Loading...
1/*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
5 */
6#include <linux/rcupdate.h>
7#include <linux/rculist.h>
8#include <linux/kernel.h>
9#include <linux/export.h>
10#include <linux/percpu.h>
11#include <linux/init.h>
12#include <linux/gfp.h>
13#include <linux/smp.h>
14#include <linux/cpu.h>
15
16#include "smpboot.h"
17
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);
1/*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
5 */
6#include <linux/rcupdate.h>
7#include <linux/rculist.h>
8#include <linux/kernel.h>
9#include <linux/export.h>
10#include <linux/percpu.h>
11#include <linux/init.h>
12#include <linux/gfp.h>
13#include <linux/smp.h>
14#include <linux/cpu.h>
15
16#include "smpboot.h"
17
18#ifdef CONFIG_USE_GENERIC_SMP_HELPERS
19static struct {
20 struct list_head queue;
21 raw_spinlock_t lock;
22} call_function __cacheline_aligned_in_smp =
23 {
24 .queue = LIST_HEAD_INIT(call_function.queue),
25 .lock = __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
26 };
27
28enum {
29 CSD_FLAG_LOCK = 0x01,
30};
31
32struct call_function_data {
33 struct call_single_data csd;
34 atomic_t refs;
35 cpumask_var_t cpumask;
36};
37
38static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
39
40struct call_single_queue {
41 struct list_head list;
42 raw_spinlock_t lock;
43};
44
45static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
46
47static int
48hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
49{
50 long cpu = (long)hcpu;
51 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
52
53 switch (action) {
54 case CPU_UP_PREPARE:
55 case CPU_UP_PREPARE_FROZEN:
56 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
57 cpu_to_node(cpu)))
58 return notifier_from_errno(-ENOMEM);
59 break;
60
61#ifdef CONFIG_HOTPLUG_CPU
62 case CPU_UP_CANCELED:
63 case CPU_UP_CANCELED_FROZEN:
64
65 case CPU_DEAD:
66 case CPU_DEAD_FROZEN:
67 free_cpumask_var(cfd->cpumask);
68 break;
69#endif
70 };
71
72 return NOTIFY_OK;
73}
74
75static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
76 .notifier_call = hotplug_cfd,
77};
78
79void __init call_function_init(void)
80{
81 void *cpu = (void *)(long)smp_processor_id();
82 int i;
83
84 for_each_possible_cpu(i) {
85 struct call_single_queue *q = &per_cpu(call_single_queue, i);
86
87 raw_spin_lock_init(&q->lock);
88 INIT_LIST_HEAD(&q->list);
89 }
90
91 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
92 register_cpu_notifier(&hotplug_cfd_notifier);
93}
94
95/*
96 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
97 *
98 * For non-synchronous ipi calls the csd can still be in use by the
99 * previous function call. For multi-cpu calls its even more interesting
100 * as we'll have to ensure no other cpu is observing our csd.
101 */
102static void csd_lock_wait(struct call_single_data *data)
103{
104 while (data->flags & CSD_FLAG_LOCK)
105 cpu_relax();
106}
107
108static void csd_lock(struct call_single_data *data)
109{
110 csd_lock_wait(data);
111 data->flags = CSD_FLAG_LOCK;
112
113 /*
114 * prevent CPU from reordering the above assignment
115 * to ->flags with any subsequent assignments to other
116 * fields of the specified call_single_data structure:
117 */
118 smp_mb();
119}
120
121static void csd_unlock(struct call_single_data *data)
122{
123 WARN_ON(!(data->flags & CSD_FLAG_LOCK));
124
125 /*
126 * ensure we're all done before releasing data:
127 */
128 smp_mb();
129
130 data->flags &= ~CSD_FLAG_LOCK;
131}
132
133/*
134 * Insert a previously allocated call_single_data element
135 * for execution on the given CPU. data must already have
136 * ->func, ->info, and ->flags set.
137 */
138static
139void generic_exec_single(int cpu, struct call_single_data *data, int wait)
140{
141 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
142 unsigned long flags;
143 int ipi;
144
145 raw_spin_lock_irqsave(&dst->lock, flags);
146 ipi = list_empty(&dst->list);
147 list_add_tail(&data->list, &dst->list);
148 raw_spin_unlock_irqrestore(&dst->lock, flags);
149
150 /*
151 * The list addition should be visible before sending the IPI
152 * handler locks the list to pull the entry off it because of
153 * normal cache coherency rules implied by spinlocks.
154 *
155 * If IPIs can go out of order to the cache coherency protocol
156 * in an architecture, sufficient synchronisation should be added
157 * to arch code to make it appear to obey cache coherency WRT
158 * locking and barrier primitives. Generic code isn't really
159 * equipped to do the right thing...
160 */
161 if (ipi)
162 arch_send_call_function_single_ipi(cpu);
163
164 if (wait)
165 csd_lock_wait(data);
166}
167
168/*
169 * Invoked by arch to handle an IPI for call function. Must be called with
170 * interrupts disabled.
171 */
172void generic_smp_call_function_interrupt(void)
173{
174 struct call_function_data *data;
175 int cpu = smp_processor_id();
176
177 /*
178 * Shouldn't receive this interrupt on a cpu that is not yet online.
179 */
180 WARN_ON_ONCE(!cpu_online(cpu));
181
182 /*
183 * Ensure entry is visible on call_function_queue after we have
184 * entered the IPI. See comment in smp_call_function_many.
185 * If we don't have this, then we may miss an entry on the list
186 * and never get another IPI to process it.
187 */
188 smp_mb();
189
190 /*
191 * It's ok to use list_for_each_rcu() here even though we may
192 * delete 'pos', since list_del_rcu() doesn't clear ->next
193 */
194 list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
195 int refs;
196 smp_call_func_t func;
197
198 /*
199 * Since we walk the list without any locks, we might
200 * see an entry that was completed, removed from the
201 * list and is in the process of being reused.
202 *
203 * We must check that the cpu is in the cpumask before
204 * checking the refs, and both must be set before
205 * executing the callback on this cpu.
206 */
207
208 if (!cpumask_test_cpu(cpu, data->cpumask))
209 continue;
210
211 smp_rmb();
212
213 if (atomic_read(&data->refs) == 0)
214 continue;
215
216 func = data->csd.func; /* save for later warn */
217 func(data->csd.info);
218
219 /*
220 * If the cpu mask is not still set then func enabled
221 * interrupts (BUG), and this cpu took another smp call
222 * function interrupt and executed func(info) twice
223 * on this cpu. That nested execution decremented refs.
224 */
225 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask)) {
226 WARN(1, "%pf enabled interrupts and double executed\n", func);
227 continue;
228 }
229
230 refs = atomic_dec_return(&data->refs);
231 WARN_ON(refs < 0);
232
233 if (refs)
234 continue;
235
236 WARN_ON(!cpumask_empty(data->cpumask));
237
238 raw_spin_lock(&call_function.lock);
239 list_del_rcu(&data->csd.list);
240 raw_spin_unlock(&call_function.lock);
241
242 csd_unlock(&data->csd);
243 }
244
245}
246
247/*
248 * Invoked by arch to handle an IPI for call function single. Must be
249 * called from the arch with interrupts disabled.
250 */
251void generic_smp_call_function_single_interrupt(void)
252{
253 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
254 unsigned int data_flags;
255 LIST_HEAD(list);
256
257 /*
258 * Shouldn't receive this interrupt on a cpu that is not yet online.
259 */
260 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
261
262 raw_spin_lock(&q->lock);
263 list_replace_init(&q->list, &list);
264 raw_spin_unlock(&q->lock);
265
266 while (!list_empty(&list)) {
267 struct call_single_data *data;
268
269 data = list_entry(list.next, struct call_single_data, list);
270 list_del(&data->list);
271
272 /*
273 * 'data' can be invalid after this call if flags == 0
274 * (when called through generic_exec_single()),
275 * so save them away before making the call:
276 */
277 data_flags = data->flags;
278
279 data->func(data->info);
280
281 /*
282 * Unlocked CSDs are valid through generic_exec_single():
283 */
284 if (data_flags & CSD_FLAG_LOCK)
285 csd_unlock(data);
286 }
287}
288
289static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
290
291/*
292 * smp_call_function_single - Run a function on a specific CPU
293 * @func: The function to run. This must be fast and non-blocking.
294 * @info: An arbitrary pointer to pass to the function.
295 * @wait: If true, wait until function has completed on other CPUs.
296 *
297 * Returns 0 on success, else a negative status code.
298 */
299int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
300 int wait)
301{
302 struct call_single_data d = {
303 .flags = 0,
304 };
305 unsigned long flags;
306 int this_cpu;
307 int err = 0;
308
309 /*
310 * prevent preemption and reschedule on another processor,
311 * as well as CPU removal
312 */
313 this_cpu = get_cpu();
314
315 /*
316 * Can deadlock when called with interrupts disabled.
317 * We allow cpu's that are not yet online though, as no one else can
318 * send smp call function interrupt to this cpu and as such deadlocks
319 * can't happen.
320 */
321 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
322 && !oops_in_progress);
323
324 if (cpu == this_cpu) {
325 local_irq_save(flags);
326 func(info);
327 local_irq_restore(flags);
328 } else {
329 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
330 struct call_single_data *data = &d;
331
332 if (!wait)
333 data = &__get_cpu_var(csd_data);
334
335 csd_lock(data);
336
337 data->func = func;
338 data->info = info;
339 generic_exec_single(cpu, data, wait);
340 } else {
341 err = -ENXIO; /* CPU not online */
342 }
343 }
344
345 put_cpu();
346
347 return err;
348}
349EXPORT_SYMBOL(smp_call_function_single);
350
351/*
352 * smp_call_function_any - Run a function on any of the given cpus
353 * @mask: The mask of cpus it can run on.
354 * @func: The function to run. This must be fast and non-blocking.
355 * @info: An arbitrary pointer to pass to the function.
356 * @wait: If true, wait until function has completed.
357 *
358 * Returns 0 on success, else a negative status code (if no cpus were online).
359 * Note that @wait will be implicitly turned on in case of allocation failures,
360 * since we fall back to on-stack allocation.
361 *
362 * Selection preference:
363 * 1) current cpu if in @mask
364 * 2) any cpu of current node if in @mask
365 * 3) any other online cpu in @mask
366 */
367int smp_call_function_any(const struct cpumask *mask,
368 smp_call_func_t func, void *info, int wait)
369{
370 unsigned int cpu;
371 const struct cpumask *nodemask;
372 int ret;
373
374 /* Try for same CPU (cheapest) */
375 cpu = get_cpu();
376 if (cpumask_test_cpu(cpu, mask))
377 goto call;
378
379 /* Try for same node. */
380 nodemask = cpumask_of_node(cpu_to_node(cpu));
381 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
382 cpu = cpumask_next_and(cpu, nodemask, mask)) {
383 if (cpu_online(cpu))
384 goto call;
385 }
386
387 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
388 cpu = cpumask_any_and(mask, cpu_online_mask);
389call:
390 ret = smp_call_function_single(cpu, func, info, wait);
391 put_cpu();
392 return ret;
393}
394EXPORT_SYMBOL_GPL(smp_call_function_any);
395
396/**
397 * __smp_call_function_single(): Run a function on a specific CPU
398 * @cpu: The CPU to run on.
399 * @data: Pre-allocated and setup data structure
400 * @wait: If true, wait until function has completed on specified CPU.
401 *
402 * Like smp_call_function_single(), but allow caller to pass in a
403 * pre-allocated data structure. Useful for embedding @data inside
404 * other structures, for instance.
405 */
406void __smp_call_function_single(int cpu, struct call_single_data *data,
407 int wait)
408{
409 unsigned int this_cpu;
410 unsigned long flags;
411
412 this_cpu = get_cpu();
413 /*
414 * Can deadlock when called with interrupts disabled.
415 * We allow cpu's that are not yet online though, as no one else can
416 * send smp call function interrupt to this cpu and as such deadlocks
417 * can't happen.
418 */
419 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
420 && !oops_in_progress);
421
422 if (cpu == this_cpu) {
423 local_irq_save(flags);
424 data->func(data->info);
425 local_irq_restore(flags);
426 } else {
427 csd_lock(data);
428 generic_exec_single(cpu, data, wait);
429 }
430 put_cpu();
431}
432
433/**
434 * smp_call_function_many(): Run a function on a set of other CPUs.
435 * @mask: The set of cpus to run on (only runs on online subset).
436 * @func: The function to run. This must be fast and non-blocking.
437 * @info: An arbitrary pointer to pass to the function.
438 * @wait: If true, wait (atomically) until function has completed
439 * on other CPUs.
440 *
441 * If @wait is true, then returns once @func has returned.
442 *
443 * You must not call this function with disabled interrupts or from a
444 * hardware interrupt handler or from a bottom half handler. Preemption
445 * must be disabled when calling this function.
446 */
447void smp_call_function_many(const struct cpumask *mask,
448 smp_call_func_t func, void *info, bool wait)
449{
450 struct call_function_data *data;
451 unsigned long flags;
452 int refs, cpu, next_cpu, this_cpu = smp_processor_id();
453
454 /*
455 * Can deadlock when called with interrupts disabled.
456 * We allow cpu's that are not yet online though, as no one else can
457 * send smp call function interrupt to this cpu and as such deadlocks
458 * can't happen.
459 */
460 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
461 && !oops_in_progress && !early_boot_irqs_disabled);
462
463 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
464 cpu = cpumask_first_and(mask, cpu_online_mask);
465 if (cpu == this_cpu)
466 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
467
468 /* No online cpus? We're done. */
469 if (cpu >= nr_cpu_ids)
470 return;
471
472 /* Do we have another CPU which isn't us? */
473 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
474 if (next_cpu == this_cpu)
475 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
476
477 /* Fastpath: do that cpu by itself. */
478 if (next_cpu >= nr_cpu_ids) {
479 smp_call_function_single(cpu, func, info, wait);
480 return;
481 }
482
483 data = &__get_cpu_var(cfd_data);
484 csd_lock(&data->csd);
485
486 /* This BUG_ON verifies our reuse assertions and can be removed */
487 BUG_ON(atomic_read(&data->refs) || !cpumask_empty(data->cpumask));
488
489 /*
490 * The global call function queue list add and delete are protected
491 * by a lock, but the list is traversed without any lock, relying
492 * on the rcu list add and delete to allow safe concurrent traversal.
493 * We reuse the call function data without waiting for any grace
494 * period after some other cpu removes it from the global queue.
495 * This means a cpu might find our data block as it is being
496 * filled out.
497 *
498 * We hold off the interrupt handler on the other cpu by
499 * ordering our writes to the cpu mask vs our setting of the
500 * refs counter. We assert only the cpu owning the data block
501 * will set a bit in cpumask, and each bit will only be cleared
502 * by the subject cpu. Each cpu must first find its bit is
503 * set and then check that refs is set indicating the element is
504 * ready to be processed, otherwise it must skip the entry.
505 *
506 * On the previous iteration refs was set to 0 by another cpu.
507 * To avoid the use of transitivity, set the counter to 0 here
508 * so the wmb will pair with the rmb in the interrupt handler.
509 */
510 atomic_set(&data->refs, 0); /* convert 3rd to 1st party write */
511
512 data->csd.func = func;
513 data->csd.info = info;
514
515 /* Ensure 0 refs is visible before mask. Also orders func and info */
516 smp_wmb();
517
518 /* We rely on the "and" being processed before the store */
519 cpumask_and(data->cpumask, mask, cpu_online_mask);
520 cpumask_clear_cpu(this_cpu, data->cpumask);
521 refs = cpumask_weight(data->cpumask);
522
523 /* Some callers race with other cpus changing the passed mask */
524 if (unlikely(!refs)) {
525 csd_unlock(&data->csd);
526 return;
527 }
528
529 raw_spin_lock_irqsave(&call_function.lock, flags);
530 /*
531 * Place entry at the _HEAD_ of the list, so that any cpu still
532 * observing the entry in generic_smp_call_function_interrupt()
533 * will not miss any other list entries:
534 */
535 list_add_rcu(&data->csd.list, &call_function.queue);
536 /*
537 * We rely on the wmb() in list_add_rcu to complete our writes
538 * to the cpumask before this write to refs, which indicates
539 * data is on the list and is ready to be processed.
540 */
541 atomic_set(&data->refs, refs);
542 raw_spin_unlock_irqrestore(&call_function.lock, flags);
543
544 /*
545 * Make the list addition visible before sending the ipi.
546 * (IPIs must obey or appear to obey normal Linux cache
547 * coherency rules -- see comment in generic_exec_single).
548 */
549 smp_mb();
550
551 /* Send a message to all CPUs in the map */
552 arch_send_call_function_ipi_mask(data->cpumask);
553
554 /* Optionally wait for the CPUs to complete */
555 if (wait)
556 csd_lock_wait(&data->csd);
557}
558EXPORT_SYMBOL(smp_call_function_many);
559
560/**
561 * smp_call_function(): Run a function on all other CPUs.
562 * @func: The function to run. This must be fast and non-blocking.
563 * @info: An arbitrary pointer to pass to the function.
564 * @wait: If true, wait (atomically) until function has completed
565 * on other CPUs.
566 *
567 * Returns 0.
568 *
569 * If @wait is true, then returns once @func has returned; otherwise
570 * it returns just before the target cpu calls @func.
571 *
572 * You must not call this function with disabled interrupts or from a
573 * hardware interrupt handler or from a bottom half handler.
574 */
575int smp_call_function(smp_call_func_t func, void *info, int wait)
576{
577 preempt_disable();
578 smp_call_function_many(cpu_online_mask, func, info, wait);
579 preempt_enable();
580
581 return 0;
582}
583EXPORT_SYMBOL(smp_call_function);
584
585void ipi_call_lock(void)
586{
587 raw_spin_lock(&call_function.lock);
588}
589
590void ipi_call_unlock(void)
591{
592 raw_spin_unlock(&call_function.lock);
593}
594
595void ipi_call_lock_irq(void)
596{
597 raw_spin_lock_irq(&call_function.lock);
598}
599
600void ipi_call_unlock_irq(void)
601{
602 raw_spin_unlock_irq(&call_function.lock);
603}
604#endif /* USE_GENERIC_SMP_HELPERS */
605
606/* Setup configured maximum number of CPUs to activate */
607unsigned int setup_max_cpus = NR_CPUS;
608EXPORT_SYMBOL(setup_max_cpus);
609
610
611/*
612 * Setup routine for controlling SMP activation
613 *
614 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
615 * activation entirely (the MPS table probe still happens, though).
616 *
617 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
618 * greater than 0, limits the maximum number of CPUs activated in
619 * SMP mode to <NUM>.
620 */
621
622void __weak arch_disable_smp_support(void) { }
623
624static int __init nosmp(char *str)
625{
626 setup_max_cpus = 0;
627 arch_disable_smp_support();
628
629 return 0;
630}
631
632early_param("nosmp", nosmp);
633
634/* this is hard limit */
635static int __init nrcpus(char *str)
636{
637 int nr_cpus;
638
639 get_option(&str, &nr_cpus);
640 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
641 nr_cpu_ids = nr_cpus;
642
643 return 0;
644}
645
646early_param("nr_cpus", nrcpus);
647
648static int __init maxcpus(char *str)
649{
650 get_option(&str, &setup_max_cpus);
651 if (setup_max_cpus == 0)
652 arch_disable_smp_support();
653
654 return 0;
655}
656
657early_param("maxcpus", maxcpus);
658
659/* Setup number of possible processor ids */
660int nr_cpu_ids __read_mostly = NR_CPUS;
661EXPORT_SYMBOL(nr_cpu_ids);
662
663/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
664void __init setup_nr_cpu_ids(void)
665{
666 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
667}
668
669/* Called by boot processor to activate the rest. */
670void __init smp_init(void)
671{
672 unsigned int cpu;
673
674 idle_threads_init();
675
676 /* FIXME: This should be done in userspace --RR */
677 for_each_present_cpu(cpu) {
678 if (num_online_cpus() >= setup_max_cpus)
679 break;
680 if (!cpu_online(cpu))
681 cpu_up(cpu);
682 }
683
684 /* Any cleanup work */
685 printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());
686 smp_cpus_done(setup_max_cpus);
687}
688
689/*
690 * Call a function on all processors. May be used during early boot while
691 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
692 * of local_irq_disable/enable().
693 */
694int on_each_cpu(void (*func) (void *info), void *info, int wait)
695{
696 unsigned long flags;
697 int ret = 0;
698
699 preempt_disable();
700 ret = smp_call_function(func, info, wait);
701 local_irq_save(flags);
702 func(info);
703 local_irq_restore(flags);
704 preempt_enable();
705 return ret;
706}
707EXPORT_SYMBOL(on_each_cpu);
708
709/**
710 * on_each_cpu_mask(): Run a function on processors specified by
711 * cpumask, which may include the local processor.
712 * @mask: The set of cpus to run on (only runs on online subset).
713 * @func: The function to run. This must be fast and non-blocking.
714 * @info: An arbitrary pointer to pass to the function.
715 * @wait: If true, wait (atomically) until function has completed
716 * on other CPUs.
717 *
718 * If @wait is true, then returns once @func has returned.
719 *
720 * You must not call this function with disabled interrupts or
721 * from a hardware interrupt handler or from a bottom half handler.
722 */
723void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
724 void *info, bool wait)
725{
726 int cpu = get_cpu();
727
728 smp_call_function_many(mask, func, info, wait);
729 if (cpumask_test_cpu(cpu, mask)) {
730 local_irq_disable();
731 func(info);
732 local_irq_enable();
733 }
734 put_cpu();
735}
736EXPORT_SYMBOL(on_each_cpu_mask);
737
738/*
739 * on_each_cpu_cond(): Call a function on each processor for which
740 * the supplied function cond_func returns true, optionally waiting
741 * for all the required CPUs to finish. This may include the local
742 * processor.
743 * @cond_func: A callback function that is passed a cpu id and
744 * the the info parameter. The function is called
745 * with preemption disabled. The function should
746 * return a blooean value indicating whether to IPI
747 * the specified CPU.
748 * @func: The function to run on all applicable CPUs.
749 * This must be fast and non-blocking.
750 * @info: An arbitrary pointer to pass to both functions.
751 * @wait: If true, wait (atomically) until function has
752 * completed on other CPUs.
753 * @gfp_flags: GFP flags to use when allocating the cpumask
754 * used internally by the function.
755 *
756 * The function might sleep if the GFP flags indicates a non
757 * atomic allocation is allowed.
758 *
759 * Preemption is disabled to protect against CPUs going offline but not online.
760 * CPUs going online during the call will not be seen or sent an IPI.
761 *
762 * You must not call this function with disabled interrupts or
763 * from a hardware interrupt handler or from a bottom half handler.
764 */
765void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
766 smp_call_func_t func, void *info, bool wait,
767 gfp_t gfp_flags)
768{
769 cpumask_var_t cpus;
770 int cpu, ret;
771
772 might_sleep_if(gfp_flags & __GFP_WAIT);
773
774 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
775 preempt_disable();
776 for_each_online_cpu(cpu)
777 if (cond_func(cpu, info))
778 cpumask_set_cpu(cpu, cpus);
779 on_each_cpu_mask(cpus, func, info, wait);
780 preempt_enable();
781 free_cpumask_var(cpus);
782 } else {
783 /*
784 * No free cpumask, bother. No matter, we'll
785 * just have to IPI them one by one.
786 */
787 preempt_disable();
788 for_each_online_cpu(cpu)
789 if (cond_func(cpu, info)) {
790 ret = smp_call_function_single(cpu, func,
791 info, wait);
792 WARN_ON_ONCE(!ret);
793 }
794 preempt_enable();
795 }
796}
797EXPORT_SYMBOL(on_each_cpu_cond);
798
799static void do_nothing(void *unused)
800{
801}
802
803/**
804 * kick_all_cpus_sync - Force all cpus out of idle
805 *
806 * Used to synchronize the update of pm_idle function pointer. It's
807 * called after the pointer is updated and returns after the dummy
808 * callback function has been executed on all cpus. The execution of
809 * the function can only happen on the remote cpus after they have
810 * left the idle function which had been called via pm_idle function
811 * pointer. So it's guaranteed that nothing uses the previous pointer
812 * anymore.
813 */
814void kick_all_cpus_sync(void)
815{
816 /* Make sure the change is visible before we kick the cpus */
817 smp_mb();
818 smp_call_function(do_nothing, NULL, 1);
819}
820EXPORT_SYMBOL_GPL(kick_all_cpus_sync);