]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - kernel/smp.c
Merge branch 'next-devicetree' of git://git.secretlab.ca/git/linux-2.6
[mirror_ubuntu-artful-kernel.git] / kernel / smp.c
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 static struct {
17 struct list_head queue;
18 raw_spinlock_t lock;
19 } call_function __cacheline_aligned_in_smp =
20 {
21 .queue = LIST_HEAD_INIT(call_function.queue),
22 .lock = __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
23 };
24
25 enum {
26 CSD_FLAG_LOCK = 0x01,
27 };
28
29 struct call_function_data {
30 struct call_single_data csd;
31 atomic_t refs;
32 cpumask_var_t cpumask;
33 };
34
35 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
36
37 struct call_single_queue {
38 struct list_head list;
39 raw_spinlock_t lock;
40 };
41
42 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
43
44 static int
45 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
46 {
47 long cpu = (long)hcpu;
48 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
49
50 switch (action) {
51 case CPU_UP_PREPARE:
52 case CPU_UP_PREPARE_FROZEN:
53 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
54 cpu_to_node(cpu)))
55 return notifier_from_errno(-ENOMEM);
56 break;
57
58 #ifdef CONFIG_HOTPLUG_CPU
59 case CPU_UP_CANCELED:
60 case CPU_UP_CANCELED_FROZEN:
61
62 case CPU_DEAD:
63 case CPU_DEAD_FROZEN:
64 free_cpumask_var(cfd->cpumask);
65 break;
66 #endif
67 };
68
69 return NOTIFY_OK;
70 }
71
72 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
73 .notifier_call = hotplug_cfd,
74 };
75
76 static int __cpuinit init_call_single_data(void)
77 {
78 void *cpu = (void *)(long)smp_processor_id();
79 int i;
80
81 for_each_possible_cpu(i) {
82 struct call_single_queue *q = &per_cpu(call_single_queue, i);
83
84 raw_spin_lock_init(&q->lock);
85 INIT_LIST_HEAD(&q->list);
86 }
87
88 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
89 register_cpu_notifier(&hotplug_cfd_notifier);
90
91 return 0;
92 }
93 early_initcall(init_call_single_data);
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 */
102 static void csd_lock_wait(struct call_single_data *data)
103 {
104 while (data->flags & CSD_FLAG_LOCK)
105 cpu_relax();
106 }
107
108 static 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
121 static 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 */
138 static
139 void 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 */
172 void 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
197 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask))
198 continue;
199
200 data->csd.func(data->csd.info);
201
202 refs = atomic_dec_return(&data->refs);
203 WARN_ON(refs < 0);
204 if (!refs) {
205 raw_spin_lock(&call_function.lock);
206 list_del_rcu(&data->csd.list);
207 raw_spin_unlock(&call_function.lock);
208 }
209
210 if (refs)
211 continue;
212
213 csd_unlock(&data->csd);
214 }
215
216 }
217
218 /*
219 * Invoked by arch to handle an IPI for call function single. Must be
220 * called from the arch with interrupts disabled.
221 */
222 void generic_smp_call_function_single_interrupt(void)
223 {
224 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
225 unsigned int data_flags;
226 LIST_HEAD(list);
227
228 /*
229 * Shouldn't receive this interrupt on a cpu that is not yet online.
230 */
231 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
232
233 raw_spin_lock(&q->lock);
234 list_replace_init(&q->list, &list);
235 raw_spin_unlock(&q->lock);
236
237 while (!list_empty(&list)) {
238 struct call_single_data *data;
239
240 data = list_entry(list.next, struct call_single_data, list);
241 list_del(&data->list);
242
243 /*
244 * 'data' can be invalid after this call if flags == 0
245 * (when called through generic_exec_single()),
246 * so save them away before making the call:
247 */
248 data_flags = data->flags;
249
250 data->func(data->info);
251
252 /*
253 * Unlocked CSDs are valid through generic_exec_single():
254 */
255 if (data_flags & CSD_FLAG_LOCK)
256 csd_unlock(data);
257 }
258 }
259
260 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
261
262 /*
263 * smp_call_function_single - Run a function on a specific CPU
264 * @func: The function to run. This must be fast and non-blocking.
265 * @info: An arbitrary pointer to pass to the function.
266 * @wait: If true, wait until function has completed on other CPUs.
267 *
268 * Returns 0 on success, else a negative status code.
269 */
270 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
271 int wait)
272 {
273 struct call_single_data d = {
274 .flags = 0,
275 };
276 unsigned long flags;
277 int this_cpu;
278 int err = 0;
279
280 /*
281 * prevent preemption and reschedule on another processor,
282 * as well as CPU removal
283 */
284 this_cpu = get_cpu();
285
286 /*
287 * Can deadlock when called with interrupts disabled.
288 * We allow cpu's that are not yet online though, as no one else can
289 * send smp call function interrupt to this cpu and as such deadlocks
290 * can't happen.
291 */
292 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
293 && !oops_in_progress);
294
295 if (cpu == this_cpu) {
296 local_irq_save(flags);
297 func(info);
298 local_irq_restore(flags);
299 } else {
300 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
301 struct call_single_data *data = &d;
302
303 if (!wait)
304 data = &__get_cpu_var(csd_data);
305
306 csd_lock(data);
307
308 data->func = func;
309 data->info = info;
310 generic_exec_single(cpu, data, wait);
311 } else {
312 err = -ENXIO; /* CPU not online */
313 }
314 }
315
316 put_cpu();
317
318 return err;
319 }
320 EXPORT_SYMBOL(smp_call_function_single);
321
322 /*
323 * smp_call_function_any - Run a function on any of the given cpus
324 * @mask: The mask of cpus it can run on.
325 * @func: The function to run. This must be fast and non-blocking.
326 * @info: An arbitrary pointer to pass to the function.
327 * @wait: If true, wait until function has completed.
328 *
329 * Returns 0 on success, else a negative status code (if no cpus were online).
330 * Note that @wait will be implicitly turned on in case of allocation failures,
331 * since we fall back to on-stack allocation.
332 *
333 * Selection preference:
334 * 1) current cpu if in @mask
335 * 2) any cpu of current node if in @mask
336 * 3) any other online cpu in @mask
337 */
338 int smp_call_function_any(const struct cpumask *mask,
339 void (*func)(void *info), void *info, int wait)
340 {
341 unsigned int cpu;
342 const struct cpumask *nodemask;
343 int ret;
344
345 /* Try for same CPU (cheapest) */
346 cpu = get_cpu();
347 if (cpumask_test_cpu(cpu, mask))
348 goto call;
349
350 /* Try for same node. */
351 nodemask = cpumask_of_node(cpu_to_node(cpu));
352 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
353 cpu = cpumask_next_and(cpu, nodemask, mask)) {
354 if (cpu_online(cpu))
355 goto call;
356 }
357
358 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
359 cpu = cpumask_any_and(mask, cpu_online_mask);
360 call:
361 ret = smp_call_function_single(cpu, func, info, wait);
362 put_cpu();
363 return ret;
364 }
365 EXPORT_SYMBOL_GPL(smp_call_function_any);
366
367 /**
368 * __smp_call_function_single(): Run a function on a specific CPU
369 * @cpu: The CPU to run on.
370 * @data: Pre-allocated and setup data structure
371 * @wait: If true, wait until function has completed on specified CPU.
372 *
373 * Like smp_call_function_single(), but allow caller to pass in a
374 * pre-allocated data structure. Useful for embedding @data inside
375 * other structures, for instance.
376 */
377 void __smp_call_function_single(int cpu, struct call_single_data *data,
378 int wait)
379 {
380 unsigned int this_cpu;
381 unsigned long flags;
382
383 this_cpu = get_cpu();
384 /*
385 * Can deadlock when called with interrupts disabled.
386 * We allow cpu's that are not yet online though, as no one else can
387 * send smp call function interrupt to this cpu and as such deadlocks
388 * can't happen.
389 */
390 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
391 && !oops_in_progress);
392
393 if (cpu == this_cpu) {
394 local_irq_save(flags);
395 data->func(data->info);
396 local_irq_restore(flags);
397 } else {
398 csd_lock(data);
399 generic_exec_single(cpu, data, wait);
400 }
401 put_cpu();
402 }
403
404 /**
405 * smp_call_function_many(): Run a function on a set of other CPUs.
406 * @mask: The set of cpus to run on (only runs on online subset).
407 * @func: The function to run. This must be fast and non-blocking.
408 * @info: An arbitrary pointer to pass to the function.
409 * @wait: If true, wait (atomically) until function has completed
410 * on other CPUs.
411 *
412 * If @wait is true, then returns once @func has returned.
413 *
414 * You must not call this function with disabled interrupts or from a
415 * hardware interrupt handler or from a bottom half handler. Preemption
416 * must be disabled when calling this function.
417 */
418 void smp_call_function_many(const struct cpumask *mask,
419 void (*func)(void *), void *info, bool wait)
420 {
421 struct call_function_data *data;
422 unsigned long flags;
423 int cpu, next_cpu, this_cpu = smp_processor_id();
424
425 /*
426 * Can deadlock when called with interrupts disabled.
427 * We allow cpu's that are not yet online though, as no one else can
428 * send smp call function interrupt to this cpu and as such deadlocks
429 * can't happen.
430 */
431 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
432 && !oops_in_progress);
433
434 /* So, what's a CPU they want? Ignoring this one. */
435 cpu = cpumask_first_and(mask, cpu_online_mask);
436 if (cpu == this_cpu)
437 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
438
439 /* No online cpus? We're done. */
440 if (cpu >= nr_cpu_ids)
441 return;
442
443 /* Do we have another CPU which isn't us? */
444 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
445 if (next_cpu == this_cpu)
446 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
447
448 /* Fastpath: do that cpu by itself. */
449 if (next_cpu >= nr_cpu_ids) {
450 smp_call_function_single(cpu, func, info, wait);
451 return;
452 }
453
454 data = &__get_cpu_var(cfd_data);
455 csd_lock(&data->csd);
456
457 data->csd.func = func;
458 data->csd.info = info;
459 cpumask_and(data->cpumask, mask, cpu_online_mask);
460 cpumask_clear_cpu(this_cpu, data->cpumask);
461 atomic_set(&data->refs, cpumask_weight(data->cpumask));
462
463 raw_spin_lock_irqsave(&call_function.lock, flags);
464 /*
465 * Place entry at the _HEAD_ of the list, so that any cpu still
466 * observing the entry in generic_smp_call_function_interrupt()
467 * will not miss any other list entries:
468 */
469 list_add_rcu(&data->csd.list, &call_function.queue);
470 raw_spin_unlock_irqrestore(&call_function.lock, flags);
471
472 /*
473 * Make the list addition visible before sending the ipi.
474 * (IPIs must obey or appear to obey normal Linux cache
475 * coherency rules -- see comment in generic_exec_single).
476 */
477 smp_mb();
478
479 /* Send a message to all CPUs in the map */
480 arch_send_call_function_ipi_mask(data->cpumask);
481
482 /* Optionally wait for the CPUs to complete */
483 if (wait)
484 csd_lock_wait(&data->csd);
485 }
486 EXPORT_SYMBOL(smp_call_function_many);
487
488 /**
489 * smp_call_function(): Run a function on all other CPUs.
490 * @func: The function to run. This must be fast and non-blocking.
491 * @info: An arbitrary pointer to pass to the function.
492 * @wait: If true, wait (atomically) until function has completed
493 * on other CPUs.
494 *
495 * Returns 0.
496 *
497 * If @wait is true, then returns once @func has returned; otherwise
498 * it returns just before the target cpu calls @func.
499 *
500 * You must not call this function with disabled interrupts or from a
501 * hardware interrupt handler or from a bottom half handler.
502 */
503 int smp_call_function(void (*func)(void *), void *info, int wait)
504 {
505 preempt_disable();
506 smp_call_function_many(cpu_online_mask, func, info, wait);
507 preempt_enable();
508
509 return 0;
510 }
511 EXPORT_SYMBOL(smp_call_function);
512
513 void ipi_call_lock(void)
514 {
515 raw_spin_lock(&call_function.lock);
516 }
517
518 void ipi_call_unlock(void)
519 {
520 raw_spin_unlock(&call_function.lock);
521 }
522
523 void ipi_call_lock_irq(void)
524 {
525 raw_spin_lock_irq(&call_function.lock);
526 }
527
528 void ipi_call_unlock_irq(void)
529 {
530 raw_spin_unlock_irq(&call_function.lock);
531 }