Loading...
1/*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
5 */
6#include <linux/rcupdate.h>
7#include <linux/rculist.h>
8#include <linux/kernel.h>
9#include <linux/export.h>
10#include <linux/percpu.h>
11#include <linux/init.h>
12#include <linux/gfp.h>
13#include <linux/smp.h>
14#include <linux/cpu.h>
15
16#include "smpboot.h"
17
18#ifdef CONFIG_USE_GENERIC_SMP_HELPERS
19static struct {
20 struct list_head queue;
21 raw_spinlock_t lock;
22} call_function __cacheline_aligned_in_smp =
23 {
24 .queue = LIST_HEAD_INIT(call_function.queue),
25 .lock = __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
26 };
27
28enum {
29 CSD_FLAG_LOCK = 0x01,
30};
31
32struct call_function_data {
33 struct call_single_data csd;
34 atomic_t refs;
35 cpumask_var_t cpumask;
36};
37
38static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
39
40struct call_single_queue {
41 struct list_head list;
42 raw_spinlock_t lock;
43};
44
45static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
46
47static int
48hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
49{
50 long cpu = (long)hcpu;
51 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
52
53 switch (action) {
54 case CPU_UP_PREPARE:
55 case CPU_UP_PREPARE_FROZEN:
56 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
57 cpu_to_node(cpu)))
58 return notifier_from_errno(-ENOMEM);
59 break;
60
61#ifdef CONFIG_HOTPLUG_CPU
62 case CPU_UP_CANCELED:
63 case CPU_UP_CANCELED_FROZEN:
64
65 case CPU_DEAD:
66 case CPU_DEAD_FROZEN:
67 free_cpumask_var(cfd->cpumask);
68 break;
69#endif
70 };
71
72 return NOTIFY_OK;
73}
74
75static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
76 .notifier_call = hotplug_cfd,
77};
78
79void __init call_function_init(void)
80{
81 void *cpu = (void *)(long)smp_processor_id();
82 int i;
83
84 for_each_possible_cpu(i) {
85 struct call_single_queue *q = &per_cpu(call_single_queue, i);
86
87 raw_spin_lock_init(&q->lock);
88 INIT_LIST_HEAD(&q->list);
89 }
90
91 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
92 register_cpu_notifier(&hotplug_cfd_notifier);
93}
94
95/*
96 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
97 *
98 * For non-synchronous ipi calls the csd can still be in use by the
99 * previous function call. For multi-cpu calls its even more interesting
100 * as we'll have to ensure no other cpu is observing our csd.
101 */
102static void csd_lock_wait(struct call_single_data *data)
103{
104 while (data->flags & CSD_FLAG_LOCK)
105 cpu_relax();
106}
107
108static void csd_lock(struct call_single_data *data)
109{
110 csd_lock_wait(data);
111 data->flags = CSD_FLAG_LOCK;
112
113 /*
114 * prevent CPU from reordering the above assignment
115 * to ->flags with any subsequent assignments to other
116 * fields of the specified call_single_data structure:
117 */
118 smp_mb();
119}
120
121static void csd_unlock(struct call_single_data *data)
122{
123 WARN_ON(!(data->flags & CSD_FLAG_LOCK));
124
125 /*
126 * ensure we're all done before releasing data:
127 */
128 smp_mb();
129
130 data->flags &= ~CSD_FLAG_LOCK;
131}
132
133/*
134 * Insert a previously allocated call_single_data element
135 * for execution on the given CPU. data must already have
136 * ->func, ->info, and ->flags set.
137 */
138static
139void generic_exec_single(int cpu, struct call_single_data *data, int wait)
140{
141 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
142 unsigned long flags;
143 int ipi;
144
145 raw_spin_lock_irqsave(&dst->lock, flags);
146 ipi = list_empty(&dst->list);
147 list_add_tail(&data->list, &dst->list);
148 raw_spin_unlock_irqrestore(&dst->lock, flags);
149
150 /*
151 * The list addition should be visible before sending the IPI
152 * handler locks the list to pull the entry off it because of
153 * normal cache coherency rules implied by spinlocks.
154 *
155 * If IPIs can go out of order to the cache coherency protocol
156 * in an architecture, sufficient synchronisation should be added
157 * to arch code to make it appear to obey cache coherency WRT
158 * locking and barrier primitives. Generic code isn't really
159 * equipped to do the right thing...
160 */
161 if (ipi)
162 arch_send_call_function_single_ipi(cpu);
163
164 if (wait)
165 csd_lock_wait(data);
166}
167
168/*
169 * Invoked by arch to handle an IPI for call function. Must be called with
170 * interrupts disabled.
171 */
172void generic_smp_call_function_interrupt(void)
173{
174 struct call_function_data *data;
175 int cpu = smp_processor_id();
176
177 /*
178 * Shouldn't receive this interrupt on a cpu that is not yet online.
179 */
180 WARN_ON_ONCE(!cpu_online(cpu));
181
182 /*
183 * Ensure entry is visible on call_function_queue after we have
184 * entered the IPI. See comment in smp_call_function_many.
185 * If we don't have this, then we may miss an entry on the list
186 * and never get another IPI to process it.
187 */
188 smp_mb();
189
190 /*
191 * It's ok to use list_for_each_rcu() here even though we may
192 * delete 'pos', since list_del_rcu() doesn't clear ->next
193 */
194 list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
195 int refs;
196 smp_call_func_t func;
197
198 /*
199 * Since we walk the list without any locks, we might
200 * see an entry that was completed, removed from the
201 * list and is in the process of being reused.
202 *
203 * We must check that the cpu is in the cpumask before
204 * checking the refs, and both must be set before
205 * executing the callback on this cpu.
206 */
207
208 if (!cpumask_test_cpu(cpu, data->cpumask))
209 continue;
210
211 smp_rmb();
212
213 if (atomic_read(&data->refs) == 0)
214 continue;
215
216 func = data->csd.func; /* save for later warn */
217 func(data->csd.info);
218
219 /*
220 * If the cpu mask is not still set then func enabled
221 * interrupts (BUG), and this cpu took another smp call
222 * function interrupt and executed func(info) twice
223 * on this cpu. That nested execution decremented refs.
224 */
225 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask)) {
226 WARN(1, "%pf enabled interrupts and double executed\n", func);
227 continue;
228 }
229
230 refs = atomic_dec_return(&data->refs);
231 WARN_ON(refs < 0);
232
233 if (refs)
234 continue;
235
236 WARN_ON(!cpumask_empty(data->cpumask));
237
238 raw_spin_lock(&call_function.lock);
239 list_del_rcu(&data->csd.list);
240 raw_spin_unlock(&call_function.lock);
241
242 csd_unlock(&data->csd);
243 }
244
245}
246
247/*
248 * Invoked by arch to handle an IPI for call function single. Must be
249 * called from the arch with interrupts disabled.
250 */
251void generic_smp_call_function_single_interrupt(void)
252{
253 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
254 unsigned int data_flags;
255 LIST_HEAD(list);
256
257 /*
258 * Shouldn't receive this interrupt on a cpu that is not yet online.
259 */
260 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
261
262 raw_spin_lock(&q->lock);
263 list_replace_init(&q->list, &list);
264 raw_spin_unlock(&q->lock);
265
266 while (!list_empty(&list)) {
267 struct call_single_data *data;
268
269 data = list_entry(list.next, struct call_single_data, list);
270 list_del(&data->list);
271
272 /*
273 * 'data' can be invalid after this call if flags == 0
274 * (when called through generic_exec_single()),
275 * so save them away before making the call:
276 */
277 data_flags = data->flags;
278
279 data->func(data->info);
280
281 /*
282 * Unlocked CSDs are valid through generic_exec_single():
283 */
284 if (data_flags & CSD_FLAG_LOCK)
285 csd_unlock(data);
286 }
287}
288
289static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
290
291/*
292 * smp_call_function_single - Run a function on a specific CPU
293 * @func: The function to run. This must be fast and non-blocking.
294 * @info: An arbitrary pointer to pass to the function.
295 * @wait: If true, wait until function has completed on other CPUs.
296 *
297 * Returns 0 on success, else a negative status code.
298 */
299int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
300 int wait)
301{
302 struct call_single_data d = {
303 .flags = 0,
304 };
305 unsigned long flags;
306 int this_cpu;
307 int err = 0;
308
309 /*
310 * prevent preemption and reschedule on another processor,
311 * as well as CPU removal
312 */
313 this_cpu = get_cpu();
314
315 /*
316 * Can deadlock when called with interrupts disabled.
317 * We allow cpu's that are not yet online though, as no one else can
318 * send smp call function interrupt to this cpu and as such deadlocks
319 * can't happen.
320 */
321 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
322 && !oops_in_progress);
323
324 if (cpu == this_cpu) {
325 local_irq_save(flags);
326 func(info);
327 local_irq_restore(flags);
328 } else {
329 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
330 struct call_single_data *data = &d;
331
332 if (!wait)
333 data = &__get_cpu_var(csd_data);
334
335 csd_lock(data);
336
337 data->func = func;
338 data->info = info;
339 generic_exec_single(cpu, data, wait);
340 } else {
341 err = -ENXIO; /* CPU not online */
342 }
343 }
344
345 put_cpu();
346
347 return err;
348}
349EXPORT_SYMBOL(smp_call_function_single);
350
351/*
352 * smp_call_function_any - Run a function on any of the given cpus
353 * @mask: The mask of cpus it can run on.
354 * @func: The function to run. This must be fast and non-blocking.
355 * @info: An arbitrary pointer to pass to the function.
356 * @wait: If true, wait until function has completed.
357 *
358 * Returns 0 on success, else a negative status code (if no cpus were online).
359 * Note that @wait will be implicitly turned on in case of allocation failures,
360 * since we fall back to on-stack allocation.
361 *
362 * Selection preference:
363 * 1) current cpu if in @mask
364 * 2) any cpu of current node if in @mask
365 * 3) any other online cpu in @mask
366 */
367int smp_call_function_any(const struct cpumask *mask,
368 smp_call_func_t func, void *info, int wait)
369{
370 unsigned int cpu;
371 const struct cpumask *nodemask;
372 int ret;
373
374 /* Try for same CPU (cheapest) */
375 cpu = get_cpu();
376 if (cpumask_test_cpu(cpu, mask))
377 goto call;
378
379 /* Try for same node. */
380 nodemask = cpumask_of_node(cpu_to_node(cpu));
381 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
382 cpu = cpumask_next_and(cpu, nodemask, mask)) {
383 if (cpu_online(cpu))
384 goto call;
385 }
386
387 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
388 cpu = cpumask_any_and(mask, cpu_online_mask);
389call:
390 ret = smp_call_function_single(cpu, func, info, wait);
391 put_cpu();
392 return ret;
393}
394EXPORT_SYMBOL_GPL(smp_call_function_any);
395
396/**
397 * __smp_call_function_single(): Run a function on a specific CPU
398 * @cpu: The CPU to run on.
399 * @data: Pre-allocated and setup data structure
400 * @wait: If true, wait until function has completed on specified CPU.
401 *
402 * Like smp_call_function_single(), but allow caller to pass in a
403 * pre-allocated data structure. Useful for embedding @data inside
404 * other structures, for instance.
405 */
406void __smp_call_function_single(int cpu, struct call_single_data *data,
407 int wait)
408{
409 unsigned int this_cpu;
410 unsigned long flags;
411
412 this_cpu = get_cpu();
413 /*
414 * Can deadlock when called with interrupts disabled.
415 * We allow cpu's that are not yet online though, as no one else can
416 * send smp call function interrupt to this cpu and as such deadlocks
417 * can't happen.
418 */
419 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
420 && !oops_in_progress);
421
422 if (cpu == this_cpu) {
423 local_irq_save(flags);
424 data->func(data->info);
425 local_irq_restore(flags);
426 } else {
427 csd_lock(data);
428 generic_exec_single(cpu, data, wait);
429 }
430 put_cpu();
431}
432
433/**
434 * smp_call_function_many(): Run a function on a set of other CPUs.
435 * @mask: The set of cpus to run on (only runs on online subset).
436 * @func: The function to run. This must be fast and non-blocking.
437 * @info: An arbitrary pointer to pass to the function.
438 * @wait: If true, wait (atomically) until function has completed
439 * on other CPUs.
440 *
441 * If @wait is true, then returns once @func has returned.
442 *
443 * You must not call this function with disabled interrupts or from a
444 * hardware interrupt handler or from a bottom half handler. Preemption
445 * must be disabled when calling this function.
446 */
447void smp_call_function_many(const struct cpumask *mask,
448 smp_call_func_t func, void *info, bool wait)
449{
450 struct call_function_data *data;
451 unsigned long flags;
452 int refs, cpu, next_cpu, this_cpu = smp_processor_id();
453
454 /*
455 * Can deadlock when called with interrupts disabled.
456 * We allow cpu's that are not yet online though, as no one else can
457 * send smp call function interrupt to this cpu and as such deadlocks
458 * can't happen.
459 */
460 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
461 && !oops_in_progress && !early_boot_irqs_disabled);
462
463 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
464 cpu = cpumask_first_and(mask, cpu_online_mask);
465 if (cpu == this_cpu)
466 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
467
468 /* No online cpus? We're done. */
469 if (cpu >= nr_cpu_ids)
470 return;
471
472 /* Do we have another CPU which isn't us? */
473 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
474 if (next_cpu == this_cpu)
475 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
476
477 /* Fastpath: do that cpu by itself. */
478 if (next_cpu >= nr_cpu_ids) {
479 smp_call_function_single(cpu, func, info, wait);
480 return;
481 }
482
483 data = &__get_cpu_var(cfd_data);
484 csd_lock(&data->csd);
485
486 /* This BUG_ON verifies our reuse assertions and can be removed */
487 BUG_ON(atomic_read(&data->refs) || !cpumask_empty(data->cpumask));
488
489 /*
490 * The global call function queue list add and delete are protected
491 * by a lock, but the list is traversed without any lock, relying
492 * on the rcu list add and delete to allow safe concurrent traversal.
493 * We reuse the call function data without waiting for any grace
494 * period after some other cpu removes it from the global queue.
495 * This means a cpu might find our data block as it is being
496 * filled out.
497 *
498 * We hold off the interrupt handler on the other cpu by
499 * ordering our writes to the cpu mask vs our setting of the
500 * refs counter. We assert only the cpu owning the data block
501 * will set a bit in cpumask, and each bit will only be cleared
502 * by the subject cpu. Each cpu must first find its bit is
503 * set and then check that refs is set indicating the element is
504 * ready to be processed, otherwise it must skip the entry.
505 *
506 * On the previous iteration refs was set to 0 by another cpu.
507 * To avoid the use of transitivity, set the counter to 0 here
508 * so the wmb will pair with the rmb in the interrupt handler.
509 */
510 atomic_set(&data->refs, 0); /* convert 3rd to 1st party write */
511
512 data->csd.func = func;
513 data->csd.info = info;
514
515 /* Ensure 0 refs is visible before mask. Also orders func and info */
516 smp_wmb();
517
518 /* We rely on the "and" being processed before the store */
519 cpumask_and(data->cpumask, mask, cpu_online_mask);
520 cpumask_clear_cpu(this_cpu, data->cpumask);
521 refs = cpumask_weight(data->cpumask);
522
523 /* Some callers race with other cpus changing the passed mask */
524 if (unlikely(!refs)) {
525 csd_unlock(&data->csd);
526 return;
527 }
528
529 raw_spin_lock_irqsave(&call_function.lock, flags);
530 /*
531 * Place entry at the _HEAD_ of the list, so that any cpu still
532 * observing the entry in generic_smp_call_function_interrupt()
533 * will not miss any other list entries:
534 */
535 list_add_rcu(&data->csd.list, &call_function.queue);
536 /*
537 * We rely on the wmb() in list_add_rcu to complete our writes
538 * to the cpumask before this write to refs, which indicates
539 * data is on the list and is ready to be processed.
540 */
541 atomic_set(&data->refs, refs);
542 raw_spin_unlock_irqrestore(&call_function.lock, flags);
543
544 /*
545 * Make the list addition visible before sending the ipi.
546 * (IPIs must obey or appear to obey normal Linux cache
547 * coherency rules -- see comment in generic_exec_single).
548 */
549 smp_mb();
550
551 /* Send a message to all CPUs in the map */
552 arch_send_call_function_ipi_mask(data->cpumask);
553
554 /* Optionally wait for the CPUs to complete */
555 if (wait)
556 csd_lock_wait(&data->csd);
557}
558EXPORT_SYMBOL(smp_call_function_many);
559
560/**
561 * smp_call_function(): Run a function on all other CPUs.
562 * @func: The function to run. This must be fast and non-blocking.
563 * @info: An arbitrary pointer to pass to the function.
564 * @wait: If true, wait (atomically) until function has completed
565 * on other CPUs.
566 *
567 * Returns 0.
568 *
569 * If @wait is true, then returns once @func has returned; otherwise
570 * it returns just before the target cpu calls @func.
571 *
572 * You must not call this function with disabled interrupts or from a
573 * hardware interrupt handler or from a bottom half handler.
574 */
575int smp_call_function(smp_call_func_t func, void *info, int wait)
576{
577 preempt_disable();
578 smp_call_function_many(cpu_online_mask, func, info, wait);
579 preempt_enable();
580
581 return 0;
582}
583EXPORT_SYMBOL(smp_call_function);
584
585void ipi_call_lock(void)
586{
587 raw_spin_lock(&call_function.lock);
588}
589
590void ipi_call_unlock(void)
591{
592 raw_spin_unlock(&call_function.lock);
593}
594
595void ipi_call_lock_irq(void)
596{
597 raw_spin_lock_irq(&call_function.lock);
598}
599
600void ipi_call_unlock_irq(void)
601{
602 raw_spin_unlock_irq(&call_function.lock);
603}
604#endif /* USE_GENERIC_SMP_HELPERS */
605
606/* Setup configured maximum number of CPUs to activate */
607unsigned int setup_max_cpus = NR_CPUS;
608EXPORT_SYMBOL(setup_max_cpus);
609
610
611/*
612 * Setup routine for controlling SMP activation
613 *
614 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
615 * activation entirely (the MPS table probe still happens, though).
616 *
617 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
618 * greater than 0, limits the maximum number of CPUs activated in
619 * SMP mode to <NUM>.
620 */
621
622void __weak arch_disable_smp_support(void) { }
623
624static int __init nosmp(char *str)
625{
626 setup_max_cpus = 0;
627 arch_disable_smp_support();
628
629 return 0;
630}
631
632early_param("nosmp", nosmp);
633
634/* this is hard limit */
635static int __init nrcpus(char *str)
636{
637 int nr_cpus;
638
639 get_option(&str, &nr_cpus);
640 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
641 nr_cpu_ids = nr_cpus;
642
643 return 0;
644}
645
646early_param("nr_cpus", nrcpus);
647
648static int __init maxcpus(char *str)
649{
650 get_option(&str, &setup_max_cpus);
651 if (setup_max_cpus == 0)
652 arch_disable_smp_support();
653
654 return 0;
655}
656
657early_param("maxcpus", maxcpus);
658
659/* Setup number of possible processor ids */
660int nr_cpu_ids __read_mostly = NR_CPUS;
661EXPORT_SYMBOL(nr_cpu_ids);
662
663/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
664void __init setup_nr_cpu_ids(void)
665{
666 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
667}
668
669/* Called by boot processor to activate the rest. */
670void __init smp_init(void)
671{
672 unsigned int cpu;
673
674 idle_threads_init();
675
676 /* FIXME: This should be done in userspace --RR */
677 for_each_present_cpu(cpu) {
678 if (num_online_cpus() >= setup_max_cpus)
679 break;
680 if (!cpu_online(cpu))
681 cpu_up(cpu);
682 }
683
684 /* Any cleanup work */
685 printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());
686 smp_cpus_done(setup_max_cpus);
687}
688
689/*
690 * Call a function on all processors. May be used during early boot while
691 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
692 * of local_irq_disable/enable().
693 */
694int on_each_cpu(void (*func) (void *info), void *info, int wait)
695{
696 unsigned long flags;
697 int ret = 0;
698
699 preempt_disable();
700 ret = smp_call_function(func, info, wait);
701 local_irq_save(flags);
702 func(info);
703 local_irq_restore(flags);
704 preempt_enable();
705 return ret;
706}
707EXPORT_SYMBOL(on_each_cpu);
708
709/**
710 * on_each_cpu_mask(): Run a function on processors specified by
711 * cpumask, which may include the local processor.
712 * @mask: The set of cpus to run on (only runs on online subset).
713 * @func: The function to run. This must be fast and non-blocking.
714 * @info: An arbitrary pointer to pass to the function.
715 * @wait: If true, wait (atomically) until function has completed
716 * on other CPUs.
717 *
718 * If @wait is true, then returns once @func has returned.
719 *
720 * You must not call this function with disabled interrupts or
721 * from a hardware interrupt handler or from a bottom half handler.
722 */
723void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
724 void *info, bool wait)
725{
726 int cpu = get_cpu();
727
728 smp_call_function_many(mask, func, info, wait);
729 if (cpumask_test_cpu(cpu, mask)) {
730 local_irq_disable();
731 func(info);
732 local_irq_enable();
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 * @gfp_flags: GFP flags to use when allocating the cpumask
754 * used internally by the function.
755 *
756 * The function might sleep if the GFP flags indicates a non
757 * atomic allocation is allowed.
758 *
759 * Preemption is disabled to protect against CPUs going offline but not online.
760 * CPUs going online during the call will not be seen or sent an IPI.
761 *
762 * You must not call this function with disabled interrupts or
763 * from a hardware interrupt handler or from a bottom half handler.
764 */
765void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
766 smp_call_func_t func, void *info, bool wait,
767 gfp_t gfp_flags)
768{
769 cpumask_var_t cpus;
770 int cpu, ret;
771
772 might_sleep_if(gfp_flags & __GFP_WAIT);
773
774 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
775 preempt_disable();
776 for_each_online_cpu(cpu)
777 if (cond_func(cpu, info))
778 cpumask_set_cpu(cpu, cpus);
779 on_each_cpu_mask(cpus, func, info, wait);
780 preempt_enable();
781 free_cpumask_var(cpus);
782 } else {
783 /*
784 * No free cpumask, bother. No matter, we'll
785 * just have to IPI them one by one.
786 */
787 preempt_disable();
788 for_each_online_cpu(cpu)
789 if (cond_func(cpu, info)) {
790 ret = smp_call_function_single(cpu, func,
791 info, wait);
792 WARN_ON_ONCE(!ret);
793 }
794 preempt_enable();
795 }
796}
797EXPORT_SYMBOL(on_each_cpu_cond);
798
799static void do_nothing(void *unused)
800{
801}
802
803/**
804 * kick_all_cpus_sync - Force all cpus out of idle
805 *
806 * Used to synchronize the update of pm_idle function pointer. It's
807 * called after the pointer is updated and returns after the dummy
808 * callback function has been executed on all cpus. The execution of
809 * the function can only happen on the remote cpus after they have
810 * left the idle function which had been called via pm_idle function
811 * pointer. So it's guaranteed that nothing uses the previous pointer
812 * anymore.
813 */
814void kick_all_cpus_sync(void)
815{
816 /* Make sure the change is visible before we kick the cpus */
817 smp_mb();
818 smp_call_function(do_nothing, NULL, 1);
819}
820EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
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);