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, bool from_periodic);
  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
  46 tick_broadcast_setup_oneshot(struct clock_event_device *bc, bool from_periodic) { BUG(); }
  47 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  48 static inline void tick_resume_broadcast_oneshot(struct clock_event_device *bc) { }
  49 # ifdef CONFIG_HOTPLUG_CPU
  50 static inline void tick_broadcast_oneshot_offline(unsigned int cpu) { }
  51 # endif
  52 #endif
  53
  54 /*
  55  * Debugging: see timer_list.c
  56  */
  57 struct tick_device *tick_get_broadcast_device(void)
  58 {
  59         return &tick_broadcast_device;
  60 }
  61
  62 struct cpumask *tick_get_broadcast_mask(void)
  63 {
  64         return tick_broadcast_mask;
  65 }
  66
  67 static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu);
  68
  69 const struct clock_event_device *tick_get_wakeup_device(int cpu)
  70 {
  71         return tick_get_oneshot_wakeup_device(cpu);
  72 }
  73
  74 /*
  75  * Start the device in periodic mode
  76  */
  77 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  78 {
  79         if (bc)
  80                 tick_setup_periodic(bc, 1);
  81 }
  82
  83 /*
  84  * Check, if the device can be utilized as broadcast device:
  85  */
  86 static bool tick_check_broadcast_device(struct clock_event_device *curdev,
  87                                         struct clock_event_device *newdev)
  88 {
  89         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
  90             (newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
  91             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
  92                 return false;
  93
  94         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT &&
  95             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
  96                 return false;
  97
  98         return !curdev || newdev->rating > curdev->rating;
  99 }
 100
 101 #ifdef CONFIG_TICK_ONESHOT
 102 static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu)
 103 {
 104         return per_cpu(tick_oneshot_wakeup_device, cpu);
 105 }
 106
 107 static void tick_oneshot_wakeup_handler(struct clock_event_device *wd)
 108 {
 109         /*
 110          * If we woke up early and the tick was reprogrammed in the
 111          * meantime then this may be spurious but harmless.
 112          */
 113         tick_receive_broadcast();
 114 }
 115
 116 static bool tick_set_oneshot_wakeup_device(struct clock_event_device *newdev,
 117                                            int cpu)
 118 {
 119         struct clock_event_device *curdev = tick_get_oneshot_wakeup_device(cpu);
 120
 121         if (!newdev)
 122                 goto set_device;
 123
 124         if ((newdev->features & CLOCK_EVT_FEAT_DUMMY) ||
 125             (newdev->features & CLOCK_EVT_FEAT_C3STOP))
 126                  return false;
 127
 128         if (!(newdev->features & CLOCK_EVT_FEAT_PERCPU) ||
 129             !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
 130                 return false;
 131
 132         if (!cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
 133                 return false;
 134
 135         if (curdev && newdev->rating <= curdev->rating)
 136                 return false;
 137
 138         if (!try_module_get(newdev->owner))
 139                 return false;
 140
 141         newdev->event_handler = tick_oneshot_wakeup_handler;
 142 set_device:
 143         clockevents_exchange_device(curdev, newdev);
 144         per_cpu(tick_oneshot_wakeup_device, cpu) = newdev;
 145         return true;
 146 }
 147 #else
 148 static struct clock_event_device *tick_get_oneshot_wakeup_device(int cpu)
 149 {
 150         return NULL;
 151 }
 152
 153 static bool tick_set_oneshot_wakeup_device(struct clock_event_device *newdev,
 154                                            int cpu)
 155 {
 156         return false;
 157 }
 158 #endif
 159
 160 /*
 161  * Conditionally install/replace broadcast device
 162  */
 163 void tick_install_broadcast_device(struct clock_event_device *dev, int cpu)
 164 {
 165         struct clock_event_device *cur = tick_broadcast_device.evtdev;
 166
 167         if (tick_set_oneshot_wakeup_device(dev, cpu))
 168                 return;
 169
 170         if (!tick_check_broadcast_device(cur, dev))
 171                 return;
 172
 173         if (!try_module_get(dev->owner))
 174                 return;
 175
 176         clockevents_exchange_device(cur, dev);
 177         if (cur)
 178                 cur->event_handler = clockevents_handle_noop;
 179         tick_broadcast_device.evtdev = dev;
 180         if (!cpumask_empty(tick_broadcast_mask))
 181                 tick_broadcast_start_periodic(dev);
 182
 183         if (!(dev->features & CLOCK_EVT_FEAT_ONESHOT))
 184                 return;
 185
 186         /*
 187          * If the system already runs in oneshot mode, switch the newly
 188          * registered broadcast device to oneshot mode explicitly.
 189          */
 190         if (tick_broadcast_oneshot_active()) {
 191                 tick_broadcast_switch_to_oneshot();
 192                 return;
 193         }
 194
 195         /*
 196          * Inform all cpus about this. We might be in a situation
 197          * where we did not switch to oneshot mode because the per cpu
 198          * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
 199          * of a oneshot capable broadcast device. Without that
 200          * notification the systems stays stuck in periodic mode
 201          * forever.
 202          */
 203         tick_clock_notify();
 204 }
 205
 206 /*
 207  * Check, if the device is the broadcast device
 208  */
 209 int tick_is_broadcast_device(struct clock_event_device *dev)
 210 {
 211         return (dev && tick_broadcast_device.evtdev == dev);
 212 }
 213
 214 int tick_broadcast_update_freq(struct clock_event_device *dev, u32 freq)
 215 {
 216         int ret = -ENODEV;
 217
 218         if (tick_is_broadcast_device(dev)) {
 219                 raw_spin_lock(&tick_broadcast_lock);
 220                 ret = __clockevents_update_freq(dev, freq);
 221                 raw_spin_unlock(&tick_broadcast_lock);
 222         }
 223         return ret;
 224 }
 225
 226
 227 static void err_broadcast(const struct cpumask *mask)
 228 {
 229         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
 230 }
 231
 232 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
 233 {
 234         if (!dev->broadcast)
 235                 dev->broadcast = tick_broadcast;
 236         if (!dev->broadcast) {
 237                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 238                              dev->name);
 239                 dev->broadcast = err_broadcast;
 240         }
 241 }
 242
 243 /*
 244  * Check, if the device is dysfunctional and a placeholder, which
 245  * needs to be handled by the broadcast device.
 246  */
 247 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 248 {
 249         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 250         unsigned long flags;
 251         int ret = 0;
 252
 253         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 254
 255         /*
 256          * Devices might be registered with both periodic and oneshot
 257          * mode disabled. This signals, that the device needs to be
 258          * operated from the broadcast device and is a placeholder for
 259          * the cpu local device.
 260          */
 261         if (!tick_device_is_functional(dev)) {
 262                 dev->event_handler = tick_handle_periodic;
 263                 tick_device_setup_broadcast_func(dev);
 264                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 265                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 266                         tick_broadcast_start_periodic(bc);
 267                 else
 268                         tick_broadcast_setup_oneshot(bc, false);
 269                 ret = 1;
 270         } else {
 271                 /*
 272                  * Clear the broadcast bit for this cpu if the
 273                  * device is not power state affected.
 274                  */
 275                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 276                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 277                 else
 278                         tick_device_setup_broadcast_func(dev);
 279
 280                 /*
 281                  * Clear the broadcast bit if the CPU is not in
 282                  * periodic broadcast on state.
 283                  */
 284                 if (!cpumask_test_cpu(cpu, tick_broadcast_on))
 285                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 286
 287                 switch (tick_broadcast_device.mode) {
 288                 case TICKDEV_MODE_ONESHOT:
 289                         /*
 290                          * If the system is in oneshot mode we can
 291                          * unconditionally clear the oneshot mask bit,
 292                          * because the CPU is running and therefore
 293                          * not in an idle state which causes the power
 294                          * state affected device to stop. Let the
 295                          * caller initialize the device.
 296                          */
 297                         tick_broadcast_clear_oneshot(cpu);
 298                         ret = 0;
 299                         break;
 300
 301                 case TICKDEV_MODE_PERIODIC:
 302                         /*
 303                          * If the system is in periodic mode, check
 304                          * whether the broadcast device can be
 305                          * switched off now.
 306                          */
 307                         if (cpumask_empty(tick_broadcast_mask) && bc)
 308                                 clockevents_shutdown(bc);
 309                         /*
 310                          * If we kept the cpu in the broadcast mask,
 311                          * tell the caller to leave the per cpu device
 312                          * in shutdown state. The periodic interrupt
 313                          * is delivered by the broadcast device, if
 314                          * the broadcast device exists and is not
 315                          * hrtimer based.
 316                          */
 317                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 318                                 ret = cpumask_test_cpu(cpu, tick_broadcast_mask);
 319                         break;
 320                 default:
 321                         break;
 322                 }
 323         }
 324         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 325         return ret;
 326 }
 327
 328 int tick_receive_broadcast(void)
 329 {
 330         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 331         struct clock_event_device *evt = td->evtdev;
 332
 333         if (!evt)
 334                 return -ENODEV;
 335
 336         if (!evt->event_handler)
 337                 return -EINVAL;
 338
 339         evt->event_handler(evt);
 340         return 0;
 341 }
 342
 343 /*
 344  * Broadcast the event to the cpus, which are set in the mask (mangled).
 345  */
 346 static bool tick_do_broadcast(struct cpumask *mask)
 347 {
 348         int cpu = smp_processor_id();
 349         struct tick_device *td;
 350         bool local = false;
 351
 352         /*
 353          * Check, if the current cpu is in the mask
 354          */
 355         if (cpumask_test_cpu(cpu, mask)) {
 356                 struct clock_event_device *bc = tick_broadcast_device.evtdev;
 357
 358                 cpumask_clear_cpu(cpu, mask);
 359                 /*
 360                  * We only run the local handler, if the broadcast
 361                  * device is not hrtimer based. Otherwise we run into
 362                  * a hrtimer recursion.
 363                  *
 364                  * local timer_interrupt()
 365                  *   local_handler()
 366                  *     expire_hrtimers()
 367                  *       bc_handler()
 368                  *         local_handler()
 369                  *           expire_hrtimers()
 370                  */
 371                 local = !(bc->features & CLOCK_EVT_FEAT_HRTIMER);
 372         }
 373
 374         if (!cpumask_empty(mask)) {
 375                 /*
 376                  * It might be necessary to actually check whether the devices
 377                  * have different broadcast functions. For now, just use the
 378                  * one of the first device. This works as long as we have this
 379                  * misfeature only on x86 (lapic)
 380                  */
 381                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 382                 td->evtdev->broadcast(mask);
 383         }
 384         return local;
 385 }
 386
 387 /*
 388  * Periodic broadcast:
 389  * - invoke the broadcast handlers
 390  */
 391 static bool tick_do_periodic_broadcast(void)
 392 {
 393         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 394         return tick_do_broadcast(tmpmask);
 395 }
 396
 397 /*
 398  * Event handler for periodic broadcast ticks
 399  */
 400 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 401 {
 402         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 403         bool bc_local;
 404
 405         raw_spin_lock(&tick_broadcast_lock);
 406
 407         /* Handle spurious interrupts gracefully */
 408         if (clockevent_state_shutdown(tick_broadcast_device.evtdev)) {
 409                 raw_spin_unlock(&tick_broadcast_lock);
 410                 return;
 411         }
 412
 413         bc_local = tick_do_periodic_broadcast();
 414
 415         if (clockevent_state_oneshot(dev)) {
 416                 ktime_t next = ktime_add_ns(dev->next_event, TICK_NSEC);
 417
 418                 clockevents_program_event(dev, next, true);
 419         }
 420         raw_spin_unlock(&tick_broadcast_lock);
 421
 422         /*
 423          * We run the handler of the local cpu after dropping
 424          * tick_broadcast_lock because the handler might deadlock when
 425          * trying to switch to oneshot mode.
 426          */
 427         if (bc_local)
 428                 td->evtdev->event_handler(td->evtdev);
 429 }
 430
 431 /**
 432  * tick_broadcast_control - Enable/disable or force broadcast mode
 433  * @mode:       The selected broadcast mode
 434  *
 435  * Called when the system enters a state where affected tick devices
 436  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
 437  */
 438 void tick_broadcast_control(enum tick_broadcast_mode mode)
 439 {
 440         struct clock_event_device *bc, *dev;
 441         struct tick_device *td;
 442         int cpu, bc_stopped;
 443         unsigned long flags;
 444
 445         /* Protects also the local clockevent device. */
 446         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 447         td = this_cpu_ptr(&tick_cpu_device);
 448         dev = td->evtdev;
 449
 450         /*
 451          * Is the device not affected by the powerstate ?
 452          */
 453         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 454                 goto out;
 455
 456         if (!tick_device_is_functional(dev))
 457                 goto out;
 458
 459         cpu = smp_processor_id();
 460         bc = tick_broadcast_device.evtdev;
 461         bc_stopped = cpumask_empty(tick_broadcast_mask);
 462
 463         switch (mode) {
 464         case TICK_BROADCAST_FORCE:
 465                 tick_broadcast_forced = 1;
 466                 fallthrough;
 467         case TICK_BROADCAST_ON:
 468                 cpumask_set_cpu(cpu, tick_broadcast_on);
 469                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 470                         /*
 471                          * Only shutdown the cpu local device, if:
 472                          *
 473                          * - the broadcast device exists
 474                          * - the broadcast device is not a hrtimer based one
 475                          * - the broadcast device is in periodic mode to
 476                          *   avoid a hiccup during switch to oneshot mode
 477                          */
 478                         if (bc && !(bc->features & CLOCK_EVT_FEAT_HRTIMER) &&
 479                             tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 480                                 clockevents_shutdown(dev);
 481                 }
 482                 break;
 483
 484         case TICK_BROADCAST_OFF:
 485                 if (tick_broadcast_forced)
 486                         break;
 487                 cpumask_clear_cpu(cpu, tick_broadcast_on);
 488                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 489                         if (tick_broadcast_device.mode ==
 490                             TICKDEV_MODE_PERIODIC)
 491                                 tick_setup_periodic(dev, 0);
 492                 }
 493                 break;
 494         }
 495
 496         if (bc) {
 497                 if (cpumask_empty(tick_broadcast_mask)) {
 498                         if (!bc_stopped)
 499                                 clockevents_shutdown(bc);
 500                 } else if (bc_stopped) {
 501                         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 502                                 tick_broadcast_start_periodic(bc);
 503                         else
 504                                 tick_broadcast_setup_oneshot(bc, false);
 505                 }
 506         }
 507 out:
 508         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 509 }
 510 EXPORT_SYMBOL_GPL(tick_broadcast_control);
 511
 512 /*
 513  * Set the periodic handler depending on broadcast on/off
 514  */
 515 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 516 {
 517         if (!broadcast)
 518                 dev->event_handler = tick_handle_periodic;
 519         else
 520                 dev->event_handler = tick_handle_periodic_broadcast;
 521 }
 522
 523 #ifdef CONFIG_HOTPLUG_CPU
 524 static void tick_shutdown_broadcast(void)
 525 {
 526         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 527
 528         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 529                 if (bc && cpumask_empty(tick_broadcast_mask))
 530                         clockevents_shutdown(bc);
 531         }
 532 }
 533
 534 /*
 535  * Remove a CPU from broadcasting
 536  */
 537 void tick_broadcast_offline(unsigned int cpu)
 538 {
 539         raw_spin_lock(&tick_broadcast_lock);
 540         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 541         cpumask_clear_cpu(cpu, tick_broadcast_on);
 542         tick_broadcast_oneshot_offline(cpu);
 543         tick_shutdown_broadcast();
 544         raw_spin_unlock(&tick_broadcast_lock);
 545 }
 546
 547 #endif
 548
 549 void tick_suspend_broadcast(void)
 550 {
 551         struct clock_event_device *bc;
 552         unsigned long flags;
 553
 554         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 555
 556         bc = tick_broadcast_device.evtdev;
 557         if (bc)
 558                 clockevents_shutdown(bc);
 559
 560         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 561 }
 562
 563 /*
 564  * This is called from tick_resume_local() on a resuming CPU. That's
 565  * called from the core resume function, tick_unfreeze() and the magic XEN
 566  * resume hackery.
 567  *
 568  * In none of these cases the broadcast device mode can change and the
 569  * bit of the resuming CPU in the broadcast mask is safe as well.
 570  */
 571 bool tick_resume_check_broadcast(void)
 572 {
 573         if (tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT)
 574                 return false;
 575         else
 576                 return cpumask_test_cpu(smp_processor_id(), tick_broadcast_mask);
 577 }
 578
 579 void tick_resume_broadcast(void)
 580 {
 581         struct clock_event_device *bc;
 582         unsigned long flags;
 583
 584         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 585
 586         bc = tick_broadcast_device.evtdev;
 587
 588         if (bc) {
 589                 clockevents_tick_resume(bc);
 590
 591                 switch (tick_broadcast_device.mode) {
 592                 case TICKDEV_MODE_PERIODIC:
 593                         if (!cpumask_empty(tick_broadcast_mask))
 594                                 tick_broadcast_start_periodic(bc);
 595                         break;
 596                 case TICKDEV_MODE_ONESHOT:
 597                         if (!cpumask_empty(tick_broadcast_mask))
 598                                 tick_resume_broadcast_oneshot(bc);
 599                         break;
 600                 }
 601         }
 602         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 603 }
 604
 605 #ifdef CONFIG_TICK_ONESHOT
 606
 607 static cpumask_var_t tick_broadcast_oneshot_mask __cpumask_var_read_mostly;
 608 static cpumask_var_t tick_broadcast_pending_mask __cpumask_var_read_mostly;
 609 static cpumask_var_t tick_broadcast_force_mask __cpumask_var_read_mostly;
 610
 611 /*
 612  * Exposed for debugging: see timer_list.c
 613  */
 614 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 615 {
 616         return tick_broadcast_oneshot_mask;
 617 }
 618
 619 /*
 620  * Called before going idle with interrupts disabled. Checks whether a
 621  * broadcast event from the other core is about to happen. We detected
 622  * that in tick_broadcast_oneshot_control(). The callsite can use this
 623  * to avoid a deep idle transition as we are about to get the
 624  * broadcast IPI right away.
 625  */
 626 int tick_check_broadcast_expired(void)
 627 {
 628         return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 629 }
 630
 631 /*
 632  * Set broadcast interrupt affinity
 633  */
 634 static void tick_broadcast_set_affinity(struct clock_event_device *bc,
 635                                         const struct cpumask *cpumask)
 636 {
 637         if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
 638                 return;
 639
 640         if (cpumask_equal(bc->cpumask, cpumask))
 641                 return;
 642
 643         bc->cpumask = cpumask;
 644         irq_set_affinity(bc->irq, bc->cpumask);
 645 }
 646
 647 static void tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
 648                                      ktime_t expires)
 649 {
 650         if (!clockevent_state_oneshot(bc))
 651                 clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 652
 653         clockevents_program_event(bc, expires, 1);
 654         tick_broadcast_set_affinity(bc, cpumask_of(cpu));
 655 }
 656
 657 static void tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 658 {
 659         clockevents_switch_state(bc, CLOCK_EVT_STATE_ONESHOT);
 660 }
 661
 662 /*
 663  * Called from irq_enter() when idle was interrupted to reenable the
 664  * per cpu device.
 665  */
 666 void tick_check_oneshot_broadcast_this_cpu(void)
 667 {
 668         if (cpumask_test_cpu(smp_processor_id(), tick_broadcast_oneshot_mask)) {
 669                 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 670
 671                 /*
 672                  * We might be in the middle of switching over from
 673                  * periodic to oneshot. If the CPU has not yet
 674                  * switched over, leave the device alone.
 675                  */
 676                 if (td->mode == TICKDEV_MODE_ONESHOT) {
 677                         clockevents_switch_state(td->evtdev,
 678                                               CLOCK_EVT_STATE_ONESHOT);
 679                 }
 680         }
 681 }
 682
 683 /*
 684  * Handle oneshot mode broadcasting
 685  */
 686 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 687 {
 688         struct tick_device *td;
 689         ktime_t now, next_event;
 690         int cpu, next_cpu = 0;
 691         bool bc_local;
 692
 693         raw_spin_lock(&tick_broadcast_lock);
 694         dev->next_event = KTIME_MAX;
 695         next_event = KTIME_MAX;
 696         cpumask_clear(tmpmask);
 697         now = ktime_get();
 698         /* Find all expired events */
 699         for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
 700                 /*
 701                  * Required for !SMP because for_each_cpu() reports
 702                  * unconditionally CPU0 as set on UP kernels.
 703                  */
 704                 if (!IS_ENABLED(CONFIG_SMP) &&
 705                     cpumask_empty(tick_broadcast_oneshot_mask))
 706                         break;
 707
 708                 td = &per_cpu(tick_cpu_device, cpu);
 709                 if (td->evtdev->next_event <= now) {
 710                         cpumask_set_cpu(cpu, tmpmask);
 711                         /*
 712                          * Mark the remote cpu in the pending mask, so
 713                          * it can avoid reprogramming the cpu local
 714                          * timer in tick_broadcast_oneshot_control().
 715                          */
 716                         cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
 717                 } else if (td->evtdev->next_event < next_event) {
 718                         next_event = td->evtdev->next_event;
 719                         next_cpu = cpu;
 720                 }
 721         }
 722
 723         /*
 724          * Remove the current cpu from the pending mask. The event is
 725          * delivered immediately in tick_do_broadcast() !
 726          */
 727         cpumask_clear_cpu(smp_processor_id(), tick_broadcast_pending_mask);
 728
 729         /* Take care of enforced broadcast requests */
 730         cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
 731         cpumask_clear(tick_broadcast_force_mask);
 732
 733         /*
 734          * Sanity check. Catch the case where we try to broadcast to
 735          * offline cpus.
 736          */
 737         if (WARN_ON_ONCE(!cpumask_subset(tmpmask, cpu_online_mask)))
 738                 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
 739
 740         /*
 741          * Wakeup the cpus which have an expired event.
 742          */
 743         bc_local = tick_do_broadcast(tmpmask);
 744
 745         /*
 746          * Two reasons for reprogram:
 747          *
 748          * - The global event did not expire any CPU local
 749          * events. This happens in dyntick mode, as the maximum PIT
 750          * delta is quite small.
 751          *
 752          * - There are pending events on sleeping CPUs which were not
 753          * in the event mask
 754          */
 755         if (next_event != KTIME_MAX)
 756                 tick_broadcast_set_event(dev, next_cpu, next_event);
 757
 758         raw_spin_unlock(&tick_broadcast_lock);
 759
 760         if (bc_local) {
 761                 td = this_cpu_ptr(&tick_cpu_device);
 762                 td->evtdev->event_handler(td->evtdev);
 763         }
 764 }
 765
 766 static int broadcast_needs_cpu(struct clock_event_device *bc, int cpu)
 767 {
 768         if (!(bc->features & CLOCK_EVT_FEAT_HRTIMER))
 769                 return 0;
 770         if (bc->next_event == KTIME_MAX)
 771                 return 0;
 772         return bc->bound_on == cpu ? -EBUSY : 0;
 773 }
 774
 775 static void broadcast_shutdown_local(struct clock_event_device *bc,
 776                                      struct clock_event_device *dev)
 777 {
 778         /*
 779          * For hrtimer based broadcasting we cannot shutdown the cpu
 780          * local device if our own event is the first one to expire or
 781          * if we own the broadcast timer.
 782          */
 783         if (bc->features & CLOCK_EVT_FEAT_HRTIMER) {
 784                 if (broadcast_needs_cpu(bc, smp_processor_id()))
 785                         return;
 786                 if (dev->next_event < bc->next_event)
 787                         return;
 788         }
 789         clockevents_switch_state(dev, CLOCK_EVT_STATE_SHUTDOWN);
 790 }
 791
 792 static int ___tick_broadcast_oneshot_control(enum tick_broadcast_state state,
 793                                              struct tick_device *td,
 794                                              int cpu)
 795 {
 796         struct clock_event_device *bc, *dev = td->evtdev;
 797         int ret = 0;
 798         ktime_t now;
 799
 800         raw_spin_lock(&tick_broadcast_lock);
 801         bc = tick_broadcast_device.evtdev;
 802
 803         if (state == TICK_BROADCAST_ENTER) {
 804                 /*
 805                  * If the current CPU owns the hrtimer broadcast
 806                  * mechanism, it cannot go deep idle and we do not add
 807                  * the CPU to the broadcast mask. We don't have to go
 808                  * through the EXIT path as the local timer is not
 809                  * shutdown.
 810                  */
 811                 ret = broadcast_needs_cpu(bc, cpu);
 812                 if (ret)
 813                         goto out;
 814
 815                 /*
 816                  * If the broadcast device is in periodic mode, we
 817                  * return.
 818                  */
 819                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 820                         /* If it is a hrtimer based broadcast, return busy */
 821                         if (bc->features & CLOCK_EVT_FEAT_HRTIMER)
 822                                 ret = -EBUSY;
 823                         goto out;
 824                 }
 825
 826                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 827                         WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
 828
 829                         /* Conditionally shut down the local timer. */
 830                         broadcast_shutdown_local(bc, dev);
 831
 832                         /*
 833                          * We only reprogram the broadcast timer if we
 834                          * did not mark ourself in the force mask and
 835                          * if the cpu local event is earlier than the
 836                          * broadcast event. If the current CPU is in
 837                          * the force mask, then we are going to be
 838                          * woken by the IPI right away; we return
 839                          * busy, so the CPU does not try to go deep
 840                          * idle.
 841                          */
 842                         if (cpumask_test_cpu(cpu, tick_broadcast_force_mask)) {
 843                                 ret = -EBUSY;
 844                         } else if (dev->next_event < bc->next_event) {
 845                                 tick_broadcast_set_event(bc, cpu, dev->next_event);
 846                                 /*
 847                                  * In case of hrtimer broadcasts the
 848                                  * programming might have moved the
 849                                  * timer to this cpu. If yes, remove
 850                                  * us from the broadcast mask and
 851                                  * return busy.
 852                                  */
 853                                 ret = broadcast_needs_cpu(bc, cpu);
 854                                 if (ret) {
 855                                         cpumask_clear_cpu(cpu,
 856                                                 tick_broadcast_oneshot_mask);
 857                                 }
 858                         }
 859                 }
 860         } else {
 861                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 862                         clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT);
 863                         /*
 864                          * The cpu which was handling the broadcast
 865                          * timer marked this cpu in the broadcast
 866                          * pending mask and fired the broadcast
 867                          * IPI. So we are going to handle the expired
 868                          * event anyway via the broadcast IPI
 869                          * handler. No need to reprogram the timer
 870                          * with an already expired event.
 871                          */
 872                         if (cpumask_test_and_clear_cpu(cpu,
 873                                        tick_broadcast_pending_mask))
 874                                 goto out;
 875
 876                         /*
 877                          * Bail out if there is no next event.
 878                          */
 879                         if (dev->next_event == KTIME_MAX)
 880                                 goto out;
 881                         /*
 882                          * If the pending bit is not set, then we are
 883                          * either the CPU handling the broadcast
 884                          * interrupt or we got woken by something else.
 885                          *
 886                          * We are no longer in the broadcast mask, so
 887                          * if the cpu local expiry time is already
 888                          * reached, we would reprogram the cpu local
 889                          * timer with an already expired event.
 890                          *
 891                          * This can lead to a ping-pong when we return
 892                          * to idle and therefore rearm the broadcast
 893                          * timer before the cpu local timer was able
 894                          * to fire. This happens because the forced
 895                          * reprogramming makes sure that the event
 896                          * will happen in the future and depending on
 897                          * the min_delta setting this might be far
 898                          * enough out that the ping-pong starts.
 899                          *
 900                          * If the cpu local next_event has expired
 901                          * then we know that the broadcast timer
 902                          * next_event has expired as well and
 903                          * broadcast is about to be handled. So we
 904                          * avoid reprogramming and enforce that the
 905                          * broadcast handler, which did not run yet,
 906                          * will invoke the cpu local handler.
 907                          *
 908                          * We cannot call the handler directly from
 909                          * here, because we might be in a NOHZ phase
 910                          * and we did not go through the irq_enter()
 911                          * nohz fixups.
 912                          */
 913                         now = ktime_get();
 914                         if (dev->next_event <= now) {
 915                                 cpumask_set_cpu(cpu, tick_broadcast_force_mask);
 916                                 goto out;
 917                         }
 918                         /*
 919                          * We got woken by something else. Reprogram
 920                          * the cpu local timer device.
 921                          */
 922                         tick_program_event(dev->next_event, 1);
 923                 }
 924         }
 925 out:
 926         raw_spin_unlock(&tick_broadcast_lock);
 927         return ret;
 928 }
 929
 930 static int tick_oneshot_wakeup_control(enum tick_broadcast_state state,
 931                                        struct tick_device *td,
 932                                        int cpu)
 933 {
 934         struct clock_event_device *dev, *wd;
 935
 936         dev = td->evtdev;
 937         if (td->mode != TICKDEV_MODE_ONESHOT)
 938                 return -EINVAL;
 939
 940         wd = tick_get_oneshot_wakeup_device(cpu);
 941         if (!wd)
 942                 return -ENODEV;
 943
 944         switch (state) {
 945         case TICK_BROADCAST_ENTER:
 946                 clockevents_switch_state(dev, CLOCK_EVT_STATE_ONESHOT_STOPPED);
 947                 clockevents_switch_state(wd, CLOCK_EVT_STATE_ONESHOT);
 948                 clockevents_program_event(wd, dev->next_event, 1);
 949                 break;
 950         case TICK_BROADCAST_EXIT:
 951                 /* We may have transitioned to oneshot mode while idle */
 952                 if (clockevent_get_state(wd) != CLOCK_EVT_STATE_ONESHOT)
 953                         return -ENODEV;
 954         }
 955
 956         return 0;
 957 }
 958
 959 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 960 {
 961         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 962         int cpu = smp_processor_id();
 963
 964         if (!tick_oneshot_wakeup_control(state, td, cpu))
 965                 return 0;
 966
 967         if (tick_broadcast_device.evtdev)
 968                 return ___tick_broadcast_oneshot_control(state, td, cpu);
 969
 970         /*
 971          * If there is no broadcast or wakeup device, tell the caller not
 972          * to go into deep idle.
 973          */
 974         return -EBUSY;
 975 }
 976
 977 /*
 978  * Reset the one shot broadcast for a cpu
 979  *
 980  * Called with tick_broadcast_lock held
 981  */
 982 static void tick_broadcast_clear_oneshot(int cpu)
 983 {
 984         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 985         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
 986 }
 987
 988 static void tick_broadcast_init_next_event(struct cpumask *mask,
 989                                            ktime_t expires)
 990 {
 991         struct tick_device *td;
 992         int cpu;
 993
 994         for_each_cpu(cpu, mask) {
 995                 td = &per_cpu(tick_cpu_device, cpu);
 996                 if (td->evtdev)
 997                         td->evtdev->next_event = expires;
 998         }
 999 }
1000
1001 static inline ktime_t tick_get_next_period(void)
1002 {
1003         ktime_t next;
1004
1005         /*
1006          * Protect against concurrent updates (store /load tearing on
1007          * 32bit). It does not matter if the time is already in the
1008          * past. The broadcast device which is about to be programmed will
1009          * fire in any case.
1010          */
1011         raw_spin_lock(&jiffies_lock);
1012         next = tick_next_period;
1013         raw_spin_unlock(&jiffies_lock);
1014         return next;
1015 }
1016
1017 /**
1018  * tick_broadcast_setup_oneshot - setup the broadcast device
1019  */
1020 static void tick_broadcast_setup_oneshot(struct clock_event_device *bc,
1021                                          bool from_periodic)
1022 {
1023         int cpu = smp_processor_id();
1024         ktime_t nexttick = 0;
1025
1026         if (!bc)
1027                 return;
1028
1029         /*
1030          * When the broadcast device was switched to oneshot by the first
1031          * CPU handling the NOHZ change, the other CPUs will reach this
1032          * code via hrtimer_run_queues() -> tick_check_oneshot_change()
1033          * too. Set up the broadcast device only once!
1034          */
1035         if (bc->event_handler == tick_handle_oneshot_broadcast) {
1036                 /*
1037                  * The CPU which switched from periodic to oneshot mode
1038                  * set the broadcast oneshot bit for all other CPUs which
1039                  * are in the general (periodic) broadcast mask to ensure
1040                  * that CPUs which wait for the periodic broadcast are
1041                  * woken up.
1042                  *
1043                  * Clear the bit for the local CPU as the set bit would
1044                  * prevent the first tick_broadcast_enter() after this CPU
1045                  * switched to oneshot state to program the broadcast
1046                  * device.
1047                  *
1048                  * This code can also be reached via tick_broadcast_control(),
1049                  * but this cannot avoid the tick_broadcast_clear_oneshot()
1050                  * as that would break the periodic to oneshot transition of
1051                  * secondary CPUs. But that's harmless as the below only
1052                  * clears already cleared bits.
1053                  */
1054                 tick_broadcast_clear_oneshot(cpu);
1055                 return;
1056         }
1057
1058
1059         bc->event_handler = tick_handle_oneshot_broadcast;
1060         bc->next_event = KTIME_MAX;
1061
1062         /*
1063          * When the tick mode is switched from periodic to oneshot it must
1064          * be ensured that CPUs which are waiting for periodic broadcast
1065          * get their wake-up at the next tick.  This is achieved by ORing
1066          * tick_broadcast_mask into tick_broadcast_oneshot_mask.
1067          *
1068          * For other callers, e.g. broadcast device replacement,
1069          * tick_broadcast_oneshot_mask must not be touched as this would
1070          * set bits for CPUs which are already NOHZ, but not idle. Their
1071          * next tick_broadcast_enter() would observe the bit set and fail
1072          * to update the expiry time and the broadcast event device.
1073          */
1074         if (from_periodic) {
1075                 cpumask_copy(tmpmask, tick_broadcast_mask);
1076                 /* Remove the local CPU as it is obviously not idle */
1077                 cpumask_clear_cpu(cpu, tmpmask);
1078                 cpumask_or(tick_broadcast_oneshot_mask, tick_broadcast_oneshot_mask, tmpmask);
1079
1080                 /*
1081                  * Ensure that the oneshot broadcast handler will wake the
1082                  * CPUs which are still waiting for periodic broadcast.
1083                  */
1084                 nexttick = tick_get_next_period();
1085                 tick_broadcast_init_next_event(tmpmask, nexttick);
1086
1087                 /*
1088                  * If the underlying broadcast clock event device is
1089                  * already in oneshot state, then there is nothing to do.
1090                  * The device was already armed for the next tick
1091                  * in tick_handle_broadcast_periodic()
1092                  */
1093                 if (clockevent_state_oneshot(bc))
1094                         return;
1095         }
1096
1097         /*
1098          * When switching from periodic to oneshot mode arm the broadcast
1099          * device for the next tick.
1100          *
1101          * If the broadcast device has been replaced in oneshot mode and
1102          * the oneshot broadcast mask is not empty, then arm it to expire
1103          * immediately in order to reevaluate the next expiring timer.
1104          * @nexttick is 0 and therefore in the past which will cause the
1105          * clockevent code to force an event.
1106          *
1107          * For both cases the programming can be avoided when the oneshot
1108          * broadcast mask is empty.
1109          *
1110          * tick_broadcast_set_event() implicitly switches the broadcast
1111          * device to oneshot state.
1112          */
1113         if (!cpumask_empty(tick_broadcast_oneshot_mask))
1114                 tick_broadcast_set_event(bc, cpu, nexttick);
1115 }
1116
1117 /*
1118  * Select oneshot operating mode for the broadcast device
1119  */
1120 void tick_broadcast_switch_to_oneshot(void)
1121 {
1122         struct clock_event_device *bc;
1123         enum tick_device_mode oldmode;
1124         unsigned long flags;
1125
1126         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
1127
1128         oldmode = tick_broadcast_device.mode;
1129         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
1130         bc = tick_broadcast_device.evtdev;
1131         if (bc)
1132                 tick_broadcast_setup_oneshot(bc, oldmode == TICKDEV_MODE_PERIODIC);
1133
1134         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
1135 }
1136
1137 #ifdef CONFIG_HOTPLUG_CPU
1138 void hotplug_cpu__broadcast_tick_pull(int deadcpu)
1139 {
1140         struct clock_event_device *bc;
1141         unsigned long flags;
1142
1143         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
1144         bc = tick_broadcast_device.evtdev;
1145
1146         if (bc && broadcast_needs_cpu(bc, deadcpu)) {
1147                 /* This moves the broadcast assignment to this CPU: */
1148                 clockevents_program_event(bc, bc->next_event, 1);
1149         }
1150         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
1151 }
1152
1153 /*
1154  * Remove a dying CPU from broadcasting
1155  */
1156 static void tick_broadcast_oneshot_offline(unsigned int cpu)
1157 {
1158         if (tick_get_oneshot_wakeup_device(cpu))
1159                 tick_set_oneshot_wakeup_device(NULL, cpu);
1160
1161         /*
1162          * Clear the broadcast masks for the dead cpu, but do not stop
1163          * the broadcast device!
1164          */
1165         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
1166         cpumask_clear_cpu(cpu, tick_broadcast_pending_mask);
1167         cpumask_clear_cpu(cpu, tick_broadcast_force_mask);
1168 }
1169 #endif
1170
1171 /*
1172  * Check, whether the broadcast device is in one shot mode
1173  */
1174 int tick_broadcast_oneshot_active(void)
1175 {
1176         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
1177 }
1178
1179 /*
1180  * Check whether the broadcast device supports oneshot.
1181  */
1182 bool tick_broadcast_oneshot_available(void)
1183 {
1184         struct clock_event_device *bc = tick_broadcast_device.evtdev;
1185
1186         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
1187 }
1188
1189 #else
1190 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
1191 {
1192         struct clock_event_device *bc = tick_broadcast_device.evtdev;
1193
1194         if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
1195                 return -EBUSY;
1196
1197         return 0;
1198 }
1199 #endif
1200
1201 void __init tick_broadcast_init(void)
1202 {
1203         zalloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
1204         zalloc_cpumask_var(&tick_broadcast_on, GFP_NOWAIT);
1205         zalloc_cpumask_var(&tmpmask, GFP_NOWAIT);
1206 #ifdef CONFIG_TICK_ONESHOT
1207         zalloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
1208         zalloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
1209         zalloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
1210 #endif
1211 }