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