1 /*
2 * kernel/stop_machine.c
3 *
4 * Copyright (C) 2008, 2005 IBM Corporation.
5 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au
6 * Copyright (C) 2010 SUSE Linux Products GmbH
7 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
8 *
9 * This file is released under the GPLv2 and any later version.
10 */
11 #include <linux/completion.h>
12 #include <linux/cpu.h>
13 #include <linux/init.h>
14 #include <linux/kthread.h>
15 #include <linux/export.h>
16 #include <linux/percpu.h>
17 #include <linux/sched.h>
18 #include <linux/stop_machine.h>
19 #include <linux/interrupt.h>
20 #include <linux/kallsyms.h>
21 #include <linux/smpboot.h>
22 #include <linux/atomic.h>
23
24 /*
25 * Structure to determine completion condition and record errors. May
26 * be shared by works on different cpus.
27 */
28 struct cpu_stop_done {
29 atomic_t nr_todo; /* nr left to execute */
30 bool executed; /* actually executed? */
31 int ret; /* collected return value */
32 struct completion completion; /* fired if nr_todo reaches 0 */
33 };
34
35 /* the actual stopper, one per every possible cpu, enabled on online cpus */
36 struct cpu_stopper {
37 spinlock_t lock;
38 bool enabled; /* is this stopper enabled? */
39 struct list_head works; /* list of pending works */
40 };
41
42 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
43 static DEFINE_PER_CPU(struct task_struct *, cpu_stopper_task);
44 static bool stop_machine_initialized = false;
45
cpu_stop_init_done(struct cpu_stop_done * done,unsigned int nr_todo)46 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
47 {
48 memset(done, 0, sizeof(*done));
49 atomic_set(&done->nr_todo, nr_todo);
50 init_completion(&done->completion);
51 }
52
53 /* signal completion unless @done is NULL */
cpu_stop_signal_done(struct cpu_stop_done * done,bool executed)54 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
55 {
56 if (done) {
57 if (executed)
58 done->executed = true;
59 if (atomic_dec_and_test(&done->nr_todo))
60 complete(&done->completion);
61 }
62 }
63
64 /* queue @work to @stopper. if offline, @work is completed immediately */
cpu_stop_queue_work(unsigned int cpu,struct cpu_stop_work * work)65 static void cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
66 {
67 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
68 struct task_struct *p = per_cpu(cpu_stopper_task, cpu);
69
70 unsigned long flags;
71
72 spin_lock_irqsave(&stopper->lock, flags);
73
74 if (stopper->enabled) {
75 list_add_tail(&work->list, &stopper->works);
76 wake_up_process(p);
77 } else
78 cpu_stop_signal_done(work->done, false);
79
80 spin_unlock_irqrestore(&stopper->lock, flags);
81 }
82
83 /**
84 * stop_one_cpu - stop a cpu
85 * @cpu: cpu to stop
86 * @fn: function to execute
87 * @arg: argument to @fn
88 *
89 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
90 * the highest priority preempting any task on the cpu and
91 * monopolizing it. This function returns after the execution is
92 * complete.
93 *
94 * This function doesn't guarantee @cpu stays online till @fn
95 * completes. If @cpu goes down in the middle, execution may happen
96 * partially or fully on different cpus. @fn should either be ready
97 * for that or the caller should ensure that @cpu stays online until
98 * this function completes.
99 *
100 * CONTEXT:
101 * Might sleep.
102 *
103 * RETURNS:
104 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
105 * otherwise, the return value of @fn.
106 */
stop_one_cpu(unsigned int cpu,cpu_stop_fn_t fn,void * arg)107 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
108 {
109 struct cpu_stop_done done;
110 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
111
112 cpu_stop_init_done(&done, 1);
113 cpu_stop_queue_work(cpu, &work);
114 wait_for_completion(&done.completion);
115 return done.executed ? done.ret : -ENOENT;
116 }
117
118 /**
119 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
120 * @cpu: cpu to stop
121 * @fn: function to execute
122 * @arg: argument to @fn
123 *
124 * Similar to stop_one_cpu() but doesn't wait for completion. The
125 * caller is responsible for ensuring @work_buf is currently unused
126 * and will remain untouched until stopper starts executing @fn.
127 *
128 * CONTEXT:
129 * Don't care.
130 */
stop_one_cpu_nowait(unsigned int cpu,cpu_stop_fn_t fn,void * arg,struct cpu_stop_work * work_buf)131 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
132 struct cpu_stop_work *work_buf)
133 {
134 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
135 cpu_stop_queue_work(cpu, work_buf);
136 }
137
138 /* static data for stop_cpus */
139 static DEFINE_MUTEX(stop_cpus_mutex);
140 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
141
queue_stop_cpus_work(const struct cpumask * cpumask,cpu_stop_fn_t fn,void * arg,struct cpu_stop_done * done)142 static void queue_stop_cpus_work(const struct cpumask *cpumask,
143 cpu_stop_fn_t fn, void *arg,
144 struct cpu_stop_done *done)
145 {
146 struct cpu_stop_work *work;
147 unsigned int cpu;
148
149 /* initialize works and done */
150 for_each_cpu(cpu, cpumask) {
151 work = &per_cpu(stop_cpus_work, cpu);
152 work->fn = fn;
153 work->arg = arg;
154 work->done = done;
155 }
156
157 /*
158 * Disable preemption while queueing to avoid getting
159 * preempted by a stopper which might wait for other stoppers
160 * to enter @fn which can lead to deadlock.
161 */
162 preempt_disable();
163 for_each_cpu(cpu, cpumask)
164 cpu_stop_queue_work(cpu, &per_cpu(stop_cpus_work, cpu));
165 preempt_enable();
166 }
167
__stop_cpus(const struct cpumask * cpumask,cpu_stop_fn_t fn,void * arg)168 static int __stop_cpus(const struct cpumask *cpumask,
169 cpu_stop_fn_t fn, void *arg)
170 {
171 struct cpu_stop_done done;
172
173 cpu_stop_init_done(&done, cpumask_weight(cpumask));
174 queue_stop_cpus_work(cpumask, fn, arg, &done);
175 wait_for_completion(&done.completion);
176 return done.executed ? done.ret : -ENOENT;
177 }
178
179 /**
180 * stop_cpus - stop multiple cpus
181 * @cpumask: cpus to stop
182 * @fn: function to execute
183 * @arg: argument to @fn
184 *
185 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
186 * @fn is run in a process context with the highest priority
187 * preempting any task on the cpu and monopolizing it. This function
188 * returns after all executions are complete.
189 *
190 * This function doesn't guarantee the cpus in @cpumask stay online
191 * till @fn completes. If some cpus go down in the middle, execution
192 * on the cpu may happen partially or fully on different cpus. @fn
193 * should either be ready for that or the caller should ensure that
194 * the cpus stay online until this function completes.
195 *
196 * All stop_cpus() calls are serialized making it safe for @fn to wait
197 * for all cpus to start executing it.
198 *
199 * CONTEXT:
200 * Might sleep.
201 *
202 * RETURNS:
203 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
204 * @cpumask were offline; otherwise, 0 if all executions of @fn
205 * returned 0, any non zero return value if any returned non zero.
206 */
stop_cpus(const struct cpumask * cpumask,cpu_stop_fn_t fn,void * arg)207 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
208 {
209 int ret;
210
211 /* static works are used, process one request at a time */
212 mutex_lock(&stop_cpus_mutex);
213 ret = __stop_cpus(cpumask, fn, arg);
214 mutex_unlock(&stop_cpus_mutex);
215 return ret;
216 }
217
218 /**
219 * try_stop_cpus - try to stop multiple cpus
220 * @cpumask: cpus to stop
221 * @fn: function to execute
222 * @arg: argument to @fn
223 *
224 * Identical to stop_cpus() except that it fails with -EAGAIN if
225 * someone else is already using the facility.
226 *
227 * CONTEXT:
228 * Might sleep.
229 *
230 * RETURNS:
231 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
232 * @fn(@arg) was not executed at all because all cpus in @cpumask were
233 * offline; otherwise, 0 if all executions of @fn returned 0, any non
234 * zero return value if any returned non zero.
235 */
try_stop_cpus(const struct cpumask * cpumask,cpu_stop_fn_t fn,void * arg)236 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
237 {
238 int ret;
239
240 /* static works are used, process one request at a time */
241 if (!mutex_trylock(&stop_cpus_mutex))
242 return -EAGAIN;
243 ret = __stop_cpus(cpumask, fn, arg);
244 mutex_unlock(&stop_cpus_mutex);
245 return ret;
246 }
247
cpu_stop_should_run(unsigned int cpu)248 static int cpu_stop_should_run(unsigned int cpu)
249 {
250 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
251 unsigned long flags;
252 int run;
253
254 spin_lock_irqsave(&stopper->lock, flags);
255 run = !list_empty(&stopper->works);
256 spin_unlock_irqrestore(&stopper->lock, flags);
257 return run;
258 }
259
cpu_stopper_thread(unsigned int cpu)260 static void cpu_stopper_thread(unsigned int cpu)
261 {
262 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
263 struct cpu_stop_work *work;
264 int ret;
265
266 repeat:
267 work = NULL;
268 spin_lock_irq(&stopper->lock);
269 if (!list_empty(&stopper->works)) {
270 work = list_first_entry(&stopper->works,
271 struct cpu_stop_work, list);
272 list_del_init(&work->list);
273 }
274 spin_unlock_irq(&stopper->lock);
275
276 if (work) {
277 cpu_stop_fn_t fn = work->fn;
278 void *arg = work->arg;
279 struct cpu_stop_done *done = work->done;
280 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
281
282 /* cpu stop callbacks are not allowed to sleep */
283 preempt_disable();
284
285 ret = fn(arg);
286 if (ret)
287 done->ret = ret;
288
289 /* restore preemption and check it's still balanced */
290 preempt_enable();
291 WARN_ONCE(preempt_count(),
292 "cpu_stop: %s(%p) leaked preempt count\n",
293 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
294 ksym_buf), arg);
295
296 cpu_stop_signal_done(done, true);
297 goto repeat;
298 }
299 }
300
301 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
302
cpu_stop_create(unsigned int cpu)303 static void cpu_stop_create(unsigned int cpu)
304 {
305 sched_set_stop_task(cpu, per_cpu(cpu_stopper_task, cpu));
306 }
307
cpu_stop_park(unsigned int cpu)308 static void cpu_stop_park(unsigned int cpu)
309 {
310 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
311 struct cpu_stop_work *work;
312 unsigned long flags;
313
314 /* drain remaining works */
315 spin_lock_irqsave(&stopper->lock, flags);
316 list_for_each_entry(work, &stopper->works, list)
317 cpu_stop_signal_done(work->done, false);
318 stopper->enabled = false;
319 spin_unlock_irqrestore(&stopper->lock, flags);
320 }
321
cpu_stop_unpark(unsigned int cpu)322 static void cpu_stop_unpark(unsigned int cpu)
323 {
324 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
325
326 spin_lock_irq(&stopper->lock);
327 stopper->enabled = true;
328 spin_unlock_irq(&stopper->lock);
329 }
330
331 static struct smp_hotplug_thread cpu_stop_threads = {
332 .store = &cpu_stopper_task,
333 .thread_should_run = cpu_stop_should_run,
334 .thread_fn = cpu_stopper_thread,
335 .thread_comm = "migration/%u",
336 .create = cpu_stop_create,
337 .setup = cpu_stop_unpark,
338 .park = cpu_stop_park,
339 .pre_unpark = cpu_stop_unpark,
340 .selfparking = true,
341 };
342
cpu_stop_init(void)343 static int __init cpu_stop_init(void)
344 {
345 unsigned int cpu;
346
347 for_each_possible_cpu(cpu) {
348 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
349
350 spin_lock_init(&stopper->lock);
351 INIT_LIST_HEAD(&stopper->works);
352 }
353
354 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
355 stop_machine_initialized = true;
356 return 0;
357 }
358 early_initcall(cpu_stop_init);
359
360 #ifdef CONFIG_STOP_MACHINE
361
362 /* This controls the threads on each CPU. */
363 enum stopmachine_state {
364 /* Dummy starting state for thread. */
365 STOPMACHINE_NONE,
366 /* Awaiting everyone to be scheduled. */
367 STOPMACHINE_PREPARE,
368 /* Disable interrupts. */
369 STOPMACHINE_DISABLE_IRQ,
370 /* Run the function */
371 STOPMACHINE_RUN,
372 /* Exit */
373 STOPMACHINE_EXIT,
374 };
375
376 struct stop_machine_data {
377 int (*fn)(void *);
378 void *data;
379 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
380 unsigned int num_threads;
381 const struct cpumask *active_cpus;
382
383 enum stopmachine_state state;
384 atomic_t thread_ack;
385 };
386
set_state(struct stop_machine_data * smdata,enum stopmachine_state newstate)387 static void set_state(struct stop_machine_data *smdata,
388 enum stopmachine_state newstate)
389 {
390 /* Reset ack counter. */
391 atomic_set(&smdata->thread_ack, smdata->num_threads);
392 smp_wmb();
393 smdata->state = newstate;
394 }
395
396 /* Last one to ack a state moves to the next state. */
ack_state(struct stop_machine_data * smdata)397 static void ack_state(struct stop_machine_data *smdata)
398 {
399 if (atomic_dec_and_test(&smdata->thread_ack))
400 set_state(smdata, smdata->state + 1);
401 }
402
403 /* This is the cpu_stop function which stops the CPU. */
stop_machine_cpu_stop(void * data)404 static int stop_machine_cpu_stop(void *data)
405 {
406 struct stop_machine_data *smdata = data;
407 enum stopmachine_state curstate = STOPMACHINE_NONE;
408 int cpu = smp_processor_id(), err = 0;
409 unsigned long flags;
410 bool is_active;
411
412 /*
413 * When called from stop_machine_from_inactive_cpu(), irq might
414 * already be disabled. Save the state and restore it on exit.
415 */
416 local_save_flags(flags);
417
418 if (!smdata->active_cpus)
419 is_active = cpu == cpumask_first(cpu_online_mask);
420 else
421 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
422
423 /* Simple state machine */
424 do {
425 /* Chill out and ensure we re-read stopmachine_state. */
426 cpu_relax();
427 if (smdata->state != curstate) {
428 curstate = smdata->state;
429 switch (curstate) {
430 case STOPMACHINE_DISABLE_IRQ:
431 local_irq_disable();
432 hard_irq_disable();
433 break;
434 case STOPMACHINE_RUN:
435 if (is_active)
436 err = smdata->fn(smdata->data);
437 break;
438 default:
439 break;
440 }
441 ack_state(smdata);
442 }
443 } while (curstate != STOPMACHINE_EXIT);
444
445 local_irq_restore(flags);
446 return err;
447 }
448
__stop_machine(int (* fn)(void *),void * data,const struct cpumask * cpus)449 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
450 {
451 struct stop_machine_data smdata = { .fn = fn, .data = data,
452 .num_threads = num_online_cpus(),
453 .active_cpus = cpus };
454
455 if (!stop_machine_initialized) {
456 /*
457 * Handle the case where stop_machine() is called
458 * early in boot before stop_machine() has been
459 * initialized.
460 */
461 unsigned long flags;
462 int ret;
463
464 WARN_ON_ONCE(smdata.num_threads != 1);
465
466 local_irq_save(flags);
467 hard_irq_disable();
468 ret = (*fn)(data);
469 local_irq_restore(flags);
470
471 return ret;
472 }
473
474 /* Set the initial state and stop all online cpus. */
475 set_state(&smdata, STOPMACHINE_PREPARE);
476 return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
477 }
478
stop_machine(int (* fn)(void *),void * data,const struct cpumask * cpus)479 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
480 {
481 int ret;
482
483 /* No CPUs can come up or down during this. */
484 get_online_cpus();
485 ret = __stop_machine(fn, data, cpus);
486 put_online_cpus();
487 return ret;
488 }
489 EXPORT_SYMBOL_GPL(stop_machine);
490
491 /**
492 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
493 * @fn: the function to run
494 * @data: the data ptr for the @fn()
495 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
496 *
497 * This is identical to stop_machine() but can be called from a CPU which
498 * is not active. The local CPU is in the process of hotplug (so no other
499 * CPU hotplug can start) and not marked active and doesn't have enough
500 * context to sleep.
501 *
502 * This function provides stop_machine() functionality for such state by
503 * using busy-wait for synchronization and executing @fn directly for local
504 * CPU.
505 *
506 * CONTEXT:
507 * Local CPU is inactive. Temporarily stops all active CPUs.
508 *
509 * RETURNS:
510 * 0 if all executions of @fn returned 0, any non zero return value if any
511 * returned non zero.
512 */
stop_machine_from_inactive_cpu(int (* fn)(void *),void * data,const struct cpumask * cpus)513 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
514 const struct cpumask *cpus)
515 {
516 struct stop_machine_data smdata = { .fn = fn, .data = data,
517 .active_cpus = cpus };
518 struct cpu_stop_done done;
519 int ret;
520
521 /* Local CPU must be inactive and CPU hotplug in progress. */
522 BUG_ON(cpu_active(raw_smp_processor_id()));
523 smdata.num_threads = num_active_cpus() + 1; /* +1 for local */
524
525 /* No proper task established and can't sleep - busy wait for lock. */
526 while (!mutex_trylock(&stop_cpus_mutex))
527 cpu_relax();
528
529 /* Schedule work on other CPUs and execute directly for local CPU */
530 set_state(&smdata, STOPMACHINE_PREPARE);
531 cpu_stop_init_done(&done, num_active_cpus());
532 queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
533 &done);
534 ret = stop_machine_cpu_stop(&smdata);
535
536 /* Busy wait for completion. */
537 while (!completion_done(&done.completion))
538 cpu_relax();
539
540 mutex_unlock(&stop_cpus_mutex);
541 return ret ?: done.ret;
542 }
543
544 #endif /* CONFIG_STOP_MACHINE */
545