xref: /linux/kernel/smp.c (revision d53b8e36925256097a08d7cb749198d85cbf9b2b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic helpers for smp ipi calls
4  *
5  * (C) Jens Axboe <jens.axboe@oracle.com> 2008
6  */
7 
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 
10 #include <linux/irq_work.h>
11 #include <linux/rcupdate.h>
12 #include <linux/rculist.h>
13 #include <linux/kernel.h>
14 #include <linux/export.h>
15 #include <linux/percpu.h>
16 #include <linux/init.h>
17 #include <linux/interrupt.h>
18 #include <linux/gfp.h>
19 #include <linux/smp.h>
20 #include <linux/cpu.h>
21 #include <linux/sched.h>
22 #include <linux/sched/idle.h>
23 #include <linux/hypervisor.h>
24 #include <linux/sched/clock.h>
25 #include <linux/nmi.h>
26 #include <linux/sched/debug.h>
27 #include <linux/jump_label.h>
28 #include <linux/string_choices.h>
29 
30 #include <trace/events/ipi.h>
31 #define CREATE_TRACE_POINTS
32 #include <trace/events/csd.h>
33 #undef CREATE_TRACE_POINTS
34 
35 #include "smpboot.h"
36 #include "sched/smp.h"
37 
38 #define CSD_TYPE(_csd)	((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK)
39 
40 struct call_function_data {
41 	call_single_data_t	__percpu *csd;
42 	cpumask_var_t		cpumask;
43 	cpumask_var_t		cpumask_ipi;
44 };
45 
46 static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
47 
48 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
49 
50 static DEFINE_PER_CPU(atomic_t, trigger_backtrace) = ATOMIC_INIT(1);
51 
52 static void __flush_smp_call_function_queue(bool warn_cpu_offline);
53 
54 int smpcfd_prepare_cpu(unsigned int cpu)
55 {
56 	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
57 
58 	if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
59 				     cpu_to_node(cpu)))
60 		return -ENOMEM;
61 	if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
62 				     cpu_to_node(cpu))) {
63 		free_cpumask_var(cfd->cpumask);
64 		return -ENOMEM;
65 	}
66 	cfd->csd = alloc_percpu(call_single_data_t);
67 	if (!cfd->csd) {
68 		free_cpumask_var(cfd->cpumask);
69 		free_cpumask_var(cfd->cpumask_ipi);
70 		return -ENOMEM;
71 	}
72 
73 	return 0;
74 }
75 
76 int smpcfd_dead_cpu(unsigned int cpu)
77 {
78 	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
79 
80 	free_cpumask_var(cfd->cpumask);
81 	free_cpumask_var(cfd->cpumask_ipi);
82 	free_percpu(cfd->csd);
83 	return 0;
84 }
85 
86 int smpcfd_dying_cpu(unsigned int cpu)
87 {
88 	/*
89 	 * The IPIs for the smp-call-function callbacks queued by other
90 	 * CPUs might arrive late, either due to hardware latencies or
91 	 * because this CPU disabled interrupts (inside stop-machine)
92 	 * before the IPIs were sent. So flush out any pending callbacks
93 	 * explicitly (without waiting for the IPIs to arrive), to
94 	 * ensure that the outgoing CPU doesn't go offline with work
95 	 * still pending.
96 	 */
97 	__flush_smp_call_function_queue(false);
98 	irq_work_run();
99 	return 0;
100 }
101 
102 void __init call_function_init(void)
103 {
104 	int i;
105 
106 	for_each_possible_cpu(i)
107 		init_llist_head(&per_cpu(call_single_queue, i));
108 
109 	smpcfd_prepare_cpu(smp_processor_id());
110 }
111 
112 static __always_inline void
113 send_call_function_single_ipi(int cpu)
114 {
115 	if (call_function_single_prep_ipi(cpu)) {
116 		trace_ipi_send_cpu(cpu, _RET_IP_,
117 				   generic_smp_call_function_single_interrupt);
118 		arch_send_call_function_single_ipi(cpu);
119 	}
120 }
121 
122 static __always_inline void
123 send_call_function_ipi_mask(struct cpumask *mask)
124 {
125 	trace_ipi_send_cpumask(mask, _RET_IP_,
126 			       generic_smp_call_function_single_interrupt);
127 	arch_send_call_function_ipi_mask(mask);
128 }
129 
130 static __always_inline void
131 csd_do_func(smp_call_func_t func, void *info, call_single_data_t *csd)
132 {
133 	trace_csd_function_entry(func, csd);
134 	func(info);
135 	trace_csd_function_exit(func, csd);
136 }
137 
138 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
139 
140 static DEFINE_STATIC_KEY_MAYBE(CONFIG_CSD_LOCK_WAIT_DEBUG_DEFAULT, csdlock_debug_enabled);
141 
142 /*
143  * Parse the csdlock_debug= kernel boot parameter.
144  *
145  * If you need to restore the old "ext" value that once provided
146  * additional debugging information, reapply the following commits:
147  *
148  * de7b09ef658d ("locking/csd_lock: Prepare more CSD lock debugging")
149  * a5aabace5fb8 ("locking/csd_lock: Add more data to CSD lock debugging")
150  */
151 static int __init csdlock_debug(char *str)
152 {
153 	int ret;
154 	unsigned int val = 0;
155 
156 	ret = get_option(&str, &val);
157 	if (ret) {
158 		if (val)
159 			static_branch_enable(&csdlock_debug_enabled);
160 		else
161 			static_branch_disable(&csdlock_debug_enabled);
162 	}
163 
164 	return 1;
165 }
166 __setup("csdlock_debug=", csdlock_debug);
167 
168 static DEFINE_PER_CPU(call_single_data_t *, cur_csd);
169 static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func);
170 static DEFINE_PER_CPU(void *, cur_csd_info);
171 
172 static ulong csd_lock_timeout = 5000;  /* CSD lock timeout in milliseconds. */
173 module_param(csd_lock_timeout, ulong, 0444);
174 static int panic_on_ipistall;  /* CSD panic timeout in milliseconds, 300000 for five minutes. */
175 module_param(panic_on_ipistall, int, 0444);
176 
177 static atomic_t csd_bug_count = ATOMIC_INIT(0);
178 
179 /* Record current CSD work for current CPU, NULL to erase. */
180 static void __csd_lock_record(call_single_data_t *csd)
181 {
182 	if (!csd) {
183 		smp_mb(); /* NULL cur_csd after unlock. */
184 		__this_cpu_write(cur_csd, NULL);
185 		return;
186 	}
187 	__this_cpu_write(cur_csd_func, csd->func);
188 	__this_cpu_write(cur_csd_info, csd->info);
189 	smp_wmb(); /* func and info before csd. */
190 	__this_cpu_write(cur_csd, csd);
191 	smp_mb(); /* Update cur_csd before function call. */
192 		  /* Or before unlock, as the case may be. */
193 }
194 
195 static __always_inline void csd_lock_record(call_single_data_t *csd)
196 {
197 	if (static_branch_unlikely(&csdlock_debug_enabled))
198 		__csd_lock_record(csd);
199 }
200 
201 static int csd_lock_wait_getcpu(call_single_data_t *csd)
202 {
203 	unsigned int csd_type;
204 
205 	csd_type = CSD_TYPE(csd);
206 	if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC)
207 		return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */
208 	return -1;
209 }
210 
211 /*
212  * Complain if too much time spent waiting.  Note that only
213  * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
214  * so waiting on other types gets much less information.
215  */
216 static bool csd_lock_wait_toolong(call_single_data_t *csd, u64 ts0, u64 *ts1, int *bug_id)
217 {
218 	int cpu = -1;
219 	int cpux;
220 	bool firsttime;
221 	u64 ts2, ts_delta;
222 	call_single_data_t *cpu_cur_csd;
223 	unsigned int flags = READ_ONCE(csd->node.u_flags);
224 	unsigned long long csd_lock_timeout_ns = csd_lock_timeout * NSEC_PER_MSEC;
225 
226 	if (!(flags & CSD_FLAG_LOCK)) {
227 		if (!unlikely(*bug_id))
228 			return true;
229 		cpu = csd_lock_wait_getcpu(csd);
230 		pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
231 			 *bug_id, raw_smp_processor_id(), cpu);
232 		return true;
233 	}
234 
235 	ts2 = sched_clock();
236 	/* How long since we last checked for a stuck CSD lock.*/
237 	ts_delta = ts2 - *ts1;
238 	if (likely(ts_delta <= csd_lock_timeout_ns || csd_lock_timeout_ns == 0))
239 		return false;
240 
241 	firsttime = !*bug_id;
242 	if (firsttime)
243 		*bug_id = atomic_inc_return(&csd_bug_count);
244 	cpu = csd_lock_wait_getcpu(csd);
245 	if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu))
246 		cpux = 0;
247 	else
248 		cpux = cpu;
249 	cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */
250 	/* How long since this CSD lock was stuck. */
251 	ts_delta = ts2 - ts0;
252 	pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n",
253 		 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts_delta,
254 		 cpu, csd->func, csd->info);
255 	/*
256 	 * If the CSD lock is still stuck after 5 minutes, it is unlikely
257 	 * to become unstuck. Use a signed comparison to avoid triggering
258 	 * on underflows when the TSC is out of sync between sockets.
259 	 */
260 	BUG_ON(panic_on_ipistall > 0 && (s64)ts_delta > ((s64)panic_on_ipistall * NSEC_PER_MSEC));
261 	if (cpu_cur_csd && csd != cpu_cur_csd) {
262 		pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
263 			 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)),
264 			 READ_ONCE(per_cpu(cur_csd_info, cpux)));
265 	} else {
266 		pr_alert("\tcsd: CSD lock (#%d) %s.\n",
267 			 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request");
268 	}
269 	if (cpu >= 0) {
270 		if (atomic_cmpxchg_acquire(&per_cpu(trigger_backtrace, cpu), 1, 0))
271 			dump_cpu_task(cpu);
272 		if (!cpu_cur_csd) {
273 			pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu);
274 			arch_send_call_function_single_ipi(cpu);
275 		}
276 	}
277 	if (firsttime)
278 		dump_stack();
279 	*ts1 = ts2;
280 
281 	return false;
282 }
283 
284 /*
285  * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
286  *
287  * For non-synchronous ipi calls the csd can still be in use by the
288  * previous function call. For multi-cpu calls its even more interesting
289  * as we'll have to ensure no other cpu is observing our csd.
290  */
291 static void __csd_lock_wait(call_single_data_t *csd)
292 {
293 	int bug_id = 0;
294 	u64 ts0, ts1;
295 
296 	ts1 = ts0 = sched_clock();
297 	for (;;) {
298 		if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id))
299 			break;
300 		cpu_relax();
301 	}
302 	smp_acquire__after_ctrl_dep();
303 }
304 
305 static __always_inline void csd_lock_wait(call_single_data_t *csd)
306 {
307 	if (static_branch_unlikely(&csdlock_debug_enabled)) {
308 		__csd_lock_wait(csd);
309 		return;
310 	}
311 
312 	smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
313 }
314 #else
315 static void csd_lock_record(call_single_data_t *csd)
316 {
317 }
318 
319 static __always_inline void csd_lock_wait(call_single_data_t *csd)
320 {
321 	smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
322 }
323 #endif
324 
325 static __always_inline void csd_lock(call_single_data_t *csd)
326 {
327 	csd_lock_wait(csd);
328 	csd->node.u_flags |= CSD_FLAG_LOCK;
329 
330 	/*
331 	 * prevent CPU from reordering the above assignment
332 	 * to ->flags with any subsequent assignments to other
333 	 * fields of the specified call_single_data_t structure:
334 	 */
335 	smp_wmb();
336 }
337 
338 static __always_inline void csd_unlock(call_single_data_t *csd)
339 {
340 	WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK));
341 
342 	/*
343 	 * ensure we're all done before releasing data:
344 	 */
345 	smp_store_release(&csd->node.u_flags, 0);
346 }
347 
348 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
349 
350 void __smp_call_single_queue(int cpu, struct llist_node *node)
351 {
352 	/*
353 	 * We have to check the type of the CSD before queueing it, because
354 	 * once queued it can have its flags cleared by
355 	 *   flush_smp_call_function_queue()
356 	 * even if we haven't sent the smp_call IPI yet (e.g. the stopper
357 	 * executes migration_cpu_stop() on the remote CPU).
358 	 */
359 	if (trace_csd_queue_cpu_enabled()) {
360 		call_single_data_t *csd;
361 		smp_call_func_t func;
362 
363 		csd = container_of(node, call_single_data_t, node.llist);
364 		func = CSD_TYPE(csd) == CSD_TYPE_TTWU ?
365 			sched_ttwu_pending : csd->func;
366 
367 		trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
368 	}
369 
370 	/*
371 	 * The list addition should be visible to the target CPU when it pops
372 	 * the head of the list to pull the entry off it in the IPI handler
373 	 * because of normal cache coherency rules implied by the underlying
374 	 * llist ops.
375 	 *
376 	 * If IPIs can go out of order to the cache coherency protocol
377 	 * in an architecture, sufficient synchronisation should be added
378 	 * to arch code to make it appear to obey cache coherency WRT
379 	 * locking and barrier primitives. Generic code isn't really
380 	 * equipped to do the right thing...
381 	 */
382 	if (llist_add(node, &per_cpu(call_single_queue, cpu)))
383 		send_call_function_single_ipi(cpu);
384 }
385 
386 /*
387  * Insert a previously allocated call_single_data_t element
388  * for execution on the given CPU. data must already have
389  * ->func, ->info, and ->flags set.
390  */
391 static int generic_exec_single(int cpu, call_single_data_t *csd)
392 {
393 	if (cpu == smp_processor_id()) {
394 		smp_call_func_t func = csd->func;
395 		void *info = csd->info;
396 		unsigned long flags;
397 
398 		/*
399 		 * We can unlock early even for the synchronous on-stack case,
400 		 * since we're doing this from the same CPU..
401 		 */
402 		csd_lock_record(csd);
403 		csd_unlock(csd);
404 		local_irq_save(flags);
405 		csd_do_func(func, info, NULL);
406 		csd_lock_record(NULL);
407 		local_irq_restore(flags);
408 		return 0;
409 	}
410 
411 	if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
412 		csd_unlock(csd);
413 		return -ENXIO;
414 	}
415 
416 	__smp_call_single_queue(cpu, &csd->node.llist);
417 
418 	return 0;
419 }
420 
421 /**
422  * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
423  *
424  * Invoked by arch to handle an IPI for call function single.
425  * Must be called with interrupts disabled.
426  */
427 void generic_smp_call_function_single_interrupt(void)
428 {
429 	__flush_smp_call_function_queue(true);
430 }
431 
432 /**
433  * __flush_smp_call_function_queue - Flush pending smp-call-function callbacks
434  *
435  * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
436  *		      offline CPU. Skip this check if set to 'false'.
437  *
438  * Flush any pending smp-call-function callbacks queued on this CPU. This is
439  * invoked by the generic IPI handler, as well as by a CPU about to go offline,
440  * to ensure that all pending IPI callbacks are run before it goes completely
441  * offline.
442  *
443  * Loop through the call_single_queue and run all the queued callbacks.
444  * Must be called with interrupts disabled.
445  */
446 static void __flush_smp_call_function_queue(bool warn_cpu_offline)
447 {
448 	call_single_data_t *csd, *csd_next;
449 	struct llist_node *entry, *prev;
450 	struct llist_head *head;
451 	static bool warned;
452 	atomic_t *tbt;
453 
454 	lockdep_assert_irqs_disabled();
455 
456 	/* Allow waiters to send backtrace NMI from here onwards */
457 	tbt = this_cpu_ptr(&trigger_backtrace);
458 	atomic_set_release(tbt, 1);
459 
460 	head = this_cpu_ptr(&call_single_queue);
461 	entry = llist_del_all(head);
462 	entry = llist_reverse_order(entry);
463 
464 	/* There shouldn't be any pending callbacks on an offline CPU. */
465 	if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
466 		     !warned && entry != NULL)) {
467 		warned = true;
468 		WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
469 
470 		/*
471 		 * We don't have to use the _safe() variant here
472 		 * because we are not invoking the IPI handlers yet.
473 		 */
474 		llist_for_each_entry(csd, entry, node.llist) {
475 			switch (CSD_TYPE(csd)) {
476 			case CSD_TYPE_ASYNC:
477 			case CSD_TYPE_SYNC:
478 			case CSD_TYPE_IRQ_WORK:
479 				pr_warn("IPI callback %pS sent to offline CPU\n",
480 					csd->func);
481 				break;
482 
483 			case CSD_TYPE_TTWU:
484 				pr_warn("IPI task-wakeup sent to offline CPU\n");
485 				break;
486 
487 			default:
488 				pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
489 					CSD_TYPE(csd));
490 				break;
491 			}
492 		}
493 	}
494 
495 	/*
496 	 * First; run all SYNC callbacks, people are waiting for us.
497 	 */
498 	prev = NULL;
499 	llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
500 		/* Do we wait until *after* callback? */
501 		if (CSD_TYPE(csd) == CSD_TYPE_SYNC) {
502 			smp_call_func_t func = csd->func;
503 			void *info = csd->info;
504 
505 			if (prev) {
506 				prev->next = &csd_next->node.llist;
507 			} else {
508 				entry = &csd_next->node.llist;
509 			}
510 
511 			csd_lock_record(csd);
512 			csd_do_func(func, info, csd);
513 			csd_unlock(csd);
514 			csd_lock_record(NULL);
515 		} else {
516 			prev = &csd->node.llist;
517 		}
518 	}
519 
520 	if (!entry)
521 		return;
522 
523 	/*
524 	 * Second; run all !SYNC callbacks.
525 	 */
526 	prev = NULL;
527 	llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
528 		int type = CSD_TYPE(csd);
529 
530 		if (type != CSD_TYPE_TTWU) {
531 			if (prev) {
532 				prev->next = &csd_next->node.llist;
533 			} else {
534 				entry = &csd_next->node.llist;
535 			}
536 
537 			if (type == CSD_TYPE_ASYNC) {
538 				smp_call_func_t func = csd->func;
539 				void *info = csd->info;
540 
541 				csd_lock_record(csd);
542 				csd_unlock(csd);
543 				csd_do_func(func, info, csd);
544 				csd_lock_record(NULL);
545 			} else if (type == CSD_TYPE_IRQ_WORK) {
546 				irq_work_single(csd);
547 			}
548 
549 		} else {
550 			prev = &csd->node.llist;
551 		}
552 	}
553 
554 	/*
555 	 * Third; only CSD_TYPE_TTWU is left, issue those.
556 	 */
557 	if (entry) {
558 		csd = llist_entry(entry, typeof(*csd), node.llist);
559 		csd_do_func(sched_ttwu_pending, entry, csd);
560 	}
561 }
562 
563 
564 /**
565  * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
566  *				   from task context (idle, migration thread)
567  *
568  * When TIF_POLLING_NRFLAG is supported and a CPU is in idle and has it
569  * set, then remote CPUs can avoid sending IPIs and wake the idle CPU by
570  * setting TIF_NEED_RESCHED. The idle task on the woken up CPU has to
571  * handle queued SMP function calls before scheduling.
572  *
573  * The migration thread has to ensure that an eventually pending wakeup has
574  * been handled before it migrates a task.
575  */
576 void flush_smp_call_function_queue(void)
577 {
578 	unsigned int was_pending;
579 	unsigned long flags;
580 
581 	if (llist_empty(this_cpu_ptr(&call_single_queue)))
582 		return;
583 
584 	local_irq_save(flags);
585 	/* Get the already pending soft interrupts for RT enabled kernels */
586 	was_pending = local_softirq_pending();
587 	__flush_smp_call_function_queue(true);
588 	if (local_softirq_pending())
589 		do_softirq_post_smp_call_flush(was_pending);
590 
591 	local_irq_restore(flags);
592 }
593 
594 /*
595  * smp_call_function_single - Run a function on a specific CPU
596  * @func: The function to run. This must be fast and non-blocking.
597  * @info: An arbitrary pointer to pass to the function.
598  * @wait: If true, wait until function has completed on other CPUs.
599  *
600  * Returns 0 on success, else a negative status code.
601  */
602 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
603 			     int wait)
604 {
605 	call_single_data_t *csd;
606 	call_single_data_t csd_stack = {
607 		.node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, },
608 	};
609 	int this_cpu;
610 	int err;
611 
612 	/*
613 	 * prevent preemption and reschedule on another processor,
614 	 * as well as CPU removal
615 	 */
616 	this_cpu = get_cpu();
617 
618 	/*
619 	 * Can deadlock when called with interrupts disabled.
620 	 * We allow cpu's that are not yet online though, as no one else can
621 	 * send smp call function interrupt to this cpu and as such deadlocks
622 	 * can't happen.
623 	 */
624 	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
625 		     && !oops_in_progress);
626 
627 	/*
628 	 * When @wait we can deadlock when we interrupt between llist_add() and
629 	 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
630 	 * csd_lock() on because the interrupt context uses the same csd
631 	 * storage.
632 	 */
633 	WARN_ON_ONCE(!in_task());
634 
635 	csd = &csd_stack;
636 	if (!wait) {
637 		csd = this_cpu_ptr(&csd_data);
638 		csd_lock(csd);
639 	}
640 
641 	csd->func = func;
642 	csd->info = info;
643 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
644 	csd->node.src = smp_processor_id();
645 	csd->node.dst = cpu;
646 #endif
647 
648 	err = generic_exec_single(cpu, csd);
649 
650 	if (wait)
651 		csd_lock_wait(csd);
652 
653 	put_cpu();
654 
655 	return err;
656 }
657 EXPORT_SYMBOL(smp_call_function_single);
658 
659 /**
660  * smp_call_function_single_async() - Run an asynchronous function on a
661  * 			         specific CPU.
662  * @cpu: The CPU to run on.
663  * @csd: Pre-allocated and setup data structure
664  *
665  * Like smp_call_function_single(), but the call is asynchonous and
666  * can thus be done from contexts with disabled interrupts.
667  *
668  * The caller passes his own pre-allocated data structure
669  * (ie: embedded in an object) and is responsible for synchronizing it
670  * such that the IPIs performed on the @csd are strictly serialized.
671  *
672  * If the function is called with one csd which has not yet been
673  * processed by previous call to smp_call_function_single_async(), the
674  * function will return immediately with -EBUSY showing that the csd
675  * object is still in progress.
676  *
677  * NOTE: Be careful, there is unfortunately no current debugging facility to
678  * validate the correctness of this serialization.
679  *
680  * Return: %0 on success or negative errno value on error
681  */
682 int smp_call_function_single_async(int cpu, call_single_data_t *csd)
683 {
684 	int err = 0;
685 
686 	preempt_disable();
687 
688 	if (csd->node.u_flags & CSD_FLAG_LOCK) {
689 		err = -EBUSY;
690 		goto out;
691 	}
692 
693 	csd->node.u_flags = CSD_FLAG_LOCK;
694 	smp_wmb();
695 
696 	err = generic_exec_single(cpu, csd);
697 
698 out:
699 	preempt_enable();
700 
701 	return err;
702 }
703 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
704 
705 /*
706  * smp_call_function_any - Run a function on any of the given cpus
707  * @mask: The mask of cpus it can run on.
708  * @func: The function to run. This must be fast and non-blocking.
709  * @info: An arbitrary pointer to pass to the function.
710  * @wait: If true, wait until function has completed.
711  *
712  * Returns 0 on success, else a negative status code (if no cpus were online).
713  *
714  * Selection preference:
715  *	1) current cpu if in @mask
716  *	2) any cpu of current node if in @mask
717  *	3) any other online cpu in @mask
718  */
719 int smp_call_function_any(const struct cpumask *mask,
720 			  smp_call_func_t func, void *info, int wait)
721 {
722 	unsigned int cpu;
723 	const struct cpumask *nodemask;
724 	int ret;
725 
726 	/* Try for same CPU (cheapest) */
727 	cpu = get_cpu();
728 	if (cpumask_test_cpu(cpu, mask))
729 		goto call;
730 
731 	/* Try for same node. */
732 	nodemask = cpumask_of_node(cpu_to_node(cpu));
733 	for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
734 	     cpu = cpumask_next_and(cpu, nodemask, mask)) {
735 		if (cpu_online(cpu))
736 			goto call;
737 	}
738 
739 	/* Any online will do: smp_call_function_single handles nr_cpu_ids. */
740 	cpu = cpumask_any_and(mask, cpu_online_mask);
741 call:
742 	ret = smp_call_function_single(cpu, func, info, wait);
743 	put_cpu();
744 	return ret;
745 }
746 EXPORT_SYMBOL_GPL(smp_call_function_any);
747 
748 /*
749  * Flags to be used as scf_flags argument of smp_call_function_many_cond().
750  *
751  * %SCF_WAIT:		Wait until function execution is completed
752  * %SCF_RUN_LOCAL:	Run also locally if local cpu is set in cpumask
753  */
754 #define SCF_WAIT	(1U << 0)
755 #define SCF_RUN_LOCAL	(1U << 1)
756 
757 static void smp_call_function_many_cond(const struct cpumask *mask,
758 					smp_call_func_t func, void *info,
759 					unsigned int scf_flags,
760 					smp_cond_func_t cond_func)
761 {
762 	int cpu, last_cpu, this_cpu = smp_processor_id();
763 	struct call_function_data *cfd;
764 	bool wait = scf_flags & SCF_WAIT;
765 	int nr_cpus = 0;
766 	bool run_remote = false;
767 	bool run_local = false;
768 
769 	lockdep_assert_preemption_disabled();
770 
771 	/*
772 	 * Can deadlock when called with interrupts disabled.
773 	 * We allow cpu's that are not yet online though, as no one else can
774 	 * send smp call function interrupt to this cpu and as such deadlocks
775 	 * can't happen.
776 	 */
777 	if (cpu_online(this_cpu) && !oops_in_progress &&
778 	    !early_boot_irqs_disabled)
779 		lockdep_assert_irqs_enabled();
780 
781 	/*
782 	 * When @wait we can deadlock when we interrupt between llist_add() and
783 	 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
784 	 * csd_lock() on because the interrupt context uses the same csd
785 	 * storage.
786 	 */
787 	WARN_ON_ONCE(!in_task());
788 
789 	/* Check if we need local execution. */
790 	if ((scf_flags & SCF_RUN_LOCAL) && cpumask_test_cpu(this_cpu, mask))
791 		run_local = true;
792 
793 	/* Check if we need remote execution, i.e., any CPU excluding this one. */
794 	cpu = cpumask_first_and(mask, cpu_online_mask);
795 	if (cpu == this_cpu)
796 		cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
797 	if (cpu < nr_cpu_ids)
798 		run_remote = true;
799 
800 	if (run_remote) {
801 		cfd = this_cpu_ptr(&cfd_data);
802 		cpumask_and(cfd->cpumask, mask, cpu_online_mask);
803 		__cpumask_clear_cpu(this_cpu, cfd->cpumask);
804 
805 		cpumask_clear(cfd->cpumask_ipi);
806 		for_each_cpu(cpu, cfd->cpumask) {
807 			call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
808 
809 			if (cond_func && !cond_func(cpu, info)) {
810 				__cpumask_clear_cpu(cpu, cfd->cpumask);
811 				continue;
812 			}
813 
814 			csd_lock(csd);
815 			if (wait)
816 				csd->node.u_flags |= CSD_TYPE_SYNC;
817 			csd->func = func;
818 			csd->info = info;
819 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
820 			csd->node.src = smp_processor_id();
821 			csd->node.dst = cpu;
822 #endif
823 			trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
824 
825 			if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
826 				__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
827 				nr_cpus++;
828 				last_cpu = cpu;
829 			}
830 		}
831 
832 		/*
833 		 * Choose the most efficient way to send an IPI. Note that the
834 		 * number of CPUs might be zero due to concurrent changes to the
835 		 * provided mask.
836 		 */
837 		if (nr_cpus == 1)
838 			send_call_function_single_ipi(last_cpu);
839 		else if (likely(nr_cpus > 1))
840 			send_call_function_ipi_mask(cfd->cpumask_ipi);
841 	}
842 
843 	if (run_local && (!cond_func || cond_func(this_cpu, info))) {
844 		unsigned long flags;
845 
846 		local_irq_save(flags);
847 		csd_do_func(func, info, NULL);
848 		local_irq_restore(flags);
849 	}
850 
851 	if (run_remote && wait) {
852 		for_each_cpu(cpu, cfd->cpumask) {
853 			call_single_data_t *csd;
854 
855 			csd = per_cpu_ptr(cfd->csd, cpu);
856 			csd_lock_wait(csd);
857 		}
858 	}
859 }
860 
861 /**
862  * smp_call_function_many(): Run a function on a set of CPUs.
863  * @mask: The set of cpus to run on (only runs on online subset).
864  * @func: The function to run. This must be fast and non-blocking.
865  * @info: An arbitrary pointer to pass to the function.
866  * @wait: Bitmask that controls the operation. If %SCF_WAIT is set, wait
867  *        (atomically) until function has completed on other CPUs. If
868  *        %SCF_RUN_LOCAL is set, the function will also be run locally
869  *        if the local CPU is set in the @cpumask.
870  *
871  * If @wait is true, then returns once @func has returned.
872  *
873  * You must not call this function with disabled interrupts or from a
874  * hardware interrupt handler or from a bottom half handler. Preemption
875  * must be disabled when calling this function.
876  */
877 void smp_call_function_many(const struct cpumask *mask,
878 			    smp_call_func_t func, void *info, bool wait)
879 {
880 	smp_call_function_many_cond(mask, func, info, wait * SCF_WAIT, NULL);
881 }
882 EXPORT_SYMBOL(smp_call_function_many);
883 
884 /**
885  * smp_call_function(): Run a function on all other CPUs.
886  * @func: The function to run. This must be fast and non-blocking.
887  * @info: An arbitrary pointer to pass to the function.
888  * @wait: If true, wait (atomically) until function has completed
889  *        on other CPUs.
890  *
891  * Returns 0.
892  *
893  * If @wait is true, then returns once @func has returned; otherwise
894  * it returns just before the target cpu calls @func.
895  *
896  * You must not call this function with disabled interrupts or from a
897  * hardware interrupt handler or from a bottom half handler.
898  */
899 void smp_call_function(smp_call_func_t func, void *info, int wait)
900 {
901 	preempt_disable();
902 	smp_call_function_many(cpu_online_mask, func, info, wait);
903 	preempt_enable();
904 }
905 EXPORT_SYMBOL(smp_call_function);
906 
907 /* Setup configured maximum number of CPUs to activate */
908 unsigned int setup_max_cpus = NR_CPUS;
909 EXPORT_SYMBOL(setup_max_cpus);
910 
911 
912 /*
913  * Setup routine for controlling SMP activation
914  *
915  * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
916  * activation entirely (the MPS table probe still happens, though).
917  *
918  * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
919  * greater than 0, limits the maximum number of CPUs activated in
920  * SMP mode to <NUM>.
921  */
922 
923 void __weak __init arch_disable_smp_support(void) { }
924 
925 static int __init nosmp(char *str)
926 {
927 	setup_max_cpus = 0;
928 	arch_disable_smp_support();
929 
930 	return 0;
931 }
932 
933 early_param("nosmp", nosmp);
934 
935 /* this is hard limit */
936 static int __init nrcpus(char *str)
937 {
938 	int nr_cpus;
939 
940 	if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids)
941 		set_nr_cpu_ids(nr_cpus);
942 
943 	return 0;
944 }
945 
946 early_param("nr_cpus", nrcpus);
947 
948 static int __init maxcpus(char *str)
949 {
950 	get_option(&str, &setup_max_cpus);
951 	if (setup_max_cpus == 0)
952 		arch_disable_smp_support();
953 
954 	return 0;
955 }
956 
957 early_param("maxcpus", maxcpus);
958 
959 #if (NR_CPUS > 1) && !defined(CONFIG_FORCE_NR_CPUS)
960 /* Setup number of possible processor ids */
961 unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
962 EXPORT_SYMBOL(nr_cpu_ids);
963 #endif
964 
965 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
966 void __init setup_nr_cpu_ids(void)
967 {
968 	set_nr_cpu_ids(find_last_bit(cpumask_bits(cpu_possible_mask), NR_CPUS) + 1);
969 }
970 
971 /* Called by boot processor to activate the rest. */
972 void __init smp_init(void)
973 {
974 	int num_nodes, num_cpus;
975 
976 	idle_threads_init();
977 	cpuhp_threads_init();
978 
979 	pr_info("Bringing up secondary CPUs ...\n");
980 
981 	bringup_nonboot_cpus(setup_max_cpus);
982 
983 	num_nodes = num_online_nodes();
984 	num_cpus  = num_online_cpus();
985 	pr_info("Brought up %d node%s, %d CPU%s\n",
986 		num_nodes, str_plural(num_nodes), num_cpus, str_plural(num_cpus));
987 
988 	/* Any cleanup work */
989 	smp_cpus_done(setup_max_cpus);
990 }
991 
992 /*
993  * on_each_cpu_cond(): Call a function on each processor for which
994  * the supplied function cond_func returns true, optionally waiting
995  * for all the required CPUs to finish. This may include the local
996  * processor.
997  * @cond_func:	A callback function that is passed a cpu id and
998  *		the info parameter. The function is called
999  *		with preemption disabled. The function should
1000  *		return a blooean value indicating whether to IPI
1001  *		the specified CPU.
1002  * @func:	The function to run on all applicable CPUs.
1003  *		This must be fast and non-blocking.
1004  * @info:	An arbitrary pointer to pass to both functions.
1005  * @wait:	If true, wait (atomically) until function has
1006  *		completed on other CPUs.
1007  *
1008  * Preemption is disabled to protect against CPUs going offline but not online.
1009  * CPUs going online during the call will not be seen or sent an IPI.
1010  *
1011  * You must not call this function with disabled interrupts or
1012  * from a hardware interrupt handler or from a bottom half handler.
1013  */
1014 void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func,
1015 			   void *info, bool wait, const struct cpumask *mask)
1016 {
1017 	unsigned int scf_flags = SCF_RUN_LOCAL;
1018 
1019 	if (wait)
1020 		scf_flags |= SCF_WAIT;
1021 
1022 	preempt_disable();
1023 	smp_call_function_many_cond(mask, func, info, scf_flags, cond_func);
1024 	preempt_enable();
1025 }
1026 EXPORT_SYMBOL(on_each_cpu_cond_mask);
1027 
1028 static void do_nothing(void *unused)
1029 {
1030 }
1031 
1032 /**
1033  * kick_all_cpus_sync - Force all cpus out of idle
1034  *
1035  * Used to synchronize the update of pm_idle function pointer. It's
1036  * called after the pointer is updated and returns after the dummy
1037  * callback function has been executed on all cpus. The execution of
1038  * the function can only happen on the remote cpus after they have
1039  * left the idle function which had been called via pm_idle function
1040  * pointer. So it's guaranteed that nothing uses the previous pointer
1041  * anymore.
1042  */
1043 void kick_all_cpus_sync(void)
1044 {
1045 	/* Make sure the change is visible before we kick the cpus */
1046 	smp_mb();
1047 	smp_call_function(do_nothing, NULL, 1);
1048 }
1049 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
1050 
1051 /**
1052  * wake_up_all_idle_cpus - break all cpus out of idle
1053  * wake_up_all_idle_cpus try to break all cpus which is in idle state even
1054  * including idle polling cpus, for non-idle cpus, we will do nothing
1055  * for them.
1056  */
1057 void wake_up_all_idle_cpus(void)
1058 {
1059 	int cpu;
1060 
1061 	for_each_possible_cpu(cpu) {
1062 		preempt_disable();
1063 		if (cpu != smp_processor_id() && cpu_online(cpu))
1064 			wake_up_if_idle(cpu);
1065 		preempt_enable();
1066 	}
1067 }
1068 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
1069 
1070 /**
1071  * struct smp_call_on_cpu_struct - Call a function on a specific CPU
1072  * @work: &work_struct
1073  * @done: &completion to signal
1074  * @func: function to call
1075  * @data: function's data argument
1076  * @ret: return value from @func
1077  * @cpu: target CPU (%-1 for any CPU)
1078  *
1079  * Used to call a function on a specific cpu and wait for it to return.
1080  * Optionally make sure the call is done on a specified physical cpu via vcpu
1081  * pinning in order to support virtualized environments.
1082  */
1083 struct smp_call_on_cpu_struct {
1084 	struct work_struct	work;
1085 	struct completion	done;
1086 	int			(*func)(void *);
1087 	void			*data;
1088 	int			ret;
1089 	int			cpu;
1090 };
1091 
1092 static void smp_call_on_cpu_callback(struct work_struct *work)
1093 {
1094 	struct smp_call_on_cpu_struct *sscs;
1095 
1096 	sscs = container_of(work, struct smp_call_on_cpu_struct, work);
1097 	if (sscs->cpu >= 0)
1098 		hypervisor_pin_vcpu(sscs->cpu);
1099 	sscs->ret = sscs->func(sscs->data);
1100 	if (sscs->cpu >= 0)
1101 		hypervisor_pin_vcpu(-1);
1102 
1103 	complete(&sscs->done);
1104 }
1105 
1106 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
1107 {
1108 	struct smp_call_on_cpu_struct sscs = {
1109 		.done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
1110 		.func = func,
1111 		.data = par,
1112 		.cpu  = phys ? cpu : -1,
1113 	};
1114 
1115 	INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
1116 
1117 	if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1118 		return -ENXIO;
1119 
1120 	queue_work_on(cpu, system_wq, &sscs.work);
1121 	wait_for_completion(&sscs.done);
1122 	destroy_work_on_stack(&sscs.work);
1123 
1124 	return sscs.ret;
1125 }
1126 EXPORT_SYMBOL_GPL(smp_call_on_cpu);
1127