Commit | Line | Data |
---|---|---|
3d442233 JA |
1 | /* |
2 | * Generic helpers for smp ipi calls | |
3 | * | |
4 | * (C) Jens Axboe <jens.axboe@oracle.com> 2008 | |
5 | * | |
6 | */ | |
7 | #include <linux/init.h> | |
8 | #include <linux/module.h> | |
9 | #include <linux/percpu.h> | |
10 | #include <linux/rcupdate.h> | |
59190f42 | 11 | #include <linux/rculist.h> |
3d442233 JA |
12 | #include <linux/smp.h> |
13 | ||
14 | static DEFINE_PER_CPU(struct call_single_queue, call_single_queue); | |
15 | static LIST_HEAD(call_function_queue); | |
16 | __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_function_lock); | |
17 | ||
18 | enum { | |
19 | CSD_FLAG_WAIT = 0x01, | |
20 | CSD_FLAG_ALLOC = 0x02, | |
21 | }; | |
22 | ||
23 | struct call_function_data { | |
24 | struct call_single_data csd; | |
25 | spinlock_t lock; | |
26 | unsigned int refs; | |
3d442233 | 27 | struct rcu_head rcu_head; |
54b11e6d | 28 | unsigned long cpumask_bits[]; |
3d442233 JA |
29 | }; |
30 | ||
31 | struct call_single_queue { | |
32 | struct list_head list; | |
33 | spinlock_t lock; | |
34 | }; | |
35 | ||
7babe8db | 36 | static int __cpuinit init_call_single_data(void) |
3d442233 JA |
37 | { |
38 | int i; | |
39 | ||
40 | for_each_possible_cpu(i) { | |
41 | struct call_single_queue *q = &per_cpu(call_single_queue, i); | |
42 | ||
43 | spin_lock_init(&q->lock); | |
44 | INIT_LIST_HEAD(&q->list); | |
45 | } | |
7babe8db | 46 | return 0; |
3d442233 | 47 | } |
7babe8db | 48 | early_initcall(init_call_single_data); |
3d442233 JA |
49 | |
50 | static void csd_flag_wait(struct call_single_data *data) | |
51 | { | |
52 | /* Wait for response */ | |
53 | do { | |
3d442233 JA |
54 | if (!(data->flags & CSD_FLAG_WAIT)) |
55 | break; | |
56 | cpu_relax(); | |
57 | } while (1); | |
58 | } | |
59 | ||
60 | /* | |
61 | * Insert a previously allocated call_single_data element for execution | |
62 | * on the given CPU. data must already have ->func, ->info, and ->flags set. | |
63 | */ | |
64 | static void generic_exec_single(int cpu, struct call_single_data *data) | |
65 | { | |
66 | struct call_single_queue *dst = &per_cpu(call_single_queue, cpu); | |
67 | int wait = data->flags & CSD_FLAG_WAIT, ipi; | |
68 | unsigned long flags; | |
69 | ||
70 | spin_lock_irqsave(&dst->lock, flags); | |
71 | ipi = list_empty(&dst->list); | |
72 | list_add_tail(&data->list, &dst->list); | |
73 | spin_unlock_irqrestore(&dst->lock, flags); | |
74 | ||
561920a0 SS |
75 | /* |
76 | * Make the list addition visible before sending the ipi. | |
77 | */ | |
78 | smp_mb(); | |
79 | ||
3d442233 JA |
80 | if (ipi) |
81 | arch_send_call_function_single_ipi(cpu); | |
82 | ||
83 | if (wait) | |
84 | csd_flag_wait(data); | |
85 | } | |
86 | ||
87 | static void rcu_free_call_data(struct rcu_head *head) | |
88 | { | |
89 | struct call_function_data *data; | |
90 | ||
91 | data = container_of(head, struct call_function_data, rcu_head); | |
92 | ||
93 | kfree(data); | |
94 | } | |
95 | ||
96 | /* | |
97 | * Invoked by arch to handle an IPI for call function. Must be called with | |
98 | * interrupts disabled. | |
99 | */ | |
100 | void generic_smp_call_function_interrupt(void) | |
101 | { | |
102 | struct call_function_data *data; | |
103 | int cpu = get_cpu(); | |
104 | ||
105 | /* | |
106 | * It's ok to use list_for_each_rcu() here even though we may delete | |
107 | * 'pos', since list_del_rcu() doesn't clear ->next | |
108 | */ | |
109 | rcu_read_lock(); | |
110 | list_for_each_entry_rcu(data, &call_function_queue, csd.list) { | |
111 | int refs; | |
112 | ||
54b11e6d | 113 | if (!cpumask_test_cpu(cpu, to_cpumask(data->cpumask_bits))) |
3d442233 JA |
114 | continue; |
115 | ||
116 | data->csd.func(data->csd.info); | |
117 | ||
118 | spin_lock(&data->lock); | |
54b11e6d | 119 | cpumask_clear_cpu(cpu, to_cpumask(data->cpumask_bits)); |
3d442233 JA |
120 | WARN_ON(data->refs == 0); |
121 | data->refs--; | |
122 | refs = data->refs; | |
123 | spin_unlock(&data->lock); | |
124 | ||
125 | if (refs) | |
126 | continue; | |
127 | ||
128 | spin_lock(&call_function_lock); | |
129 | list_del_rcu(&data->csd.list); | |
130 | spin_unlock(&call_function_lock); | |
131 | ||
132 | if (data->csd.flags & CSD_FLAG_WAIT) { | |
133 | /* | |
134 | * serialize stores to data with the flag clear | |
135 | * and wakeup | |
136 | */ | |
137 | smp_wmb(); | |
138 | data->csd.flags &= ~CSD_FLAG_WAIT; | |
c2fc1198 NP |
139 | } |
140 | if (data->csd.flags & CSD_FLAG_ALLOC) | |
3d442233 JA |
141 | call_rcu(&data->rcu_head, rcu_free_call_data); |
142 | } | |
143 | rcu_read_unlock(); | |
144 | ||
145 | put_cpu(); | |
146 | } | |
147 | ||
148 | /* | |
149 | * Invoked by arch to handle an IPI for call function single. Must be called | |
150 | * from the arch with interrupts disabled. | |
151 | */ | |
152 | void generic_smp_call_function_single_interrupt(void) | |
153 | { | |
154 | struct call_single_queue *q = &__get_cpu_var(call_single_queue); | |
155 | LIST_HEAD(list); | |
156 | ||
157 | /* | |
158 | * Need to see other stores to list head for checking whether | |
159 | * list is empty without holding q->lock | |
160 | */ | |
561920a0 | 161 | smp_read_barrier_depends(); |
3d442233 JA |
162 | while (!list_empty(&q->list)) { |
163 | unsigned int data_flags; | |
164 | ||
165 | spin_lock(&q->lock); | |
166 | list_replace_init(&q->list, &list); | |
167 | spin_unlock(&q->lock); | |
168 | ||
169 | while (!list_empty(&list)) { | |
170 | struct call_single_data *data; | |
171 | ||
172 | data = list_entry(list.next, struct call_single_data, | |
173 | list); | |
174 | list_del(&data->list); | |
175 | ||
176 | /* | |
177 | * 'data' can be invalid after this call if | |
178 | * flags == 0 (when called through | |
179 | * generic_exec_single(), so save them away before | |
180 | * making the call. | |
181 | */ | |
182 | data_flags = data->flags; | |
183 | ||
184 | data->func(data->info); | |
185 | ||
186 | if (data_flags & CSD_FLAG_WAIT) { | |
187 | smp_wmb(); | |
188 | data->flags &= ~CSD_FLAG_WAIT; | |
189 | } else if (data_flags & CSD_FLAG_ALLOC) | |
190 | kfree(data); | |
191 | } | |
192 | /* | |
193 | * See comment on outer loop | |
194 | */ | |
561920a0 | 195 | smp_read_barrier_depends(); |
3d442233 JA |
196 | } |
197 | } | |
198 | ||
199 | /* | |
200 | * smp_call_function_single - Run a function on a specific CPU | |
201 | * @func: The function to run. This must be fast and non-blocking. | |
202 | * @info: An arbitrary pointer to pass to the function. | |
3d442233 JA |
203 | * @wait: If true, wait until function has completed on other CPUs. |
204 | * | |
205 | * Returns 0 on success, else a negative status code. Note that @wait | |
206 | * will be implicitly turned on in case of allocation failures, since | |
207 | * we fall back to on-stack allocation. | |
208 | */ | |
209 | int smp_call_function_single(int cpu, void (*func) (void *info), void *info, | |
8691e5a8 | 210 | int wait) |
3d442233 JA |
211 | { |
212 | struct call_single_data d; | |
213 | unsigned long flags; | |
f73be6de PA |
214 | /* prevent preemption and reschedule on another processor, |
215 | as well as CPU removal */ | |
3d442233 | 216 | int me = get_cpu(); |
f73be6de | 217 | int err = 0; |
3d442233 JA |
218 | |
219 | /* Can deadlock when called with interrupts disabled */ | |
220 | WARN_ON(irqs_disabled()); | |
221 | ||
222 | if (cpu == me) { | |
223 | local_irq_save(flags); | |
224 | func(info); | |
225 | local_irq_restore(flags); | |
f73be6de | 226 | } else if ((unsigned)cpu < NR_CPUS && cpu_online(cpu)) { |
3d442233 JA |
227 | struct call_single_data *data = NULL; |
228 | ||
229 | if (!wait) { | |
230 | data = kmalloc(sizeof(*data), GFP_ATOMIC); | |
231 | if (data) | |
232 | data->flags = CSD_FLAG_ALLOC; | |
233 | } | |
234 | if (!data) { | |
235 | data = &d; | |
236 | data->flags = CSD_FLAG_WAIT; | |
237 | } | |
238 | ||
239 | data->func = func; | |
240 | data->info = info; | |
241 | generic_exec_single(cpu, data); | |
f73be6de PA |
242 | } else { |
243 | err = -ENXIO; /* CPU not online */ | |
3d442233 JA |
244 | } |
245 | ||
246 | put_cpu(); | |
f73be6de | 247 | return err; |
3d442233 JA |
248 | } |
249 | EXPORT_SYMBOL(smp_call_function_single); | |
250 | ||
251 | /** | |
252 | * __smp_call_function_single(): Run a function on another CPU | |
253 | * @cpu: The CPU to run on. | |
254 | * @data: Pre-allocated and setup data structure | |
255 | * | |
256 | * Like smp_call_function_single(), but allow caller to pass in a pre-allocated | |
257 | * data structure. Useful for embedding @data inside other structures, for | |
258 | * instance. | |
259 | * | |
260 | */ | |
261 | void __smp_call_function_single(int cpu, struct call_single_data *data) | |
262 | { | |
263 | /* Can deadlock when called with interrupts disabled */ | |
264 | WARN_ON((data->flags & CSD_FLAG_WAIT) && irqs_disabled()); | |
265 | ||
266 | generic_exec_single(cpu, data); | |
267 | } | |
268 | ||
269 | /** | |
54b11e6d RR |
270 | * smp_call_function_many(): Run a function on a set of other CPUs. |
271 | * @mask: The set of cpus to run on (only runs on online subset). | |
3d442233 JA |
272 | * @func: The function to run. This must be fast and non-blocking. |
273 | * @info: An arbitrary pointer to pass to the function. | |
274 | * @wait: If true, wait (atomically) until function has completed on other CPUs. | |
275 | * | |
3d442233 JA |
276 | * If @wait is true, then returns once @func has returned. Note that @wait |
277 | * will be implicitly turned on in case of allocation failures, since | |
278 | * we fall back to on-stack allocation. | |
279 | * | |
280 | * You must not call this function with disabled interrupts or from a | |
281 | * hardware interrupt handler or from a bottom half handler. Preemption | |
282 | * must be disabled when calling this function. | |
283 | */ | |
54b11e6d RR |
284 | void smp_call_function_many(const struct cpumask *mask, |
285 | void (*func)(void *), void *info, | |
286 | bool wait) | |
3d442233 | 287 | { |
54b11e6d | 288 | struct call_function_data *data; |
3d442233 | 289 | unsigned long flags; |
54b11e6d | 290 | int cpu, next_cpu; |
3d442233 JA |
291 | |
292 | /* Can deadlock when called with interrupts disabled */ | |
293 | WARN_ON(irqs_disabled()); | |
294 | ||
54b11e6d RR |
295 | /* So, what's a CPU they want? Ignoring this one. */ |
296 | cpu = cpumask_first_and(mask, cpu_online_mask); | |
297 | if (cpu == smp_processor_id()) | |
298 | cpu = cpumask_next_and(cpu, mask, cpu_online_mask); | |
299 | /* No online cpus? We're done. */ | |
300 | if (cpu >= nr_cpu_ids) | |
301 | return; | |
302 | ||
303 | /* Do we have another CPU which isn't us? */ | |
304 | next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask); | |
305 | if (next_cpu == smp_processor_id()) | |
306 | next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask); | |
307 | ||
308 | /* Fastpath: do that cpu by itself. */ | |
309 | if (next_cpu >= nr_cpu_ids) { | |
310 | smp_call_function_single(cpu, func, info, wait); | |
311 | return; | |
3d442233 JA |
312 | } |
313 | ||
54b11e6d RR |
314 | data = kmalloc(sizeof(*data) + cpumask_size(), GFP_ATOMIC); |
315 | if (unlikely(!data)) { | |
316 | /* Slow path. */ | |
317 | for_each_online_cpu(cpu) { | |
318 | if (cpu == smp_processor_id()) | |
319 | continue; | |
320 | if (cpumask_test_cpu(cpu, mask)) | |
321 | smp_call_function_single(cpu, func, info, wait); | |
322 | } | |
323 | return; | |
3d442233 JA |
324 | } |
325 | ||
326 | spin_lock_init(&data->lock); | |
54b11e6d RR |
327 | data->csd.flags = CSD_FLAG_ALLOC; |
328 | if (wait) | |
329 | data->csd.flags |= CSD_FLAG_WAIT; | |
3d442233 JA |
330 | data->csd.func = func; |
331 | data->csd.info = info; | |
54b11e6d RR |
332 | cpumask_and(to_cpumask(data->cpumask_bits), mask, cpu_online_mask); |
333 | cpumask_clear_cpu(smp_processor_id(), to_cpumask(data->cpumask_bits)); | |
334 | data->refs = cpumask_weight(to_cpumask(data->cpumask_bits)); | |
3d442233 JA |
335 | |
336 | spin_lock_irqsave(&call_function_lock, flags); | |
337 | list_add_tail_rcu(&data->csd.list, &call_function_queue); | |
338 | spin_unlock_irqrestore(&call_function_lock, flags); | |
339 | ||
561920a0 SS |
340 | /* |
341 | * Make the list addition visible before sending the ipi. | |
342 | */ | |
343 | smp_mb(); | |
344 | ||
3d442233 | 345 | /* Send a message to all CPUs in the map */ |
54b11e6d | 346 | arch_send_call_function_ipi(*to_cpumask(data->cpumask_bits)); |
3d442233 JA |
347 | |
348 | /* optionally wait for the CPUs to complete */ | |
54b11e6d | 349 | if (wait) |
3d442233 | 350 | csd_flag_wait(&data->csd); |
3d442233 | 351 | } |
54b11e6d | 352 | EXPORT_SYMBOL(smp_call_function_many); |
3d442233 JA |
353 | |
354 | /** | |
355 | * smp_call_function(): Run a function on all other CPUs. | |
356 | * @func: The function to run. This must be fast and non-blocking. | |
357 | * @info: An arbitrary pointer to pass to the function. | |
3d442233 JA |
358 | * @wait: If true, wait (atomically) until function has completed on other CPUs. |
359 | * | |
54b11e6d | 360 | * Returns 0. |
3d442233 JA |
361 | * |
362 | * If @wait is true, then returns once @func has returned; otherwise | |
363 | * it returns just before the target cpu calls @func. In case of allocation | |
364 | * failure, @wait will be implicitly turned on. | |
365 | * | |
366 | * You must not call this function with disabled interrupts or from a | |
367 | * hardware interrupt handler or from a bottom half handler. | |
368 | */ | |
8691e5a8 | 369 | int smp_call_function(void (*func)(void *), void *info, int wait) |
3d442233 | 370 | { |
3d442233 | 371 | preempt_disable(); |
54b11e6d | 372 | smp_call_function_many(cpu_online_mask, func, info, wait); |
3d442233 | 373 | preempt_enable(); |
54b11e6d | 374 | return 0; |
3d442233 JA |
375 | } |
376 | EXPORT_SYMBOL(smp_call_function); | |
377 | ||
378 | void ipi_call_lock(void) | |
379 | { | |
380 | spin_lock(&call_function_lock); | |
381 | } | |
382 | ||
383 | void ipi_call_unlock(void) | |
384 | { | |
385 | spin_unlock(&call_function_lock); | |
386 | } | |
387 | ||
388 | void ipi_call_lock_irq(void) | |
389 | { | |
390 | spin_lock_irq(&call_function_lock); | |
391 | } | |
392 | ||
393 | void ipi_call_unlock_irq(void) | |
394 | { | |
395 | spin_unlock_irq(&call_function_lock); | |
396 | } |