sched/loadavg: Fix loadavg artifacts on fully idle and on fully loaded systems
[pandora-kernel.git] / kernel / stop_machine.c
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
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         struct task_struct      *thread;        /* stopper thread */
41 };
42
43 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
44 static bool stop_machine_initialized = false;
45
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 */
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 */
65 static void cpu_stop_queue_work(struct cpu_stopper *stopper,
66                                 struct cpu_stop_work *work)
67 {
68         unsigned long flags;
69
70         spin_lock_irqsave(&stopper->lock, flags);
71
72         if (stopper->enabled) {
73                 list_add_tail(&work->list, &stopper->works);
74                 wake_up_process(stopper->thread);
75         } else
76                 cpu_stop_signal_done(work->done, false);
77
78         spin_unlock_irqrestore(&stopper->lock, flags);
79 }
80
81 /**
82  * stop_one_cpu - stop a cpu
83  * @cpu: cpu to stop
84  * @fn: function to execute
85  * @arg: argument to @fn
86  *
87  * Execute @fn(@arg) on @cpu.  @fn is run in a process context with
88  * the highest priority preempting any task on the cpu and
89  * monopolizing it.  This function returns after the execution is
90  * complete.
91  *
92  * This function doesn't guarantee @cpu stays online till @fn
93  * completes.  If @cpu goes down in the middle, execution may happen
94  * partially or fully on different cpus.  @fn should either be ready
95  * for that or the caller should ensure that @cpu stays online until
96  * this function completes.
97  *
98  * CONTEXT:
99  * Might sleep.
100  *
101  * RETURNS:
102  * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
103  * otherwise, the return value of @fn.
104  */
105 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
106 {
107         struct cpu_stop_done done;
108         struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
109
110         cpu_stop_init_done(&done, 1);
111         cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), &work);
112         wait_for_completion(&done.completion);
113         return done.executed ? done.ret : -ENOENT;
114 }
115
116 /**
117  * stop_one_cpu_nowait - stop a cpu but don't wait for completion
118  * @cpu: cpu to stop
119  * @fn: function to execute
120  * @arg: argument to @fn
121  *
122  * Similar to stop_one_cpu() but doesn't wait for completion.  The
123  * caller is responsible for ensuring @work_buf is currently unused
124  * and will remain untouched until stopper starts executing @fn.
125  *
126  * CONTEXT:
127  * Don't care.
128  */
129 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
130                         struct cpu_stop_work *work_buf)
131 {
132         *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
133         cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
134 }
135
136 /* static data for stop_cpus */
137 static DEFINE_MUTEX(stop_cpus_mutex);
138 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
139
140 static void queue_stop_cpus_work(const struct cpumask *cpumask,
141                                  cpu_stop_fn_t fn, void *arg,
142                                  struct cpu_stop_done *done)
143 {
144         struct cpu_stop_work *work;
145         unsigned int cpu;
146
147         /* initialize works and done */
148         for_each_cpu(cpu, cpumask) {
149                 work = &per_cpu(stop_cpus_work, cpu);
150                 work->fn = fn;
151                 work->arg = arg;
152                 work->done = done;
153         }
154
155         /*
156          * Disable preemption while queueing to avoid getting
157          * preempted by a stopper which might wait for other stoppers
158          * to enter @fn which can lead to deadlock.
159          */
160         preempt_disable();
161         for_each_cpu(cpu, cpumask)
162                 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
163                                     &per_cpu(stop_cpus_work, cpu));
164         preempt_enable();
165 }
166
167 static int __stop_cpus(const struct cpumask *cpumask,
168                        cpu_stop_fn_t fn, void *arg)
169 {
170         struct cpu_stop_done done;
171
172         cpu_stop_init_done(&done, cpumask_weight(cpumask));
173         queue_stop_cpus_work(cpumask, fn, arg, &done);
174         wait_for_completion(&done.completion);
175         return done.executed ? done.ret : -ENOENT;
176 }
177
178 /**
179  * stop_cpus - stop multiple cpus
180  * @cpumask: cpus to stop
181  * @fn: function to execute
182  * @arg: argument to @fn
183  *
184  * Execute @fn(@arg) on online cpus in @cpumask.  On each target cpu,
185  * @fn is run in a process context with the highest priority
186  * preempting any task on the cpu and monopolizing it.  This function
187  * returns after all executions are complete.
188  *
189  * This function doesn't guarantee the cpus in @cpumask stay online
190  * till @fn completes.  If some cpus go down in the middle, execution
191  * on the cpu may happen partially or fully on different cpus.  @fn
192  * should either be ready for that or the caller should ensure that
193  * the cpus stay online until this function completes.
194  *
195  * All stop_cpus() calls are serialized making it safe for @fn to wait
196  * for all cpus to start executing it.
197  *
198  * CONTEXT:
199  * Might sleep.
200  *
201  * RETURNS:
202  * -ENOENT if @fn(@arg) was not executed at all because all cpus in
203  * @cpumask were offline; otherwise, 0 if all executions of @fn
204  * returned 0, any non zero return value if any returned non zero.
205  */
206 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
207 {
208         int ret;
209
210         /* static works are used, process one request at a time */
211         mutex_lock(&stop_cpus_mutex);
212         ret = __stop_cpus(cpumask, fn, arg);
213         mutex_unlock(&stop_cpus_mutex);
214         return ret;
215 }
216
217 /**
218  * try_stop_cpus - try to stop multiple cpus
219  * @cpumask: cpus to stop
220  * @fn: function to execute
221  * @arg: argument to @fn
222  *
223  * Identical to stop_cpus() except that it fails with -EAGAIN if
224  * someone else is already using the facility.
225  *
226  * CONTEXT:
227  * Might sleep.
228  *
229  * RETURNS:
230  * -EAGAIN if someone else is already stopping cpus, -ENOENT if
231  * @fn(@arg) was not executed at all because all cpus in @cpumask were
232  * offline; otherwise, 0 if all executions of @fn returned 0, any non
233  * zero return value if any returned non zero.
234  */
235 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
236 {
237         int ret;
238
239         /* static works are used, process one request at a time */
240         if (!mutex_trylock(&stop_cpus_mutex))
241                 return -EAGAIN;
242         ret = __stop_cpus(cpumask, fn, arg);
243         mutex_unlock(&stop_cpus_mutex);
244         return ret;
245 }
246
247 static int cpu_stopper_thread(void *data)
248 {
249         struct cpu_stopper *stopper = data;
250         struct cpu_stop_work *work;
251         int ret;
252
253 repeat:
254         set_current_state(TASK_INTERRUPTIBLE);  /* mb paired w/ kthread_stop */
255
256         if (kthread_should_stop()) {
257                 __set_current_state(TASK_RUNNING);
258                 return 0;
259         }
260
261         work = NULL;
262         spin_lock_irq(&stopper->lock);
263         if (!list_empty(&stopper->works)) {
264                 work = list_first_entry(&stopper->works,
265                                         struct cpu_stop_work, list);
266                 list_del_init(&work->list);
267         }
268         spin_unlock_irq(&stopper->lock);
269
270         if (work) {
271                 cpu_stop_fn_t fn = work->fn;
272                 void *arg = work->arg;
273                 struct cpu_stop_done *done = work->done;
274                 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
275
276                 __set_current_state(TASK_RUNNING);
277
278                 /* cpu stop callbacks are not allowed to sleep */
279                 preempt_disable();
280
281                 ret = fn(arg);
282                 if (ret)
283                         done->ret = ret;
284
285                 /* restore preemption and check it's still balanced */
286                 preempt_enable();
287                 WARN_ONCE(preempt_count(),
288                           "cpu_stop: %s(%p) leaked preempt count\n",
289                           kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
290                                           ksym_buf), arg);
291
292                 cpu_stop_signal_done(done, true);
293         } else
294                 schedule();
295
296         goto repeat;
297 }
298
299 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
300
301 /* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
302 static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
303                                            unsigned long action, void *hcpu)
304 {
305         unsigned int cpu = (unsigned long)hcpu;
306         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
307         struct task_struct *p;
308
309         switch (action & ~CPU_TASKS_FROZEN) {
310         case CPU_UP_PREPARE:
311                 BUG_ON(stopper->thread || stopper->enabled ||
312                        !list_empty(&stopper->works));
313                 p = kthread_create_on_node(cpu_stopper_thread,
314                                            stopper,
315                                            cpu_to_node(cpu),
316                                            "migration/%d", cpu);
317                 if (IS_ERR(p))
318                         return notifier_from_errno(PTR_ERR(p));
319                 get_task_struct(p);
320                 kthread_bind(p, cpu);
321                 sched_set_stop_task(cpu, p);
322                 stopper->thread = p;
323                 break;
324
325         case CPU_ONLINE:
326                 /* strictly unnecessary, as first user will wake it */
327                 wake_up_process(stopper->thread);
328                 /* mark enabled */
329                 spin_lock_irq(&stopper->lock);
330                 stopper->enabled = true;
331                 spin_unlock_irq(&stopper->lock);
332                 break;
333
334 #ifdef CONFIG_HOTPLUG_CPU
335         case CPU_UP_CANCELED:
336         case CPU_POST_DEAD:
337         {
338                 struct cpu_stop_work *work;
339
340                 sched_set_stop_task(cpu, NULL);
341                 /* kill the stopper */
342                 kthread_stop(stopper->thread);
343                 /* drain remaining works */
344                 spin_lock_irq(&stopper->lock);
345                 list_for_each_entry(work, &stopper->works, list)
346                         cpu_stop_signal_done(work->done, false);
347                 stopper->enabled = false;
348                 spin_unlock_irq(&stopper->lock);
349                 /* release the stopper */
350                 put_task_struct(stopper->thread);
351                 stopper->thread = NULL;
352                 break;
353         }
354 #endif
355         }
356
357         return NOTIFY_OK;
358 }
359
360 /*
361  * Give it a higher priority so that cpu stopper is available to other
362  * cpu notifiers.  It currently shares the same priority as sched
363  * migration_notifier.
364  */
365 static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
366         .notifier_call  = cpu_stop_cpu_callback,
367         .priority       = 10,
368 };
369
370 static int __init cpu_stop_init(void)
371 {
372         void *bcpu = (void *)(long)smp_processor_id();
373         unsigned int cpu;
374         int err;
375
376         for_each_possible_cpu(cpu) {
377                 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
378
379                 spin_lock_init(&stopper->lock);
380                 INIT_LIST_HEAD(&stopper->works);
381         }
382
383         /* start one for the boot cpu */
384         err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
385                                     bcpu);
386         BUG_ON(err != NOTIFY_OK);
387         cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_ONLINE, bcpu);
388         register_cpu_notifier(&cpu_stop_cpu_notifier);
389
390         stop_machine_initialized = true;
391
392         return 0;
393 }
394 early_initcall(cpu_stop_init);
395
396 #ifdef CONFIG_STOP_MACHINE
397
398 /* This controls the threads on each CPU. */
399 enum stopmachine_state {
400         /* Dummy starting state for thread. */
401         STOPMACHINE_NONE,
402         /* Awaiting everyone to be scheduled. */
403         STOPMACHINE_PREPARE,
404         /* Disable interrupts. */
405         STOPMACHINE_DISABLE_IRQ,
406         /* Run the function */
407         STOPMACHINE_RUN,
408         /* Exit */
409         STOPMACHINE_EXIT,
410 };
411
412 struct stop_machine_data {
413         int                     (*fn)(void *);
414         void                    *data;
415         /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
416         unsigned int            num_threads;
417         const struct cpumask    *active_cpus;
418
419         enum stopmachine_state  state;
420         atomic_t                thread_ack;
421 };
422
423 static void set_state(struct stop_machine_data *smdata,
424                       enum stopmachine_state newstate)
425 {
426         /* Reset ack counter. */
427         atomic_set(&smdata->thread_ack, smdata->num_threads);
428         smp_wmb();
429         smdata->state = newstate;
430 }
431
432 /* Last one to ack a state moves to the next state. */
433 static void ack_state(struct stop_machine_data *smdata)
434 {
435         if (atomic_dec_and_test(&smdata->thread_ack))
436                 set_state(smdata, smdata->state + 1);
437 }
438
439 /* This is the cpu_stop function which stops the CPU. */
440 static int stop_machine_cpu_stop(void *data)
441 {
442         struct stop_machine_data *smdata = data;
443         enum stopmachine_state curstate = STOPMACHINE_NONE;
444         int cpu = smp_processor_id(), err = 0;
445         unsigned long flags;
446         bool is_active;
447
448         /*
449          * When called from stop_machine_from_inactive_cpu(), irq might
450          * already be disabled.  Save the state and restore it on exit.
451          */
452         local_save_flags(flags);
453
454         if (!smdata->active_cpus)
455                 is_active = cpu == cpumask_first(cpu_online_mask);
456         else
457                 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
458
459         /* Simple state machine */
460         do {
461                 /* Chill out and ensure we re-read stopmachine_state. */
462                 cpu_relax();
463                 if (smdata->state != curstate) {
464                         curstate = smdata->state;
465                         switch (curstate) {
466                         case STOPMACHINE_DISABLE_IRQ:
467                                 local_irq_disable();
468                                 hard_irq_disable();
469                                 break;
470                         case STOPMACHINE_RUN:
471                                 if (is_active)
472                                         err = smdata->fn(smdata->data);
473                                 break;
474                         default:
475                                 break;
476                         }
477                         ack_state(smdata);
478                 }
479         } while (curstate != STOPMACHINE_EXIT);
480
481         local_irq_restore(flags);
482         return err;
483 }
484
485 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
486 {
487         struct stop_machine_data smdata = { .fn = fn, .data = data,
488                                             .num_threads = num_online_cpus(),
489                                             .active_cpus = cpus };
490
491         if (!stop_machine_initialized) {
492                 /*
493                  * Handle the case where stop_machine() is called
494                  * early in boot before stop_machine() has been
495                  * initialized.
496                  */
497                 unsigned long flags;
498                 int ret;
499
500                 WARN_ON_ONCE(smdata.num_threads != 1);
501
502                 local_irq_save(flags);
503                 hard_irq_disable();
504                 ret = (*fn)(data);
505                 local_irq_restore(flags);
506
507                 return ret;
508         }
509
510         /* Set the initial state and stop all online cpus. */
511         set_state(&smdata, STOPMACHINE_PREPARE);
512         return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
513 }
514
515 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
516 {
517         int ret;
518
519         /* No CPUs can come up or down during this. */
520         get_online_cpus();
521         ret = __stop_machine(fn, data, cpus);
522         put_online_cpus();
523         return ret;
524 }
525 EXPORT_SYMBOL_GPL(stop_machine);
526
527 /**
528  * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
529  * @fn: the function to run
530  * @data: the data ptr for the @fn()
531  * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
532  *
533  * This is identical to stop_machine() but can be called from a CPU which
534  * is not active.  The local CPU is in the process of hotplug (so no other
535  * CPU hotplug can start) and not marked active and doesn't have enough
536  * context to sleep.
537  *
538  * This function provides stop_machine() functionality for such state by
539  * using busy-wait for synchronization and executing @fn directly for local
540  * CPU.
541  *
542  * CONTEXT:
543  * Local CPU is inactive.  Temporarily stops all active CPUs.
544  *
545  * RETURNS:
546  * 0 if all executions of @fn returned 0, any non zero return value if any
547  * returned non zero.
548  */
549 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
550                                   const struct cpumask *cpus)
551 {
552         struct stop_machine_data smdata = { .fn = fn, .data = data,
553                                             .active_cpus = cpus };
554         struct cpu_stop_done done;
555         int ret;
556
557         /* Local CPU must be inactive and CPU hotplug in progress. */
558         BUG_ON(cpu_active(raw_smp_processor_id()));
559         smdata.num_threads = num_active_cpus() + 1;     /* +1 for local */
560
561         /* No proper task established and can't sleep - busy wait for lock. */
562         while (!mutex_trylock(&stop_cpus_mutex))
563                 cpu_relax();
564
565         /* Schedule work on other CPUs and execute directly for local CPU */
566         set_state(&smdata, STOPMACHINE_PREPARE);
567         cpu_stop_init_done(&done, num_active_cpus());
568         queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
569                              &done);
570         ret = stop_machine_cpu_stop(&smdata);
571
572         /* Busy wait for completion. */
573         while (!completion_done(&done.completion))
574                 cpu_relax();
575
576         mutex_unlock(&stop_cpus_mutex);
577         return ret ?: done.ret;
578 }
579
580 #endif  /* CONFIG_STOP_MACHINE */