xref: /openbmc/linux/kernel/smp.c (revision 4800cd83)
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/module.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 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
17 static struct {
18 	struct list_head	queue;
19 	raw_spinlock_t		lock;
20 } call_function __cacheline_aligned_in_smp =
21 	{
22 		.queue		= LIST_HEAD_INIT(call_function.queue),
23 		.lock		= __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
24 	};
25 
26 enum {
27 	CSD_FLAG_LOCK		= 0x01,
28 };
29 
30 struct call_function_data {
31 	struct call_single_data	csd;
32 	atomic_t		refs;
33 	cpumask_var_t		cpumask;
34 };
35 
36 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
37 
38 struct call_single_queue {
39 	struct list_head	list;
40 	raw_spinlock_t		lock;
41 };
42 
43 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
44 
45 static int
46 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
47 {
48 	long cpu = (long)hcpu;
49 	struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
50 
51 	switch (action) {
52 	case CPU_UP_PREPARE:
53 	case CPU_UP_PREPARE_FROZEN:
54 		if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
55 				cpu_to_node(cpu)))
56 			return notifier_from_errno(-ENOMEM);
57 		break;
58 
59 #ifdef CONFIG_HOTPLUG_CPU
60 	case CPU_UP_CANCELED:
61 	case CPU_UP_CANCELED_FROZEN:
62 
63 	case CPU_DEAD:
64 	case CPU_DEAD_FROZEN:
65 		free_cpumask_var(cfd->cpumask);
66 		break;
67 #endif
68 	};
69 
70 	return NOTIFY_OK;
71 }
72 
73 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
74 	.notifier_call		= hotplug_cfd,
75 };
76 
77 static int __cpuinit init_call_single_data(void)
78 {
79 	void *cpu = (void *)(long)smp_processor_id();
80 	int i;
81 
82 	for_each_possible_cpu(i) {
83 		struct call_single_queue *q = &per_cpu(call_single_queue, i);
84 
85 		raw_spin_lock_init(&q->lock);
86 		INIT_LIST_HEAD(&q->list);
87 	}
88 
89 	hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
90 	register_cpu_notifier(&hotplug_cfd_notifier);
91 
92 	return 0;
93 }
94 early_initcall(init_call_single_data);
95 
96 /*
97  * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
98  *
99  * For non-synchronous ipi calls the csd can still be in use by the
100  * previous function call. For multi-cpu calls its even more interesting
101  * as we'll have to ensure no other cpu is observing our csd.
102  */
103 static void csd_lock_wait(struct call_single_data *data)
104 {
105 	while (data->flags & CSD_FLAG_LOCK)
106 		cpu_relax();
107 }
108 
109 static void csd_lock(struct call_single_data *data)
110 {
111 	csd_lock_wait(data);
112 	data->flags = CSD_FLAG_LOCK;
113 
114 	/*
115 	 * prevent CPU from reordering the above assignment
116 	 * to ->flags with any subsequent assignments to other
117 	 * fields of the specified call_single_data structure:
118 	 */
119 	smp_mb();
120 }
121 
122 static void csd_unlock(struct call_single_data *data)
123 {
124 	WARN_ON(!(data->flags & CSD_FLAG_LOCK));
125 
126 	/*
127 	 * ensure we're all done before releasing data:
128 	 */
129 	smp_mb();
130 
131 	data->flags &= ~CSD_FLAG_LOCK;
132 }
133 
134 /*
135  * Insert a previously allocated call_single_data element
136  * for execution on the given CPU. data must already have
137  * ->func, ->info, and ->flags set.
138  */
139 static
140 void generic_exec_single(int cpu, struct call_single_data *data, int wait)
141 {
142 	struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
143 	unsigned long flags;
144 	int ipi;
145 
146 	raw_spin_lock_irqsave(&dst->lock, flags);
147 	ipi = list_empty(&dst->list);
148 	list_add_tail(&data->list, &dst->list);
149 	raw_spin_unlock_irqrestore(&dst->lock, flags);
150 
151 	/*
152 	 * The list addition should be visible before sending the IPI
153 	 * handler locks the list to pull the entry off it because of
154 	 * normal cache coherency rules implied by spinlocks.
155 	 *
156 	 * If IPIs can go out of order to the cache coherency protocol
157 	 * in an architecture, sufficient synchronisation should be added
158 	 * to arch code to make it appear to obey cache coherency WRT
159 	 * locking and barrier primitives. Generic code isn't really
160 	 * equipped to do the right thing...
161 	 */
162 	if (ipi)
163 		arch_send_call_function_single_ipi(cpu);
164 
165 	if (wait)
166 		csd_lock_wait(data);
167 }
168 
169 /*
170  * Invoked by arch to handle an IPI for call function. Must be called with
171  * interrupts disabled.
172  */
173 void generic_smp_call_function_interrupt(void)
174 {
175 	struct call_function_data *data;
176 	int cpu = smp_processor_id();
177 
178 	/*
179 	 * Shouldn't receive this interrupt on a cpu that is not yet online.
180 	 */
181 	WARN_ON_ONCE(!cpu_online(cpu));
182 
183 	/*
184 	 * Ensure entry is visible on call_function_queue after we have
185 	 * entered the IPI. See comment in smp_call_function_many.
186 	 * If we don't have this, then we may miss an entry on the list
187 	 * and never get another IPI to process it.
188 	 */
189 	smp_mb();
190 
191 	/*
192 	 * It's ok to use list_for_each_rcu() here even though we may
193 	 * delete 'pos', since list_del_rcu() doesn't clear ->next
194 	 */
195 	list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
196 		int refs;
197 		void (*func) (void *info);
198 
199 		/*
200 		 * Since we walk the list without any locks, we might
201 		 * see an entry that was completed, removed from the
202 		 * list and is in the process of being reused.
203 		 *
204 		 * We must check that the cpu is in the cpumask before
205 		 * checking the refs, and both must be set before
206 		 * executing the callback on this cpu.
207 		 */
208 
209 		if (!cpumask_test_cpu(cpu, data->cpumask))
210 			continue;
211 
212 		smp_rmb();
213 
214 		if (atomic_read(&data->refs) == 0)
215 			continue;
216 
217 		func = data->csd.func;			/* for later warn */
218 		data->csd.func(data->csd.info);
219 
220 		/*
221 		 * If the cpu mask is not still set then it enabled interrupts,
222 		 * we took another smp interrupt, and executed the function
223 		 * twice on this cpu.  In theory that copy decremented refs.
224 		 */
225 		if (!cpumask_test_and_clear_cpu(cpu, data->cpumask)) {
226 			WARN(1, "%pS enabled interrupts and double executed\n",
227 			     func);
228 			continue;
229 		}
230 
231 		refs = atomic_dec_return(&data->refs);
232 		WARN_ON(refs < 0);
233 
234 		if (refs)
235 			continue;
236 
237 		WARN_ON(!cpumask_empty(data->cpumask));
238 
239 		raw_spin_lock(&call_function.lock);
240 		list_del_rcu(&data->csd.list);
241 		raw_spin_unlock(&call_function.lock);
242 
243 		csd_unlock(&data->csd);
244 	}
245 
246 }
247 
248 /*
249  * Invoked by arch to handle an IPI for call function single. Must be
250  * called from the arch with interrupts disabled.
251  */
252 void generic_smp_call_function_single_interrupt(void)
253 {
254 	struct call_single_queue *q = &__get_cpu_var(call_single_queue);
255 	unsigned int data_flags;
256 	LIST_HEAD(list);
257 
258 	/*
259 	 * Shouldn't receive this interrupt on a cpu that is not yet online.
260 	 */
261 	WARN_ON_ONCE(!cpu_online(smp_processor_id()));
262 
263 	raw_spin_lock(&q->lock);
264 	list_replace_init(&q->list, &list);
265 	raw_spin_unlock(&q->lock);
266 
267 	while (!list_empty(&list)) {
268 		struct call_single_data *data;
269 
270 		data = list_entry(list.next, struct call_single_data, list);
271 		list_del(&data->list);
272 
273 		/*
274 		 * 'data' can be invalid after this call if flags == 0
275 		 * (when called through generic_exec_single()),
276 		 * so save them away before making the call:
277 		 */
278 		data_flags = data->flags;
279 
280 		data->func(data->info);
281 
282 		/*
283 		 * Unlocked CSDs are valid through generic_exec_single():
284 		 */
285 		if (data_flags & CSD_FLAG_LOCK)
286 			csd_unlock(data);
287 	}
288 }
289 
290 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
291 
292 /*
293  * smp_call_function_single - Run a function on a specific CPU
294  * @func: The function to run. This must be fast and non-blocking.
295  * @info: An arbitrary pointer to pass to the function.
296  * @wait: If true, wait until function has completed on other CPUs.
297  *
298  * Returns 0 on success, else a negative status code.
299  */
300 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
301 			     int wait)
302 {
303 	struct call_single_data d = {
304 		.flags = 0,
305 	};
306 	unsigned long flags;
307 	int this_cpu;
308 	int err = 0;
309 
310 	/*
311 	 * prevent preemption and reschedule on another processor,
312 	 * as well as CPU removal
313 	 */
314 	this_cpu = get_cpu();
315 
316 	/*
317 	 * Can deadlock when called with interrupts disabled.
318 	 * We allow cpu's that are not yet online though, as no one else can
319 	 * send smp call function interrupt to this cpu and as such deadlocks
320 	 * can't happen.
321 	 */
322 	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
323 		     && !oops_in_progress);
324 
325 	if (cpu == this_cpu) {
326 		local_irq_save(flags);
327 		func(info);
328 		local_irq_restore(flags);
329 	} else {
330 		if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
331 			struct call_single_data *data = &d;
332 
333 			if (!wait)
334 				data = &__get_cpu_var(csd_data);
335 
336 			csd_lock(data);
337 
338 			data->func = func;
339 			data->info = info;
340 			generic_exec_single(cpu, data, wait);
341 		} else {
342 			err = -ENXIO;	/* CPU not online */
343 		}
344 	}
345 
346 	put_cpu();
347 
348 	return err;
349 }
350 EXPORT_SYMBOL(smp_call_function_single);
351 
352 /*
353  * smp_call_function_any - Run a function on any of the given cpus
354  * @mask: The mask of cpus it can run on.
355  * @func: The function to run. This must be fast and non-blocking.
356  * @info: An arbitrary pointer to pass to the function.
357  * @wait: If true, wait until function has completed.
358  *
359  * Returns 0 on success, else a negative status code (if no cpus were online).
360  * Note that @wait will be implicitly turned on in case of allocation failures,
361  * since we fall back to on-stack allocation.
362  *
363  * Selection preference:
364  *	1) current cpu if in @mask
365  *	2) any cpu of current node if in @mask
366  *	3) any other online cpu in @mask
367  */
368 int smp_call_function_any(const struct cpumask *mask,
369 			  smp_call_func_t func, void *info, int wait)
370 {
371 	unsigned int cpu;
372 	const struct cpumask *nodemask;
373 	int ret;
374 
375 	/* Try for same CPU (cheapest) */
376 	cpu = get_cpu();
377 	if (cpumask_test_cpu(cpu, mask))
378 		goto call;
379 
380 	/* Try for same node. */
381 	nodemask = cpumask_of_node(cpu_to_node(cpu));
382 	for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
383 	     cpu = cpumask_next_and(cpu, nodemask, mask)) {
384 		if (cpu_online(cpu))
385 			goto call;
386 	}
387 
388 	/* Any online will do: smp_call_function_single handles nr_cpu_ids. */
389 	cpu = cpumask_any_and(mask, cpu_online_mask);
390 call:
391 	ret = smp_call_function_single(cpu, func, info, wait);
392 	put_cpu();
393 	return ret;
394 }
395 EXPORT_SYMBOL_GPL(smp_call_function_any);
396 
397 /**
398  * __smp_call_function_single(): Run a function on a specific CPU
399  * @cpu: The CPU to run on.
400  * @data: Pre-allocated and setup data structure
401  * @wait: If true, wait until function has completed on specified CPU.
402  *
403  * Like smp_call_function_single(), but allow caller to pass in a
404  * pre-allocated data structure. Useful for embedding @data inside
405  * other structures, for instance.
406  */
407 void __smp_call_function_single(int cpu, struct call_single_data *data,
408 				int wait)
409 {
410 	unsigned int this_cpu;
411 	unsigned long flags;
412 
413 	this_cpu = get_cpu();
414 	/*
415 	 * Can deadlock when called with interrupts disabled.
416 	 * We allow cpu's that are not yet online though, as no one else can
417 	 * send smp call function interrupt to this cpu and as such deadlocks
418 	 * can't happen.
419 	 */
420 	WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
421 		     && !oops_in_progress);
422 
423 	if (cpu == this_cpu) {
424 		local_irq_save(flags);
425 		data->func(data->info);
426 		local_irq_restore(flags);
427 	} else {
428 		csd_lock(data);
429 		generic_exec_single(cpu, data, wait);
430 	}
431 	put_cpu();
432 }
433 
434 /**
435  * smp_call_function_many(): Run a function on a set of other CPUs.
436  * @mask: The set of cpus to run on (only runs on online subset).
437  * @func: The function to run. This must be fast and non-blocking.
438  * @info: An arbitrary pointer to pass to the function.
439  * @wait: If true, wait (atomically) until function has completed
440  *        on other CPUs.
441  *
442  * If @wait is true, then returns once @func has returned.
443  *
444  * You must not call this function with disabled interrupts or from a
445  * hardware interrupt handler or from a bottom half handler. Preemption
446  * must be disabled when calling this function.
447  */
448 void smp_call_function_many(const struct cpumask *mask,
449 			    smp_call_func_t func, void *info, bool wait)
450 {
451 	struct call_function_data *data;
452 	unsigned long flags;
453 	int cpu, next_cpu, this_cpu = smp_processor_id();
454 
455 	/*
456 	 * Can deadlock when called with interrupts disabled.
457 	 * We allow cpu's that are not yet online though, as no one else can
458 	 * send smp call function interrupt to this cpu and as such deadlocks
459 	 * can't happen.
460 	 */
461 	WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
462 		     && !oops_in_progress && !early_boot_irqs_disabled);
463 
464 	/* So, what's a CPU they want? Ignoring this one. */
465 	cpu = cpumask_first_and(mask, cpu_online_mask);
466 	if (cpu == this_cpu)
467 		cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
468 
469 	/* No online cpus?  We're done. */
470 	if (cpu >= nr_cpu_ids)
471 		return;
472 
473 	/* Do we have another CPU which isn't us? */
474 	next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
475 	if (next_cpu == this_cpu)
476 		next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
477 
478 	/* Fastpath: do that cpu by itself. */
479 	if (next_cpu >= nr_cpu_ids) {
480 		smp_call_function_single(cpu, func, info, wait);
481 		return;
482 	}
483 
484 	data = &__get_cpu_var(cfd_data);
485 	csd_lock(&data->csd);
486 	BUG_ON(atomic_read(&data->refs) || !cpumask_empty(data->cpumask));
487 
488 	data->csd.func = func;
489 	data->csd.info = info;
490 	cpumask_and(data->cpumask, mask, cpu_online_mask);
491 	cpumask_clear_cpu(this_cpu, data->cpumask);
492 
493 	/*
494 	 * To ensure the interrupt handler gets an complete view
495 	 * we order the cpumask and refs writes and order the read
496 	 * of them in the interrupt handler.  In addition we may
497 	 * only clear our own cpu bit from the mask.
498 	 */
499 	smp_wmb();
500 
501 	atomic_set(&data->refs, cpumask_weight(data->cpumask));
502 
503 	raw_spin_lock_irqsave(&call_function.lock, flags);
504 	/*
505 	 * Place entry at the _HEAD_ of the list, so that any cpu still
506 	 * observing the entry in generic_smp_call_function_interrupt()
507 	 * will not miss any other list entries:
508 	 */
509 	list_add_rcu(&data->csd.list, &call_function.queue);
510 	raw_spin_unlock_irqrestore(&call_function.lock, flags);
511 
512 	/*
513 	 * Make the list addition visible before sending the ipi.
514 	 * (IPIs must obey or appear to obey normal Linux cache
515 	 * coherency rules -- see comment in generic_exec_single).
516 	 */
517 	smp_mb();
518 
519 	/* Send a message to all CPUs in the map */
520 	arch_send_call_function_ipi_mask(data->cpumask);
521 
522 	/* Optionally wait for the CPUs to complete */
523 	if (wait)
524 		csd_lock_wait(&data->csd);
525 }
526 EXPORT_SYMBOL(smp_call_function_many);
527 
528 /**
529  * smp_call_function(): Run a function on all other CPUs.
530  * @func: The function to run. This must be fast and non-blocking.
531  * @info: An arbitrary pointer to pass to the function.
532  * @wait: If true, wait (atomically) until function has completed
533  *        on other CPUs.
534  *
535  * Returns 0.
536  *
537  * If @wait is true, then returns once @func has returned; otherwise
538  * it returns just before the target cpu calls @func.
539  *
540  * You must not call this function with disabled interrupts or from a
541  * hardware interrupt handler or from a bottom half handler.
542  */
543 int smp_call_function(smp_call_func_t func, void *info, int wait)
544 {
545 	preempt_disable();
546 	smp_call_function_many(cpu_online_mask, func, info, wait);
547 	preempt_enable();
548 
549 	return 0;
550 }
551 EXPORT_SYMBOL(smp_call_function);
552 
553 void ipi_call_lock(void)
554 {
555 	raw_spin_lock(&call_function.lock);
556 }
557 
558 void ipi_call_unlock(void)
559 {
560 	raw_spin_unlock(&call_function.lock);
561 }
562 
563 void ipi_call_lock_irq(void)
564 {
565 	raw_spin_lock_irq(&call_function.lock);
566 }
567 
568 void ipi_call_unlock_irq(void)
569 {
570 	raw_spin_unlock_irq(&call_function.lock);
571 }
572 #endif /* USE_GENERIC_SMP_HELPERS */
573 
574 /*
575  * Call a function on all processors.  May be used during early boot while
576  * early_boot_irqs_disabled is set.  Use local_irq_save/restore() instead
577  * of local_irq_disable/enable().
578  */
579 int on_each_cpu(void (*func) (void *info), void *info, int wait)
580 {
581 	unsigned long flags;
582 	int ret = 0;
583 
584 	preempt_disable();
585 	ret = smp_call_function(func, info, wait);
586 	local_irq_save(flags);
587 	func(info);
588 	local_irq_restore(flags);
589 	preempt_enable();
590 	return ret;
591 }
592 EXPORT_SYMBOL(on_each_cpu);
593