smp.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Generic helpers for smp ipi calls
   4 *
   5 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
   6 */
   7
   8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10#include <linux/irq_work.h>
  11#include <linux/rcupdate.h>
  12#include <linux/rculist.h>
  13#include <linux/kernel.h>
  14#include <linux/export.h>
  15#include <linux/percpu.h>
  16#include <linux/init.h>
  17#include <linux/interrupt.h>
  18#include <linux/gfp.h>
  19#include <linux/smp.h>
  20#include <linux/cpu.h>
  21#include <linux/sched.h>
  22#include <linux/sched/idle.h>
  23#include <linux/hypervisor.h>
  24#include <linux/sched/clock.h>
  25#include <linux/nmi.h>
  26#include <linux/sched/debug.h>
  27#include <linux/jump_label.h>
  28#include <linux/string_choices.h>
  29
  30#include <trace/events/ipi.h>
  31#define CREATE_TRACE_POINTS
  32#include <trace/events/csd.h>
  33#undef CREATE_TRACE_POINTS
  34
  35#include "smpboot.h"
  36#include "sched/smp.h"
  37
  38#define CSD_TYPE(_csd)	((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK)
  39
  40struct call_function_data {
  41	call_single_data_t	__percpu *csd;
  42	cpumask_var_t		cpumask;
  43	cpumask_var_t		cpumask_ipi;
  44};
  45
  46static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
  47
  48static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
  49
  50static DEFINE_PER_CPU(atomic_t, trigger_backtrace) = ATOMIC_INIT(1);
  51
  52static void __flush_smp_call_function_queue(bool warn_cpu_offline);
  53
  54int smpcfd_prepare_cpu(unsigned int cpu)
  55{
  56	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
  57
  58	if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
  59				     cpu_to_node(cpu)))
  60		return -ENOMEM;
  61	if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
  62				     cpu_to_node(cpu))) {
  63		free_cpumask_var(cfd->cpumask);
  64		return -ENOMEM;
  65	}
  66	cfd->csd = alloc_percpu(call_single_data_t);
  67	if (!cfd->csd) {
  68		free_cpumask_var(cfd->cpumask);
  69		free_cpumask_var(cfd->cpumask_ipi);
  70		return -ENOMEM;
  71	}
  72
  73	return 0;
  74}
  75
  76int smpcfd_dead_cpu(unsigned int cpu)
  77{
  78	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
  79
  80	free_cpumask_var(cfd->cpumask);
  81	free_cpumask_var(cfd->cpumask_ipi);
  82	free_percpu(cfd->csd);
  83	return 0;
  84}
  85
  86int smpcfd_dying_cpu(unsigned int cpu)
  87{
  88	/*
  89	 * The IPIs for the smp-call-function callbacks queued by other CPUs
  90	 * might arrive late, either due to hardware latencies or because this
  91	 * CPU disabled interrupts (inside stop-machine) before the IPIs were
  92	 * sent. So flush out any pending callbacks explicitly (without waiting
  93	 * for the IPIs to arrive), to ensure that the outgoing CPU doesn't go
  94	 * offline with work still pending.
  95	 *
  96	 * This runs with interrupts disabled inside the stopper task invoked by
  97	 * stop_machine(), ensuring mutually exclusive CPU offlining and IPI flush.
  98	 */
  99	__flush_smp_call_function_queue(false);
 100	irq_work_run();
 101	return 0;
 102}
 103
 104void __init call_function_init(void)
 105{
 106	int i;
 107
 108	for_each_possible_cpu(i)
 109		init_llist_head(&per_cpu(call_single_queue, i));
 110
 111	smpcfd_prepare_cpu(smp_processor_id());
 112}
 113
 114static __always_inline void
 115send_call_function_single_ipi(int cpu)
 116{
 117	if (call_function_single_prep_ipi(cpu)) {
 118		trace_ipi_send_cpu(cpu, _RET_IP_,
 119				   generic_smp_call_function_single_interrupt);
 120		arch_send_call_function_single_ipi(cpu);
 121	}
 122}
 123
 124static __always_inline void
 125send_call_function_ipi_mask(struct cpumask *mask)
 126{
 127	trace_ipi_send_cpumask(mask, _RET_IP_,
 128			       generic_smp_call_function_single_interrupt);
 129	arch_send_call_function_ipi_mask(mask);
 130}
 131
 132static __always_inline void
 133csd_do_func(smp_call_func_t func, void *info, call_single_data_t *csd)
 134{
 135	trace_csd_function_entry(func, csd);
 136	func(info);
 137	trace_csd_function_exit(func, csd);
 138}
 139
 140#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 141
 142static DEFINE_STATIC_KEY_MAYBE(CONFIG_CSD_LOCK_WAIT_DEBUG_DEFAULT, csdlock_debug_enabled);
 143
 144/*
 145 * Parse the csdlock_debug= kernel boot parameter.
 146 *
 147 * If you need to restore the old "ext" value that once provided
 148 * additional debugging information, reapply the following commits:
 149 *
 150 * de7b09ef658d ("locking/csd_lock: Prepare more CSD lock debugging")
 151 * a5aabace5fb8 ("locking/csd_lock: Add more data to CSD lock debugging")
 152 */
 153static int __init csdlock_debug(char *str)
 154{
 155	int ret;
 156	unsigned int val = 0;
 157
 158	ret = get_option(&str, &val);
 159	if (ret) {
 160		if (val)
 161			static_branch_enable(&csdlock_debug_enabled);
 162		else
 163			static_branch_disable(&csdlock_debug_enabled);
 164	}
 165
 166	return 1;
 167}
 168__setup("csdlock_debug=", csdlock_debug);
 169
 170static DEFINE_PER_CPU(call_single_data_t *, cur_csd);
 171static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func);
 172static DEFINE_PER_CPU(void *, cur_csd_info);
 173
 174static ulong csd_lock_timeout = 5000;  /* CSD lock timeout in milliseconds. */
 175module_param(csd_lock_timeout, ulong, 0644);
 176static int panic_on_ipistall;  /* CSD panic timeout in milliseconds, 300000 for five minutes. */
 177module_param(panic_on_ipistall, int, 0644);
 178
 179static atomic_t csd_bug_count = ATOMIC_INIT(0);
 180
 181/* Record current CSD work for current CPU, NULL to erase. */
 182static void __csd_lock_record(call_single_data_t *csd)
 183{
 184	if (!csd) {
 185		smp_mb(); /* NULL cur_csd after unlock. */
 186		__this_cpu_write(cur_csd, NULL);
 187		return;
 188	}
 189	__this_cpu_write(cur_csd_func, csd->func);
 190	__this_cpu_write(cur_csd_info, csd->info);
 191	smp_wmb(); /* func and info before csd. */
 192	__this_cpu_write(cur_csd, csd);
 193	smp_mb(); /* Update cur_csd before function call. */
 194		  /* Or before unlock, as the case may be. */
 195}
 196
 197static __always_inline void csd_lock_record(call_single_data_t *csd)
 198{
 199	if (static_branch_unlikely(&csdlock_debug_enabled))
 200		__csd_lock_record(csd);
 201}
 202
 203static int csd_lock_wait_getcpu(call_single_data_t *csd)
 204{
 205	unsigned int csd_type;
 206
 207	csd_type = CSD_TYPE(csd);
 208	if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC)
 209		return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */
 210	return -1;
 211}
 212
 213static atomic_t n_csd_lock_stuck;
 214
 215/**
 216 * csd_lock_is_stuck - Has a CSD-lock acquisition been stuck too long?
 217 *
 218 * Returns @true if a CSD-lock acquisition is stuck and has been stuck
 219 * long enough for a "non-responsive CSD lock" message to be printed.
 220 */
 221bool csd_lock_is_stuck(void)
 222{
 223	return !!atomic_read(&n_csd_lock_stuck);
 224}
 225
 226/*
 227 * Complain if too much time spent waiting.  Note that only
 228 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
 229 * so waiting on other types gets much less information.
 230 */
 231static bool csd_lock_wait_toolong(call_single_data_t *csd, u64 ts0, u64 *ts1, int *bug_id, unsigned long *nmessages)
 232{
 233	int cpu = -1;
 234	int cpux;
 235	bool firsttime;
 236	u64 ts2, ts_delta;
 237	call_single_data_t *cpu_cur_csd;
 238	unsigned int flags = READ_ONCE(csd->node.u_flags);
 239	unsigned long long csd_lock_timeout_ns = csd_lock_timeout * NSEC_PER_MSEC;
 240
 241	if (!(flags & CSD_FLAG_LOCK)) {
 242		if (!unlikely(*bug_id))
 243			return true;
 244		cpu = csd_lock_wait_getcpu(csd);
 245		pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
 246			 *bug_id, raw_smp_processor_id(), cpu);
 247		atomic_dec(&n_csd_lock_stuck);
 248		return true;
 249	}
 250
 251	ts2 = ktime_get_mono_fast_ns();
 252	/* How long since we last checked for a stuck CSD lock.*/
 253	ts_delta = ts2 - *ts1;
 254	if (likely(ts_delta <= csd_lock_timeout_ns * (*nmessages + 1) *
 255			       (!*nmessages ? 1 : (ilog2(num_online_cpus()) / 2 + 1)) ||
 256		   csd_lock_timeout_ns == 0))
 257		return false;
 258
 259	if (ts0 > ts2) {
 260		/* Our own sched_clock went backward; don't blame another CPU. */
 261		ts_delta = ts0 - ts2;
 262		pr_alert("sched_clock on CPU %d went backward by %llu ns\n", raw_smp_processor_id(), ts_delta);
 263		*ts1 = ts2;
 264		return false;
 265	}
 266
 267	firsttime = !*bug_id;
 268	if (firsttime)
 269		*bug_id = atomic_inc_return(&csd_bug_count);
 270	cpu = csd_lock_wait_getcpu(csd);
 271	if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu))
 272		cpux = 0;
 273	else
 274		cpux = cpu;
 275	cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */
 276	/* How long since this CSD lock was stuck. */
 277	ts_delta = ts2 - ts0;
 278	pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %lld ns for CPU#%02d %pS(%ps).\n",
 279		 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), (s64)ts_delta,
 280		 cpu, csd->func, csd->info);
 281	(*nmessages)++;
 282	if (firsttime)
 283		atomic_inc(&n_csd_lock_stuck);
 284	/*
 285	 * If the CSD lock is still stuck after 5 minutes, it is unlikely
 286	 * to become unstuck. Use a signed comparison to avoid triggering
 287	 * on underflows when the TSC is out of sync between sockets.
 288	 */
 289	BUG_ON(panic_on_ipistall > 0 && (s64)ts_delta > ((s64)panic_on_ipistall * NSEC_PER_MSEC));
 290	if (cpu_cur_csd && csd != cpu_cur_csd) {
 291		pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
 292			 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)),
 293			 READ_ONCE(per_cpu(cur_csd_info, cpux)));
 294	} else {
 295		pr_alert("\tcsd: CSD lock (#%d) %s.\n",
 296			 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request");
 297	}
 298	if (cpu >= 0) {
 299		if (atomic_cmpxchg_acquire(&per_cpu(trigger_backtrace, cpu), 1, 0))
 300			dump_cpu_task(cpu);
 301		if (!cpu_cur_csd) {
 302			pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu);
 303			arch_send_call_function_single_ipi(cpu);
 304		}
 305	}
 306	if (firsttime)
 307		dump_stack();
 308	*ts1 = ts2;
 309
 310	return false;
 311}
 312
 313/*
 314 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
 315 *
 316 * For non-synchronous ipi calls the csd can still be in use by the
 317 * previous function call. For multi-cpu calls its even more interesting
 318 * as we'll have to ensure no other cpu is observing our csd.
 319 */
 320static void __csd_lock_wait(call_single_data_t *csd)
 321{
 322	unsigned long nmessages = 0;
 323	int bug_id = 0;
 324	u64 ts0, ts1;
 325
 326	ts1 = ts0 = ktime_get_mono_fast_ns();
 327	for (;;) {
 328		if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id, &nmessages))
 329			break;
 330		cpu_relax();
 331	}
 332	smp_acquire__after_ctrl_dep();
 333}
 334
 335static __always_inline void csd_lock_wait(call_single_data_t *csd)
 336{
 337	if (static_branch_unlikely(&csdlock_debug_enabled)) {
 338		__csd_lock_wait(csd);
 339		return;
 340	}
 341
 342	smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
 343}
 344#else
 345static void csd_lock_record(call_single_data_t *csd)
 346{
 347}
 348
 349static __always_inline void csd_lock_wait(call_single_data_t *csd)
 350{
 351	smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
 352}
 353#endif
 354
 355static __always_inline void csd_lock(call_single_data_t *csd)
 356{
 357	csd_lock_wait(csd);
 358	csd->node.u_flags |= CSD_FLAG_LOCK;
 359
 360	/*
 361	 * prevent CPU from reordering the above assignment
 362	 * to ->flags with any subsequent assignments to other
 363	 * fields of the specified call_single_data_t structure:
 364	 */
 365	smp_wmb();
 366}
 367
 368static __always_inline void csd_unlock(call_single_data_t *csd)
 369{
 370	WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK));
 371
 372	/*
 373	 * ensure we're all done before releasing data:
 374	 */
 375	smp_store_release(&csd->node.u_flags, 0);
 376}
 377
 378static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
 379
 380void __smp_call_single_queue(int cpu, struct llist_node *node)
 381{
 382	/*
 383	 * We have to check the type of the CSD before queueing it, because
 384	 * once queued it can have its flags cleared by
 385	 *   flush_smp_call_function_queue()
 386	 * even if we haven't sent the smp_call IPI yet (e.g. the stopper
 387	 * executes migration_cpu_stop() on the remote CPU).
 388	 */
 389	if (trace_csd_queue_cpu_enabled()) {
 390		call_single_data_t *csd;
 391		smp_call_func_t func;
 392
 393		csd = container_of(node, call_single_data_t, node.llist);
 394		func = CSD_TYPE(csd) == CSD_TYPE_TTWU ?
 395			sched_ttwu_pending : csd->func;
 396
 397		trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
 398	}
 399
 400	/*
 401	 * The list addition should be visible to the target CPU when it pops
 402	 * the head of the list to pull the entry off it in the IPI handler
 403	 * because of normal cache coherency rules implied by the underlying
 404	 * llist ops.
 405	 *
 406	 * If IPIs can go out of order to the cache coherency protocol
 407	 * in an architecture, sufficient synchronisation should be added
 408	 * to arch code to make it appear to obey cache coherency WRT
 409	 * locking and barrier primitives. Generic code isn't really
 410	 * equipped to do the right thing...
 411	 */
 412	if (llist_add(node, &per_cpu(call_single_queue, cpu)))
 413		send_call_function_single_ipi(cpu);
 414}
 415
 416/*
 417 * Insert a previously allocated call_single_data_t element
 418 * for execution on the given CPU. data must already have
 419 * ->func, ->info, and ->flags set.
 420 */
 421static int generic_exec_single(int cpu, call_single_data_t *csd)
 422{
 423	/*
 424	 * Preemption already disabled here so stopper cannot run on this CPU,
 425	 * ensuring mutually exclusive CPU offlining and last IPI flush.
 426	 */
 427	if (cpu == smp_processor_id()) {
 428		smp_call_func_t func = csd->func;
 429		void *info = csd->info;
 430		unsigned long flags;
 431
 432		/*
 433		 * We can unlock early even for the synchronous on-stack case,
 434		 * since we're doing this from the same CPU..
 435		 */
 436		csd_lock_record(csd);
 437		csd_unlock(csd);
 438		local_irq_save(flags);
 439		csd_do_func(func, info, NULL);
 440		csd_lock_record(NULL);
 441		local_irq_restore(flags);
 442		return 0;
 443	}
 444
 445	if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
 446		csd_unlock(csd);
 447		return -ENXIO;
 448	}
 449
 450	__smp_call_single_queue(cpu, &csd->node.llist);
 451
 452	return 0;
 453}
 454
 455/**
 456 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
 457 *
 458 * Invoked by arch to handle an IPI for call function single.
 459 * Must be called with interrupts disabled.
 460 */
 461void generic_smp_call_function_single_interrupt(void)
 462{
 463	__flush_smp_call_function_queue(true);
 464}
 465
 466/**
 467 * __flush_smp_call_function_queue - Flush pending smp-call-function callbacks
 468 *
 469 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
 470 *		      offline CPU. Skip this check if set to 'false'.
 471 *
 472 * Flush any pending smp-call-function callbacks queued on this CPU. This is
 473 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
 474 * to ensure that all pending IPI callbacks are run before it goes completely
 475 * offline.
 476 *
 477 * Loop through the call_single_queue and run all the queued callbacks.
 478 * Must be called with interrupts disabled.
 479 */
 480static void __flush_smp_call_function_queue(bool warn_cpu_offline)
 481{
 482	call_single_data_t *csd, *csd_next;
 483	struct llist_node *entry, *prev;
 484	struct llist_head *head;
 485	static bool warned;
 486	atomic_t *tbt;
 487
 488	lockdep_assert_irqs_disabled();
 489
 490	/* Allow waiters to send backtrace NMI from here onwards */
 491	tbt = this_cpu_ptr(&trigger_backtrace);
 492	atomic_set_release(tbt, 1);
 493
 494	head = this_cpu_ptr(&call_single_queue);
 495	entry = llist_del_all(head);
 496	entry = llist_reverse_order(entry);
 497
 498	/* There shouldn't be any pending callbacks on an offline CPU. */
 499	if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
 500		     !warned && entry != NULL)) {
 501		warned = true;
 502		WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
 503
 504		/*
 505		 * We don't have to use the _safe() variant here
 506		 * because we are not invoking the IPI handlers yet.
 507		 */
 508		llist_for_each_entry(csd, entry, node.llist) {
 509			switch (CSD_TYPE(csd)) {
 510			case CSD_TYPE_ASYNC:
 511			case CSD_TYPE_SYNC:
 512			case CSD_TYPE_IRQ_WORK:
 513				pr_warn("IPI callback %pS sent to offline CPU\n",
 514					csd->func);
 515				break;
 516
 517			case CSD_TYPE_TTWU:
 518				pr_warn("IPI task-wakeup sent to offline CPU\n");
 519				break;
 520
 521			default:
 522				pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
 523					CSD_TYPE(csd));
 524				break;
 525			}
 526		}
 527	}
 528
 529	/*
 530	 * First; run all SYNC callbacks, people are waiting for us.
 531	 */
 532	prev = NULL;
 533	llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
 534		/* Do we wait until *after* callback? */
 535		if (CSD_TYPE(csd) == CSD_TYPE_SYNC) {
 536			smp_call_func_t func = csd->func;
 537			void *info = csd->info;
 538
 539			if (prev) {
 540				prev->next = &csd_next->node.llist;
 541			} else {
 542				entry = &csd_next->node.llist;
 543			}
 544
 545			csd_lock_record(csd);
 546			csd_do_func(func, info, csd);
 547			csd_unlock(csd);
 548			csd_lock_record(NULL);
 549		} else {
 550			prev = &csd->node.llist;
 551		}
 552	}
 553
 554	if (!entry)
 555		return;
 556
 557	/*
 558	 * Second; run all !SYNC callbacks.
 559	 */
 560	prev = NULL;
 561	llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
 562		int type = CSD_TYPE(csd);
 563
 564		if (type != CSD_TYPE_TTWU) {
 565			if (prev) {
 566				prev->next = &csd_next->node.llist;
 567			} else {
 568				entry = &csd_next->node.llist;
 569			}
 570
 571			if (type == CSD_TYPE_ASYNC) {
 572				smp_call_func_t func = csd->func;
 573				void *info = csd->info;
 574
 575				csd_lock_record(csd);
 576				csd_unlock(csd);
 577				csd_do_func(func, info, csd);
 578				csd_lock_record(NULL);
 579			} else if (type == CSD_TYPE_IRQ_WORK) {
 580				irq_work_single(csd);
 581			}
 582
 583		} else {
 584			prev = &csd->node.llist;
 585		}
 586	}
 587
 588	/*
 589	 * Third; only CSD_TYPE_TTWU is left, issue those.
 590	 */
 591	if (entry) {
 592		csd = llist_entry(entry, typeof(*csd), node.llist);
 593		csd_do_func(sched_ttwu_pending, entry, csd);
 594	}
 595}
 596
 597
 598/**
 599 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
 600 *				   from task context (idle, migration thread)
 601 *
 602 * When TIF_POLLING_NRFLAG is supported and a CPU is in idle and has it
 603 * set, then remote CPUs can avoid sending IPIs and wake the idle CPU by
 604 * setting TIF_NEED_RESCHED. The idle task on the woken up CPU has to
 605 * handle queued SMP function calls before scheduling.
 606 *
 607 * The migration thread has to ensure that an eventually pending wakeup has
 608 * been handled before it migrates a task.
 609 */
 610void flush_smp_call_function_queue(void)
 611{
 612	unsigned int was_pending;
 613	unsigned long flags;
 614
 615	if (llist_empty(this_cpu_ptr(&call_single_queue)))
 616		return;
 617
 618	local_irq_save(flags);
 619	/* Get the already pending soft interrupts for RT enabled kernels */
 620	was_pending = local_softirq_pending();
 621	__flush_smp_call_function_queue(true);
 622	if (local_softirq_pending())
 623		do_softirq_post_smp_call_flush(was_pending);
 624
 625	local_irq_restore(flags);
 626}
 627
 628/*
 629 * smp_call_function_single - Run a function on a specific CPU
 630 * @func: The function to run. This must be fast and non-blocking.
 631 * @info: An arbitrary pointer to pass to the function.
 632 * @wait: If true, wait until function has completed on other CPUs.
 633 *
 634 * Returns 0 on success, else a negative status code.
 635 */
 636int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
 637			     int wait)
 638{
 639	call_single_data_t *csd;
 640	call_single_data_t csd_stack = {
 641		.node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, },
 642	};
 643	int this_cpu;
 644	int err;
 645
 646	/*
 647	 * Prevent preemption and reschedule on another CPU, as well as CPU
 648	 * removal. This prevents stopper from running on this CPU, thus
 649	 * providing mutual exclusion of the below cpu_online() check and
 650	 * IPI sending ensuring IPI are not missed by CPU going offline.
 651	 */
 652	this_cpu = get_cpu();
 653
 654	/*
 655	 * Can deadlock when called with interrupts disabled.
 656	 * We allow cpu's that are not yet online though, as no one else can
 657	 * send smp call function interrupt to this cpu and as such deadlocks
 658	 * can't happen.
 659	 */
 660	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
 661		     && !oops_in_progress);
 662
 663	/*
 664	 * When @wait we can deadlock when we interrupt between llist_add() and
 665	 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
 666	 * csd_lock() on because the interrupt context uses the same csd
 667	 * storage.
 668	 */
 669	WARN_ON_ONCE(!in_task());
 670
 671	csd = &csd_stack;
 672	if (!wait) {
 673		csd = this_cpu_ptr(&csd_data);
 674		csd_lock(csd);
 675	}
 676
 677	csd->func = func;
 678	csd->info = info;
 679#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 680	csd->node.src = smp_processor_id();
 681	csd->node.dst = cpu;
 682#endif
 683
 684	err = generic_exec_single(cpu, csd);
 685
 686	if (wait)
 687		csd_lock_wait(csd);
 688
 689	put_cpu();
 690
 691	return err;
 692}
 693EXPORT_SYMBOL(smp_call_function_single);
 694
 695/**
 696 * smp_call_function_single_async() - Run an asynchronous function on a
 697 * 			         specific CPU.
 698 * @cpu: The CPU to run on.
 699 * @csd: Pre-allocated and setup data structure
 700 *
 701 * Like smp_call_function_single(), but the call is asynchonous and
 702 * can thus be done from contexts with disabled interrupts.
 703 *
 704 * The caller passes his own pre-allocated data structure
 705 * (ie: embedded in an object) and is responsible for synchronizing it
 706 * such that the IPIs performed on the @csd are strictly serialized.
 707 *
 708 * If the function is called with one csd which has not yet been
 709 * processed by previous call to smp_call_function_single_async(), the
 710 * function will return immediately with -EBUSY showing that the csd
 711 * object is still in progress.
 712 *
 713 * NOTE: Be careful, there is unfortunately no current debugging facility to
 714 * validate the correctness of this serialization.
 715 *
 716 * Return: %0 on success or negative errno value on error
 717 */
 718int smp_call_function_single_async(int cpu, call_single_data_t *csd)
 719{
 720	int err = 0;
 721
 722	preempt_disable();
 723
 724	if (csd->node.u_flags & CSD_FLAG_LOCK) {
 725		err = -EBUSY;
 726		goto out;
 727	}
 728
 729	csd->node.u_flags = CSD_FLAG_LOCK;
 730	smp_wmb();
 731
 732	err = generic_exec_single(cpu, csd);
 733
 734out:
 735	preempt_enable();
 736
 737	return err;
 738}
 739EXPORT_SYMBOL_GPL(smp_call_function_single_async);
 740
 741/*
 742 * smp_call_function_any - Run a function on any of the given cpus
 743 * @mask: The mask of cpus it can run on.
 744 * @func: The function to run. This must be fast and non-blocking.
 745 * @info: An arbitrary pointer to pass to the function.
 746 * @wait: If true, wait until function has completed.
 747 *
 748 * Returns 0 on success, else a negative status code (if no cpus were online).
 749 *
 750 * Selection preference:
 751 *	1) current cpu if in @mask
 752 *	2) nearest cpu in @mask, based on NUMA topology
 753 */
 754int smp_call_function_any(const struct cpumask *mask,
 755			  smp_call_func_t func, void *info, int wait)
 756{
 757	unsigned int cpu;
 758	int ret;
 759
 760	/* Try for same CPU (cheapest) */
 761	cpu = get_cpu();
 762	if (!cpumask_test_cpu(cpu, mask))
 763		cpu = sched_numa_find_nth_cpu(mask, 0, cpu_to_node(cpu));
 764
 765	ret = smp_call_function_single(cpu, func, info, wait);
 766	put_cpu();
 767	return ret;
 768}
 769EXPORT_SYMBOL_GPL(smp_call_function_any);
 770
 771/*
 772 * Flags to be used as scf_flags argument of smp_call_function_many_cond().
 773 *
 774 * %SCF_WAIT:		Wait until function execution is completed
 775 * %SCF_RUN_LOCAL:	Run also locally if local cpu is set in cpumask
 776 */
 777#define SCF_WAIT	(1U << 0)
 778#define SCF_RUN_LOCAL	(1U << 1)
 779
 780static void smp_call_function_many_cond(const struct cpumask *mask,
 781					smp_call_func_t func, void *info,
 782					unsigned int scf_flags,
 783					smp_cond_func_t cond_func)
 784{
 785	int cpu, last_cpu, this_cpu = smp_processor_id();
 786	struct call_function_data *cfd;
 787	bool wait = scf_flags & SCF_WAIT;
 788	int nr_cpus = 0;
 789	bool run_remote = false;
 790
 791	lockdep_assert_preemption_disabled();
 792
 793	/*
 794	 * Can deadlock when called with interrupts disabled.
 795	 * We allow cpu's that are not yet online though, as no one else can
 796	 * send smp call function interrupt to this cpu and as such deadlocks
 797	 * can't happen.
 798	 */
 799	if (cpu_online(this_cpu) && !oops_in_progress &&
 800	    !early_boot_irqs_disabled)
 801		lockdep_assert_irqs_enabled();
 802
 803	/*
 804	 * When @wait we can deadlock when we interrupt between llist_add() and
 805	 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
 806	 * csd_lock() on because the interrupt context uses the same csd
 807	 * storage.
 808	 */
 809	WARN_ON_ONCE(!in_task());
 810
 811	/* Check if we need remote execution, i.e., any CPU excluding this one. */
 812	if (cpumask_any_and_but(mask, cpu_online_mask, this_cpu) < nr_cpu_ids) {
 813		cfd = this_cpu_ptr(&cfd_data);
 814		cpumask_and(cfd->cpumask, mask, cpu_online_mask);
 815		__cpumask_clear_cpu(this_cpu, cfd->cpumask);
 816
 817		cpumask_clear(cfd->cpumask_ipi);
 818		for_each_cpu(cpu, cfd->cpumask) {
 819			call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
 820
 821			if (cond_func && !cond_func(cpu, info)) {
 822				__cpumask_clear_cpu(cpu, cfd->cpumask);
 823				continue;
 824			}
 825
 826			/* Work is enqueued on a remote CPU. */
 827			run_remote = true;
 828
 829			csd_lock(csd);
 830			if (wait)
 831				csd->node.u_flags |= CSD_TYPE_SYNC;
 832			csd->func = func;
 833			csd->info = info;
 834#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 835			csd->node.src = smp_processor_id();
 836			csd->node.dst = cpu;
 837#endif
 838			trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
 839
 840			/*
 841			 * Kick the remote CPU if this is the first work
 842			 * item enqueued.
 843			 */
 844			if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
 845				__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
 846				nr_cpus++;
 847				last_cpu = cpu;
 848			}
 849		}
 850
 851		/*
 852		 * Choose the most efficient way to send an IPI. Note that the
 853		 * number of CPUs might be zero due to concurrent changes to the
 854		 * provided mask.
 855		 */
 856		if (nr_cpus == 1)
 857			send_call_function_single_ipi(last_cpu);
 858		else if (likely(nr_cpus > 1))
 859			send_call_function_ipi_mask(cfd->cpumask_ipi);
 860	}
 861
 862	/* Check if we need local execution. */
 863	if ((scf_flags & SCF_RUN_LOCAL) && cpumask_test_cpu(this_cpu, mask) &&
 864	    (!cond_func || cond_func(this_cpu, info))) {
 865		unsigned long flags;
 866
 867		local_irq_save(flags);
 868		csd_do_func(func, info, NULL);
 869		local_irq_restore(flags);
 870	}
 871
 872	if (run_remote && wait) {
 873		for_each_cpu(cpu, cfd->cpumask) {
 874			call_single_data_t *csd;
 875
 876			csd = per_cpu_ptr(cfd->csd, cpu);
 877			csd_lock_wait(csd);
 878		}
 879	}
 880}
 881
 882/**
 883 * smp_call_function_many(): Run a function on a set of CPUs.
 884 * @mask: The set of cpus to run on (only runs on online subset).
 885 * @func: The function to run. This must be fast and non-blocking.
 886 * @info: An arbitrary pointer to pass to the function.
 887 * @wait: If true, wait (atomically) until function has completed
 888 *        on other CPUs.
 889 *
 890 * You must not call this function with disabled interrupts or from a
 891 * hardware interrupt handler or from a bottom half handler. Preemption
 892 * must be disabled when calling this function.
 893 *
 894 * @func is not called on the local CPU even if @mask contains it.  Consider
 895 * using on_each_cpu_cond_mask() instead if this is not desirable.
 896 */
 897void smp_call_function_many(const struct cpumask *mask,
 898			    smp_call_func_t func, void *info, bool wait)
 899{
 900	smp_call_function_many_cond(mask, func, info, wait * SCF_WAIT, NULL);
 901}
 902EXPORT_SYMBOL(smp_call_function_many);
 903
 904/**
 905 * smp_call_function(): Run a function on all other CPUs.
 906 * @func: The function to run. This must be fast and non-blocking.
 907 * @info: An arbitrary pointer to pass to the function.
 908 * @wait: If true, wait (atomically) until function has completed
 909 *        on other CPUs.
 910 *
 911 * Returns 0.
 912 *
 913 * If @wait is true, then returns once @func has returned; otherwise
 914 * it returns just before the target cpu calls @func.
 915 *
 916 * You must not call this function with disabled interrupts or from a
 917 * hardware interrupt handler or from a bottom half handler.
 918 */
 919void smp_call_function(smp_call_func_t func, void *info, int wait)
 920{
 921	preempt_disable();
 922	smp_call_function_many(cpu_online_mask, func, info, wait);
 923	preempt_enable();
 924}
 925EXPORT_SYMBOL(smp_call_function);
 926
 927/* Setup configured maximum number of CPUs to activate */
 928unsigned int setup_max_cpus = NR_CPUS;
 929EXPORT_SYMBOL(setup_max_cpus);
 930
 931
 932/*
 933 * Setup routine for controlling SMP activation
 934 *
 935 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
 936 * activation entirely (the MPS table probe still happens, though).
 937 *
 938 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
 939 * greater than 0, limits the maximum number of CPUs activated in
 940 * SMP mode to <NUM>.
 941 */
 942
 943void __weak __init arch_disable_smp_support(void) { }
 944
 945static int __init nosmp(char *str)
 946{
 947	setup_max_cpus = 0;
 948	arch_disable_smp_support();
 949
 950	return 0;
 951}
 952
 953early_param("nosmp", nosmp);
 954
 955/* this is hard limit */
 956static int __init nrcpus(char *str)
 957{
 958	int nr_cpus;
 959
 960	if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids)
 961		set_nr_cpu_ids(nr_cpus);
 962
 963	return 0;
 964}
 965
 966early_param("nr_cpus", nrcpus);
 967
 968static int __init maxcpus(char *str)
 969{
 970	get_option(&str, &setup_max_cpus);
 971	if (setup_max_cpus == 0)
 972		arch_disable_smp_support();
 973
 974	return 0;
 975}
 976
 977early_param("maxcpus", maxcpus);
 978
 979#if (NR_CPUS > 1) && !defined(CONFIG_FORCE_NR_CPUS)
 980/* Setup number of possible processor ids */
 981unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
 982EXPORT_SYMBOL(nr_cpu_ids);
 983#endif
 984
 985/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
 986void __init setup_nr_cpu_ids(void)
 987{
 988	set_nr_cpu_ids(find_last_bit(cpumask_bits(cpu_possible_mask), NR_CPUS) + 1);
 989}
 990
 991/* Called by boot processor to activate the rest. */
 992void __init smp_init(void)
 993{
 994	int num_nodes, num_cpus;
 995
 996	idle_threads_init();
 997	cpuhp_threads_init();
 998
 999	pr_info("Bringing up secondary CPUs ...\n");
1000
1001	bringup_nonboot_cpus(setup_max_cpus);
1002
1003	num_nodes = num_online_nodes();
1004	num_cpus  = num_online_cpus();
1005	pr_info("Brought up %d node%s, %d CPU%s\n",
1006		num_nodes, str_plural(num_nodes), num_cpus, str_plural(num_cpus));
1007
1008	/* Any cleanup work */
1009	smp_cpus_done(setup_max_cpus);
1010}
1011
1012/*
1013 * on_each_cpu_cond(): Call a function on each processor for which
1014 * the supplied function cond_func returns true, optionally waiting
1015 * for all the required CPUs to finish. This may include the local
1016 * processor.
1017 * @cond_func:	A callback function that is passed a cpu id and
1018 *		the info parameter. The function is called
1019 *		with preemption disabled. The function should
1020 *		return a boolean value indicating whether to IPI
1021 *		the specified CPU.
1022 * @func:	The function to run on all applicable CPUs.
1023 *		This must be fast and non-blocking.
1024 * @info:	An arbitrary pointer to pass to both functions.
1025 * @wait:	If true, wait (atomically) until function has
1026 *		completed on other CPUs.
1027 *
1028 * Preemption is disabled to protect against CPUs going offline but not online.
1029 * CPUs going online during the call will not be seen or sent an IPI.
1030 *
1031 * You must not call this function with disabled interrupts or
1032 * from a hardware interrupt handler or from a bottom half handler.
1033 */
1034void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func,
1035			   void *info, bool wait, const struct cpumask *mask)
1036{
1037	unsigned int scf_flags = SCF_RUN_LOCAL;
1038
1039	if (wait)
1040		scf_flags |= SCF_WAIT;
1041
1042	preempt_disable();
1043	smp_call_function_many_cond(mask, func, info, scf_flags, cond_func);
1044	preempt_enable();
1045}
1046EXPORT_SYMBOL(on_each_cpu_cond_mask);
1047
1048static void do_nothing(void *unused)
1049{
1050}
1051
1052/**
1053 * kick_all_cpus_sync - Force all cpus out of idle
1054 *
1055 * Used to synchronize the update of pm_idle function pointer. It's
1056 * called after the pointer is updated and returns after the dummy
1057 * callback function has been executed on all cpus. The execution of
1058 * the function can only happen on the remote cpus after they have
1059 * left the idle function which had been called via pm_idle function
1060 * pointer. So it's guaranteed that nothing uses the previous pointer
1061 * anymore.
1062 */
1063void kick_all_cpus_sync(void)
1064{
1065	/* Make sure the change is visible before we kick the cpus */
1066	smp_mb();
1067	smp_call_function(do_nothing, NULL, 1);
1068}
1069EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
1070
1071/**
1072 * wake_up_all_idle_cpus - break all cpus out of idle
1073 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
1074 * including idle polling cpus, for non-idle cpus, we will do nothing
1075 * for them.
1076 */
1077void wake_up_all_idle_cpus(void)
1078{
1079	int cpu;
1080
1081	for_each_possible_cpu(cpu) {
1082		preempt_disable();
1083		if (cpu != smp_processor_id() && cpu_online(cpu))
1084			wake_up_if_idle(cpu);
1085		preempt_enable();
1086	}
1087}
1088EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
1089
1090/**
1091 * cpus_peek_for_pending_ipi - Check for pending IPI for CPUs
1092 * @mask: The CPU mask for the CPUs to check.
1093 *
1094 * This function walks through the @mask to check if there are any pending IPIs
1095 * scheduled, for any of the CPUs in the @mask. It does not guarantee
1096 * correctness as it only provides a racy snapshot.
1097 *
1098 * Returns true if there is a pending IPI scheduled and false otherwise.
1099 */
1100bool cpus_peek_for_pending_ipi(const struct cpumask *mask)
1101{
1102	unsigned int cpu;
1103
1104	for_each_cpu(cpu, mask) {
1105		if (!llist_empty(per_cpu_ptr(&call_single_queue, cpu)))
1106			return true;
1107	}
1108
1109	return false;
1110}
1111
1112/**
1113 * struct smp_call_on_cpu_struct - Call a function on a specific CPU
1114 * @work: &work_struct
1115 * @done: &completion to signal
1116 * @func: function to call
1117 * @data: function's data argument
1118 * @ret: return value from @func
1119 * @cpu: target CPU (%-1 for any CPU)
1120 *
1121 * Used to call a function on a specific cpu and wait for it to return.
1122 * Optionally make sure the call is done on a specified physical cpu via vcpu
1123 * pinning in order to support virtualized environments.
1124 */
1125struct smp_call_on_cpu_struct {
1126	struct work_struct	work;
1127	struct completion	done;
1128	int			(*func)(void *);
1129	void			*data;
1130	int			ret;
1131	int			cpu;
1132};
1133
1134static void smp_call_on_cpu_callback(struct work_struct *work)
1135{
1136	struct smp_call_on_cpu_struct *sscs;
1137
1138	sscs = container_of(work, struct smp_call_on_cpu_struct, work);
1139	if (sscs->cpu >= 0)
1140		hypervisor_pin_vcpu(sscs->cpu);
1141	sscs->ret = sscs->func(sscs->data);
1142	if (sscs->cpu >= 0)
1143		hypervisor_pin_vcpu(-1);
1144
1145	complete(&sscs->done);
1146}
1147
1148int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
1149{
1150	struct smp_call_on_cpu_struct sscs = {
1151		.done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
1152		.func = func,
1153		.data = par,
1154		.cpu  = phys ? cpu : -1,
1155	};
1156
1157	INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
1158
1159	if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1160		return -ENXIO;
1161
1162	queue_work_on(cpu, system_wq, &sscs.work);
1163	wait_for_completion(&sscs.done);
1164	destroy_work_on_stack(&sscs.work);
1165
1166	return sscs.ret;
1167}
1168EXPORT_SYMBOL_GPL(smp_call_on_cpu);