blob: 399905fdfa3f85bbc167fd468d56166fb7b12b5c
1 | /* |
2 | * Generic helpers for smp ipi calls |
3 | * |
4 | * (C) Jens Axboe <jens.axboe@oracle.com> 2008 |
5 | */ |
6 | #include <linux/irq_work.h> |
7 | #include <linux/rcupdate.h> |
8 | #include <linux/rculist.h> |
9 | #include <linux/kernel.h> |
10 | #include <linux/export.h> |
11 | #include <linux/percpu.h> |
12 | #include <linux/init.h> |
13 | #include <linux/gfp.h> |
14 | #include <linux/smp.h> |
15 | #include <linux/cpu.h> |
16 | #include <linux/sched.h> |
17 | #include <linux/hypervisor.h> |
18 | |
19 | #include "smpboot.h" |
20 | |
21 | enum { |
22 | CSD_FLAG_LOCK = 0x01, |
23 | CSD_FLAG_SYNCHRONOUS = 0x02, |
24 | }; |
25 | |
26 | struct call_function_data { |
27 | struct call_single_data __percpu *csd; |
28 | cpumask_var_t cpumask; |
29 | }; |
30 | |
31 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data); |
32 | |
33 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue); |
34 | |
35 | static void flush_smp_call_function_queue(bool warn_cpu_offline); |
36 | |
37 | int smpcfd_prepare_cpu(unsigned int cpu) |
38 | { |
39 | struct call_function_data *cfd = &per_cpu(cfd_data, cpu); |
40 | |
41 | if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL, |
42 | cpu_to_node(cpu))) |
43 | return -ENOMEM; |
44 | cfd->csd = alloc_percpu(struct call_single_data); |
45 | if (!cfd->csd) { |
46 | free_cpumask_var(cfd->cpumask); |
47 | return -ENOMEM; |
48 | } |
49 | |
50 | return 0; |
51 | } |
52 | |
53 | int smpcfd_dead_cpu(unsigned int cpu) |
54 | { |
55 | struct call_function_data *cfd = &per_cpu(cfd_data, cpu); |
56 | |
57 | free_cpumask_var(cfd->cpumask); |
58 | free_percpu(cfd->csd); |
59 | return 0; |
60 | } |
61 | |
62 | int smpcfd_dying_cpu(unsigned int cpu) |
63 | { |
64 | /* |
65 | * The IPIs for the smp-call-function callbacks queued by other |
66 | * CPUs might arrive late, either due to hardware latencies or |
67 | * because this CPU disabled interrupts (inside stop-machine) |
68 | * before the IPIs were sent. So flush out any pending callbacks |
69 | * explicitly (without waiting for the IPIs to arrive), to |
70 | * ensure that the outgoing CPU doesn't go offline with work |
71 | * still pending. |
72 | */ |
73 | flush_smp_call_function_queue(false); |
74 | return 0; |
75 | } |
76 | |
77 | void __init call_function_init(void) |
78 | { |
79 | int i; |
80 | |
81 | for_each_possible_cpu(i) |
82 | init_llist_head(&per_cpu(call_single_queue, i)); |
83 | |
84 | smpcfd_prepare_cpu(smp_processor_id()); |
85 | } |
86 | |
87 | /* |
88 | * csd_lock/csd_unlock used to serialize access to per-cpu csd resources |
89 | * |
90 | * For non-synchronous ipi calls the csd can still be in use by the |
91 | * previous function call. For multi-cpu calls its even more interesting |
92 | * as we'll have to ensure no other cpu is observing our csd. |
93 | */ |
94 | static __always_inline void csd_lock_wait(struct call_single_data *csd) |
95 | { |
96 | smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK)); |
97 | } |
98 | |
99 | static __always_inline void csd_lock(struct call_single_data *csd) |
100 | { |
101 | csd_lock_wait(csd); |
102 | csd->flags |= CSD_FLAG_LOCK; |
103 | |
104 | /* |
105 | * prevent CPU from reordering the above assignment |
106 | * to ->flags with any subsequent assignments to other |
107 | * fields of the specified call_single_data structure: |
108 | */ |
109 | smp_wmb(); |
110 | } |
111 | |
112 | static __always_inline void csd_unlock(struct call_single_data *csd) |
113 | { |
114 | WARN_ON(!(csd->flags & CSD_FLAG_LOCK)); |
115 | |
116 | /* |
117 | * ensure we're all done before releasing data: |
118 | */ |
119 | smp_store_release(&csd->flags, 0); |
120 | } |
121 | |
122 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data); |
123 | |
124 | /* |
125 | * Insert a previously allocated call_single_data element |
126 | * for execution on the given CPU. data must already have |
127 | * ->func, ->info, and ->flags set. |
128 | */ |
129 | static int generic_exec_single(int cpu, struct call_single_data *csd, |
130 | smp_call_func_t func, void *info) |
131 | { |
132 | if (cpu == smp_processor_id()) { |
133 | unsigned long flags; |
134 | |
135 | /* |
136 | * We can unlock early even for the synchronous on-stack case, |
137 | * since we're doing this from the same CPU.. |
138 | */ |
139 | csd_unlock(csd); |
140 | local_irq_save(flags); |
141 | func(info); |
142 | local_irq_restore(flags); |
143 | return 0; |
144 | } |
145 | |
146 | |
147 | if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) { |
148 | csd_unlock(csd); |
149 | return -ENXIO; |
150 | } |
151 | |
152 | csd->func = func; |
153 | csd->info = info; |
154 | |
155 | /* |
156 | * The list addition should be visible before sending the IPI |
157 | * handler locks the list to pull the entry off it because of |
158 | * normal cache coherency rules implied by spinlocks. |
159 | * |
160 | * If IPIs can go out of order to the cache coherency protocol |
161 | * in an architecture, sufficient synchronisation should be added |
162 | * to arch code to make it appear to obey cache coherency WRT |
163 | * locking and barrier primitives. Generic code isn't really |
164 | * equipped to do the right thing... |
165 | */ |
166 | if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu))) |
167 | arch_send_call_function_single_ipi(cpu); |
168 | |
169 | return 0; |
170 | } |
171 | |
172 | /** |
173 | * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks |
174 | * |
175 | * Invoked by arch to handle an IPI for call function single. |
176 | * Must be called with interrupts disabled. |
177 | */ |
178 | void generic_smp_call_function_single_interrupt(void) |
179 | { |
180 | flush_smp_call_function_queue(true); |
181 | } |
182 | |
183 | /** |
184 | * flush_smp_call_function_queue - Flush pending smp-call-function callbacks |
185 | * |
186 | * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an |
187 | * offline CPU. Skip this check if set to 'false'. |
188 | * |
189 | * Flush any pending smp-call-function callbacks queued on this CPU. This is |
190 | * invoked by the generic IPI handler, as well as by a CPU about to go offline, |
191 | * to ensure that all pending IPI callbacks are run before it goes completely |
192 | * offline. |
193 | * |
194 | * Loop through the call_single_queue and run all the queued callbacks. |
195 | * Must be called with interrupts disabled. |
196 | */ |
197 | static void flush_smp_call_function_queue(bool warn_cpu_offline) |
198 | { |
199 | struct llist_head *head; |
200 | struct llist_node *entry; |
201 | struct call_single_data *csd, *csd_next; |
202 | static bool warned; |
203 | |
204 | WARN_ON(!irqs_disabled()); |
205 | |
206 | head = this_cpu_ptr(&call_single_queue); |
207 | entry = llist_del_all(head); |
208 | entry = llist_reverse_order(entry); |
209 | |
210 | /* There shouldn't be any pending callbacks on an offline CPU. */ |
211 | if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) && |
212 | !warned && !llist_empty(head))) { |
213 | warned = true; |
214 | WARN(1, "IPI on offline CPU %d\n", smp_processor_id()); |
215 | |
216 | /* |
217 | * We don't have to use the _safe() variant here |
218 | * because we are not invoking the IPI handlers yet. |
219 | */ |
220 | llist_for_each_entry(csd, entry, llist) |
221 | pr_warn("IPI callback %pS sent to offline CPU\n", |
222 | csd->func); |
223 | } |
224 | |
225 | llist_for_each_entry_safe(csd, csd_next, entry, llist) { |
226 | smp_call_func_t func = csd->func; |
227 | void *info = csd->info; |
228 | |
229 | /* Do we wait until *after* callback? */ |
230 | if (csd->flags & CSD_FLAG_SYNCHRONOUS) { |
231 | func(info); |
232 | csd_unlock(csd); |
233 | } else { |
234 | csd_unlock(csd); |
235 | func(info); |
236 | } |
237 | } |
238 | |
239 | /* |
240 | * Handle irq works queued remotely by irq_work_queue_on(). |
241 | * Smp functions above are typically synchronous so they |
242 | * better run first since some other CPUs may be busy waiting |
243 | * for them. |
244 | */ |
245 | irq_work_run(); |
246 | } |
247 | |
248 | /* |
249 | * smp_call_function_single - Run a function on a specific CPU |
250 | * @func: The function to run. This must be fast and non-blocking. |
251 | * @info: An arbitrary pointer to pass to the function. |
252 | * @wait: If true, wait until function has completed on other CPUs. |
253 | * |
254 | * Returns 0 on success, else a negative status code. |
255 | */ |
256 | int smp_call_function_single(int cpu, smp_call_func_t func, void *info, |
257 | int wait) |
258 | { |
259 | struct call_single_data *csd; |
260 | struct call_single_data csd_stack = { .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS }; |
261 | int this_cpu; |
262 | int err; |
263 | |
264 | /* |
265 | * prevent preemption and reschedule on another processor, |
266 | * as well as CPU removal |
267 | */ |
268 | this_cpu = get_cpu(); |
269 | |
270 | /* |
271 | * Can deadlock when called with interrupts disabled. |
272 | * We allow cpu's that are not yet online though, as no one else can |
273 | * send smp call function interrupt to this cpu and as such deadlocks |
274 | * can't happen. |
275 | */ |
276 | WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() |
277 | && !oops_in_progress); |
278 | |
279 | csd = &csd_stack; |
280 | if (!wait) { |
281 | csd = this_cpu_ptr(&csd_data); |
282 | csd_lock(csd); |
283 | } |
284 | |
285 | err = generic_exec_single(cpu, csd, func, info); |
286 | |
287 | if (wait) |
288 | csd_lock_wait(csd); |
289 | |
290 | put_cpu(); |
291 | |
292 | return err; |
293 | } |
294 | EXPORT_SYMBOL(smp_call_function_single); |
295 | |
296 | /** |
297 | * smp_call_function_single_async(): Run an asynchronous function on a |
298 | * specific CPU. |
299 | * @cpu: The CPU to run on. |
300 | * @csd: Pre-allocated and setup data structure |
301 | * |
302 | * Like smp_call_function_single(), but the call is asynchonous and |
303 | * can thus be done from contexts with disabled interrupts. |
304 | * |
305 | * The caller passes his own pre-allocated data structure |
306 | * (ie: embedded in an object) and is responsible for synchronizing it |
307 | * such that the IPIs performed on the @csd are strictly serialized. |
308 | * |
309 | * NOTE: Be careful, there is unfortunately no current debugging facility to |
310 | * validate the correctness of this serialization. |
311 | */ |
312 | int smp_call_function_single_async(int cpu, struct call_single_data *csd) |
313 | { |
314 | int err = 0; |
315 | |
316 | preempt_disable(); |
317 | |
318 | /* We could deadlock if we have to wait here with interrupts disabled! */ |
319 | if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK)) |
320 | csd_lock_wait(csd); |
321 | |
322 | csd->flags = CSD_FLAG_LOCK; |
323 | smp_wmb(); |
324 | |
325 | err = generic_exec_single(cpu, csd, csd->func, csd->info); |
326 | preempt_enable(); |
327 | |
328 | return err; |
329 | } |
330 | EXPORT_SYMBOL_GPL(smp_call_function_single_async); |
331 | |
332 | /* |
333 | * smp_call_function_any - Run a function on any of the given cpus |
334 | * @mask: The mask of cpus it can run on. |
335 | * @func: The function to run. This must be fast and non-blocking. |
336 | * @info: An arbitrary pointer to pass to the function. |
337 | * @wait: If true, wait until function has completed. |
338 | * |
339 | * Returns 0 on success, else a negative status code (if no cpus were online). |
340 | * |
341 | * Selection preference: |
342 | * 1) current cpu if in @mask |
343 | * 2) any cpu of current node if in @mask |
344 | * 3) any other online cpu in @mask |
345 | */ |
346 | int smp_call_function_any(const struct cpumask *mask, |
347 | smp_call_func_t func, void *info, int wait) |
348 | { |
349 | unsigned int cpu; |
350 | const struct cpumask *nodemask; |
351 | int ret; |
352 | |
353 | /* Try for same CPU (cheapest) */ |
354 | cpu = get_cpu(); |
355 | if (cpumask_test_cpu(cpu, mask)) |
356 | goto call; |
357 | |
358 | /* Try for same node. */ |
359 | nodemask = cpumask_of_node(cpu_to_node(cpu)); |
360 | for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids; |
361 | cpu = cpumask_next_and(cpu, nodemask, mask)) { |
362 | if (cpu_online(cpu)) |
363 | goto call; |
364 | } |
365 | |
366 | /* Any online will do: smp_call_function_single handles nr_cpu_ids. */ |
367 | cpu = cpumask_any_and(mask, cpu_online_mask); |
368 | call: |
369 | ret = smp_call_function_single(cpu, func, info, wait); |
370 | put_cpu(); |
371 | return ret; |
372 | } |
373 | EXPORT_SYMBOL_GPL(smp_call_function_any); |
374 | |
375 | /** |
376 | * smp_call_function_many(): Run a function on a set of other CPUs. |
377 | * @mask: The set of cpus to run on (only runs on online subset). |
378 | * @func: The function to run. This must be fast and non-blocking. |
379 | * @info: An arbitrary pointer to pass to the function. |
380 | * @wait: If true, wait (atomically) until function has completed |
381 | * on other CPUs. |
382 | * |
383 | * If @wait is true, then returns once @func has returned. |
384 | * |
385 | * You must not call this function with disabled interrupts or from a |
386 | * hardware interrupt handler or from a bottom half handler. Preemption |
387 | * must be disabled when calling this function. |
388 | */ |
389 | void smp_call_function_many(const struct cpumask *mask, |
390 | smp_call_func_t func, void *info, bool wait) |
391 | { |
392 | struct call_function_data *cfd; |
393 | int cpu, next_cpu, this_cpu = smp_processor_id(); |
394 | |
395 | /* |
396 | * Can deadlock when called with interrupts disabled. |
397 | * We allow cpu's that are not yet online though, as no one else can |
398 | * send smp call function interrupt to this cpu and as such deadlocks |
399 | * can't happen. |
400 | */ |
401 | WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() |
402 | && !oops_in_progress && !early_boot_irqs_disabled); |
403 | |
404 | /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */ |
405 | cpu = cpumask_first_and(mask, cpu_online_mask); |
406 | if (cpu == this_cpu) |
407 | cpu = cpumask_next_and(cpu, mask, cpu_online_mask); |
408 | |
409 | /* No online cpus? We're done. */ |
410 | if (cpu >= nr_cpu_ids) |
411 | return; |
412 | |
413 | /* Do we have another CPU which isn't us? */ |
414 | next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask); |
415 | if (next_cpu == this_cpu) |
416 | next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask); |
417 | |
418 | /* Fastpath: do that cpu by itself. */ |
419 | if (next_cpu >= nr_cpu_ids) { |
420 | smp_call_function_single(cpu, func, info, wait); |
421 | return; |
422 | } |
423 | |
424 | cfd = this_cpu_ptr(&cfd_data); |
425 | |
426 | cpumask_and(cfd->cpumask, mask, cpu_online_mask); |
427 | cpumask_clear_cpu(this_cpu, cfd->cpumask); |
428 | |
429 | /* Some callers race with other cpus changing the passed mask */ |
430 | if (unlikely(!cpumask_weight(cfd->cpumask))) |
431 | return; |
432 | |
433 | for_each_cpu(cpu, cfd->cpumask) { |
434 | struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu); |
435 | |
436 | csd_lock(csd); |
437 | if (wait) |
438 | csd->flags |= CSD_FLAG_SYNCHRONOUS; |
439 | csd->func = func; |
440 | csd->info = info; |
441 | llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)); |
442 | } |
443 | |
444 | /* Send a message to all CPUs in the map */ |
445 | arch_send_call_function_ipi_mask(cfd->cpumask); |
446 | |
447 | if (wait) { |
448 | for_each_cpu(cpu, cfd->cpumask) { |
449 | struct call_single_data *csd; |
450 | |
451 | csd = per_cpu_ptr(cfd->csd, cpu); |
452 | csd_lock_wait(csd); |
453 | } |
454 | } |
455 | } |
456 | EXPORT_SYMBOL(smp_call_function_many); |
457 | |
458 | /** |
459 | * smp_call_function(): Run a function on all other CPUs. |
460 | * @func: The function to run. This must be fast and non-blocking. |
461 | * @info: An arbitrary pointer to pass to the function. |
462 | * @wait: If true, wait (atomically) until function has completed |
463 | * on other CPUs. |
464 | * |
465 | * Returns 0. |
466 | * |
467 | * If @wait is true, then returns once @func has returned; otherwise |
468 | * it returns just before the target cpu calls @func. |
469 | * |
470 | * You must not call this function with disabled interrupts or from a |
471 | * hardware interrupt handler or from a bottom half handler. |
472 | */ |
473 | int smp_call_function(smp_call_func_t func, void *info, int wait) |
474 | { |
475 | preempt_disable(); |
476 | smp_call_function_many(cpu_online_mask, func, info, wait); |
477 | preempt_enable(); |
478 | |
479 | return 0; |
480 | } |
481 | EXPORT_SYMBOL(smp_call_function); |
482 | |
483 | /* Setup configured maximum number of CPUs to activate */ |
484 | unsigned int setup_max_cpus = NR_CPUS; |
485 | EXPORT_SYMBOL(setup_max_cpus); |
486 | |
487 | |
488 | /* |
489 | * Setup routine for controlling SMP activation |
490 | * |
491 | * Command-line option of "nosmp" or "maxcpus=0" will disable SMP |
492 | * activation entirely (the MPS table probe still happens, though). |
493 | * |
494 | * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer |
495 | * greater than 0, limits the maximum number of CPUs activated in |
496 | * SMP mode to <NUM>. |
497 | */ |
498 | |
499 | void __weak arch_disable_smp_support(void) { } |
500 | |
501 | static int __init nosmp(char *str) |
502 | { |
503 | setup_max_cpus = 0; |
504 | arch_disable_smp_support(); |
505 | |
506 | return 0; |
507 | } |
508 | |
509 | early_param("nosmp", nosmp); |
510 | |
511 | /* this is hard limit */ |
512 | static int __init nrcpus(char *str) |
513 | { |
514 | int nr_cpus; |
515 | |
516 | get_option(&str, &nr_cpus); |
517 | if (nr_cpus > 0 && nr_cpus < nr_cpu_ids) |
518 | nr_cpu_ids = nr_cpus; |
519 | |
520 | return 0; |
521 | } |
522 | |
523 | early_param("nr_cpus", nrcpus); |
524 | |
525 | static int __init maxcpus(char *str) |
526 | { |
527 | get_option(&str, &setup_max_cpus); |
528 | if (setup_max_cpus == 0) |
529 | arch_disable_smp_support(); |
530 | |
531 | return 0; |
532 | } |
533 | |
534 | early_param("maxcpus", maxcpus); |
535 | |
536 | /* Setup number of possible processor ids */ |
537 | int nr_cpu_ids __read_mostly = NR_CPUS; |
538 | EXPORT_SYMBOL(nr_cpu_ids); |
539 | |
540 | /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */ |
541 | void __init setup_nr_cpu_ids(void) |
542 | { |
543 | nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1; |
544 | } |
545 | |
546 | void __weak smp_announce(void) |
547 | { |
548 | printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus()); |
549 | } |
550 | |
551 | /* Called by boot processor to activate the rest. */ |
552 | void __init smp_init(void) |
553 | { |
554 | unsigned int cpu; |
555 | |
556 | idle_threads_init(); |
557 | cpuhp_threads_init(); |
558 | |
559 | /* FIXME: This should be done in userspace --RR */ |
560 | for_each_present_cpu(cpu) { |
561 | if (num_online_cpus() >= setup_max_cpus) |
562 | break; |
563 | if (!cpu_online(cpu)) |
564 | cpu_up(cpu); |
565 | } |
566 | |
567 | /* Final decision about SMT support */ |
568 | cpu_smt_check_topology(); |
569 | /* Any cleanup work */ |
570 | smp_announce(); |
571 | smp_cpus_done(setup_max_cpus); |
572 | } |
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 | |
594 | /** |
595 | * on_each_cpu_mask(): Run a function on processors specified by |
596 | * cpumask, which may include the local processor. |
597 | * @mask: The set of cpus to run on (only runs on online subset). |
598 | * @func: The function to run. This must be fast and non-blocking. |
599 | * @info: An arbitrary pointer to pass to the function. |
600 | * @wait: If true, wait (atomically) until function has completed |
601 | * on other CPUs. |
602 | * |
603 | * If @wait is true, then returns once @func has returned. |
604 | * |
605 | * You must not call this function with disabled interrupts or from a |
606 | * hardware interrupt handler or from a bottom half handler. The |
607 | * exception is that it may be used during early boot while |
608 | * early_boot_irqs_disabled is set. |
609 | */ |
610 | void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func, |
611 | void *info, bool wait) |
612 | { |
613 | int cpu = get_cpu(); |
614 | |
615 | smp_call_function_many(mask, func, info, wait); |
616 | if (cpumask_test_cpu(cpu, mask)) { |
617 | unsigned long flags; |
618 | local_irq_save(flags); |
619 | func(info); |
620 | local_irq_restore(flags); |
621 | } |
622 | put_cpu(); |
623 | } |
624 | EXPORT_SYMBOL(on_each_cpu_mask); |
625 | |
626 | /* |
627 | * on_each_cpu_cond(): Call a function on each processor for which |
628 | * the supplied function cond_func returns true, optionally waiting |
629 | * for all the required CPUs to finish. This may include the local |
630 | * processor. |
631 | * @cond_func: A callback function that is passed a cpu id and |
632 | * the the info parameter. The function is called |
633 | * with preemption disabled. The function should |
634 | * return a blooean value indicating whether to IPI |
635 | * the specified CPU. |
636 | * @func: The function to run on all applicable CPUs. |
637 | * This must be fast and non-blocking. |
638 | * @info: An arbitrary pointer to pass to both functions. |
639 | * @wait: If true, wait (atomically) until function has |
640 | * completed on other CPUs. |
641 | * @gfp_flags: GFP flags to use when allocating the cpumask |
642 | * used internally by the function. |
643 | * |
644 | * The function might sleep if the GFP flags indicates a non |
645 | * atomic allocation is allowed. |
646 | * |
647 | * Preemption is disabled to protect against CPUs going offline but not online. |
648 | * CPUs going online during the call will not be seen or sent an IPI. |
649 | * |
650 | * You must not call this function with disabled interrupts or |
651 | * from a hardware interrupt handler or from a bottom half handler. |
652 | */ |
653 | void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info), |
654 | smp_call_func_t func, void *info, bool wait, |
655 | gfp_t gfp_flags) |
656 | { |
657 | cpumask_var_t cpus; |
658 | int cpu, ret; |
659 | |
660 | might_sleep_if(gfpflags_allow_blocking(gfp_flags)); |
661 | |
662 | if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) { |
663 | preempt_disable(); |
664 | for_each_online_cpu(cpu) |
665 | if (cond_func(cpu, info)) |
666 | cpumask_set_cpu(cpu, cpus); |
667 | on_each_cpu_mask(cpus, func, info, wait); |
668 | preempt_enable(); |
669 | free_cpumask_var(cpus); |
670 | } else { |
671 | /* |
672 | * No free cpumask, bother. No matter, we'll |
673 | * just have to IPI them one by one. |
674 | */ |
675 | preempt_disable(); |
676 | for_each_online_cpu(cpu) |
677 | if (cond_func(cpu, info)) { |
678 | ret = smp_call_function_single(cpu, func, |
679 | info, wait); |
680 | WARN_ON_ONCE(ret); |
681 | } |
682 | preempt_enable(); |
683 | } |
684 | } |
685 | EXPORT_SYMBOL(on_each_cpu_cond); |
686 | |
687 | static void do_nothing(void *unused) |
688 | { |
689 | } |
690 | |
691 | /** |
692 | * kick_all_cpus_sync - Force all cpus out of idle |
693 | * |
694 | * Used to synchronize the update of pm_idle function pointer. It's |
695 | * called after the pointer is updated and returns after the dummy |
696 | * callback function has been executed on all cpus. The execution of |
697 | * the function can only happen on the remote cpus after they have |
698 | * left the idle function which had been called via pm_idle function |
699 | * pointer. So it's guaranteed that nothing uses the previous pointer |
700 | * anymore. |
701 | */ |
702 | void kick_all_cpus_sync(void) |
703 | { |
704 | /* Make sure the change is visible before we kick the cpus */ |
705 | smp_mb(); |
706 | smp_call_function(do_nothing, NULL, 1); |
707 | } |
708 | EXPORT_SYMBOL_GPL(kick_all_cpus_sync); |
709 | |
710 | /** |
711 | * wake_up_all_idle_cpus - break all cpus out of idle |
712 | * wake_up_all_idle_cpus try to break all cpus which is in idle state even |
713 | * including idle polling cpus, for non-idle cpus, we will do nothing |
714 | * for them. |
715 | */ |
716 | void wake_up_all_idle_cpus(void) |
717 | { |
718 | int cpu; |
719 | |
720 | preempt_disable(); |
721 | for_each_online_cpu(cpu) { |
722 | if (cpu == smp_processor_id()) |
723 | continue; |
724 | |
725 | wake_up_if_idle(cpu); |
726 | } |
727 | preempt_enable(); |
728 | } |
729 | EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus); |
730 | |
731 | /** |
732 | * smp_call_on_cpu - Call a function on a specific cpu |
733 | * |
734 | * Used to call a function on a specific cpu and wait for it to return. |
735 | * Optionally make sure the call is done on a specified physical cpu via vcpu |
736 | * pinning in order to support virtualized environments. |
737 | */ |
738 | struct smp_call_on_cpu_struct { |
739 | struct work_struct work; |
740 | struct completion done; |
741 | int (*func)(void *); |
742 | void *data; |
743 | int ret; |
744 | int cpu; |
745 | }; |
746 | |
747 | static void smp_call_on_cpu_callback(struct work_struct *work) |
748 | { |
749 | struct smp_call_on_cpu_struct *sscs; |
750 | |
751 | sscs = container_of(work, struct smp_call_on_cpu_struct, work); |
752 | if (sscs->cpu >= 0) |
753 | hypervisor_pin_vcpu(sscs->cpu); |
754 | sscs->ret = sscs->func(sscs->data); |
755 | if (sscs->cpu >= 0) |
756 | hypervisor_pin_vcpu(-1); |
757 | |
758 | complete(&sscs->done); |
759 | } |
760 | |
761 | int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys) |
762 | { |
763 | struct smp_call_on_cpu_struct sscs = { |
764 | .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done), |
765 | .func = func, |
766 | .data = par, |
767 | .cpu = phys ? cpu : -1, |
768 | }; |
769 | |
770 | INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback); |
771 | |
772 | if (cpu >= nr_cpu_ids || !cpu_online(cpu)) |
773 | return -ENXIO; |
774 | |
775 | queue_work_on(cpu, system_wq, &sscs.work); |
776 | wait_for_completion(&sscs.done); |
777 | |
778 | return sscs.ret; |
779 | } |
780 | EXPORT_SYMBOL_GPL(smp_call_on_cpu); |
781 |