Linux Audio

Check our new training course

Loading...
v4.10.11
 
  1/*
  2 * Generic helpers for smp ipi calls
  3 *
  4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
  5 */
  6
  7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  8
  9#include <linux/irq_work.h>
 10#include <linux/rcupdate.h>
 11#include <linux/rculist.h>
 12#include <linux/kernel.h>
 13#include <linux/export.h>
 14#include <linux/percpu.h>
 15#include <linux/init.h>
 
 16#include <linux/gfp.h>
 17#include <linux/smp.h>
 18#include <linux/cpu.h>
 19#include <linux/sched.h>
 
 20#include <linux/hypervisor.h>
 
 
 
 
 21
 22#include "smpboot.h"
 
 23
 24enum {
 25	CSD_FLAG_LOCK		= 0x01,
 26	CSD_FLAG_SYNCHRONOUS	= 0x02,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 27};
 28
 29struct call_function_data {
 30	struct call_single_data	__percpu *csd;
 31	cpumask_var_t		cpumask;
 
 32};
 33
 34static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
 35
 36static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
 37
 38static void flush_smp_call_function_queue(bool warn_cpu_offline);
 39
 40int smpcfd_prepare_cpu(unsigned int cpu)
 41{
 42	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
 43
 44	if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
 45				     cpu_to_node(cpu)))
 46		return -ENOMEM;
 47	cfd->csd = alloc_percpu(struct call_single_data);
 48	if (!cfd->csd) {
 
 
 
 
 
 49		free_cpumask_var(cfd->cpumask);
 
 50		return -ENOMEM;
 51	}
 52
 53	return 0;
 54}
 55
 56int smpcfd_dead_cpu(unsigned int cpu)
 57{
 58	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
 59
 60	free_cpumask_var(cfd->cpumask);
 61	free_percpu(cfd->csd);
 
 62	return 0;
 63}
 64
 65int smpcfd_dying_cpu(unsigned int cpu)
 66{
 67	/*
 68	 * The IPIs for the smp-call-function callbacks queued by other
 69	 * CPUs might arrive late, either due to hardware latencies or
 70	 * because this CPU disabled interrupts (inside stop-machine)
 71	 * before the IPIs were sent. So flush out any pending callbacks
 72	 * explicitly (without waiting for the IPIs to arrive), to
 73	 * ensure that the outgoing CPU doesn't go offline with work
 74	 * still pending.
 75	 */
 76	flush_smp_call_function_queue(false);
 
 77	return 0;
 78}
 79
 80void __init call_function_init(void)
 81{
 82	int i;
 83
 84	for_each_possible_cpu(i)
 85		init_llist_head(&per_cpu(call_single_queue, i));
 86
 87	smpcfd_prepare_cpu(smp_processor_id());
 88}
 89
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 90/*
 91 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
 92 *
 93 * For non-synchronous ipi calls the csd can still be in use by the
 94 * previous function call. For multi-cpu calls its even more interesting
 95 * as we'll have to ensure no other cpu is observing our csd.
 96 */
 97static __always_inline void csd_lock_wait(struct call_single_data *csd)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 98{
 99	smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
 
 
 
 
 
100}
101
102static __always_inline void csd_lock(struct call_single_data *csd)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
103{
104	csd_lock_wait(csd);
105	csd->flags |= CSD_FLAG_LOCK;
106
107	/*
108	 * prevent CPU from reordering the above assignment
109	 * to ->flags with any subsequent assignments to other
110	 * fields of the specified call_single_data structure:
111	 */
112	smp_wmb();
113}
114
115static __always_inline void csd_unlock(struct call_single_data *csd)
116{
117	WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
118
119	/*
120	 * ensure we're all done before releasing data:
121	 */
122	smp_store_release(&csd->flags, 0);
123}
124
125static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
127/*
128 * Insert a previously allocated call_single_data element
129 * for execution on the given CPU. data must already have
130 * ->func, ->info, and ->flags set.
131 */
132static int generic_exec_single(int cpu, struct call_single_data *csd,
133			       smp_call_func_t func, void *info)
134{
135	if (cpu == smp_processor_id()) {
 
 
136		unsigned long flags;
137
138		/*
139		 * We can unlock early even for the synchronous on-stack case,
140		 * since we're doing this from the same CPU..
141		 */
 
142		csd_unlock(csd);
143		local_irq_save(flags);
144		func(info);
 
145		local_irq_restore(flags);
146		return 0;
147	}
148
149
150	if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
151		csd_unlock(csd);
152		return -ENXIO;
153	}
154
155	csd->func = func;
156	csd->info = info;
157
158	/*
159	 * The list addition should be visible before sending the IPI
160	 * handler locks the list to pull the entry off it because of
161	 * normal cache coherency rules implied by spinlocks.
162	 *
163	 * If IPIs can go out of order to the cache coherency protocol
164	 * in an architecture, sufficient synchronisation should be added
165	 * to arch code to make it appear to obey cache coherency WRT
166	 * locking and barrier primitives. Generic code isn't really
167	 * equipped to do the right thing...
168	 */
169	if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
170		arch_send_call_function_single_ipi(cpu);
171
172	return 0;
173}
174
175/**
176 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
177 *
178 * Invoked by arch to handle an IPI for call function single.
179 * Must be called with interrupts disabled.
180 */
181void generic_smp_call_function_single_interrupt(void)
182{
 
 
183	flush_smp_call_function_queue(true);
184}
185
186/**
187 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
188 *
189 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
190 *		      offline CPU. Skip this check if set to 'false'.
191 *
192 * Flush any pending smp-call-function callbacks queued on this CPU. This is
193 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
194 * to ensure that all pending IPI callbacks are run before it goes completely
195 * offline.
196 *
197 * Loop through the call_single_queue and run all the queued callbacks.
198 * Must be called with interrupts disabled.
199 */
200static void flush_smp_call_function_queue(bool warn_cpu_offline)
201{
 
 
202	struct llist_head *head;
203	struct llist_node *entry;
204	struct call_single_data *csd, *csd_next;
205	static bool warned;
206
207	WARN_ON(!irqs_disabled());
208
209	head = this_cpu_ptr(&call_single_queue);
 
 
210	entry = llist_del_all(head);
 
 
 
 
211	entry = llist_reverse_order(entry);
212
213	/* There shouldn't be any pending callbacks on an offline CPU. */
214	if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
215		     !warned && !llist_empty(head))) {
216		warned = true;
217		WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
218
219		/*
220		 * We don't have to use the _safe() variant here
221		 * because we are not invoking the IPI handlers yet.
222		 */
223		llist_for_each_entry(csd, entry, llist)
224			pr_warn("IPI callback %pS sent to offline CPU\n",
225				csd->func);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
226	}
227
228	llist_for_each_entry_safe(csd, csd_next, entry, llist) {
229		smp_call_func_t func = csd->func;
230		void *info = csd->info;
231
 
232		/* Do we wait until *after* callback? */
233		if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
 
 
 
 
 
 
 
 
 
 
234			func(info);
235			csd_unlock(csd);
 
236		} else {
237			csd_unlock(csd);
238			func(info);
239		}
240	}
241
 
 
 
 
 
 
 
242	/*
243	 * Handle irq works queued remotely by irq_work_queue_on().
244	 * Smp functions above are typically synchronous so they
245	 * better run first since some other CPUs may be busy waiting
246	 * for them.
247	 */
248	irq_work_run();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
249}
250
251/*
252 * smp_call_function_single - Run a function on a specific CPU
253 * @func: The function to run. This must be fast and non-blocking.
254 * @info: An arbitrary pointer to pass to the function.
255 * @wait: If true, wait until function has completed on other CPUs.
256 *
257 * Returns 0 on success, else a negative status code.
258 */
259int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
260			     int wait)
261{
262	struct call_single_data *csd;
263	struct call_single_data csd_stack = { .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS };
 
 
264	int this_cpu;
265	int err;
266
267	/*
268	 * prevent preemption and reschedule on another processor,
269	 * as well as CPU removal
270	 */
271	this_cpu = get_cpu();
272
273	/*
274	 * Can deadlock when called with interrupts disabled.
275	 * We allow cpu's that are not yet online though, as no one else can
276	 * send smp call function interrupt to this cpu and as such deadlocks
277	 * can't happen.
278	 */
279	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
280		     && !oops_in_progress);
281
 
 
 
 
 
 
 
 
282	csd = &csd_stack;
283	if (!wait) {
284		csd = this_cpu_ptr(&csd_data);
285		csd_lock(csd);
286	}
287
288	err = generic_exec_single(cpu, csd, func, info);
 
 
 
 
 
 
 
289
290	if (wait)
291		csd_lock_wait(csd);
292
293	put_cpu();
294
295	return err;
296}
297EXPORT_SYMBOL(smp_call_function_single);
298
299/**
300 * smp_call_function_single_async(): Run an asynchronous function on a
301 * 			         specific CPU.
302 * @cpu: The CPU to run on.
303 * @csd: Pre-allocated and setup data structure
304 *
305 * Like smp_call_function_single(), but the call is asynchonous and
306 * can thus be done from contexts with disabled interrupts.
307 *
308 * The caller passes his own pre-allocated data structure
309 * (ie: embedded in an object) and is responsible for synchronizing it
310 * such that the IPIs performed on the @csd are strictly serialized.
311 *
 
 
 
 
 
312 * NOTE: Be careful, there is unfortunately no current debugging facility to
313 * validate the correctness of this serialization.
314 */
315int smp_call_function_single_async(int cpu, struct call_single_data *csd)
316{
317	int err = 0;
318
319	preempt_disable();
320
321	/* We could deadlock if we have to wait here with interrupts disabled! */
322	if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
323		csd_lock_wait(csd);
 
324
325	csd->flags = CSD_FLAG_LOCK;
326	smp_wmb();
327
328	err = generic_exec_single(cpu, csd, csd->func, csd->info);
 
 
329	preempt_enable();
330
331	return err;
332}
333EXPORT_SYMBOL_GPL(smp_call_function_single_async);
334
335/*
336 * smp_call_function_any - Run a function on any of the given cpus
337 * @mask: The mask of cpus it can run on.
338 * @func: The function to run. This must be fast and non-blocking.
339 * @info: An arbitrary pointer to pass to the function.
340 * @wait: If true, wait until function has completed.
341 *
342 * Returns 0 on success, else a negative status code (if no cpus were online).
343 *
344 * Selection preference:
345 *	1) current cpu if in @mask
346 *	2) any cpu of current node if in @mask
347 *	3) any other online cpu in @mask
348 */
349int smp_call_function_any(const struct cpumask *mask,
350			  smp_call_func_t func, void *info, int wait)
351{
352	unsigned int cpu;
353	const struct cpumask *nodemask;
354	int ret;
355
356	/* Try for same CPU (cheapest) */
357	cpu = get_cpu();
358	if (cpumask_test_cpu(cpu, mask))
359		goto call;
360
361	/* Try for same node. */
362	nodemask = cpumask_of_node(cpu_to_node(cpu));
363	for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
364	     cpu = cpumask_next_and(cpu, nodemask, mask)) {
365		if (cpu_online(cpu))
366			goto call;
367	}
368
369	/* Any online will do: smp_call_function_single handles nr_cpu_ids. */
370	cpu = cpumask_any_and(mask, cpu_online_mask);
371call:
372	ret = smp_call_function_single(cpu, func, info, wait);
373	put_cpu();
374	return ret;
375}
376EXPORT_SYMBOL_GPL(smp_call_function_any);
377
378/**
379 * smp_call_function_many(): Run a function on a set of other CPUs.
380 * @mask: The set of cpus to run on (only runs on online subset).
381 * @func: The function to run. This must be fast and non-blocking.
382 * @info: An arbitrary pointer to pass to the function.
383 * @wait: If true, wait (atomically) until function has completed
384 *        on other CPUs.
385 *
386 * If @wait is true, then returns once @func has returned.
387 *
388 * You must not call this function with disabled interrupts or from a
389 * hardware interrupt handler or from a bottom half handler. Preemption
390 * must be disabled when calling this function.
391 */
392void smp_call_function_many(const struct cpumask *mask,
393			    smp_call_func_t func, void *info, bool wait)
 
 
 
 
 
394{
 
395	struct call_function_data *cfd;
396	int cpu, next_cpu, this_cpu = smp_processor_id();
 
 
 
 
 
397
398	/*
399	 * Can deadlock when called with interrupts disabled.
400	 * We allow cpu's that are not yet online though, as no one else can
401	 * send smp call function interrupt to this cpu and as such deadlocks
402	 * can't happen.
403	 */
404	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
405		     && !oops_in_progress && !early_boot_irqs_disabled);
 
 
 
 
 
 
 
 
 
 
 
 
 
406
407	/* Try to fastpath.  So, what's a CPU they want? Ignoring this one. */
408	cpu = cpumask_first_and(mask, cpu_online_mask);
409	if (cpu == this_cpu)
410		cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
 
 
411
412	/* No online cpus?  We're done. */
413	if (cpu >= nr_cpu_ids)
414		return;
 
415
416	/* Do we have another CPU which isn't us? */
417	next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
418	if (next_cpu == this_cpu)
419		next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
420
421	/* Fastpath: do that cpu by itself. */
422	if (next_cpu >= nr_cpu_ids) {
423		smp_call_function_single(cpu, func, info, wait);
424		return;
425	}
426
427	cfd = this_cpu_ptr(&cfd_data);
 
428
429	cpumask_and(cfd->cpumask, mask, cpu_online_mask);
430	cpumask_clear_cpu(this_cpu, cfd->cpumask);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
431
432	/* Some callers race with other cpus changing the passed mask */
433	if (unlikely(!cpumask_weight(cfd->cpumask)))
434		return;
435
436	for_each_cpu(cpu, cfd->cpumask) {
437		struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
 
 
 
 
 
 
 
438
439		csd_lock(csd);
440		if (wait)
441			csd->flags |= CSD_FLAG_SYNCHRONOUS;
442		csd->func = func;
443		csd->info = info;
444		llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
445	}
446
447	/* Send a message to all CPUs in the map */
448	arch_send_call_function_ipi_mask(cfd->cpumask);
 
 
 
 
 
449
450	if (wait) {
451		for_each_cpu(cpu, cfd->cpumask) {
452			struct call_single_data *csd;
453
454			csd = per_cpu_ptr(cfd->csd, cpu);
455			csd_lock_wait(csd);
456		}
457	}
458}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
459EXPORT_SYMBOL(smp_call_function_many);
460
461/**
462 * smp_call_function(): Run a function on all other CPUs.
463 * @func: The function to run. This must be fast and non-blocking.
464 * @info: An arbitrary pointer to pass to the function.
465 * @wait: If true, wait (atomically) until function has completed
466 *        on other CPUs.
467 *
468 * Returns 0.
469 *
470 * If @wait is true, then returns once @func has returned; otherwise
471 * it returns just before the target cpu calls @func.
472 *
473 * You must not call this function with disabled interrupts or from a
474 * hardware interrupt handler or from a bottom half handler.
475 */
476int smp_call_function(smp_call_func_t func, void *info, int wait)
477{
478	preempt_disable();
479	smp_call_function_many(cpu_online_mask, func, info, wait);
480	preempt_enable();
481
482	return 0;
483}
484EXPORT_SYMBOL(smp_call_function);
485
486/* Setup configured maximum number of CPUs to activate */
487unsigned int setup_max_cpus = NR_CPUS;
488EXPORT_SYMBOL(setup_max_cpus);
489
490
491/*
492 * Setup routine for controlling SMP activation
493 *
494 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
495 * activation entirely (the MPS table probe still happens, though).
496 *
497 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
498 * greater than 0, limits the maximum number of CPUs activated in
499 * SMP mode to <NUM>.
500 */
501
502void __weak arch_disable_smp_support(void) { }
503
504static int __init nosmp(char *str)
505{
506	setup_max_cpus = 0;
507	arch_disable_smp_support();
508
509	return 0;
510}
511
512early_param("nosmp", nosmp);
513
514/* this is hard limit */
515static int __init nrcpus(char *str)
516{
517	int nr_cpus;
518
519	get_option(&str, &nr_cpus);
520	if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
521		nr_cpu_ids = nr_cpus;
522
523	return 0;
524}
525
526early_param("nr_cpus", nrcpus);
527
528static int __init maxcpus(char *str)
529{
530	get_option(&str, &setup_max_cpus);
531	if (setup_max_cpus == 0)
532		arch_disable_smp_support();
533
534	return 0;
535}
536
537early_param("maxcpus", maxcpus);
538
539/* Setup number of possible processor ids */
540int nr_cpu_ids __read_mostly = NR_CPUS;
541EXPORT_SYMBOL(nr_cpu_ids);
542
543/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
544void __init setup_nr_cpu_ids(void)
545{
546	nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
547}
548
549/* Called by boot processor to activate the rest. */
550void __init smp_init(void)
551{
552	int num_nodes, num_cpus;
553	unsigned int cpu;
554
555	idle_threads_init();
556	cpuhp_threads_init();
557
558	pr_info("Bringing up secondary CPUs ...\n");
559
560	/* FIXME: This should be done in userspace --RR */
561	for_each_present_cpu(cpu) {
562		if (num_online_cpus() >= setup_max_cpus)
563			break;
564		if (!cpu_online(cpu))
565			cpu_up(cpu);
566	}
567
568	num_nodes = num_online_nodes();
569	num_cpus  = num_online_cpus();
570	pr_info("Brought up %d node%s, %d CPU%s\n",
571		num_nodes, (num_nodes > 1 ? "s" : ""),
572		num_cpus,  (num_cpus  > 1 ? "s" : ""));
573
574	/* Any cleanup work */
575	smp_cpus_done(setup_max_cpus);
576}
577
578/*
579 * Call a function on all processors.  May be used during early boot while
580 * early_boot_irqs_disabled is set.  Use local_irq_save/restore() instead
581 * of local_irq_disable/enable().
582 */
583int on_each_cpu(void (*func) (void *info), void *info, int wait)
584{
585	unsigned long flags;
586	int ret = 0;
587
588	preempt_disable();
589	ret = smp_call_function(func, info, wait);
590	local_irq_save(flags);
591	func(info);
592	local_irq_restore(flags);
593	preempt_enable();
594	return ret;
595}
596EXPORT_SYMBOL(on_each_cpu);
597
598/**
599 * on_each_cpu_mask(): Run a function on processors specified by
600 * cpumask, which may include the local processor.
601 * @mask: The set of cpus to run on (only runs on online subset).
602 * @func: The function to run. This must be fast and non-blocking.
603 * @info: An arbitrary pointer to pass to the function.
604 * @wait: If true, wait (atomically) until function has completed
605 *        on other CPUs.
606 *
607 * If @wait is true, then returns once @func has returned.
608 *
609 * You must not call this function with disabled interrupts or from a
610 * hardware interrupt handler or from a bottom half handler.  The
611 * exception is that it may be used during early boot while
612 * early_boot_irqs_disabled is set.
613 */
614void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
615			void *info, bool wait)
616{
617	int cpu = get_cpu();
618
619	smp_call_function_many(mask, func, info, wait);
620	if (cpumask_test_cpu(cpu, mask)) {
621		unsigned long flags;
622		local_irq_save(flags);
623		func(info);
624		local_irq_restore(flags);
625	}
626	put_cpu();
627}
628EXPORT_SYMBOL(on_each_cpu_mask);
629
630/*
631 * on_each_cpu_cond(): Call a function on each processor for which
632 * the supplied function cond_func returns true, optionally waiting
633 * for all the required CPUs to finish. This may include the local
634 * processor.
635 * @cond_func:	A callback function that is passed a cpu id and
636 *		the the info parameter. The function is called
637 *		with preemption disabled. The function should
638 *		return a blooean value indicating whether to IPI
639 *		the specified CPU.
640 * @func:	The function to run on all applicable CPUs.
641 *		This must be fast and non-blocking.
642 * @info:	An arbitrary pointer to pass to both functions.
643 * @wait:	If true, wait (atomically) until function has
644 *		completed on other CPUs.
645 * @gfp_flags:	GFP flags to use when allocating the cpumask
646 *		used internally by the function.
647 *
648 * The function might sleep if the GFP flags indicates a non
649 * atomic allocation is allowed.
650 *
651 * Preemption is disabled to protect against CPUs going offline but not online.
652 * CPUs going online during the call will not be seen or sent an IPI.
653 *
654 * You must not call this function with disabled interrupts or
655 * from a hardware interrupt handler or from a bottom half handler.
656 */
657void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
658			smp_call_func_t func, void *info, bool wait,
659			gfp_t gfp_flags)
660{
661	cpumask_var_t cpus;
662	int cpu, ret;
663
664	might_sleep_if(gfpflags_allow_blocking(gfp_flags));
665
666	if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
667		preempt_disable();
668		for_each_online_cpu(cpu)
669			if (cond_func(cpu, info))
670				cpumask_set_cpu(cpu, cpus);
671		on_each_cpu_mask(cpus, func, info, wait);
672		preempt_enable();
673		free_cpumask_var(cpus);
674	} else {
675		/*
676		 * No free cpumask, bother. No matter, we'll
677		 * just have to IPI them one by one.
678		 */
679		preempt_disable();
680		for_each_online_cpu(cpu)
681			if (cond_func(cpu, info)) {
682				ret = smp_call_function_single(cpu, func,
683								info, wait);
684				WARN_ON_ONCE(ret);
685			}
686		preempt_enable();
687	}
688}
689EXPORT_SYMBOL(on_each_cpu_cond);
690
691static void do_nothing(void *unused)
692{
693}
694
695/**
696 * kick_all_cpus_sync - Force all cpus out of idle
697 *
698 * Used to synchronize the update of pm_idle function pointer. It's
699 * called after the pointer is updated and returns after the dummy
700 * callback function has been executed on all cpus. The execution of
701 * the function can only happen on the remote cpus after they have
702 * left the idle function which had been called via pm_idle function
703 * pointer. So it's guaranteed that nothing uses the previous pointer
704 * anymore.
705 */
706void kick_all_cpus_sync(void)
707{
708	/* Make sure the change is visible before we kick the cpus */
709	smp_mb();
710	smp_call_function(do_nothing, NULL, 1);
711}
712EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
713
714/**
715 * wake_up_all_idle_cpus - break all cpus out of idle
716 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
717 * including idle polling cpus, for non-idle cpus, we will do nothing
718 * for them.
719 */
720void wake_up_all_idle_cpus(void)
721{
722	int cpu;
723
724	preempt_disable();
725	for_each_online_cpu(cpu) {
726		if (cpu == smp_processor_id())
727			continue;
728
729		wake_up_if_idle(cpu);
730	}
731	preempt_enable();
732}
733EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
734
735/**
736 * smp_call_on_cpu - Call a function on a specific cpu
737 *
738 * Used to call a function on a specific cpu and wait for it to return.
739 * Optionally make sure the call is done on a specified physical cpu via vcpu
740 * pinning in order to support virtualized environments.
741 */
742struct smp_call_on_cpu_struct {
743	struct work_struct	work;
744	struct completion	done;
745	int			(*func)(void *);
746	void			*data;
747	int			ret;
748	int			cpu;
749};
750
751static void smp_call_on_cpu_callback(struct work_struct *work)
752{
753	struct smp_call_on_cpu_struct *sscs;
754
755	sscs = container_of(work, struct smp_call_on_cpu_struct, work);
756	if (sscs->cpu >= 0)
757		hypervisor_pin_vcpu(sscs->cpu);
758	sscs->ret = sscs->func(sscs->data);
759	if (sscs->cpu >= 0)
760		hypervisor_pin_vcpu(-1);
761
762	complete(&sscs->done);
763}
764
765int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
766{
767	struct smp_call_on_cpu_struct sscs = {
768		.done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
769		.func = func,
770		.data = par,
771		.cpu  = phys ? cpu : -1,
772	};
773
774	INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
775
776	if (cpu >= nr_cpu_ids || !cpu_online(cpu))
777		return -ENXIO;
778
779	queue_work_on(cpu, system_wq, &sscs.work);
780	wait_for_completion(&sscs.done);
781
782	return sscs.ret;
783}
784EXPORT_SYMBOL_GPL(smp_call_on_cpu);
v5.14.15
   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
  29#include "smpboot.h"
  30#include "sched/smp.h"
  31
  32#define CSD_TYPE(_csd)	((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK)
  33
  34#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
  35union cfd_seq_cnt {
  36	u64		val;
  37	struct {
  38		u64	src:16;
  39		u64	dst:16;
  40#define CFD_SEQ_NOCPU	0xffff
  41		u64	type:4;
  42#define CFD_SEQ_QUEUE	0
  43#define CFD_SEQ_IPI	1
  44#define CFD_SEQ_NOIPI	2
  45#define CFD_SEQ_PING	3
  46#define CFD_SEQ_PINGED	4
  47#define CFD_SEQ_HANDLE	5
  48#define CFD_SEQ_DEQUEUE	6
  49#define CFD_SEQ_IDLE	7
  50#define CFD_SEQ_GOTIPI	8
  51#define CFD_SEQ_HDLEND	9
  52		u64	cnt:28;
  53	}		u;
  54};
  55
  56static char *seq_type[] = {
  57	[CFD_SEQ_QUEUE]		= "queue",
  58	[CFD_SEQ_IPI]		= "ipi",
  59	[CFD_SEQ_NOIPI]		= "noipi",
  60	[CFD_SEQ_PING]		= "ping",
  61	[CFD_SEQ_PINGED]	= "pinged",
  62	[CFD_SEQ_HANDLE]	= "handle",
  63	[CFD_SEQ_DEQUEUE]	= "dequeue (src CPU 0 == empty)",
  64	[CFD_SEQ_IDLE]		= "idle",
  65	[CFD_SEQ_GOTIPI]	= "gotipi",
  66	[CFD_SEQ_HDLEND]	= "hdlend (src CPU 0 == early)",
  67};
  68
  69struct cfd_seq_local {
  70	u64	ping;
  71	u64	pinged;
  72	u64	handle;
  73	u64	dequeue;
  74	u64	idle;
  75	u64	gotipi;
  76	u64	hdlend;
  77};
  78#endif
  79
  80struct cfd_percpu {
  81	call_single_data_t	csd;
  82#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
  83	u64	seq_queue;
  84	u64	seq_ipi;
  85	u64	seq_noipi;
  86#endif
  87};
  88
  89struct call_function_data {
  90	struct cfd_percpu	__percpu *pcpu;
  91	cpumask_var_t		cpumask;
  92	cpumask_var_t		cpumask_ipi;
  93};
  94
  95static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
  96
  97static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
  98
  99static void flush_smp_call_function_queue(bool warn_cpu_offline);
 100
 101int smpcfd_prepare_cpu(unsigned int cpu)
 102{
 103	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
 104
 105	if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
 106				     cpu_to_node(cpu)))
 107		return -ENOMEM;
 108	if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
 109				     cpu_to_node(cpu))) {
 110		free_cpumask_var(cfd->cpumask);
 111		return -ENOMEM;
 112	}
 113	cfd->pcpu = alloc_percpu(struct cfd_percpu);
 114	if (!cfd->pcpu) {
 115		free_cpumask_var(cfd->cpumask);
 116		free_cpumask_var(cfd->cpumask_ipi);
 117		return -ENOMEM;
 118	}
 119
 120	return 0;
 121}
 122
 123int smpcfd_dead_cpu(unsigned int cpu)
 124{
 125	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
 126
 127	free_cpumask_var(cfd->cpumask);
 128	free_cpumask_var(cfd->cpumask_ipi);
 129	free_percpu(cfd->pcpu);
 130	return 0;
 131}
 132
 133int smpcfd_dying_cpu(unsigned int cpu)
 134{
 135	/*
 136	 * The IPIs for the smp-call-function callbacks queued by other
 137	 * CPUs might arrive late, either due to hardware latencies or
 138	 * because this CPU disabled interrupts (inside stop-machine)
 139	 * before the IPIs were sent. So flush out any pending callbacks
 140	 * explicitly (without waiting for the IPIs to arrive), to
 141	 * ensure that the outgoing CPU doesn't go offline with work
 142	 * still pending.
 143	 */
 144	flush_smp_call_function_queue(false);
 145	irq_work_run();
 146	return 0;
 147}
 148
 149void __init call_function_init(void)
 150{
 151	int i;
 152
 153	for_each_possible_cpu(i)
 154		init_llist_head(&per_cpu(call_single_queue, i));
 155
 156	smpcfd_prepare_cpu(smp_processor_id());
 157}
 158
 159#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 160
 161static DEFINE_STATIC_KEY_FALSE(csdlock_debug_enabled);
 162static DEFINE_STATIC_KEY_FALSE(csdlock_debug_extended);
 163
 164static int __init csdlock_debug(char *str)
 165{
 166	unsigned int val = 0;
 167
 168	if (str && !strcmp(str, "ext")) {
 169		val = 1;
 170		static_branch_enable(&csdlock_debug_extended);
 171	} else
 172		get_option(&str, &val);
 173
 174	if (val)
 175		static_branch_enable(&csdlock_debug_enabled);
 176
 177	return 0;
 178}
 179early_param("csdlock_debug", csdlock_debug);
 180
 181static DEFINE_PER_CPU(call_single_data_t *, cur_csd);
 182static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func);
 183static DEFINE_PER_CPU(void *, cur_csd_info);
 184static DEFINE_PER_CPU(struct cfd_seq_local, cfd_seq_local);
 185
 186#define CSD_LOCK_TIMEOUT (5ULL * NSEC_PER_SEC)
 187static atomic_t csd_bug_count = ATOMIC_INIT(0);
 188static u64 cfd_seq;
 189
 190#define CFD_SEQ(s, d, t, c)	\
 191	(union cfd_seq_cnt){ .u.src = s, .u.dst = d, .u.type = t, .u.cnt = c }
 192
 193static u64 cfd_seq_inc(unsigned int src, unsigned int dst, unsigned int type)
 194{
 195	union cfd_seq_cnt new, old;
 196
 197	new = CFD_SEQ(src, dst, type, 0);
 198
 199	do {
 200		old.val = READ_ONCE(cfd_seq);
 201		new.u.cnt = old.u.cnt + 1;
 202	} while (cmpxchg(&cfd_seq, old.val, new.val) != old.val);
 203
 204	return old.val;
 205}
 206
 207#define cfd_seq_store(var, src, dst, type)				\
 208	do {								\
 209		if (static_branch_unlikely(&csdlock_debug_extended))	\
 210			var = cfd_seq_inc(src, dst, type);		\
 211	} while (0)
 212
 213/* Record current CSD work for current CPU, NULL to erase. */
 214static void __csd_lock_record(struct __call_single_data *csd)
 215{
 216	if (!csd) {
 217		smp_mb(); /* NULL cur_csd after unlock. */
 218		__this_cpu_write(cur_csd, NULL);
 219		return;
 220	}
 221	__this_cpu_write(cur_csd_func, csd->func);
 222	__this_cpu_write(cur_csd_info, csd->info);
 223	smp_wmb(); /* func and info before csd. */
 224	__this_cpu_write(cur_csd, csd);
 225	smp_mb(); /* Update cur_csd before function call. */
 226		  /* Or before unlock, as the case may be. */
 227}
 228
 229static __always_inline void csd_lock_record(struct __call_single_data *csd)
 230{
 231	if (static_branch_unlikely(&csdlock_debug_enabled))
 232		__csd_lock_record(csd);
 233}
 234
 235static int csd_lock_wait_getcpu(struct __call_single_data *csd)
 236{
 237	unsigned int csd_type;
 238
 239	csd_type = CSD_TYPE(csd);
 240	if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC)
 241		return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */
 242	return -1;
 243}
 244
 245static void cfd_seq_data_add(u64 val, unsigned int src, unsigned int dst,
 246			     unsigned int type, union cfd_seq_cnt *data,
 247			     unsigned int *n_data, unsigned int now)
 248{
 249	union cfd_seq_cnt new[2];
 250	unsigned int i, j, k;
 251
 252	new[0].val = val;
 253	new[1] = CFD_SEQ(src, dst, type, new[0].u.cnt + 1);
 254
 255	for (i = 0; i < 2; i++) {
 256		if (new[i].u.cnt <= now)
 257			new[i].u.cnt |= 0x80000000U;
 258		for (j = 0; j < *n_data; j++) {
 259			if (new[i].u.cnt == data[j].u.cnt) {
 260				/* Direct read value trumps generated one. */
 261				if (i == 0)
 262					data[j].val = new[i].val;
 263				break;
 264			}
 265			if (new[i].u.cnt < data[j].u.cnt) {
 266				for (k = *n_data; k > j; k--)
 267					data[k].val = data[k - 1].val;
 268				data[j].val = new[i].val;
 269				(*n_data)++;
 270				break;
 271			}
 272		}
 273		if (j == *n_data) {
 274			data[j].val = new[i].val;
 275			(*n_data)++;
 276		}
 277	}
 278}
 279
 280static const char *csd_lock_get_type(unsigned int type)
 281{
 282	return (type >= ARRAY_SIZE(seq_type)) ? "?" : seq_type[type];
 283}
 284
 285static void csd_lock_print_extended(struct __call_single_data *csd, int cpu)
 286{
 287	struct cfd_seq_local *seq = &per_cpu(cfd_seq_local, cpu);
 288	unsigned int srccpu = csd->node.src;
 289	struct call_function_data *cfd = per_cpu_ptr(&cfd_data, srccpu);
 290	struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu);
 291	unsigned int now;
 292	union cfd_seq_cnt data[2 * ARRAY_SIZE(seq_type)];
 293	unsigned int n_data = 0, i;
 294
 295	data[0].val = READ_ONCE(cfd_seq);
 296	now = data[0].u.cnt;
 297
 298	cfd_seq_data_add(pcpu->seq_queue,			srccpu, cpu,	       CFD_SEQ_QUEUE,  data, &n_data, now);
 299	cfd_seq_data_add(pcpu->seq_ipi,				srccpu, cpu,	       CFD_SEQ_IPI,    data, &n_data, now);
 300	cfd_seq_data_add(pcpu->seq_noipi,			srccpu, cpu,	       CFD_SEQ_NOIPI,  data, &n_data, now);
 301
 302	cfd_seq_data_add(per_cpu(cfd_seq_local.ping, srccpu),	srccpu, CFD_SEQ_NOCPU, CFD_SEQ_PING,   data, &n_data, now);
 303	cfd_seq_data_add(per_cpu(cfd_seq_local.pinged, srccpu), srccpu, CFD_SEQ_NOCPU, CFD_SEQ_PINGED, data, &n_data, now);
 304
 305	cfd_seq_data_add(seq->idle,    CFD_SEQ_NOCPU, cpu, CFD_SEQ_IDLE,    data, &n_data, now);
 306	cfd_seq_data_add(seq->gotipi,  CFD_SEQ_NOCPU, cpu, CFD_SEQ_GOTIPI,  data, &n_data, now);
 307	cfd_seq_data_add(seq->handle,  CFD_SEQ_NOCPU, cpu, CFD_SEQ_HANDLE,  data, &n_data, now);
 308	cfd_seq_data_add(seq->dequeue, CFD_SEQ_NOCPU, cpu, CFD_SEQ_DEQUEUE, data, &n_data, now);
 309	cfd_seq_data_add(seq->hdlend,  CFD_SEQ_NOCPU, cpu, CFD_SEQ_HDLEND,  data, &n_data, now);
 310
 311	for (i = 0; i < n_data; i++) {
 312		pr_alert("\tcsd: cnt(%07x): %04x->%04x %s\n",
 313			 data[i].u.cnt & ~0x80000000U, data[i].u.src,
 314			 data[i].u.dst, csd_lock_get_type(data[i].u.type));
 315	}
 316	pr_alert("\tcsd: cnt now: %07x\n", now);
 317}
 318
 319/*
 320 * Complain if too much time spent waiting.  Note that only
 321 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
 322 * so waiting on other types gets much less information.
 323 */
 324static bool csd_lock_wait_toolong(struct __call_single_data *csd, u64 ts0, u64 *ts1, int *bug_id)
 325{
 326	int cpu = -1;
 327	int cpux;
 328	bool firsttime;
 329	u64 ts2, ts_delta;
 330	call_single_data_t *cpu_cur_csd;
 331	unsigned int flags = READ_ONCE(csd->node.u_flags);
 332
 333	if (!(flags & CSD_FLAG_LOCK)) {
 334		if (!unlikely(*bug_id))
 335			return true;
 336		cpu = csd_lock_wait_getcpu(csd);
 337		pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
 338			 *bug_id, raw_smp_processor_id(), cpu);
 339		return true;
 340	}
 341
 342	ts2 = sched_clock();
 343	ts_delta = ts2 - *ts1;
 344	if (likely(ts_delta <= CSD_LOCK_TIMEOUT))
 345		return false;
 346
 347	firsttime = !*bug_id;
 348	if (firsttime)
 349		*bug_id = atomic_inc_return(&csd_bug_count);
 350	cpu = csd_lock_wait_getcpu(csd);
 351	if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu))
 352		cpux = 0;
 353	else
 354		cpux = cpu;
 355	cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */
 356	pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n",
 357		 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0,
 358		 cpu, csd->func, csd->info);
 359	if (cpu_cur_csd && csd != cpu_cur_csd) {
 360		pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
 361			 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)),
 362			 READ_ONCE(per_cpu(cur_csd_info, cpux)));
 363	} else {
 364		pr_alert("\tcsd: CSD lock (#%d) %s.\n",
 365			 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request");
 366	}
 367	if (cpu >= 0) {
 368		if (static_branch_unlikely(&csdlock_debug_extended))
 369			csd_lock_print_extended(csd, cpu);
 370		if (!trigger_single_cpu_backtrace(cpu))
 371			dump_cpu_task(cpu);
 372		if (!cpu_cur_csd) {
 373			pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu);
 374			arch_send_call_function_single_ipi(cpu);
 375		}
 376	}
 377	dump_stack();
 378	*ts1 = ts2;
 379
 380	return false;
 381}
 382
 383/*
 384 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
 385 *
 386 * For non-synchronous ipi calls the csd can still be in use by the
 387 * previous function call. For multi-cpu calls its even more interesting
 388 * as we'll have to ensure no other cpu is observing our csd.
 389 */
 390static void __csd_lock_wait(struct __call_single_data *csd)
 391{
 392	int bug_id = 0;
 393	u64 ts0, ts1;
 394
 395	ts1 = ts0 = sched_clock();
 396	for (;;) {
 397		if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id))
 398			break;
 399		cpu_relax();
 400	}
 401	smp_acquire__after_ctrl_dep();
 402}
 403
 404static __always_inline void csd_lock_wait(struct __call_single_data *csd)
 405{
 406	if (static_branch_unlikely(&csdlock_debug_enabled)) {
 407		__csd_lock_wait(csd);
 408		return;
 409	}
 410
 411	smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
 412}
 413
 414static void __smp_call_single_queue_debug(int cpu, struct llist_node *node)
 415{
 416	unsigned int this_cpu = smp_processor_id();
 417	struct cfd_seq_local *seq = this_cpu_ptr(&cfd_seq_local);
 418	struct call_function_data *cfd = this_cpu_ptr(&cfd_data);
 419	struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu);
 420
 421	cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE);
 422	if (llist_add(node, &per_cpu(call_single_queue, cpu))) {
 423		cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI);
 424		cfd_seq_store(seq->ping, this_cpu, cpu, CFD_SEQ_PING);
 425		send_call_function_single_ipi(cpu);
 426		cfd_seq_store(seq->pinged, this_cpu, cpu, CFD_SEQ_PINGED);
 427	} else {
 428		cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI);
 429	}
 430}
 431#else
 432#define cfd_seq_store(var, src, dst, type)
 433
 434static void csd_lock_record(struct __call_single_data *csd)
 435{
 436}
 437
 438static __always_inline void csd_lock_wait(struct __call_single_data *csd)
 439{
 440	smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
 441}
 442#endif
 443
 444static __always_inline void csd_lock(struct __call_single_data *csd)
 445{
 446	csd_lock_wait(csd);
 447	csd->node.u_flags |= CSD_FLAG_LOCK;
 448
 449	/*
 450	 * prevent CPU from reordering the above assignment
 451	 * to ->flags with any subsequent assignments to other
 452	 * fields of the specified call_single_data_t structure:
 453	 */
 454	smp_wmb();
 455}
 456
 457static __always_inline void csd_unlock(struct __call_single_data *csd)
 458{
 459	WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK));
 460
 461	/*
 462	 * ensure we're all done before releasing data:
 463	 */
 464	smp_store_release(&csd->node.u_flags, 0);
 465}
 466
 467static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
 468
 469void __smp_call_single_queue(int cpu, struct llist_node *node)
 470{
 471#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 472	if (static_branch_unlikely(&csdlock_debug_extended)) {
 473		unsigned int type;
 474
 475		type = CSD_TYPE(container_of(node, call_single_data_t,
 476					     node.llist));
 477		if (type == CSD_TYPE_SYNC || type == CSD_TYPE_ASYNC) {
 478			__smp_call_single_queue_debug(cpu, node);
 479			return;
 480		}
 481	}
 482#endif
 483
 484	/*
 485	 * The list addition should be visible before sending the IPI
 486	 * handler locks the list to pull the entry off it because of
 487	 * normal cache coherency rules implied by spinlocks.
 488	 *
 489	 * If IPIs can go out of order to the cache coherency protocol
 490	 * in an architecture, sufficient synchronisation should be added
 491	 * to arch code to make it appear to obey cache coherency WRT
 492	 * locking and barrier primitives. Generic code isn't really
 493	 * equipped to do the right thing...
 494	 */
 495	if (llist_add(node, &per_cpu(call_single_queue, cpu)))
 496		send_call_function_single_ipi(cpu);
 497}
 498
 499/*
 500 * Insert a previously allocated call_single_data_t element
 501 * for execution on the given CPU. data must already have
 502 * ->func, ->info, and ->flags set.
 503 */
 504static int generic_exec_single(int cpu, struct __call_single_data *csd)
 
 505{
 506	if (cpu == smp_processor_id()) {
 507		smp_call_func_t func = csd->func;
 508		void *info = csd->info;
 509		unsigned long flags;
 510
 511		/*
 512		 * We can unlock early even for the synchronous on-stack case,
 513		 * since we're doing this from the same CPU..
 514		 */
 515		csd_lock_record(csd);
 516		csd_unlock(csd);
 517		local_irq_save(flags);
 518		func(info);
 519		csd_lock_record(NULL);
 520		local_irq_restore(flags);
 521		return 0;
 522	}
 523
 
 524	if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
 525		csd_unlock(csd);
 526		return -ENXIO;
 527	}
 528
 529	__smp_call_single_queue(cpu, &csd->node.llist);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 530
 531	return 0;
 532}
 533
 534/**
 535 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
 536 *
 537 * Invoked by arch to handle an IPI for call function single.
 538 * Must be called with interrupts disabled.
 539 */
 540void generic_smp_call_function_single_interrupt(void)
 541{
 542	cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->gotipi, CFD_SEQ_NOCPU,
 543		      smp_processor_id(), CFD_SEQ_GOTIPI);
 544	flush_smp_call_function_queue(true);
 545}
 546
 547/**
 548 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
 549 *
 550 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
 551 *		      offline CPU. Skip this check if set to 'false'.
 552 *
 553 * Flush any pending smp-call-function callbacks queued on this CPU. This is
 554 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
 555 * to ensure that all pending IPI callbacks are run before it goes completely
 556 * offline.
 557 *
 558 * Loop through the call_single_queue and run all the queued callbacks.
 559 * Must be called with interrupts disabled.
 560 */
 561static void flush_smp_call_function_queue(bool warn_cpu_offline)
 562{
 563	call_single_data_t *csd, *csd_next;
 564	struct llist_node *entry, *prev;
 565	struct llist_head *head;
 
 
 566	static bool warned;
 567
 568	lockdep_assert_irqs_disabled();
 569
 570	head = this_cpu_ptr(&call_single_queue);
 571	cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->handle, CFD_SEQ_NOCPU,
 572		      smp_processor_id(), CFD_SEQ_HANDLE);
 573	entry = llist_del_all(head);
 574	cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->dequeue,
 575		      /* Special meaning of source cpu: 0 == queue empty */
 576		      entry ? CFD_SEQ_NOCPU : 0,
 577		      smp_processor_id(), CFD_SEQ_DEQUEUE);
 578	entry = llist_reverse_order(entry);
 579
 580	/* There shouldn't be any pending callbacks on an offline CPU. */
 581	if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
 582		     !warned && !llist_empty(head))) {
 583		warned = true;
 584		WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
 585
 586		/*
 587		 * We don't have to use the _safe() variant here
 588		 * because we are not invoking the IPI handlers yet.
 589		 */
 590		llist_for_each_entry(csd, entry, node.llist) {
 591			switch (CSD_TYPE(csd)) {
 592			case CSD_TYPE_ASYNC:
 593			case CSD_TYPE_SYNC:
 594			case CSD_TYPE_IRQ_WORK:
 595				pr_warn("IPI callback %pS sent to offline CPU\n",
 596					csd->func);
 597				break;
 598
 599			case CSD_TYPE_TTWU:
 600				pr_warn("IPI task-wakeup sent to offline CPU\n");
 601				break;
 602
 603			default:
 604				pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
 605					CSD_TYPE(csd));
 606				break;
 607			}
 608		}
 609	}
 610
 611	/*
 612	 * First; run all SYNC callbacks, people are waiting for us.
 613	 */
 614	prev = NULL;
 615	llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
 616		/* Do we wait until *after* callback? */
 617		if (CSD_TYPE(csd) == CSD_TYPE_SYNC) {
 618			smp_call_func_t func = csd->func;
 619			void *info = csd->info;
 620
 621			if (prev) {
 622				prev->next = &csd_next->node.llist;
 623			} else {
 624				entry = &csd_next->node.llist;
 625			}
 626
 627			csd_lock_record(csd);
 628			func(info);
 629			csd_unlock(csd);
 630			csd_lock_record(NULL);
 631		} else {
 632			prev = &csd->node.llist;
 
 633		}
 634	}
 635
 636	if (!entry) {
 637		cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->hdlend,
 638			      0, smp_processor_id(),
 639			      CFD_SEQ_HDLEND);
 640		return;
 641	}
 642
 643	/*
 644	 * Second; run all !SYNC callbacks.
 
 
 
 645	 */
 646	prev = NULL;
 647	llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
 648		int type = CSD_TYPE(csd);
 649
 650		if (type != CSD_TYPE_TTWU) {
 651			if (prev) {
 652				prev->next = &csd_next->node.llist;
 653			} else {
 654				entry = &csd_next->node.llist;
 655			}
 656
 657			if (type == CSD_TYPE_ASYNC) {
 658				smp_call_func_t func = csd->func;
 659				void *info = csd->info;
 660
 661				csd_lock_record(csd);
 662				csd_unlock(csd);
 663				func(info);
 664				csd_lock_record(NULL);
 665			} else if (type == CSD_TYPE_IRQ_WORK) {
 666				irq_work_single(csd);
 667			}
 668
 669		} else {
 670			prev = &csd->node.llist;
 671		}
 672	}
 673
 674	/*
 675	 * Third; only CSD_TYPE_TTWU is left, issue those.
 676	 */
 677	if (entry)
 678		sched_ttwu_pending(entry);
 679
 680	cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->hdlend, CFD_SEQ_NOCPU,
 681		      smp_processor_id(), CFD_SEQ_HDLEND);
 682}
 683
 684void flush_smp_call_function_from_idle(void)
 685{
 686	unsigned long flags;
 687
 688	if (llist_empty(this_cpu_ptr(&call_single_queue)))
 689		return;
 690
 691	cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->idle, CFD_SEQ_NOCPU,
 692		      smp_processor_id(), CFD_SEQ_IDLE);
 693	local_irq_save(flags);
 694	flush_smp_call_function_queue(true);
 695	if (local_softirq_pending())
 696		do_softirq();
 697
 698	local_irq_restore(flags);
 699}
 700
 701/*
 702 * smp_call_function_single - Run a function on a specific CPU
 703 * @func: The function to run. This must be fast and non-blocking.
 704 * @info: An arbitrary pointer to pass to the function.
 705 * @wait: If true, wait until function has completed on other CPUs.
 706 *
 707 * Returns 0 on success, else a negative status code.
 708 */
 709int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
 710			     int wait)
 711{
 712	call_single_data_t *csd;
 713	call_single_data_t csd_stack = {
 714		.node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, },
 715	};
 716	int this_cpu;
 717	int err;
 718
 719	/*
 720	 * prevent preemption and reschedule on another processor,
 721	 * as well as CPU removal
 722	 */
 723	this_cpu = get_cpu();
 724
 725	/*
 726	 * Can deadlock when called with interrupts disabled.
 727	 * We allow cpu's that are not yet online though, as no one else can
 728	 * send smp call function interrupt to this cpu and as such deadlocks
 729	 * can't happen.
 730	 */
 731	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
 732		     && !oops_in_progress);
 733
 734	/*
 735	 * When @wait we can deadlock when we interrupt between llist_add() and
 736	 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
 737	 * csd_lock() on because the interrupt context uses the same csd
 738	 * storage.
 739	 */
 740	WARN_ON_ONCE(!in_task());
 741
 742	csd = &csd_stack;
 743	if (!wait) {
 744		csd = this_cpu_ptr(&csd_data);
 745		csd_lock(csd);
 746	}
 747
 748	csd->func = func;
 749	csd->info = info;
 750#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 751	csd->node.src = smp_processor_id();
 752	csd->node.dst = cpu;
 753#endif
 754
 755	err = generic_exec_single(cpu, csd);
 756
 757	if (wait)
 758		csd_lock_wait(csd);
 759
 760	put_cpu();
 761
 762	return err;
 763}
 764EXPORT_SYMBOL(smp_call_function_single);
 765
 766/**
 767 * smp_call_function_single_async(): Run an asynchronous function on a
 768 * 			         specific CPU.
 769 * @cpu: The CPU to run on.
 770 * @csd: Pre-allocated and setup data structure
 771 *
 772 * Like smp_call_function_single(), but the call is asynchonous and
 773 * can thus be done from contexts with disabled interrupts.
 774 *
 775 * The caller passes his own pre-allocated data structure
 776 * (ie: embedded in an object) and is responsible for synchronizing it
 777 * such that the IPIs performed on the @csd are strictly serialized.
 778 *
 779 * If the function is called with one csd which has not yet been
 780 * processed by previous call to smp_call_function_single_async(), the
 781 * function will return immediately with -EBUSY showing that the csd
 782 * object is still in progress.
 783 *
 784 * NOTE: Be careful, there is unfortunately no current debugging facility to
 785 * validate the correctness of this serialization.
 786 */
 787int smp_call_function_single_async(int cpu, struct __call_single_data *csd)
 788{
 789	int err = 0;
 790
 791	preempt_disable();
 792
 793	if (csd->node.u_flags & CSD_FLAG_LOCK) {
 794		err = -EBUSY;
 795		goto out;
 796	}
 797
 798	csd->node.u_flags = CSD_FLAG_LOCK;
 799	smp_wmb();
 800
 801	err = generic_exec_single(cpu, csd);
 802
 803out:
 804	preempt_enable();
 805
 806	return err;
 807}
 808EXPORT_SYMBOL_GPL(smp_call_function_single_async);
 809
 810/*
 811 * smp_call_function_any - Run a function on any of the given cpus
 812 * @mask: The mask of cpus it can run on.
 813 * @func: The function to run. This must be fast and non-blocking.
 814 * @info: An arbitrary pointer to pass to the function.
 815 * @wait: If true, wait until function has completed.
 816 *
 817 * Returns 0 on success, else a negative status code (if no cpus were online).
 818 *
 819 * Selection preference:
 820 *	1) current cpu if in @mask
 821 *	2) any cpu of current node if in @mask
 822 *	3) any other online cpu in @mask
 823 */
 824int smp_call_function_any(const struct cpumask *mask,
 825			  smp_call_func_t func, void *info, int wait)
 826{
 827	unsigned int cpu;
 828	const struct cpumask *nodemask;
 829	int ret;
 830
 831	/* Try for same CPU (cheapest) */
 832	cpu = get_cpu();
 833	if (cpumask_test_cpu(cpu, mask))
 834		goto call;
 835
 836	/* Try for same node. */
 837	nodemask = cpumask_of_node(cpu_to_node(cpu));
 838	for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
 839	     cpu = cpumask_next_and(cpu, nodemask, mask)) {
 840		if (cpu_online(cpu))
 841			goto call;
 842	}
 843
 844	/* Any online will do: smp_call_function_single handles nr_cpu_ids. */
 845	cpu = cpumask_any_and(mask, cpu_online_mask);
 846call:
 847	ret = smp_call_function_single(cpu, func, info, wait);
 848	put_cpu();
 849	return ret;
 850}
 851EXPORT_SYMBOL_GPL(smp_call_function_any);
 852
 853/*
 854 * Flags to be used as scf_flags argument of smp_call_function_many_cond().
 
 
 
 
 
 
 
 855 *
 856 * %SCF_WAIT:		Wait until function execution is completed
 857 * %SCF_RUN_LOCAL:	Run also locally if local cpu is set in cpumask
 
 858 */
 859#define SCF_WAIT	(1U << 0)
 860#define SCF_RUN_LOCAL	(1U << 1)
 861
 862static void smp_call_function_many_cond(const struct cpumask *mask,
 863					smp_call_func_t func, void *info,
 864					unsigned int scf_flags,
 865					smp_cond_func_t cond_func)
 866{
 867	int cpu, last_cpu, this_cpu = smp_processor_id();
 868	struct call_function_data *cfd;
 869	bool wait = scf_flags & SCF_WAIT;
 870	bool run_remote = false;
 871	bool run_local = false;
 872	int nr_cpus = 0;
 873
 874	lockdep_assert_preemption_disabled();
 875
 876	/*
 877	 * Can deadlock when called with interrupts disabled.
 878	 * We allow cpu's that are not yet online though, as no one else can
 879	 * send smp call function interrupt to this cpu and as such deadlocks
 880	 * can't happen.
 881	 */
 882	if (cpu_online(this_cpu) && !oops_in_progress &&
 883	    !early_boot_irqs_disabled)
 884		lockdep_assert_irqs_enabled();
 885
 886	/*
 887	 * When @wait we can deadlock when we interrupt between llist_add() and
 888	 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
 889	 * csd_lock() on because the interrupt context uses the same csd
 890	 * storage.
 891	 */
 892	WARN_ON_ONCE(!in_task());
 893
 894	/* Check if we need local execution. */
 895	if ((scf_flags & SCF_RUN_LOCAL) && cpumask_test_cpu(this_cpu, mask))
 896		run_local = true;
 897
 898	/* Check if we need remote execution, i.e., any CPU excluding this one. */
 899	cpu = cpumask_first_and(mask, cpu_online_mask);
 900	if (cpu == this_cpu)
 901		cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
 902	if (cpu < nr_cpu_ids)
 903		run_remote = true;
 904
 905	if (run_remote) {
 906		cfd = this_cpu_ptr(&cfd_data);
 907		cpumask_and(cfd->cpumask, mask, cpu_online_mask);
 908		__cpumask_clear_cpu(this_cpu, cfd->cpumask);
 909
 910		cpumask_clear(cfd->cpumask_ipi);
 911		for_each_cpu(cpu, cfd->cpumask) {
 912			struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu);
 913			call_single_data_t *csd = &pcpu->csd;
 
 
 
 
 
 
 914
 915			if (cond_func && !cond_func(cpu, info))
 916				continue;
 917
 918			csd_lock(csd);
 919			if (wait)
 920				csd->node.u_flags |= CSD_TYPE_SYNC;
 921			csd->func = func;
 922			csd->info = info;
 923#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
 924			csd->node.src = smp_processor_id();
 925			csd->node.dst = cpu;
 926#endif
 927			cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE);
 928			if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
 929				__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
 930				nr_cpus++;
 931				last_cpu = cpu;
 932
 933				cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI);
 934			} else {
 935				cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI);
 936			}
 937		}
 938
 939		cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->ping, this_cpu, CFD_SEQ_NOCPU, CFD_SEQ_PING);
 
 
 940
 941		/*
 942		 * Choose the most efficient way to send an IPI. Note that the
 943		 * number of CPUs might be zero due to concurrent changes to the
 944		 * provided mask.
 945		 */
 946		if (nr_cpus == 1)
 947			send_call_function_single_ipi(last_cpu);
 948		else if (likely(nr_cpus > 1))
 949			arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
 950
 951		cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->pinged, this_cpu, CFD_SEQ_NOCPU, CFD_SEQ_PINGED);
 
 
 
 
 
 952	}
 953
 954	if (run_local && (!cond_func || cond_func(this_cpu, info))) {
 955		unsigned long flags;
 956
 957		local_irq_save(flags);
 958		func(info);
 959		local_irq_restore(flags);
 960	}
 961
 962	if (run_remote && wait) {
 963		for_each_cpu(cpu, cfd->cpumask) {
 964			call_single_data_t *csd;
 965
 966			csd = &per_cpu_ptr(cfd->pcpu, cpu)->csd;
 967			csd_lock_wait(csd);
 968		}
 969	}
 970}
 971
 972/**
 973 * smp_call_function_many(): Run a function on a set of CPUs.
 974 * @mask: The set of cpus to run on (only runs on online subset).
 975 * @func: The function to run. This must be fast and non-blocking.
 976 * @info: An arbitrary pointer to pass to the function.
 977 * @flags: Bitmask that controls the operation. If %SCF_WAIT is set, wait
 978 *        (atomically) until function has completed on other CPUs. If
 979 *        %SCF_RUN_LOCAL is set, the function will also be run locally
 980 *        if the local CPU is set in the @cpumask.
 981 *
 982 * If @wait is true, then returns once @func has returned.
 983 *
 984 * You must not call this function with disabled interrupts or from a
 985 * hardware interrupt handler or from a bottom half handler. Preemption
 986 * must be disabled when calling this function.
 987 */
 988void smp_call_function_many(const struct cpumask *mask,
 989			    smp_call_func_t func, void *info, bool wait)
 990{
 991	smp_call_function_many_cond(mask, func, info, wait * SCF_WAIT, NULL);
 992}
 993EXPORT_SYMBOL(smp_call_function_many);
 994
 995/**
 996 * smp_call_function(): Run a function on all other CPUs.
 997 * @func: The function to run. This must be fast and non-blocking.
 998 * @info: An arbitrary pointer to pass to the function.
 999 * @wait: If true, wait (atomically) until function has completed
1000 *        on other CPUs.
1001 *
1002 * Returns 0.
1003 *
1004 * If @wait is true, then returns once @func has returned; otherwise
1005 * it returns just before the target cpu calls @func.
1006 *
1007 * You must not call this function with disabled interrupts or from a
1008 * hardware interrupt handler or from a bottom half handler.
1009 */
1010void smp_call_function(smp_call_func_t func, void *info, int wait)
1011{
1012	preempt_disable();
1013	smp_call_function_many(cpu_online_mask, func, info, wait);
1014	preempt_enable();
 
 
1015}
1016EXPORT_SYMBOL(smp_call_function);
1017
1018/* Setup configured maximum number of CPUs to activate */
1019unsigned int setup_max_cpus = NR_CPUS;
1020EXPORT_SYMBOL(setup_max_cpus);
1021
1022
1023/*
1024 * Setup routine for controlling SMP activation
1025 *
1026 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
1027 * activation entirely (the MPS table probe still happens, though).
1028 *
1029 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
1030 * greater than 0, limits the maximum number of CPUs activated in
1031 * SMP mode to <NUM>.
1032 */
1033
1034void __weak arch_disable_smp_support(void) { }
1035
1036static int __init nosmp(char *str)
1037{
1038	setup_max_cpus = 0;
1039	arch_disable_smp_support();
1040
1041	return 0;
1042}
1043
1044early_param("nosmp", nosmp);
1045
1046/* this is hard limit */
1047static int __init nrcpus(char *str)
1048{
1049	int nr_cpus;
1050
1051	if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids)
 
1052		nr_cpu_ids = nr_cpus;
1053
1054	return 0;
1055}
1056
1057early_param("nr_cpus", nrcpus);
1058
1059static int __init maxcpus(char *str)
1060{
1061	get_option(&str, &setup_max_cpus);
1062	if (setup_max_cpus == 0)
1063		arch_disable_smp_support();
1064
1065	return 0;
1066}
1067
1068early_param("maxcpus", maxcpus);
1069
1070/* Setup number of possible processor ids */
1071unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
1072EXPORT_SYMBOL(nr_cpu_ids);
1073
1074/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
1075void __init setup_nr_cpu_ids(void)
1076{
1077	nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
1078}
1079
1080/* Called by boot processor to activate the rest. */
1081void __init smp_init(void)
1082{
1083	int num_nodes, num_cpus;
 
1084
1085	idle_threads_init();
1086	cpuhp_threads_init();
1087
1088	pr_info("Bringing up secondary CPUs ...\n");
1089
1090	bringup_nonboot_cpus(setup_max_cpus);
 
 
 
 
 
 
1091
1092	num_nodes = num_online_nodes();
1093	num_cpus  = num_online_cpus();
1094	pr_info("Brought up %d node%s, %d CPU%s\n",
1095		num_nodes, (num_nodes > 1 ? "s" : ""),
1096		num_cpus,  (num_cpus  > 1 ? "s" : ""));
1097
1098	/* Any cleanup work */
1099	smp_cpus_done(setup_max_cpus);
1100}
1101
1102/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1103 * on_each_cpu_cond(): Call a function on each processor for which
1104 * the supplied function cond_func returns true, optionally waiting
1105 * for all the required CPUs to finish. This may include the local
1106 * processor.
1107 * @cond_func:	A callback function that is passed a cpu id and
1108 *		the info parameter. The function is called
1109 *		with preemption disabled. The function should
1110 *		return a blooean value indicating whether to IPI
1111 *		the specified CPU.
1112 * @func:	The function to run on all applicable CPUs.
1113 *		This must be fast and non-blocking.
1114 * @info:	An arbitrary pointer to pass to both functions.
1115 * @wait:	If true, wait (atomically) until function has
1116 *		completed on other CPUs.
 
 
 
 
 
1117 *
1118 * Preemption is disabled to protect against CPUs going offline but not online.
1119 * CPUs going online during the call will not be seen or sent an IPI.
1120 *
1121 * You must not call this function with disabled interrupts or
1122 * from a hardware interrupt handler or from a bottom half handler.
1123 */
1124void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func,
1125			   void *info, bool wait, const struct cpumask *mask)
1126{
1127	unsigned int scf_flags = SCF_RUN_LOCAL;
1128
1129	if (wait)
1130		scf_flags |= SCF_WAIT;
1131
1132	preempt_disable();
1133	smp_call_function_many_cond(mask, func, info, scf_flags, cond_func);
1134	preempt_enable();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1135}
1136EXPORT_SYMBOL(on_each_cpu_cond_mask);
1137
1138static void do_nothing(void *unused)
1139{
1140}
1141
1142/**
1143 * kick_all_cpus_sync - Force all cpus out of idle
1144 *
1145 * Used to synchronize the update of pm_idle function pointer. It's
1146 * called after the pointer is updated and returns after the dummy
1147 * callback function has been executed on all cpus. The execution of
1148 * the function can only happen on the remote cpus after they have
1149 * left the idle function which had been called via pm_idle function
1150 * pointer. So it's guaranteed that nothing uses the previous pointer
1151 * anymore.
1152 */
1153void kick_all_cpus_sync(void)
1154{
1155	/* Make sure the change is visible before we kick the cpus */
1156	smp_mb();
1157	smp_call_function(do_nothing, NULL, 1);
1158}
1159EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
1160
1161/**
1162 * wake_up_all_idle_cpus - break all cpus out of idle
1163 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
1164 * including idle polling cpus, for non-idle cpus, we will do nothing
1165 * for them.
1166 */
1167void wake_up_all_idle_cpus(void)
1168{
1169	int cpu;
1170
1171	preempt_disable();
1172	for_each_online_cpu(cpu) {
1173		if (cpu == smp_processor_id())
1174			continue;
1175
1176		wake_up_if_idle(cpu);
1177	}
1178	preempt_enable();
1179}
1180EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
1181
1182/**
1183 * smp_call_on_cpu - Call a function on a specific cpu
1184 *
1185 * Used to call a function on a specific cpu and wait for it to return.
1186 * Optionally make sure the call is done on a specified physical cpu via vcpu
1187 * pinning in order to support virtualized environments.
1188 */
1189struct smp_call_on_cpu_struct {
1190	struct work_struct	work;
1191	struct completion	done;
1192	int			(*func)(void *);
1193	void			*data;
1194	int			ret;
1195	int			cpu;
1196};
1197
1198static void smp_call_on_cpu_callback(struct work_struct *work)
1199{
1200	struct smp_call_on_cpu_struct *sscs;
1201
1202	sscs = container_of(work, struct smp_call_on_cpu_struct, work);
1203	if (sscs->cpu >= 0)
1204		hypervisor_pin_vcpu(sscs->cpu);
1205	sscs->ret = sscs->func(sscs->data);
1206	if (sscs->cpu >= 0)
1207		hypervisor_pin_vcpu(-1);
1208
1209	complete(&sscs->done);
1210}
1211
1212int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
1213{
1214	struct smp_call_on_cpu_struct sscs = {
1215		.done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
1216		.func = func,
1217		.data = par,
1218		.cpu  = phys ? cpu : -1,
1219	};
1220
1221	INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
1222
1223	if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1224		return -ENXIO;
1225
1226	queue_work_on(cpu, system_wq, &sscs.work);
1227	wait_for_completion(&sscs.done);
1228
1229	return sscs.ret;
1230}
1231EXPORT_SYMBOL_GPL(smp_call_on_cpu);