cpu)
 794 {
 795         struct clock_event_device *bc, *dev = td->evtdev;
 796         int ret = 0;
 797         ktime_t now;
 798
 799         raw_spin_lock(&tick_broadcast_lock);
 800         bc = tick_broadcast_device.evtdev;
 801
 802         if (state == TICK_BROADCAST_ENTER) {
 803                 /*
 804                  * If the current CPU owns the hrtimer broadcast
 805                  * mechanism, it cannot go deep idle and we do not add
 806                  * the CPU to the broadcast mask. We don't have to go
 807                  * through the EXIT path as the local timer is not
 808                  * shutdown.
 809                  */
 810                 ret = broadcast_needs_cpu(bc, cpu);
 811                 if (ret)
 812                         goto out;
 813
 814                 /*
 815                  * If the broadcast device is in periodic mode, we
 816                  * return.
 817                  */
 818                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 819                         /* If it is a hrtimer based broadcast, return busy */
 820                         if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
 821                                 ret = -EBUSY;
 822                         goto out;
 823                 }
 824
 825                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 826                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 827
 828                         /* Conditionally shut down the local timer. */
 829                         broadcast_shutdown_local(bc, dev);
 830
 831                         /*
 832                          * We only reprogram the broadcast timer if we
 833                          * did not mark ourself in the force mask and
 834                          * if the cpu local event is earlier than the
 835                          * broadcast event. If the current CPU is in
 836                          * the force mask, then we are going to be
 837                          * woken by the IPI right away; we return
 838                          * busy, so the CPU does not try to go deep
 839                          * idle.
 840                          */
 841                         if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
 842                                 ret = -EBUSY;
 843                         } else if (dev->next_event < bc->next_event) {
 844                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 845                                 /*
 846                                  * In case of hrtimer broadcasts the
 847                                  * programming might have moved the
 848                                  * timer to this cpu. If yes, remove
 849                                  * us from the broadcast mask and
 850                                  * return busy.
 851                                  */
 852                                 ret = broadcast_needs_cpu(bc, cpu);
 853                                 if (ret) {
 854                                         cpumask_clear_cpu(cpu,
 855                                                 tick_broadcast_oneshot_mask);
 856                                 }
 857                         }
 858                 }
 859         } else {
 860                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 861                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 862                         /*
 863                          * The cpu which was handling the broadcast
 864                          * timer marked this cpu in the broadcast
 865                          * pending mask and fired the broadcast
 866                          * IPI. So we are going to handle the expired
 867                          * event anyway via the broadcast IPI
 868                          * handler. No need to reprogram the timer
 869                          * with an already expired event.
 870                          */
 871                         if (cpumask_test_and_clear_cpu(cpu,
 872                                        tick_broadcast_pending_mask))
 873                                 goto out;
 874
 875                         /*
 876                          * Bail out if there is no next event.
 877                          */
 878                         if (dev->next_event == KTIME_MAX)
 879                                 goto out;
 880                         /*
 881                          * If the pending bit is not set, then we are
 882                          * either the CPU handling the broadcast
 883                          * interrupt or we got woken by something else.
 884                          *
 885                          * We are no longer in the broadcast mask, so
 886                          * if the cpu local expiry time is already
 887                          * reached, we would reprogram the cpu local
 888                          * timer with an already expired event.
 889                          *
 890                          * This can lead to a ping-pong when we return
 891                          * to idle and therefore rearm the broadcast
 892                          * timer before the cpu local timer was able
 893                          * to fire. This happens because the forced
 894                          * reprogramming makes sure that the event
 895                          * will happen in the future and depending on
 896                          * the min_delta setting this might be far
 897                          * enough out that the ping-pong starts.
 898                          *
 899                          * If the cpu local next_event has expired
 900                          * then we know that the broadcast timer
 901                          * next_event has expired as well and
 902                          * broadcast is about to be handled. So we
 903                          * avoid reprogramming and enforce that the
 904                          * broadcast handler, which did not run yet,
 905                          * will invoke the cpu local handler.
 906                          *
 907                          * We cannot call the handler directly from
 908                          * here, because we might be in a NOHZ phase
 909                          * and we did not go through the irq_enter()
 910                          * nohz fixups.
 911                          */
 912                         now = ktime_get();
 913                         if (dev->next_event <= now) {
 914                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 915                                 goto out;
 916                         }
 917                         /*
 918                          * We got woken by something else. Reprogram
 919                          * the cpu local timer device.
 920                          */
 921                         tick_program_event(dev->next_event, 1);
 922                 }
 923         }
 924 out:
 925         raw_spin_unlock(&tick_broadcast_lock);
 926         return ret;
 927 }
 928
 929 static int tick_oneshot_wakeup_control(enum tick_broadcast_state state,
 930                                        struct tick_device *td,
 931                                        int cpu)
 932 {
 933         struct clock_event_device *dev, *wd;
 934
 935         dev = td->evtdev;
 936         if (td->mode != TICKDEV_MODE_ONESHOT)
 937                 return -EINVAL;
 938
 939         wd = tick_get_oneshot_wakeup_device(cpu);
 940         if (!wd)
 941                 return -ENODEV;
 942
 943         switch (state) {
 944         case TICK_BROADCAST_ENTER:
 945                 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT_STOPPED);
 946                 clockevents_switch_state(wd, CLOCK_EVT_STATE_ONESHOT);
 947                 clockevents_program_event(wd, dev->next_event, 1);
 948                 break;
 949         case TICK_BROADCAST_EXIT:
 950                 /* We may have transitioned to oneshot mode while idle */
 951                 if (clockevent_get_state(wd) != CLOCK_EVT_STATE_ONESHOT)
 952                         return -ENODEV;
 953         }
 954
 955         return 0;
 956 }
 957
 958 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 959 {
 960         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 961         int cpu = smp_processor_id();
 962
 963         if (!tick_oneshot_wakeup_control(state, td, cpu))
 964                 return 0;
 965
 966         if (tick_broadcast_device.evtdev)
 967                 return ___tick_broadcast_oneshot_control(state, td, cpu);
 968
 969         /*
 970          * If there is no broadcast or wakeup device, tell the caller not
 971          * to go into deep idle.
 972          */
 973         return -EBUSY;
 974 }
 975
 976 /*
 977  * Reset the one shot broadcast for a cpu
 978  *
 979  * Called with tick_broadcast_lock held
 980  */
 981 static void tick_broadcast_clear_oneshot(int cpu)
 982 {
 983         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 984         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 985 }
 986
 987 static void tick_broadcast_init_next_event(struct cpumask *mask,
 988                                            ktime_t expires)
 989 {
 990         struct tick_device *td;
 991         int cpu;
 992
 993         for_each_cpu(cpu, mask) {
 994                 td = &per_cpu(tick_cpu_device, cpu);
 995                 if (td->evtdev)
 996                         td->evtdev->next_event = expires;
 997         }
 998 }
 999
1000 static inline ktime_t tick_get_next_period(void)
1001 {
1002         ktime_t next;
1003
1004         /*
1005          * Protect against concurrent updates (store /load tearing on
1006          * 32bit). It does not matter if the time is already in the
1007          * past. The broadcast device which is about to be programmed will
1008          * fire in any case.
1009          */
1010         raw_spin_lock(&jiffies_lock);
1011         next = tick_next_period;
1012         raw_spin_unlock(&jiffies_lock);
1013         return next;
1014 }
1015
1016 /**
1017  * tick_broadcast_setup_oneshot - setup the broadcast device
1018  */
1019 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
1020 {
1021         int cpu = smp_processor_id();
1022
1023         if (!bc)
1024                 return;
1025
1026         /* Set it up only once ! */
1027         if (bc->event_handler != tick_handle_oneshot_broadcast) {
1028                 int was_periodic = clockevent_state_periodic(bc);
1029
1030                 bc->event_handler = tick_handle_oneshot_broadcast;
1031
1032                 /*
1033                  * We must be careful here. There might be other CPUs
1034                  * waiting for periodic broadcast. We need to set the
1035                  * oneshot_mask bits for those and program the
1036                  * broadcast device to fire.
1037                  */
1038                 cpumask_copy(tmpmask, tick_broadcast_mask);
1039                 cpumask_clear_cpu(cpu, tmpmask);
1040                 cpumask_or(tick_broadcast_oneshot_mask,
1041                            tick_broadcast_oneshot_mask, tmpmask);
1042
1043                 if (was_periodic && !cpumask_empty(tmpmask)) {
1044                         ktime_t nextevt = tick_get_next_period();
1045
1046                         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
1047                         tick_broadcast_init_next_event(tmpmask, nextevt);
1048                         tick_broadcast_set_event(bc, cpu, nextevt);
1049                 } else
1050                         bc->next_event = KTIME_MAX;
1051         } else {
1052                 /*
1053                  * The first cpu which switches to oneshot mode sets
1054                  * the bit for all other cpus which are in the general
1055                  * (periodic) broadcast mask. So the bit is set and
1056                  * would prevent the first broadcast enter after this
1057                  * to program the bc device.
1058                  */
1059                 tick_broadcast_clear_oneshot(cpu);
1060         }
1061 }
1062
1063 /*
1064  * Select oneshot operating mode for the broadcast device
1065  */
1066 void tick_broadcast_switch_to_oneshot(void)
1067 {
1068         struct clock_event_device *bc;
1069         unsigned long flags;
1070
1071         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
1072
1073         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
1074         bc = tick_broadcast_device.evtdev;
1075         if (bc)
1076                 tick_broadcast_setup_oneshot(bc);
1077
1078         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
1079 }
1080
1081 #ifdef CONFIG_HOTPLUG_CPU
1082 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
1083 {
1084         struct clock_event_device *bc;
1085         unsigned long flags;
1086
1087         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
1088         bc = tick_broadcast_device.evtdev;
1089
1090         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
1091                 /* This moves the broadcast assignment to this CPU: */
1092                 clockevents_program_event(bc, bc->next_event, 1);
1093         }
1094         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
1095 }
1096
1097 /*
1098  * Remove a dying CPU from broadcasting
1099  */
1100 static void tick_broadcast_oneshot_offline(unsigned int cpu)
1101 {
1102         if (tick_get_oneshot_wakeup_device(cpu))
1103                 tick_set_oneshot_wakeup_device(NULL, cpu);
1104
1105         /*
1106          * Clear the broadcast masks for the dead cpu, but do not stop
1107          * the broadcast device!
1108          */
1109         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
1110         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
1111         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
1112 }
1113 #endif
1114
1115 /*
1116  * Check, whether the broadcast device is in one shot mode
1117  */
1118 int tick_broadcast_oneshot_active(void)
1119 {
1120         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
1121 }
1122
1123 /*
1124  * Check whether the broadcast device supports oneshot.
1125  */
1126 bool tick_broadcast_oneshot_available(void)
1127 {
1128         struct clock_event_device *bc = tick_broadcast_device.evtdev;
1129
1130         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
1131 }
1132
1133 #else
1134 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
1135 {
1136         struct clock_event_device *bc = tick_broadcast_device.evtdev;
1137
1138         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
1139                 return -EBUSY;
1140
1141         return 0;
1142 }
1143 #endif
1144
1145 void __init tick_broadcast_init(void)
1146 {
1147         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
1148         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
1149         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1150 #ifdef CONFIG_TICK_ONESHOT
1151         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1152         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1153         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1154 #endif
1155 }