Loading...
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
26enum {
27 CSD_FLAG_LOCK = 0x01,
28 CSD_FLAG_SYNCHRONOUS = 0x02,
29};
30
31struct call_function_data {
32 call_single_data_t __percpu *csd;
33 cpumask_var_t cpumask;
34 cpumask_var_t cpumask_ipi;
35};
36
37static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
38
39static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
40
41static void flush_smp_call_function_queue(bool warn_cpu_offline);
42
43int smpcfd_prepare_cpu(unsigned int cpu)
44{
45 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
46
47 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
48 cpu_to_node(cpu)))
49 return -ENOMEM;
50 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
51 cpu_to_node(cpu))) {
52 free_cpumask_var(cfd->cpumask);
53 return -ENOMEM;
54 }
55 cfd->csd = alloc_percpu(call_single_data_t);
56 if (!cfd->csd) {
57 free_cpumask_var(cfd->cpumask);
58 free_cpumask_var(cfd->cpumask_ipi);
59 return -ENOMEM;
60 }
61
62 return 0;
63}
64
65int smpcfd_dead_cpu(unsigned int cpu)
66{
67 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
68
69 free_cpumask_var(cfd->cpumask);
70 free_cpumask_var(cfd->cpumask_ipi);
71 free_percpu(cfd->csd);
72 return 0;
73}
74
75int smpcfd_dying_cpu(unsigned int cpu)
76{
77 /*
78 * The IPIs for the smp-call-function callbacks queued by other
79 * CPUs might arrive late, either due to hardware latencies or
80 * because this CPU disabled interrupts (inside stop-machine)
81 * before the IPIs were sent. So flush out any pending callbacks
82 * explicitly (without waiting for the IPIs to arrive), to
83 * ensure that the outgoing CPU doesn't go offline with work
84 * still pending.
85 */
86 flush_smp_call_function_queue(false);
87 return 0;
88}
89
90void __init call_function_init(void)
91{
92 int i;
93
94 for_each_possible_cpu(i)
95 init_llist_head(&per_cpu(call_single_queue, i));
96
97 smpcfd_prepare_cpu(smp_processor_id());
98}
99
100/*
101 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
102 *
103 * For non-synchronous ipi calls the csd can still be in use by the
104 * previous function call. For multi-cpu calls its even more interesting
105 * as we'll have to ensure no other cpu is observing our csd.
106 */
107static __always_inline void csd_lock_wait(call_single_data_t *csd)
108{
109 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
110}
111
112static __always_inline void csd_lock(call_single_data_t *csd)
113{
114 csd_lock_wait(csd);
115 csd->flags |= CSD_FLAG_LOCK;
116
117 /*
118 * prevent CPU from reordering the above assignment
119 * to ->flags with any subsequent assignments to other
120 * fields of the specified call_single_data_t structure:
121 */
122 smp_wmb();
123}
124
125static __always_inline void csd_unlock(call_single_data_t *csd)
126{
127 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
128
129 /*
130 * ensure we're all done before releasing data:
131 */
132 smp_store_release(&csd->flags, 0);
133}
134
135static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
136
137/*
138 * Insert a previously allocated call_single_data_t element
139 * for execution on the given CPU. data must already have
140 * ->func, ->info, and ->flags set.
141 */
142static int generic_exec_single(int cpu, call_single_data_t *csd,
143 smp_call_func_t func, void *info)
144{
145 if (cpu == smp_processor_id()) {
146 unsigned long flags;
147
148 /*
149 * We can unlock early even for the synchronous on-stack case,
150 * since we're doing this from the same CPU..
151 */
152 csd_unlock(csd);
153 local_irq_save(flags);
154 func(info);
155 local_irq_restore(flags);
156 return 0;
157 }
158
159
160 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
161 csd_unlock(csd);
162 return -ENXIO;
163 }
164
165 csd->func = func;
166 csd->info = info;
167
168 /*
169 * The list addition should be visible before sending the IPI
170 * handler locks the list to pull the entry off it because of
171 * normal cache coherency rules implied by spinlocks.
172 *
173 * If IPIs can go out of order to the cache coherency protocol
174 * in an architecture, sufficient synchronisation should be added
175 * to arch code to make it appear to obey cache coherency WRT
176 * locking and barrier primitives. Generic code isn't really
177 * equipped to do the right thing...
178 */
179 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
180 arch_send_call_function_single_ipi(cpu);
181
182 return 0;
183}
184
185/**
186 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
187 *
188 * Invoked by arch to handle an IPI for call function single.
189 * Must be called with interrupts disabled.
190 */
191void generic_smp_call_function_single_interrupt(void)
192{
193 flush_smp_call_function_queue(true);
194}
195
196/**
197 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
198 *
199 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
200 * offline CPU. Skip this check if set to 'false'.
201 *
202 * Flush any pending smp-call-function callbacks queued on this CPU. This is
203 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
204 * to ensure that all pending IPI callbacks are run before it goes completely
205 * offline.
206 *
207 * Loop through the call_single_queue and run all the queued callbacks.
208 * Must be called with interrupts disabled.
209 */
210static void flush_smp_call_function_queue(bool warn_cpu_offline)
211{
212 struct llist_head *head;
213 struct llist_node *entry;
214 call_single_data_t *csd, *csd_next;
215 static bool warned;
216
217 lockdep_assert_irqs_disabled();
218
219 head = this_cpu_ptr(&call_single_queue);
220 entry = llist_del_all(head);
221 entry = llist_reverse_order(entry);
222
223 /* There shouldn't be any pending callbacks on an offline CPU. */
224 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
225 !warned && !llist_empty(head))) {
226 warned = true;
227 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
228
229 /*
230 * We don't have to use the _safe() variant here
231 * because we are not invoking the IPI handlers yet.
232 */
233 llist_for_each_entry(csd, entry, llist)
234 pr_warn("IPI callback %pS sent to offline CPU\n",
235 csd->func);
236 }
237
238 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
239 smp_call_func_t func = csd->func;
240 void *info = csd->info;
241
242 /* Do we wait until *after* callback? */
243 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
244 func(info);
245 csd_unlock(csd);
246 } else {
247 csd_unlock(csd);
248 func(info);
249 }
250 }
251
252 /*
253 * Handle irq works queued remotely by irq_work_queue_on().
254 * Smp functions above are typically synchronous so they
255 * better run first since some other CPUs may be busy waiting
256 * for them.
257 */
258 irq_work_run();
259}
260
261/*
262 * smp_call_function_single - Run a function on a specific CPU
263 * @func: The function to run. This must be fast and non-blocking.
264 * @info: An arbitrary pointer to pass to the function.
265 * @wait: If true, wait until function has completed on other CPUs.
266 *
267 * Returns 0 on success, else a negative status code.
268 */
269int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
270 int wait)
271{
272 call_single_data_t *csd;
273 call_single_data_t csd_stack = {
274 .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS,
275 };
276 int this_cpu;
277 int err;
278
279 /*
280 * prevent preemption and reschedule on another processor,
281 * as well as CPU removal
282 */
283 this_cpu = get_cpu();
284
285 /*
286 * Can deadlock when called with interrupts disabled.
287 * We allow cpu's that are not yet online though, as no one else can
288 * send smp call function interrupt to this cpu and as such deadlocks
289 * can't happen.
290 */
291 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
292 && !oops_in_progress);
293
294 /*
295 * When @wait we can deadlock when we interrupt between llist_add() and
296 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
297 * csd_lock() on because the interrupt context uses the same csd
298 * storage.
299 */
300 WARN_ON_ONCE(!in_task());
301
302 csd = &csd_stack;
303 if (!wait) {
304 csd = this_cpu_ptr(&csd_data);
305 csd_lock(csd);
306 }
307
308 err = generic_exec_single(cpu, csd, func, info);
309
310 if (wait)
311 csd_lock_wait(csd);
312
313 put_cpu();
314
315 return err;
316}
317EXPORT_SYMBOL(smp_call_function_single);
318
319/**
320 * smp_call_function_single_async(): Run an asynchronous function on a
321 * specific CPU.
322 * @cpu: The CPU to run on.
323 * @csd: Pre-allocated and setup data structure
324 *
325 * Like smp_call_function_single(), but the call is asynchonous and
326 * can thus be done from contexts with disabled interrupts.
327 *
328 * The caller passes his own pre-allocated data structure
329 * (ie: embedded in an object) and is responsible for synchronizing it
330 * such that the IPIs performed on the @csd are strictly serialized.
331 *
332 * NOTE: Be careful, there is unfortunately no current debugging facility to
333 * validate the correctness of this serialization.
334 */
335int smp_call_function_single_async(int cpu, call_single_data_t *csd)
336{
337 int err = 0;
338
339 preempt_disable();
340
341 /* We could deadlock if we have to wait here with interrupts disabled! */
342 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
343 csd_lock_wait(csd);
344
345 csd->flags = CSD_FLAG_LOCK;
346 smp_wmb();
347
348 err = generic_exec_single(cpu, csd, csd->func, csd->info);
349 preempt_enable();
350
351 return err;
352}
353EXPORT_SYMBOL_GPL(smp_call_function_single_async);
354
355/*
356 * smp_call_function_any - Run a function on any of the given cpus
357 * @mask: The mask of cpus it can run on.
358 * @func: The function to run. This must be fast and non-blocking.
359 * @info: An arbitrary pointer to pass to the function.
360 * @wait: If true, wait until function has completed.
361 *
362 * Returns 0 on success, else a negative status code (if no cpus were online).
363 *
364 * Selection preference:
365 * 1) current cpu if in @mask
366 * 2) any cpu of current node if in @mask
367 * 3) any other online cpu in @mask
368 */
369int smp_call_function_any(const struct cpumask *mask,
370 smp_call_func_t func, void *info, int wait)
371{
372 unsigned int cpu;
373 const struct cpumask *nodemask;
374 int ret;
375
376 /* Try for same CPU (cheapest) */
377 cpu = get_cpu();
378 if (cpumask_test_cpu(cpu, mask))
379 goto call;
380
381 /* Try for same node. */
382 nodemask = cpumask_of_node(cpu_to_node(cpu));
383 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
384 cpu = cpumask_next_and(cpu, nodemask, mask)) {
385 if (cpu_online(cpu))
386 goto call;
387 }
388
389 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
390 cpu = cpumask_any_and(mask, cpu_online_mask);
391call:
392 ret = smp_call_function_single(cpu, func, info, wait);
393 put_cpu();
394 return ret;
395}
396EXPORT_SYMBOL_GPL(smp_call_function_any);
397
398/**
399 * smp_call_function_many(): Run a function on a set of other CPUs.
400 * @mask: The set of cpus to run on (only runs on online subset).
401 * @func: The function to run. This must be fast and non-blocking.
402 * @info: An arbitrary pointer to pass to the function.
403 * @wait: If true, wait (atomically) until function has completed
404 * on other CPUs.
405 *
406 * If @wait is true, then returns once @func has returned.
407 *
408 * You must not call this function with disabled interrupts or from a
409 * hardware interrupt handler or from a bottom half handler. Preemption
410 * must be disabled when calling this function.
411 */
412void smp_call_function_many(const struct cpumask *mask,
413 smp_call_func_t func, void *info, bool wait)
414{
415 struct call_function_data *cfd;
416 int cpu, next_cpu, this_cpu = smp_processor_id();
417
418 /*
419 * Can deadlock when called with interrupts disabled.
420 * We allow cpu's that are not yet online though, as no one else can
421 * send smp call function interrupt to this cpu and as such deadlocks
422 * can't happen.
423 */
424 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
425 && !oops_in_progress && !early_boot_irqs_disabled);
426
427 /*
428 * When @wait we can deadlock when we interrupt between llist_add() and
429 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
430 * csd_lock() on because the interrupt context uses the same csd
431 * storage.
432 */
433 WARN_ON_ONCE(!in_task());
434
435 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
436 cpu = cpumask_first_and(mask, cpu_online_mask);
437 if (cpu == this_cpu)
438 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
439
440 /* No online cpus? We're done. */
441 if (cpu >= nr_cpu_ids)
442 return;
443
444 /* Do we have another CPU which isn't us? */
445 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
446 if (next_cpu == this_cpu)
447 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
448
449 /* Fastpath: do that cpu by itself. */
450 if (next_cpu >= nr_cpu_ids) {
451 smp_call_function_single(cpu, func, info, wait);
452 return;
453 }
454
455 cfd = this_cpu_ptr(&cfd_data);
456
457 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
458 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
459
460 /* Some callers race with other cpus changing the passed mask */
461 if (unlikely(!cpumask_weight(cfd->cpumask)))
462 return;
463
464 cpumask_clear(cfd->cpumask_ipi);
465 for_each_cpu(cpu, cfd->cpumask) {
466 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
467
468 csd_lock(csd);
469 if (wait)
470 csd->flags |= CSD_FLAG_SYNCHRONOUS;
471 csd->func = func;
472 csd->info = info;
473 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
474 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
475 }
476
477 /* Send a message to all CPUs in the map */
478 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
479
480 if (wait) {
481 for_each_cpu(cpu, cfd->cpumask) {
482 call_single_data_t *csd;
483
484 csd = per_cpu_ptr(cfd->csd, cpu);
485 csd_lock_wait(csd);
486 }
487 }
488}
489EXPORT_SYMBOL(smp_call_function_many);
490
491/**
492 * smp_call_function(): Run a function on all other CPUs.
493 * @func: The function to run. This must be fast and non-blocking.
494 * @info: An arbitrary pointer to pass to the function.
495 * @wait: If true, wait (atomically) until function has completed
496 * on other CPUs.
497 *
498 * Returns 0.
499 *
500 * If @wait is true, then returns once @func has returned; otherwise
501 * it returns just before the target cpu calls @func.
502 *
503 * You must not call this function with disabled interrupts or from a
504 * hardware interrupt handler or from a bottom half handler.
505 */
506void smp_call_function(smp_call_func_t func, void *info, int wait)
507{
508 preempt_disable();
509 smp_call_function_many(cpu_online_mask, func, info, wait);
510 preempt_enable();
511}
512EXPORT_SYMBOL(smp_call_function);
513
514/* Setup configured maximum number of CPUs to activate */
515unsigned int setup_max_cpus = NR_CPUS;
516EXPORT_SYMBOL(setup_max_cpus);
517
518
519/*
520 * Setup routine for controlling SMP activation
521 *
522 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
523 * activation entirely (the MPS table probe still happens, though).
524 *
525 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
526 * greater than 0, limits the maximum number of CPUs activated in
527 * SMP mode to <NUM>.
528 */
529
530void __weak arch_disable_smp_support(void) { }
531
532static int __init nosmp(char *str)
533{
534 setup_max_cpus = 0;
535 arch_disable_smp_support();
536
537 return 0;
538}
539
540early_param("nosmp", nosmp);
541
542/* this is hard limit */
543static int __init nrcpus(char *str)
544{
545 int nr_cpus;
546
547 get_option(&str, &nr_cpus);
548 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
549 nr_cpu_ids = nr_cpus;
550
551 return 0;
552}
553
554early_param("nr_cpus", nrcpus);
555
556static int __init maxcpus(char *str)
557{
558 get_option(&str, &setup_max_cpus);
559 if (setup_max_cpus == 0)
560 arch_disable_smp_support();
561
562 return 0;
563}
564
565early_param("maxcpus", maxcpus);
566
567/* Setup number of possible processor ids */
568unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
569EXPORT_SYMBOL(nr_cpu_ids);
570
571/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
572void __init setup_nr_cpu_ids(void)
573{
574 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
575}
576
577/* Called by boot processor to activate the rest. */
578void __init smp_init(void)
579{
580 int num_nodes, num_cpus;
581 unsigned int cpu;
582
583 idle_threads_init();
584 cpuhp_threads_init();
585
586 pr_info("Bringing up secondary CPUs ...\n");
587
588 /* FIXME: This should be done in userspace --RR */
589 for_each_present_cpu(cpu) {
590 if (num_online_cpus() >= setup_max_cpus)
591 break;
592 if (!cpu_online(cpu))
593 cpu_up(cpu);
594 }
595
596 num_nodes = num_online_nodes();
597 num_cpus = num_online_cpus();
598 pr_info("Brought up %d node%s, %d CPU%s\n",
599 num_nodes, (num_nodes > 1 ? "s" : ""),
600 num_cpus, (num_cpus > 1 ? "s" : ""));
601
602 /* Any cleanup work */
603 smp_cpus_done(setup_max_cpus);
604}
605
606/*
607 * Call a function on all processors. May be used during early boot while
608 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
609 * of local_irq_disable/enable().
610 */
611void on_each_cpu(void (*func) (void *info), void *info, int wait)
612{
613 unsigned long flags;
614
615 preempt_disable();
616 smp_call_function(func, info, wait);
617 local_irq_save(flags);
618 func(info);
619 local_irq_restore(flags);
620 preempt_enable();
621}
622EXPORT_SYMBOL(on_each_cpu);
623
624/**
625 * on_each_cpu_mask(): Run a function on processors specified by
626 * cpumask, which may include the local processor.
627 * @mask: The set of cpus to run on (only runs on online subset).
628 * @func: The function to run. This must be fast and non-blocking.
629 * @info: An arbitrary pointer to pass to the function.
630 * @wait: If true, wait (atomically) until function has completed
631 * on other CPUs.
632 *
633 * If @wait is true, then returns once @func has returned.
634 *
635 * You must not call this function with disabled interrupts or from a
636 * hardware interrupt handler or from a bottom half handler. The
637 * exception is that it may be used during early boot while
638 * early_boot_irqs_disabled is set.
639 */
640void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
641 void *info, bool wait)
642{
643 int cpu = get_cpu();
644
645 smp_call_function_many(mask, func, info, wait);
646 if (cpumask_test_cpu(cpu, mask)) {
647 unsigned long flags;
648 local_irq_save(flags);
649 func(info);
650 local_irq_restore(flags);
651 }
652 put_cpu();
653}
654EXPORT_SYMBOL(on_each_cpu_mask);
655
656/*
657 * on_each_cpu_cond(): Call a function on each processor for which
658 * the supplied function cond_func returns true, optionally waiting
659 * for all the required CPUs to finish. This may include the local
660 * processor.
661 * @cond_func: A callback function that is passed a cpu id and
662 * the the info parameter. The function is called
663 * with preemption disabled. The function should
664 * return a blooean value indicating whether to IPI
665 * the specified CPU.
666 * @func: The function to run on all applicable CPUs.
667 * This must be fast and non-blocking.
668 * @info: An arbitrary pointer to pass to both functions.
669 * @wait: If true, wait (atomically) until function has
670 * completed on other CPUs.
671 * @gfp_flags: GFP flags to use when allocating the cpumask
672 * used internally by the function.
673 *
674 * The function might sleep if the GFP flags indicates a non
675 * atomic allocation is allowed.
676 *
677 * Preemption is disabled to protect against CPUs going offline but not online.
678 * CPUs going online during the call will not be seen or sent an IPI.
679 *
680 * You must not call this function with disabled interrupts or
681 * from a hardware interrupt handler or from a bottom half handler.
682 */
683void on_each_cpu_cond_mask(bool (*cond_func)(int cpu, void *info),
684 smp_call_func_t func, void *info, bool wait,
685 gfp_t gfp_flags, const struct cpumask *mask)
686{
687 cpumask_var_t cpus;
688 int cpu, ret;
689
690 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
691
692 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
693 preempt_disable();
694 for_each_cpu(cpu, mask)
695 if (cond_func(cpu, info))
696 __cpumask_set_cpu(cpu, cpus);
697 on_each_cpu_mask(cpus, func, info, wait);
698 preempt_enable();
699 free_cpumask_var(cpus);
700 } else {
701 /*
702 * No free cpumask, bother. No matter, we'll
703 * just have to IPI them one by one.
704 */
705 preempt_disable();
706 for_each_cpu(cpu, mask)
707 if (cond_func(cpu, info)) {
708 ret = smp_call_function_single(cpu, func,
709 info, wait);
710 WARN_ON_ONCE(ret);
711 }
712 preempt_enable();
713 }
714}
715EXPORT_SYMBOL(on_each_cpu_cond_mask);
716
717void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
718 smp_call_func_t func, void *info, bool wait,
719 gfp_t gfp_flags)
720{
721 on_each_cpu_cond_mask(cond_func, func, info, wait, gfp_flags,
722 cpu_online_mask);
723}
724EXPORT_SYMBOL(on_each_cpu_cond);
725
726static void do_nothing(void *unused)
727{
728}
729
730/**
731 * kick_all_cpus_sync - Force all cpus out of idle
732 *
733 * Used to synchronize the update of pm_idle function pointer. It's
734 * called after the pointer is updated and returns after the dummy
735 * callback function has been executed on all cpus. The execution of
736 * the function can only happen on the remote cpus after they have
737 * left the idle function which had been called via pm_idle function
738 * pointer. So it's guaranteed that nothing uses the previous pointer
739 * anymore.
740 */
741void kick_all_cpus_sync(void)
742{
743 /* Make sure the change is visible before we kick the cpus */
744 smp_mb();
745 smp_call_function(do_nothing, NULL, 1);
746}
747EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
748
749/**
750 * wake_up_all_idle_cpus - break all cpus out of idle
751 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
752 * including idle polling cpus, for non-idle cpus, we will do nothing
753 * for them.
754 */
755void wake_up_all_idle_cpus(void)
756{
757 int cpu;
758
759 preempt_disable();
760 for_each_online_cpu(cpu) {
761 if (cpu == smp_processor_id())
762 continue;
763
764 wake_up_if_idle(cpu);
765 }
766 preempt_enable();
767}
768EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
769
770/**
771 * smp_call_on_cpu - Call a function on a specific cpu
772 *
773 * Used to call a function on a specific cpu and wait for it to return.
774 * Optionally make sure the call is done on a specified physical cpu via vcpu
775 * pinning in order to support virtualized environments.
776 */
777struct smp_call_on_cpu_struct {
778 struct work_struct work;
779 struct completion done;
780 int (*func)(void *);
781 void *data;
782 int ret;
783 int cpu;
784};
785
786static void smp_call_on_cpu_callback(struct work_struct *work)
787{
788 struct smp_call_on_cpu_struct *sscs;
789
790 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
791 if (sscs->cpu >= 0)
792 hypervisor_pin_vcpu(sscs->cpu);
793 sscs->ret = sscs->func(sscs->data);
794 if (sscs->cpu >= 0)
795 hypervisor_pin_vcpu(-1);
796
797 complete(&sscs->done);
798}
799
800int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
801{
802 struct smp_call_on_cpu_struct sscs = {
803 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
804 .func = func,
805 .data = par,
806 .cpu = phys ? cpu : -1,
807 };
808
809 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
810
811 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
812 return -ENXIO;
813
814 queue_work_on(cpu, system_wq, &sscs.work);
815 wait_for_completion(&sscs.done);
816
817 return sscs.ret;
818}
819EXPORT_SYMBOL_GPL(smp_call_on_cpu);
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
18enum {
19 CSD_FLAG_LOCK = 0x01,
20 CSD_FLAG_WAIT = 0x02,
21};
22
23struct call_function_data {
24 struct call_single_data __percpu *csd;
25 cpumask_var_t cpumask;
26};
27
28static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
29
30static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
31
32static int
33hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
34{
35 long cpu = (long)hcpu;
36 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
37
38 switch (action) {
39 case CPU_UP_PREPARE:
40 case CPU_UP_PREPARE_FROZEN:
41 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
42 cpu_to_node(cpu)))
43 return notifier_from_errno(-ENOMEM);
44 cfd->csd = alloc_percpu(struct call_single_data);
45 if (!cfd->csd) {
46 free_cpumask_var(cfd->cpumask);
47 return notifier_from_errno(-ENOMEM);
48 }
49 break;
50
51#ifdef CONFIG_HOTPLUG_CPU
52 case CPU_UP_CANCELED:
53 case CPU_UP_CANCELED_FROZEN:
54
55 case CPU_DEAD:
56 case CPU_DEAD_FROZEN:
57 free_cpumask_var(cfd->cpumask);
58 free_percpu(cfd->csd);
59 break;
60#endif
61 };
62
63 return NOTIFY_OK;
64}
65
66static struct notifier_block hotplug_cfd_notifier = {
67 .notifier_call = hotplug_cfd,
68};
69
70void __init call_function_init(void)
71{
72 void *cpu = (void *)(long)smp_processor_id();
73 int i;
74
75 for_each_possible_cpu(i)
76 init_llist_head(&per_cpu(call_single_queue, i));
77
78 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
79 register_cpu_notifier(&hotplug_cfd_notifier);
80}
81
82/*
83 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
84 *
85 * For non-synchronous ipi calls the csd can still be in use by the
86 * previous function call. For multi-cpu calls its even more interesting
87 * as we'll have to ensure no other cpu is observing our csd.
88 */
89static void csd_lock_wait(struct call_single_data *csd)
90{
91 while (csd->flags & CSD_FLAG_LOCK)
92 cpu_relax();
93}
94
95static void csd_lock(struct call_single_data *csd)
96{
97 csd_lock_wait(csd);
98 csd->flags |= CSD_FLAG_LOCK;
99
100 /*
101 * prevent CPU from reordering the above assignment
102 * to ->flags with any subsequent assignments to other
103 * fields of the specified call_single_data structure:
104 */
105 smp_mb();
106}
107
108static void csd_unlock(struct call_single_data *csd)
109{
110 WARN_ON((csd->flags & CSD_FLAG_WAIT) && !(csd->flags & CSD_FLAG_LOCK));
111
112 /*
113 * ensure we're all done before releasing data:
114 */
115 smp_mb();
116
117 csd->flags &= ~CSD_FLAG_LOCK;
118}
119
120static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
121
122/*
123 * Insert a previously allocated call_single_data element
124 * for execution on the given CPU. data must already have
125 * ->func, ->info, and ->flags set.
126 */
127static int generic_exec_single(int cpu, struct call_single_data *csd,
128 smp_call_func_t func, void *info, int wait)
129{
130 struct call_single_data csd_stack = { .flags = 0 };
131 unsigned long flags;
132
133
134 if (cpu == smp_processor_id()) {
135 local_irq_save(flags);
136 func(info);
137 local_irq_restore(flags);
138 return 0;
139 }
140
141
142 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu))
143 return -ENXIO;
144
145
146 if (!csd) {
147 csd = &csd_stack;
148 if (!wait)
149 csd = &__get_cpu_var(csd_data);
150 }
151
152 csd_lock(csd);
153
154 csd->func = func;
155 csd->info = info;
156
157 if (wait)
158 csd->flags |= CSD_FLAG_WAIT;
159
160 /*
161 * The list addition should be visible before sending the IPI
162 * handler locks the list to pull the entry off it because of
163 * normal cache coherency rules implied by spinlocks.
164 *
165 * If IPIs can go out of order to the cache coherency protocol
166 * in an architecture, sufficient synchronisation should be added
167 * to arch code to make it appear to obey cache coherency WRT
168 * locking and barrier primitives. Generic code isn't really
169 * equipped to do the right thing...
170 */
171 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
172 arch_send_call_function_single_ipi(cpu);
173
174 if (wait)
175 csd_lock_wait(csd);
176
177 return 0;
178}
179
180/*
181 * Invoked by arch to handle an IPI for call function single. Must be
182 * called from the arch with interrupts disabled.
183 */
184void generic_smp_call_function_single_interrupt(void)
185{
186 struct llist_node *entry;
187 struct call_single_data *csd, *csd_next;
188
189 /*
190 * Shouldn't receive this interrupt on a cpu that is not yet online.
191 */
192 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
193
194 entry = llist_del_all(&__get_cpu_var(call_single_queue));
195 entry = llist_reverse_order(entry);
196
197 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
198 csd->func(csd->info);
199 csd_unlock(csd);
200 }
201}
202
203/*
204 * smp_call_function_single - Run a function on a specific CPU
205 * @func: The function to run. This must be fast and non-blocking.
206 * @info: An arbitrary pointer to pass to the function.
207 * @wait: If true, wait until function has completed on other CPUs.
208 *
209 * Returns 0 on success, else a negative status code.
210 */
211int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
212 int wait)
213{
214 int this_cpu;
215 int err;
216
217 /*
218 * prevent preemption and reschedule on another processor,
219 * as well as CPU removal
220 */
221 this_cpu = get_cpu();
222
223 /*
224 * Can deadlock when called with interrupts disabled.
225 * We allow cpu's that are not yet online though, as no one else can
226 * send smp call function interrupt to this cpu and as such deadlocks
227 * can't happen.
228 */
229 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
230 && !oops_in_progress);
231
232 err = generic_exec_single(cpu, NULL, func, info, wait);
233
234 put_cpu();
235
236 return err;
237}
238EXPORT_SYMBOL(smp_call_function_single);
239
240/**
241 * smp_call_function_single_async(): Run an asynchronous function on a
242 * specific CPU.
243 * @cpu: The CPU to run on.
244 * @csd: Pre-allocated and setup data structure
245 *
246 * Like smp_call_function_single(), but the call is asynchonous and
247 * can thus be done from contexts with disabled interrupts.
248 *
249 * The caller passes his own pre-allocated data structure
250 * (ie: embedded in an object) and is responsible for synchronizing it
251 * such that the IPIs performed on the @csd are strictly serialized.
252 *
253 * NOTE: Be careful, there is unfortunately no current debugging facility to
254 * validate the correctness of this serialization.
255 */
256int smp_call_function_single_async(int cpu, struct call_single_data *csd)
257{
258 int err = 0;
259
260 preempt_disable();
261 err = generic_exec_single(cpu, csd, csd->func, csd->info, 0);
262 preempt_enable();
263
264 return err;
265}
266EXPORT_SYMBOL_GPL(smp_call_function_single_async);
267
268/*
269 * smp_call_function_any - Run a function on any of the given cpus
270 * @mask: The mask of cpus it can run on.
271 * @func: The function to run. This must be fast and non-blocking.
272 * @info: An arbitrary pointer to pass to the function.
273 * @wait: If true, wait until function has completed.
274 *
275 * Returns 0 on success, else a negative status code (if no cpus were online).
276 *
277 * Selection preference:
278 * 1) current cpu if in @mask
279 * 2) any cpu of current node if in @mask
280 * 3) any other online cpu in @mask
281 */
282int smp_call_function_any(const struct cpumask *mask,
283 smp_call_func_t func, void *info, int wait)
284{
285 unsigned int cpu;
286 const struct cpumask *nodemask;
287 int ret;
288
289 /* Try for same CPU (cheapest) */
290 cpu = get_cpu();
291 if (cpumask_test_cpu(cpu, mask))
292 goto call;
293
294 /* Try for same node. */
295 nodemask = cpumask_of_node(cpu_to_node(cpu));
296 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
297 cpu = cpumask_next_and(cpu, nodemask, mask)) {
298 if (cpu_online(cpu))
299 goto call;
300 }
301
302 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
303 cpu = cpumask_any_and(mask, cpu_online_mask);
304call:
305 ret = smp_call_function_single(cpu, func, info, wait);
306 put_cpu();
307 return ret;
308}
309EXPORT_SYMBOL_GPL(smp_call_function_any);
310
311/**
312 * smp_call_function_many(): Run a function on a set of other CPUs.
313 * @mask: The set of cpus to run on (only runs on online subset).
314 * @func: The function to run. This must be fast and non-blocking.
315 * @info: An arbitrary pointer to pass to the function.
316 * @wait: If true, wait (atomically) until function has completed
317 * on other CPUs.
318 *
319 * If @wait is true, then returns once @func has returned.
320 *
321 * You must not call this function with disabled interrupts or from a
322 * hardware interrupt handler or from a bottom half handler. Preemption
323 * must be disabled when calling this function.
324 */
325void smp_call_function_many(const struct cpumask *mask,
326 smp_call_func_t func, void *info, bool wait)
327{
328 struct call_function_data *cfd;
329 int cpu, next_cpu, this_cpu = smp_processor_id();
330
331 /*
332 * Can deadlock when called with interrupts disabled.
333 * We allow cpu's that are not yet online though, as no one else can
334 * send smp call function interrupt to this cpu and as such deadlocks
335 * can't happen.
336 */
337 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
338 && !oops_in_progress && !early_boot_irqs_disabled);
339
340 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
341 cpu = cpumask_first_and(mask, cpu_online_mask);
342 if (cpu == this_cpu)
343 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
344
345 /* No online cpus? We're done. */
346 if (cpu >= nr_cpu_ids)
347 return;
348
349 /* Do we have another CPU which isn't us? */
350 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
351 if (next_cpu == this_cpu)
352 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
353
354 /* Fastpath: do that cpu by itself. */
355 if (next_cpu >= nr_cpu_ids) {
356 smp_call_function_single(cpu, func, info, wait);
357 return;
358 }
359
360 cfd = &__get_cpu_var(cfd_data);
361
362 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
363 cpumask_clear_cpu(this_cpu, cfd->cpumask);
364
365 /* Some callers race with other cpus changing the passed mask */
366 if (unlikely(!cpumask_weight(cfd->cpumask)))
367 return;
368
369 for_each_cpu(cpu, cfd->cpumask) {
370 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
371
372 csd_lock(csd);
373 csd->func = func;
374 csd->info = info;
375 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
376 }
377
378 /* Send a message to all CPUs in the map */
379 arch_send_call_function_ipi_mask(cfd->cpumask);
380
381 if (wait) {
382 for_each_cpu(cpu, cfd->cpumask) {
383 struct call_single_data *csd;
384
385 csd = per_cpu_ptr(cfd->csd, cpu);
386 csd_lock_wait(csd);
387 }
388 }
389}
390EXPORT_SYMBOL(smp_call_function_many);
391
392/**
393 * smp_call_function(): Run a function on all other CPUs.
394 * @func: The function to run. This must be fast and non-blocking.
395 * @info: An arbitrary pointer to pass to the function.
396 * @wait: If true, wait (atomically) until function has completed
397 * on other CPUs.
398 *
399 * Returns 0.
400 *
401 * If @wait is true, then returns once @func has returned; otherwise
402 * it returns just before the target cpu calls @func.
403 *
404 * You must not call this function with disabled interrupts or from a
405 * hardware interrupt handler or from a bottom half handler.
406 */
407int smp_call_function(smp_call_func_t func, void *info, int wait)
408{
409 preempt_disable();
410 smp_call_function_many(cpu_online_mask, func, info, wait);
411 preempt_enable();
412
413 return 0;
414}
415EXPORT_SYMBOL(smp_call_function);
416
417/* Setup configured maximum number of CPUs to activate */
418unsigned int setup_max_cpus = NR_CPUS;
419EXPORT_SYMBOL(setup_max_cpus);
420
421
422/*
423 * Setup routine for controlling SMP activation
424 *
425 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
426 * activation entirely (the MPS table probe still happens, though).
427 *
428 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
429 * greater than 0, limits the maximum number of CPUs activated in
430 * SMP mode to <NUM>.
431 */
432
433void __weak arch_disable_smp_support(void) { }
434
435static int __init nosmp(char *str)
436{
437 setup_max_cpus = 0;
438 arch_disable_smp_support();
439
440 return 0;
441}
442
443early_param("nosmp", nosmp);
444
445/* this is hard limit */
446static int __init nrcpus(char *str)
447{
448 int nr_cpus;
449
450 get_option(&str, &nr_cpus);
451 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
452 nr_cpu_ids = nr_cpus;
453
454 return 0;
455}
456
457early_param("nr_cpus", nrcpus);
458
459static int __init maxcpus(char *str)
460{
461 get_option(&str, &setup_max_cpus);
462 if (setup_max_cpus == 0)
463 arch_disable_smp_support();
464
465 return 0;
466}
467
468early_param("maxcpus", maxcpus);
469
470/* Setup number of possible processor ids */
471int nr_cpu_ids __read_mostly = NR_CPUS;
472EXPORT_SYMBOL(nr_cpu_ids);
473
474/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
475void __init setup_nr_cpu_ids(void)
476{
477 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
478}
479
480void __weak smp_announce(void)
481{
482 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
483}
484
485/* Called by boot processor to activate the rest. */
486void __init smp_init(void)
487{
488 unsigned int cpu;
489
490 idle_threads_init();
491
492 /* FIXME: This should be done in userspace --RR */
493 for_each_present_cpu(cpu) {
494 if (num_online_cpus() >= setup_max_cpus)
495 break;
496 if (!cpu_online(cpu))
497 cpu_up(cpu);
498 }
499
500 /* Any cleanup work */
501 smp_announce();
502 smp_cpus_done(setup_max_cpus);
503}
504
505/*
506 * Call a function on all processors. May be used during early boot while
507 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
508 * of local_irq_disable/enable().
509 */
510int on_each_cpu(void (*func) (void *info), void *info, int wait)
511{
512 unsigned long flags;
513 int ret = 0;
514
515 preempt_disable();
516 ret = smp_call_function(func, info, wait);
517 local_irq_save(flags);
518 func(info);
519 local_irq_restore(flags);
520 preempt_enable();
521 return ret;
522}
523EXPORT_SYMBOL(on_each_cpu);
524
525/**
526 * on_each_cpu_mask(): Run a function on processors specified by
527 * cpumask, which may include the local processor.
528 * @mask: The set of cpus to run on (only runs on online subset).
529 * @func: The function to run. This must be fast and non-blocking.
530 * @info: An arbitrary pointer to pass to the function.
531 * @wait: If true, wait (atomically) until function has completed
532 * on other CPUs.
533 *
534 * If @wait is true, then returns once @func has returned.
535 *
536 * You must not call this function with disabled interrupts or from a
537 * hardware interrupt handler or from a bottom half handler. The
538 * exception is that it may be used during early boot while
539 * early_boot_irqs_disabled is set.
540 */
541void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
542 void *info, bool wait)
543{
544 int cpu = get_cpu();
545
546 smp_call_function_many(mask, func, info, wait);
547 if (cpumask_test_cpu(cpu, mask)) {
548 unsigned long flags;
549 local_irq_save(flags);
550 func(info);
551 local_irq_restore(flags);
552 }
553 put_cpu();
554}
555EXPORT_SYMBOL(on_each_cpu_mask);
556
557/*
558 * on_each_cpu_cond(): Call a function on each processor for which
559 * the supplied function cond_func returns true, optionally waiting
560 * for all the required CPUs to finish. This may include the local
561 * processor.
562 * @cond_func: A callback function that is passed a cpu id and
563 * the the info parameter. The function is called
564 * with preemption disabled. The function should
565 * return a blooean value indicating whether to IPI
566 * the specified CPU.
567 * @func: The function to run on all applicable CPUs.
568 * This must be fast and non-blocking.
569 * @info: An arbitrary pointer to pass to both functions.
570 * @wait: If true, wait (atomically) until function has
571 * completed on other CPUs.
572 * @gfp_flags: GFP flags to use when allocating the cpumask
573 * used internally by the function.
574 *
575 * The function might sleep if the GFP flags indicates a non
576 * atomic allocation is allowed.
577 *
578 * Preemption is disabled to protect against CPUs going offline but not online.
579 * CPUs going online during the call will not be seen or sent an IPI.
580 *
581 * You must not call this function with disabled interrupts or
582 * from a hardware interrupt handler or from a bottom half handler.
583 */
584void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
585 smp_call_func_t func, void *info, bool wait,
586 gfp_t gfp_flags)
587{
588 cpumask_var_t cpus;
589 int cpu, ret;
590
591 might_sleep_if(gfp_flags & __GFP_WAIT);
592
593 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
594 preempt_disable();
595 for_each_online_cpu(cpu)
596 if (cond_func(cpu, info))
597 cpumask_set_cpu(cpu, cpus);
598 on_each_cpu_mask(cpus, func, info, wait);
599 preempt_enable();
600 free_cpumask_var(cpus);
601 } else {
602 /*
603 * No free cpumask, bother. No matter, we'll
604 * just have to IPI them one by one.
605 */
606 preempt_disable();
607 for_each_online_cpu(cpu)
608 if (cond_func(cpu, info)) {
609 ret = smp_call_function_single(cpu, func,
610 info, wait);
611 WARN_ON_ONCE(!ret);
612 }
613 preempt_enable();
614 }
615}
616EXPORT_SYMBOL(on_each_cpu_cond);
617
618static void do_nothing(void *unused)
619{
620}
621
622/**
623 * kick_all_cpus_sync - Force all cpus out of idle
624 *
625 * Used to synchronize the update of pm_idle function pointer. It's
626 * called after the pointer is updated and returns after the dummy
627 * callback function has been executed on all cpus. The execution of
628 * the function can only happen on the remote cpus after they have
629 * left the idle function which had been called via pm_idle function
630 * pointer. So it's guaranteed that nothing uses the previous pointer
631 * anymore.
632 */
633void kick_all_cpus_sync(void)
634{
635 /* Make sure the change is visible before we kick the cpus */
636 smp_mb();
637 smp_call_function(do_nothing, NULL, 1);
638}
639EXPORT_SYMBOL_GPL(kick_all_cpus_sync);