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