Loading...
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);
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/gfp.h>
18#include <linux/smp.h>
19#include <linux/cpu.h>
20#include <linux/sched.h>
21#include <linux/sched/idle.h>
22#include <linux/hypervisor.h>
23
24#include "smpboot.h"
25#include "sched/smp.h"
26
27#define CSD_TYPE(_csd) ((_csd)->flags & CSD_FLAG_TYPE_MASK)
28
29struct call_function_data {
30 call_single_data_t __percpu *csd;
31 cpumask_var_t cpumask;
32 cpumask_var_t cpumask_ipi;
33};
34
35static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
36
37static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
38
39static void flush_smp_call_function_queue(bool warn_cpu_offline);
40
41int smpcfd_prepare_cpu(unsigned int cpu)
42{
43 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
44
45 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
46 cpu_to_node(cpu)))
47 return -ENOMEM;
48 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
49 cpu_to_node(cpu))) {
50 free_cpumask_var(cfd->cpumask);
51 return -ENOMEM;
52 }
53 cfd->csd = alloc_percpu(call_single_data_t);
54 if (!cfd->csd) {
55 free_cpumask_var(cfd->cpumask);
56 free_cpumask_var(cfd->cpumask_ipi);
57 return -ENOMEM;
58 }
59
60 return 0;
61}
62
63int smpcfd_dead_cpu(unsigned int cpu)
64{
65 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
66
67 free_cpumask_var(cfd->cpumask);
68 free_cpumask_var(cfd->cpumask_ipi);
69 free_percpu(cfd->csd);
70 return 0;
71}
72
73int smpcfd_dying_cpu(unsigned int cpu)
74{
75 /*
76 * The IPIs for the smp-call-function callbacks queued by other
77 * CPUs might arrive late, either due to hardware latencies or
78 * because this CPU disabled interrupts (inside stop-machine)
79 * before the IPIs were sent. So flush out any pending callbacks
80 * explicitly (without waiting for the IPIs to arrive), to
81 * ensure that the outgoing CPU doesn't go offline with work
82 * still pending.
83 */
84 flush_smp_call_function_queue(false);
85 irq_work_run();
86 return 0;
87}
88
89void __init call_function_init(void)
90{
91 int i;
92
93 for_each_possible_cpu(i)
94 init_llist_head(&per_cpu(call_single_queue, i));
95
96 smpcfd_prepare_cpu(smp_processor_id());
97}
98
99/*
100 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
101 *
102 * For non-synchronous ipi calls the csd can still be in use by the
103 * previous function call. For multi-cpu calls its even more interesting
104 * as we'll have to ensure no other cpu is observing our csd.
105 */
106static __always_inline void csd_lock_wait(call_single_data_t *csd)
107{
108 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
109}
110
111static __always_inline void csd_lock(call_single_data_t *csd)
112{
113 csd_lock_wait(csd);
114 csd->flags |= CSD_FLAG_LOCK;
115
116 /*
117 * prevent CPU from reordering the above assignment
118 * to ->flags with any subsequent assignments to other
119 * fields of the specified call_single_data_t structure:
120 */
121 smp_wmb();
122}
123
124static __always_inline void csd_unlock(call_single_data_t *csd)
125{
126 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
127
128 /*
129 * ensure we're all done before releasing data:
130 */
131 smp_store_release(&csd->flags, 0);
132}
133
134static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
135
136void __smp_call_single_queue(int cpu, struct llist_node *node)
137{
138 /*
139 * The list addition should be visible before sending the IPI
140 * handler locks the list to pull the entry off it because of
141 * normal cache coherency rules implied by spinlocks.
142 *
143 * If IPIs can go out of order to the cache coherency protocol
144 * in an architecture, sufficient synchronisation should be added
145 * to arch code to make it appear to obey cache coherency WRT
146 * locking and barrier primitives. Generic code isn't really
147 * equipped to do the right thing...
148 */
149 if (llist_add(node, &per_cpu(call_single_queue, cpu)))
150 send_call_function_single_ipi(cpu);
151}
152
153/*
154 * Insert a previously allocated call_single_data_t element
155 * for execution on the given CPU. data must already have
156 * ->func, ->info, and ->flags set.
157 */
158static int generic_exec_single(int cpu, call_single_data_t *csd)
159{
160 if (cpu == smp_processor_id()) {
161 smp_call_func_t func = csd->func;
162 void *info = csd->info;
163 unsigned long flags;
164
165 /*
166 * We can unlock early even for the synchronous on-stack case,
167 * since we're doing this from the same CPU..
168 */
169 csd_unlock(csd);
170 local_irq_save(flags);
171 func(info);
172 local_irq_restore(flags);
173 return 0;
174 }
175
176 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
177 csd_unlock(csd);
178 return -ENXIO;
179 }
180
181 __smp_call_single_queue(cpu, &csd->llist);
182
183 return 0;
184}
185
186/**
187 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
188 *
189 * Invoked by arch to handle an IPI for call function single.
190 * Must be called with interrupts disabled.
191 */
192void generic_smp_call_function_single_interrupt(void)
193{
194 flush_smp_call_function_queue(true);
195}
196
197/**
198 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
199 *
200 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
201 * offline CPU. Skip this check if set to 'false'.
202 *
203 * Flush any pending smp-call-function callbacks queued on this CPU. This is
204 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
205 * to ensure that all pending IPI callbacks are run before it goes completely
206 * offline.
207 *
208 * Loop through the call_single_queue and run all the queued callbacks.
209 * Must be called with interrupts disabled.
210 */
211static void flush_smp_call_function_queue(bool warn_cpu_offline)
212{
213 call_single_data_t *csd, *csd_next;
214 struct llist_node *entry, *prev;
215 struct llist_head *head;
216 static bool warned;
217
218 lockdep_assert_irqs_disabled();
219
220 head = this_cpu_ptr(&call_single_queue);
221 entry = llist_del_all(head);
222 entry = llist_reverse_order(entry);
223
224 /* There shouldn't be any pending callbacks on an offline CPU. */
225 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
226 !warned && !llist_empty(head))) {
227 warned = true;
228 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
229
230 /*
231 * We don't have to use the _safe() variant here
232 * because we are not invoking the IPI handlers yet.
233 */
234 llist_for_each_entry(csd, entry, llist) {
235 switch (CSD_TYPE(csd)) {
236 case CSD_TYPE_ASYNC:
237 case CSD_TYPE_SYNC:
238 case CSD_TYPE_IRQ_WORK:
239 pr_warn("IPI callback %pS sent to offline CPU\n",
240 csd->func);
241 break;
242
243 case CSD_TYPE_TTWU:
244 pr_warn("IPI task-wakeup sent to offline CPU\n");
245 break;
246
247 default:
248 pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
249 CSD_TYPE(csd));
250 break;
251 }
252 }
253 }
254
255 /*
256 * First; run all SYNC callbacks, people are waiting for us.
257 */
258 prev = NULL;
259 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
260 /* Do we wait until *after* callback? */
261 if (CSD_TYPE(csd) == CSD_TYPE_SYNC) {
262 smp_call_func_t func = csd->func;
263 void *info = csd->info;
264
265 if (prev) {
266 prev->next = &csd_next->llist;
267 } else {
268 entry = &csd_next->llist;
269 }
270
271 func(info);
272 csd_unlock(csd);
273 } else {
274 prev = &csd->llist;
275 }
276 }
277
278 if (!entry)
279 return;
280
281 /*
282 * Second; run all !SYNC callbacks.
283 */
284 prev = NULL;
285 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
286 int type = CSD_TYPE(csd);
287
288 if (type != CSD_TYPE_TTWU) {
289 if (prev) {
290 prev->next = &csd_next->llist;
291 } else {
292 entry = &csd_next->llist;
293 }
294
295 if (type == CSD_TYPE_ASYNC) {
296 smp_call_func_t func = csd->func;
297 void *info = csd->info;
298
299 csd_unlock(csd);
300 func(info);
301 } else if (type == CSD_TYPE_IRQ_WORK) {
302 irq_work_single(csd);
303 }
304
305 } else {
306 prev = &csd->llist;
307 }
308 }
309
310 /*
311 * Third; only CSD_TYPE_TTWU is left, issue those.
312 */
313 if (entry)
314 sched_ttwu_pending(entry);
315}
316
317void flush_smp_call_function_from_idle(void)
318{
319 unsigned long flags;
320
321 if (llist_empty(this_cpu_ptr(&call_single_queue)))
322 return;
323
324 local_irq_save(flags);
325 flush_smp_call_function_queue(true);
326 local_irq_restore(flags);
327}
328
329/*
330 * smp_call_function_single - Run a function on a specific CPU
331 * @func: The function to run. This must be fast and non-blocking.
332 * @info: An arbitrary pointer to pass to the function.
333 * @wait: If true, wait until function has completed on other CPUs.
334 *
335 * Returns 0 on success, else a negative status code.
336 */
337int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
338 int wait)
339{
340 call_single_data_t *csd;
341 call_single_data_t csd_stack = {
342 .flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC,
343 };
344 int this_cpu;
345 int err;
346
347 /*
348 * prevent preemption and reschedule on another processor,
349 * as well as CPU removal
350 */
351 this_cpu = get_cpu();
352
353 /*
354 * Can deadlock when called with interrupts disabled.
355 * We allow cpu's that are not yet online though, as no one else can
356 * send smp call function interrupt to this cpu and as such deadlocks
357 * can't happen.
358 */
359 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
360 && !oops_in_progress);
361
362 /*
363 * When @wait we can deadlock when we interrupt between llist_add() and
364 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
365 * csd_lock() on because the interrupt context uses the same csd
366 * storage.
367 */
368 WARN_ON_ONCE(!in_task());
369
370 csd = &csd_stack;
371 if (!wait) {
372 csd = this_cpu_ptr(&csd_data);
373 csd_lock(csd);
374 }
375
376 csd->func = func;
377 csd->info = info;
378
379 err = generic_exec_single(cpu, csd);
380
381 if (wait)
382 csd_lock_wait(csd);
383
384 put_cpu();
385
386 return err;
387}
388EXPORT_SYMBOL(smp_call_function_single);
389
390/**
391 * smp_call_function_single_async(): Run an asynchronous function on a
392 * specific CPU.
393 * @cpu: The CPU to run on.
394 * @csd: Pre-allocated and setup data structure
395 *
396 * Like smp_call_function_single(), but the call is asynchonous and
397 * can thus be done from contexts with disabled interrupts.
398 *
399 * The caller passes his own pre-allocated data structure
400 * (ie: embedded in an object) and is responsible for synchronizing it
401 * such that the IPIs performed on the @csd are strictly serialized.
402 *
403 * If the function is called with one csd which has not yet been
404 * processed by previous call to smp_call_function_single_async(), the
405 * function will return immediately with -EBUSY showing that the csd
406 * object is still in progress.
407 *
408 * NOTE: Be careful, there is unfortunately no current debugging facility to
409 * validate the correctness of this serialization.
410 */
411int smp_call_function_single_async(int cpu, call_single_data_t *csd)
412{
413 int err = 0;
414
415 preempt_disable();
416
417 if (csd->flags & CSD_FLAG_LOCK) {
418 err = -EBUSY;
419 goto out;
420 }
421
422 csd->flags = CSD_FLAG_LOCK;
423 smp_wmb();
424
425 err = generic_exec_single(cpu, csd);
426
427out:
428 preempt_enable();
429
430 return err;
431}
432EXPORT_SYMBOL_GPL(smp_call_function_single_async);
433
434/*
435 * smp_call_function_any - Run a function on any of the given cpus
436 * @mask: The mask of cpus it can run on.
437 * @func: The function to run. This must be fast and non-blocking.
438 * @info: An arbitrary pointer to pass to the function.
439 * @wait: If true, wait until function has completed.
440 *
441 * Returns 0 on success, else a negative status code (if no cpus were online).
442 *
443 * Selection preference:
444 * 1) current cpu if in @mask
445 * 2) any cpu of current node if in @mask
446 * 3) any other online cpu in @mask
447 */
448int smp_call_function_any(const struct cpumask *mask,
449 smp_call_func_t func, void *info, int wait)
450{
451 unsigned int cpu;
452 const struct cpumask *nodemask;
453 int ret;
454
455 /* Try for same CPU (cheapest) */
456 cpu = get_cpu();
457 if (cpumask_test_cpu(cpu, mask))
458 goto call;
459
460 /* Try for same node. */
461 nodemask = cpumask_of_node(cpu_to_node(cpu));
462 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
463 cpu = cpumask_next_and(cpu, nodemask, mask)) {
464 if (cpu_online(cpu))
465 goto call;
466 }
467
468 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
469 cpu = cpumask_any_and(mask, cpu_online_mask);
470call:
471 ret = smp_call_function_single(cpu, func, info, wait);
472 put_cpu();
473 return ret;
474}
475EXPORT_SYMBOL_GPL(smp_call_function_any);
476
477static void smp_call_function_many_cond(const struct cpumask *mask,
478 smp_call_func_t func, void *info,
479 bool wait, smp_cond_func_t cond_func)
480{
481 struct call_function_data *cfd;
482 int cpu, next_cpu, this_cpu = smp_processor_id();
483
484 /*
485 * Can deadlock when called with interrupts disabled.
486 * We allow cpu's that are not yet online though, as no one else can
487 * send smp call function interrupt to this cpu and as such deadlocks
488 * can't happen.
489 */
490 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
491 && !oops_in_progress && !early_boot_irqs_disabled);
492
493 /*
494 * When @wait we can deadlock when we interrupt between llist_add() and
495 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
496 * csd_lock() on because the interrupt context uses the same csd
497 * storage.
498 */
499 WARN_ON_ONCE(!in_task());
500
501 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
502 cpu = cpumask_first_and(mask, cpu_online_mask);
503 if (cpu == this_cpu)
504 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
505
506 /* No online cpus? We're done. */
507 if (cpu >= nr_cpu_ids)
508 return;
509
510 /* Do we have another CPU which isn't us? */
511 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
512 if (next_cpu == this_cpu)
513 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
514
515 /* Fastpath: do that cpu by itself. */
516 if (next_cpu >= nr_cpu_ids) {
517 if (!cond_func || cond_func(cpu, info))
518 smp_call_function_single(cpu, func, info, wait);
519 return;
520 }
521
522 cfd = this_cpu_ptr(&cfd_data);
523
524 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
525 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
526
527 /* Some callers race with other cpus changing the passed mask */
528 if (unlikely(!cpumask_weight(cfd->cpumask)))
529 return;
530
531 cpumask_clear(cfd->cpumask_ipi);
532 for_each_cpu(cpu, cfd->cpumask) {
533 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
534
535 if (cond_func && !cond_func(cpu, info))
536 continue;
537
538 csd_lock(csd);
539 if (wait)
540 csd->flags |= CSD_TYPE_SYNC;
541 csd->func = func;
542 csd->info = info;
543 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
544 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
545 }
546
547 /* Send a message to all CPUs in the map */
548 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
549
550 if (wait) {
551 for_each_cpu(cpu, cfd->cpumask) {
552 call_single_data_t *csd;
553
554 csd = per_cpu_ptr(cfd->csd, cpu);
555 csd_lock_wait(csd);
556 }
557 }
558}
559
560/**
561 * smp_call_function_many(): Run a function on a set of other CPUs.
562 * @mask: The set of cpus to run on (only runs on online subset).
563 * @func: The function to run. This must be fast and non-blocking.
564 * @info: An arbitrary pointer to pass to the function.
565 * @wait: If true, wait (atomically) until function has completed
566 * on other CPUs.
567 *
568 * If @wait is true, then returns once @func has returned.
569 *
570 * You must not call this function with disabled interrupts or from a
571 * hardware interrupt handler or from a bottom half handler. Preemption
572 * must be disabled when calling this function.
573 */
574void smp_call_function_many(const struct cpumask *mask,
575 smp_call_func_t func, void *info, bool wait)
576{
577 smp_call_function_many_cond(mask, func, info, wait, NULL);
578}
579EXPORT_SYMBOL(smp_call_function_many);
580
581/**
582 * smp_call_function(): Run a function on all other CPUs.
583 * @func: The function to run. This must be fast and non-blocking.
584 * @info: An arbitrary pointer to pass to the function.
585 * @wait: If true, wait (atomically) until function has completed
586 * on other CPUs.
587 *
588 * Returns 0.
589 *
590 * If @wait is true, then returns once @func has returned; otherwise
591 * it returns just before the target cpu calls @func.
592 *
593 * You must not call this function with disabled interrupts or from a
594 * hardware interrupt handler or from a bottom half handler.
595 */
596void smp_call_function(smp_call_func_t func, void *info, int wait)
597{
598 preempt_disable();
599 smp_call_function_many(cpu_online_mask, func, info, wait);
600 preempt_enable();
601}
602EXPORT_SYMBOL(smp_call_function);
603
604/* Setup configured maximum number of CPUs to activate */
605unsigned int setup_max_cpus = NR_CPUS;
606EXPORT_SYMBOL(setup_max_cpus);
607
608
609/*
610 * Setup routine for controlling SMP activation
611 *
612 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
613 * activation entirely (the MPS table probe still happens, though).
614 *
615 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
616 * greater than 0, limits the maximum number of CPUs activated in
617 * SMP mode to <NUM>.
618 */
619
620void __weak arch_disable_smp_support(void) { }
621
622static int __init nosmp(char *str)
623{
624 setup_max_cpus = 0;
625 arch_disable_smp_support();
626
627 return 0;
628}
629
630early_param("nosmp", nosmp);
631
632/* this is hard limit */
633static int __init nrcpus(char *str)
634{
635 int nr_cpus;
636
637 if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids)
638 nr_cpu_ids = nr_cpus;
639
640 return 0;
641}
642
643early_param("nr_cpus", nrcpus);
644
645static int __init maxcpus(char *str)
646{
647 get_option(&str, &setup_max_cpus);
648 if (setup_max_cpus == 0)
649 arch_disable_smp_support();
650
651 return 0;
652}
653
654early_param("maxcpus", maxcpus);
655
656/* Setup number of possible processor ids */
657unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
658EXPORT_SYMBOL(nr_cpu_ids);
659
660/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
661void __init setup_nr_cpu_ids(void)
662{
663 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
664}
665
666/* Called by boot processor to activate the rest. */
667void __init smp_init(void)
668{
669 int num_nodes, num_cpus;
670
671 idle_threads_init();
672 cpuhp_threads_init();
673
674 pr_info("Bringing up secondary CPUs ...\n");
675
676 bringup_nonboot_cpus(setup_max_cpus);
677
678 num_nodes = num_online_nodes();
679 num_cpus = num_online_cpus();
680 pr_info("Brought up %d node%s, %d CPU%s\n",
681 num_nodes, (num_nodes > 1 ? "s" : ""),
682 num_cpus, (num_cpus > 1 ? "s" : ""));
683
684 /* Any cleanup work */
685 smp_cpus_done(setup_max_cpus);
686}
687
688/*
689 * Call a function on all processors. May be used during early boot while
690 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
691 * of local_irq_disable/enable().
692 */
693void on_each_cpu(smp_call_func_t func, void *info, int wait)
694{
695 unsigned long flags;
696
697 preempt_disable();
698 smp_call_function(func, info, wait);
699 local_irq_save(flags);
700 func(info);
701 local_irq_restore(flags);
702 preempt_enable();
703}
704EXPORT_SYMBOL(on_each_cpu);
705
706/**
707 * on_each_cpu_mask(): Run a function on processors specified by
708 * cpumask, which may include the local processor.
709 * @mask: The set of cpus to run on (only runs on online subset).
710 * @func: The function to run. This must be fast and non-blocking.
711 * @info: An arbitrary pointer to pass to the function.
712 * @wait: If true, wait (atomically) until function has completed
713 * on other CPUs.
714 *
715 * If @wait is true, then returns once @func has returned.
716 *
717 * You must not call this function with disabled interrupts or from a
718 * hardware interrupt handler or from a bottom half handler. The
719 * exception is that it may be used during early boot while
720 * early_boot_irqs_disabled is set.
721 */
722void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
723 void *info, bool wait)
724{
725 int cpu = get_cpu();
726
727 smp_call_function_many(mask, func, info, wait);
728 if (cpumask_test_cpu(cpu, mask)) {
729 unsigned long flags;
730 local_irq_save(flags);
731 func(info);
732 local_irq_restore(flags);
733 }
734 put_cpu();
735}
736EXPORT_SYMBOL(on_each_cpu_mask);
737
738/*
739 * on_each_cpu_cond(): Call a function on each processor for which
740 * the supplied function cond_func returns true, optionally waiting
741 * for all the required CPUs to finish. This may include the local
742 * processor.
743 * @cond_func: A callback function that is passed a cpu id and
744 * the the info parameter. The function is called
745 * with preemption disabled. The function should
746 * return a blooean value indicating whether to IPI
747 * the specified CPU.
748 * @func: The function to run on all applicable CPUs.
749 * This must be fast and non-blocking.
750 * @info: An arbitrary pointer to pass to both functions.
751 * @wait: If true, wait (atomically) until function has
752 * completed on other CPUs.
753 *
754 * Preemption is disabled to protect against CPUs going offline but not online.
755 * CPUs going online during the call will not be seen or sent an IPI.
756 *
757 * You must not call this function with disabled interrupts or
758 * from a hardware interrupt handler or from a bottom half handler.
759 */
760void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func,
761 void *info, bool wait, const struct cpumask *mask)
762{
763 int cpu = get_cpu();
764
765 smp_call_function_many_cond(mask, func, info, wait, cond_func);
766 if (cpumask_test_cpu(cpu, mask) && cond_func(cpu, info)) {
767 unsigned long flags;
768
769 local_irq_save(flags);
770 func(info);
771 local_irq_restore(flags);
772 }
773 put_cpu();
774}
775EXPORT_SYMBOL(on_each_cpu_cond_mask);
776
777void on_each_cpu_cond(smp_cond_func_t cond_func, smp_call_func_t func,
778 void *info, bool wait)
779{
780 on_each_cpu_cond_mask(cond_func, func, info, wait, cpu_online_mask);
781}
782EXPORT_SYMBOL(on_each_cpu_cond);
783
784static void do_nothing(void *unused)
785{
786}
787
788/**
789 * kick_all_cpus_sync - Force all cpus out of idle
790 *
791 * Used to synchronize the update of pm_idle function pointer. It's
792 * called after the pointer is updated and returns after the dummy
793 * callback function has been executed on all cpus. The execution of
794 * the function can only happen on the remote cpus after they have
795 * left the idle function which had been called via pm_idle function
796 * pointer. So it's guaranteed that nothing uses the previous pointer
797 * anymore.
798 */
799void kick_all_cpus_sync(void)
800{
801 /* Make sure the change is visible before we kick the cpus */
802 smp_mb();
803 smp_call_function(do_nothing, NULL, 1);
804}
805EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
806
807/**
808 * wake_up_all_idle_cpus - break all cpus out of idle
809 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
810 * including idle polling cpus, for non-idle cpus, we will do nothing
811 * for them.
812 */
813void wake_up_all_idle_cpus(void)
814{
815 int cpu;
816
817 preempt_disable();
818 for_each_online_cpu(cpu) {
819 if (cpu == smp_processor_id())
820 continue;
821
822 wake_up_if_idle(cpu);
823 }
824 preempt_enable();
825}
826EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
827
828/**
829 * smp_call_on_cpu - Call a function on a specific cpu
830 *
831 * Used to call a function on a specific cpu and wait for it to return.
832 * Optionally make sure the call is done on a specified physical cpu via vcpu
833 * pinning in order to support virtualized environments.
834 */
835struct smp_call_on_cpu_struct {
836 struct work_struct work;
837 struct completion done;
838 int (*func)(void *);
839 void *data;
840 int ret;
841 int cpu;
842};
843
844static void smp_call_on_cpu_callback(struct work_struct *work)
845{
846 struct smp_call_on_cpu_struct *sscs;
847
848 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
849 if (sscs->cpu >= 0)
850 hypervisor_pin_vcpu(sscs->cpu);
851 sscs->ret = sscs->func(sscs->data);
852 if (sscs->cpu >= 0)
853 hypervisor_pin_vcpu(-1);
854
855 complete(&sscs->done);
856}
857
858int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
859{
860 struct smp_call_on_cpu_struct sscs = {
861 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
862 .func = func,
863 .data = par,
864 .cpu = phys ? cpu : -1,
865 };
866
867 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
868
869 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
870 return -ENXIO;
871
872 queue_work_on(cpu, system_wq, &sscs.work);
873 wait_for_completion(&sscs.done);
874
875 return sscs.ret;
876}
877EXPORT_SYMBOL_GPL(smp_call_on_cpu);