projects/linux·blob v4.10-rc1 · refs

1805 lines · 43.8 KiB

1/* CPU control.
2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
3 *
4 * This code is licenced under the GPL.
5 */
6#include <linux/proc_fs.h>
7#include <linux/smp.h>
8#include <linux/init.h>
9#include <linux/notifier.h>
10#include <linux/sched.h>
11#include <linux/unistd.h>
12#include <linux/cpu.h>
13#include <linux/oom.h>
14#include <linux/rcupdate.h>
15#include <linux/export.h>
16#include <linux/bug.h>
17#include <linux/kthread.h>
18#include <linux/stop_machine.h>
19#include <linux/mutex.h>
20#include <linux/gfp.h>
21#include <linux/suspend.h>
22#include <linux/lockdep.h>
23#include <linux/tick.h>
24#include <linux/irq.h>
25#include <linux/smpboot.h>
26#include <linux/relay.h>
27#include <linux/slab.h>
28
29#include <trace/events/power.h>
30#define CREATE_TRACE_POINTS
31#include <trace/events/cpuhp.h>
32
33#include "smpboot.h"
34
35/**
36 * cpuhp_cpu_state - Per cpu hotplug state storage
37 * @state: The current cpu state
38 * @target: The target state
39 * @thread: Pointer to the hotplug thread
40 * @should_run: Thread should execute
41 * @rollback: Perform a rollback
42 * @single: Single callback invocation
43 * @bringup: Single callback bringup or teardown selector
44 * @cb_state: The state for a single callback (install/uninstall)
45 * @result: Result of the operation
46 * @done: Signal completion to the issuer of the task
47 */
48struct cpuhp_cpu_state {
49 enum cpuhp_state state;
50 enum cpuhp_state target;
51#ifdef CONFIG_SMP
52 struct task_struct *thread;
53 bool should_run;
54 bool rollback;
55 bool single;
56 bool bringup;
57 struct hlist_node *node;
58 enum cpuhp_state cb_state;
59 int result;
60 struct completion done;
61#endif
62};
63
64static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
65
66/**
67 * cpuhp_step - Hotplug state machine step
68 * @name: Name of the step
69 * @startup: Startup function of the step
70 * @teardown: Teardown function of the step
71 * @skip_onerr: Do not invoke the functions on error rollback
72 * Will go away once the notifiers are gone
73 * @cant_stop: Bringup/teardown can't be stopped at this step
74 */
75struct cpuhp_step {
76 const char *name;
77 union {
78 int (*single)(unsigned int cpu);
79 int (*multi)(unsigned int cpu,
80 struct hlist_node *node);
81 } startup;
82 union {
83 int (*single)(unsigned int cpu);
84 int (*multi)(unsigned int cpu,
85 struct hlist_node *node);
86 } teardown;
87 struct hlist_head list;
88 bool skip_onerr;
89 bool cant_stop;
90 bool multi_instance;
91};
92
93static DEFINE_MUTEX(cpuhp_state_mutex);
94static struct cpuhp_step cpuhp_bp_states[];
95static struct cpuhp_step cpuhp_ap_states[];
96
97static bool cpuhp_is_ap_state(enum cpuhp_state state)
98{
99 /*
100 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
101 * purposes as that state is handled explicitly in cpu_down.
102 */
103 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
104}
105
106static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
107{
108 struct cpuhp_step *sp;
109
110 sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
111 return sp + state;
112}
113
114/**
115 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
116 * @cpu: The cpu for which the callback should be invoked
117 * @step: The step in the state machine
118 * @bringup: True if the bringup callback should be invoked
119 *
120 * Called from cpu hotplug and from the state register machinery.
121 */
122static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
123 bool bringup, struct hlist_node *node)
124{
125 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
126 struct cpuhp_step *step = cpuhp_get_step(state);
127 int (*cbm)(unsigned int cpu, struct hlist_node *node);
128 int (*cb)(unsigned int cpu);
129 int ret, cnt;
130
131 if (!step->multi_instance) {
132 cb = bringup ? step->startup.single : step->teardown.single;
133 if (!cb)
134 return 0;
135 trace_cpuhp_enter(cpu, st->target, state, cb);
136 ret = cb(cpu);
137 trace_cpuhp_exit(cpu, st->state, state, ret);
138 return ret;
139 }
140 cbm = bringup ? step->startup.multi : step->teardown.multi;
141 if (!cbm)
142 return 0;
143
144 /* Single invocation for instance add/remove */
145 if (node) {
146 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
147 ret = cbm(cpu, node);
148 trace_cpuhp_exit(cpu, st->state, state, ret);
149 return ret;
150 }
151
152 /* State transition. Invoke on all instances */
153 cnt = 0;
154 hlist_for_each(node, &step->list) {
155 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
156 ret = cbm(cpu, node);
157 trace_cpuhp_exit(cpu, st->state, state, ret);
158 if (ret)
159 goto err;
160 cnt++;
161 }
162 return 0;
163err:
164 /* Rollback the instances if one failed */
165 cbm = !bringup ? step->startup.multi : step->teardown.multi;
166 if (!cbm)
167 return ret;
168
169 hlist_for_each(node, &step->list) {
170 if (!cnt--)
171 break;
172 cbm(cpu, node);
173 }
174 return ret;
175}
176
177#ifdef CONFIG_SMP
178/* Serializes the updates to cpu_online_mask, cpu_present_mask */
179static DEFINE_MUTEX(cpu_add_remove_lock);
180bool cpuhp_tasks_frozen;
181EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
182
183/*
184 * The following two APIs (cpu_maps_update_begin/done) must be used when
185 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
186 */
187void cpu_maps_update_begin(void)
188{
189 mutex_lock(&cpu_add_remove_lock);
190}
191
192void cpu_maps_update_done(void)
193{
194 mutex_unlock(&cpu_add_remove_lock);
195}
196
197/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
198 * Should always be manipulated under cpu_add_remove_lock
199 */
200static int cpu_hotplug_disabled;
201
202#ifdef CONFIG_HOTPLUG_CPU
203
204static struct {
205 struct task_struct *active_writer;
206 /* wait queue to wake up the active_writer */
207 wait_queue_head_t wq;
208 /* verifies that no writer will get active while readers are active */
209 struct mutex lock;
210 /*
211 * Also blocks the new readers during
212 * an ongoing cpu hotplug operation.
213 */
214 atomic_t refcount;
215
216#ifdef CONFIG_DEBUG_LOCK_ALLOC
217 struct lockdep_map dep_map;
218#endif
219} cpu_hotplug = {
220 .active_writer = NULL,
221 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
222 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
223#ifdef CONFIG_DEBUG_LOCK_ALLOC
224 .dep_map = STATIC_LOCKDEP_MAP_INIT("cpu_hotplug.dep_map", &cpu_hotplug.dep_map),
225#endif
226};
227
228/* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
229#define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
230#define cpuhp_lock_acquire_tryread() \
231 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
232#define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
233#define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
234
235
236void get_online_cpus(void)
237{
238 might_sleep();
239 if (cpu_hotplug.active_writer == current)
240 return;
241 cpuhp_lock_acquire_read();
242 mutex_lock(&cpu_hotplug.lock);
243 atomic_inc(&cpu_hotplug.refcount);
244 mutex_unlock(&cpu_hotplug.lock);
245}
246EXPORT_SYMBOL_GPL(get_online_cpus);
247
248void put_online_cpus(void)
249{
250 int refcount;
251
252 if (cpu_hotplug.active_writer == current)
253 return;
254
255 refcount = atomic_dec_return(&cpu_hotplug.refcount);
256 if (WARN_ON(refcount < 0)) /* try to fix things up */
257 atomic_inc(&cpu_hotplug.refcount);
258
259 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
260 wake_up(&cpu_hotplug.wq);
261
262 cpuhp_lock_release();
263
264}
265EXPORT_SYMBOL_GPL(put_online_cpus);
266
267/*
268 * This ensures that the hotplug operation can begin only when the
269 * refcount goes to zero.
270 *
271 * Note that during a cpu-hotplug operation, the new readers, if any,
272 * will be blocked by the cpu_hotplug.lock
273 *
274 * Since cpu_hotplug_begin() is always called after invoking
275 * cpu_maps_update_begin(), we can be sure that only one writer is active.
276 *
277 * Note that theoretically, there is a possibility of a livelock:
278 * - Refcount goes to zero, last reader wakes up the sleeping
279 * writer.
280 * - Last reader unlocks the cpu_hotplug.lock.
281 * - A new reader arrives at this moment, bumps up the refcount.
282 * - The writer acquires the cpu_hotplug.lock finds the refcount
283 * non zero and goes to sleep again.
284 *
285 * However, this is very difficult to achieve in practice since
286 * get_online_cpus() not an api which is called all that often.
287 *
288 */
289void cpu_hotplug_begin(void)
290{
291 DEFINE_WAIT(wait);
292
293 cpu_hotplug.active_writer = current;
294 cpuhp_lock_acquire();
295
296 for (;;) {
297 mutex_lock(&cpu_hotplug.lock);
298 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
299 if (likely(!atomic_read(&cpu_hotplug.refcount)))
300 break;
301 mutex_unlock(&cpu_hotplug.lock);
302 schedule();
303 }
304 finish_wait(&cpu_hotplug.wq, &wait);
305}
306
307void cpu_hotplug_done(void)
308{
309 cpu_hotplug.active_writer = NULL;
310 mutex_unlock(&cpu_hotplug.lock);
311 cpuhp_lock_release();
312}
313
314/*
315 * Wait for currently running CPU hotplug operations to complete (if any) and
316 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
317 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
318 * hotplug path before performing hotplug operations. So acquiring that lock
319 * guarantees mutual exclusion from any currently running hotplug operations.
320 */
321void cpu_hotplug_disable(void)
322{
323 cpu_maps_update_begin();
324 cpu_hotplug_disabled++;
325 cpu_maps_update_done();
326}
327EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
328
329static void __cpu_hotplug_enable(void)
330{
331 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
332 return;
333 cpu_hotplug_disabled--;
334}
335
336void cpu_hotplug_enable(void)
337{
338 cpu_maps_update_begin();
339 __cpu_hotplug_enable();
340 cpu_maps_update_done();
341}
342EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
343#endif /* CONFIG_HOTPLUG_CPU */
344
345/* Notifier wrappers for transitioning to state machine */
346
347static int bringup_wait_for_ap(unsigned int cpu)
348{
349 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
350
351 wait_for_completion(&st->done);
352 return st->result;
353}
354
355static int bringup_cpu(unsigned int cpu)
356{
357 struct task_struct *idle = idle_thread_get(cpu);
358 int ret;
359
360 /*
361 * Some architectures have to walk the irq descriptors to
362 * setup the vector space for the cpu which comes online.
363 * Prevent irq alloc/free across the bringup.
364 */
365 irq_lock_sparse();
366
367 /* Arch-specific enabling code. */
368 ret = __cpu_up(cpu, idle);
369 irq_unlock_sparse();
370 if (ret)
371 return ret;
372 ret = bringup_wait_for_ap(cpu);
373 BUG_ON(!cpu_online(cpu));
374 return ret;
375}
376
377/*
378 * Hotplug state machine related functions
379 */
380static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
381{
382 for (st->state++; st->state < st->target; st->state++) {
383 struct cpuhp_step *step = cpuhp_get_step(st->state);
384
385 if (!step->skip_onerr)
386 cpuhp_invoke_callback(cpu, st->state, true, NULL);
387 }
388}
389
390static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
391 enum cpuhp_state target)
392{
393 enum cpuhp_state prev_state = st->state;
394 int ret = 0;
395
396 for (; st->state > target; st->state--) {
397 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
398 if (ret) {
399 st->target = prev_state;
400 undo_cpu_down(cpu, st);
401 break;
402 }
403 }
404 return ret;
405}
406
407static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
408{
409 for (st->state--; st->state > st->target; st->state--) {
410 struct cpuhp_step *step = cpuhp_get_step(st->state);
411
412 if (!step->skip_onerr)
413 cpuhp_invoke_callback(cpu, st->state, false, NULL);
414 }
415}
416
417static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
418 enum cpuhp_state target)
419{
420 enum cpuhp_state prev_state = st->state;
421 int ret = 0;
422
423 while (st->state < target) {
424 st->state++;
425 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
426 if (ret) {
427 st->target = prev_state;
428 undo_cpu_up(cpu, st);
429 break;
430 }
431 }
432 return ret;
433}
434
435/*
436 * The cpu hotplug threads manage the bringup and teardown of the cpus
437 */
438static void cpuhp_create(unsigned int cpu)
439{
440 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
441
442 init_completion(&st->done);
443}
444
445static int cpuhp_should_run(unsigned int cpu)
446{
447 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
448
449 return st->should_run;
450}
451
452/* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
453static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
454{
455 enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
456
457 return cpuhp_down_callbacks(cpu, st, target);
458}
459
460/* Execute the online startup callbacks. Used to be CPU_ONLINE */
461static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
462{
463 return cpuhp_up_callbacks(cpu, st, st->target);
464}
465
466/*
467 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
468 * callbacks when a state gets [un]installed at runtime.
469 */
470static void cpuhp_thread_fun(unsigned int cpu)
471{
472 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
473 int ret = 0;
474
475 /*
476 * Paired with the mb() in cpuhp_kick_ap_work and
477 * cpuhp_invoke_ap_callback, so the work set is consistent visible.
478 */
479 smp_mb();
480 if (!st->should_run)
481 return;
482
483 st->should_run = false;
484
485 /* Single callback invocation for [un]install ? */
486 if (st->single) {
487 if (st->cb_state < CPUHP_AP_ONLINE) {
488 local_irq_disable();
489 ret = cpuhp_invoke_callback(cpu, st->cb_state,
490 st->bringup, st->node);
491 local_irq_enable();
492 } else {
493 ret = cpuhp_invoke_callback(cpu, st->cb_state,
494 st->bringup, st->node);
495 }
496 } else if (st->rollback) {
497 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
498
499 undo_cpu_down(cpu, st);
500 st->rollback = false;
501 } else {
502 /* Cannot happen .... */
503 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
504
505 /* Regular hotplug work */
506 if (st->state < st->target)
507 ret = cpuhp_ap_online(cpu, st);
508 else if (st->state > st->target)
509 ret = cpuhp_ap_offline(cpu, st);
510 }
511 st->result = ret;
512 complete(&st->done);
513}
514
515/* Invoke a single callback on a remote cpu */
516static int
517cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
518 struct hlist_node *node)
519{
520 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
521
522 if (!cpu_online(cpu))
523 return 0;
524
525 /*
526 * If we are up and running, use the hotplug thread. For early calls
527 * we invoke the thread function directly.
528 */
529 if (!st->thread)
530 return cpuhp_invoke_callback(cpu, state, bringup, node);
531
532 st->cb_state = state;
533 st->single = true;
534 st->bringup = bringup;
535 st->node = node;
536
537 /*
538 * Make sure the above stores are visible before should_run becomes
539 * true. Paired with the mb() above in cpuhp_thread_fun()
540 */
541 smp_mb();
542 st->should_run = true;
543 wake_up_process(st->thread);
544 wait_for_completion(&st->done);
545 return st->result;
546}
547
548/* Regular hotplug invocation of the AP hotplug thread */
549static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
550{
551 st->result = 0;
552 st->single = false;
553 /*
554 * Make sure the above stores are visible before should_run becomes
555 * true. Paired with the mb() above in cpuhp_thread_fun()
556 */
557 smp_mb();
558 st->should_run = true;
559 wake_up_process(st->thread);
560}
561
562static int cpuhp_kick_ap_work(unsigned int cpu)
563{
564 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
565 enum cpuhp_state state = st->state;
566
567 trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
568 __cpuhp_kick_ap_work(st);
569 wait_for_completion(&st->done);
570 trace_cpuhp_exit(cpu, st->state, state, st->result);
571 return st->result;
572}
573
574static struct smp_hotplug_thread cpuhp_threads = {
575 .store = &cpuhp_state.thread,
576 .create = &cpuhp_create,
577 .thread_should_run = cpuhp_should_run,
578 .thread_fn = cpuhp_thread_fun,
579 .thread_comm = "cpuhp/%u",
580 .selfparking = true,
581};
582
583void __init cpuhp_threads_init(void)
584{
585 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
586 kthread_unpark(this_cpu_read(cpuhp_state.thread));
587}
588
589#ifdef CONFIG_HOTPLUG_CPU
590/**
591 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
592 * @cpu: a CPU id
593 *
594 * This function walks all processes, finds a valid mm struct for each one and
595 * then clears a corresponding bit in mm's cpumask. While this all sounds
596 * trivial, there are various non-obvious corner cases, which this function
597 * tries to solve in a safe manner.
598 *
599 * Also note that the function uses a somewhat relaxed locking scheme, so it may
600 * be called only for an already offlined CPU.
601 */
602void clear_tasks_mm_cpumask(int cpu)
603{
604 struct task_struct *p;
605
606 /*
607 * This function is called after the cpu is taken down and marked
608 * offline, so its not like new tasks will ever get this cpu set in
609 * their mm mask. -- Peter Zijlstra
610 * Thus, we may use rcu_read_lock() here, instead of grabbing
611 * full-fledged tasklist_lock.
612 */
613 WARN_ON(cpu_online(cpu));
614 rcu_read_lock();
615 for_each_process(p) {
616 struct task_struct *t;
617
618 /*
619 * Main thread might exit, but other threads may still have
620 * a valid mm. Find one.
621 */
622 t = find_lock_task_mm(p);
623 if (!t)
624 continue;
625 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
626 task_unlock(t);
627 }
628 rcu_read_unlock();
629}
630
631static inline void check_for_tasks(int dead_cpu)
632{
633 struct task_struct *g, *p;
634
635 read_lock(&tasklist_lock);
636 for_each_process_thread(g, p) {
637 if (!p->on_rq)
638 continue;
639 /*
640 * We do the check with unlocked task_rq(p)->lock.
641 * Order the reading to do not warn about a task,
642 * which was running on this cpu in the past, and
643 * it's just been woken on another cpu.
644 */
645 rmb();
646 if (task_cpu(p) != dead_cpu)
647 continue;
648
649 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
650 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
651 }
652 read_unlock(&tasklist_lock);
653}
654
655/* Take this CPU down. */
656static int take_cpu_down(void *_param)
657{
658 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
659 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
660 int err, cpu = smp_processor_id();
661
662 /* Ensure this CPU doesn't handle any more interrupts. */
663 err = __cpu_disable();
664 if (err < 0)
665 return err;
666
667 /*
668 * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
669 * do this step again.
670 */
671 WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
672 st->state--;
673 /* Invoke the former CPU_DYING callbacks */
674 for (; st->state > target; st->state--)
675 cpuhp_invoke_callback(cpu, st->state, false, NULL);
676
677 /* Give up timekeeping duties */
678 tick_handover_do_timer();
679 /* Park the stopper thread */
680 stop_machine_park(cpu);
681 return 0;
682}
683
684static int takedown_cpu(unsigned int cpu)
685{
686 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
687 int err;
688
689 /* Park the smpboot threads */
690 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
691 smpboot_park_threads(cpu);
692
693 /*
694 * Prevent irq alloc/free while the dying cpu reorganizes the
695 * interrupt affinities.
696 */
697 irq_lock_sparse();
698
699 /*
700 * So now all preempt/rcu users must observe !cpu_active().
701 */
702 err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
703 if (err) {
704 /* CPU refused to die */
705 irq_unlock_sparse();
706 /* Unpark the hotplug thread so we can rollback there */
707 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
708 return err;
709 }
710 BUG_ON(cpu_online(cpu));
711
712 /*
713 * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
714 * runnable tasks from the cpu, there's only the idle task left now
715 * that the migration thread is done doing the stop_machine thing.
716 *
717 * Wait for the stop thread to go away.
718 */
719 wait_for_completion(&st->done);
720 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
721
722 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
723 irq_unlock_sparse();
724
725 hotplug_cpu__broadcast_tick_pull(cpu);
726 /* This actually kills the CPU. */
727 __cpu_die(cpu);
728
729 tick_cleanup_dead_cpu(cpu);
730 return 0;
731}
732
733static void cpuhp_complete_idle_dead(void *arg)
734{
735 struct cpuhp_cpu_state *st = arg;
736
737 complete(&st->done);
738}
739
740void cpuhp_report_idle_dead(void)
741{
742 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
743
744 BUG_ON(st->state != CPUHP_AP_OFFLINE);
745 rcu_report_dead(smp_processor_id());
746 st->state = CPUHP_AP_IDLE_DEAD;
747 /*
748 * We cannot call complete after rcu_report_dead() so we delegate it
749 * to an online cpu.
750 */
751 smp_call_function_single(cpumask_first(cpu_online_mask),
752 cpuhp_complete_idle_dead, st, 0);
753}
754
755#else
756#define takedown_cpu NULL
757#endif
758
759#ifdef CONFIG_HOTPLUG_CPU
760
761/* Requires cpu_add_remove_lock to be held */
762static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
763 enum cpuhp_state target)
764{
765 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
766 int prev_state, ret = 0;
767 bool hasdied = false;
768
769 if (num_online_cpus() == 1)
770 return -EBUSY;
771
772 if (!cpu_present(cpu))
773 return -EINVAL;
774
775 cpu_hotplug_begin();
776
777 cpuhp_tasks_frozen = tasks_frozen;
778
779 prev_state = st->state;
780 st->target = target;
781 /*
782 * If the current CPU state is in the range of the AP hotplug thread,
783 * then we need to kick the thread.
784 */
785 if (st->state > CPUHP_TEARDOWN_CPU) {
786 ret = cpuhp_kick_ap_work(cpu);
787 /*
788 * The AP side has done the error rollback already. Just
789 * return the error code..
790 */
791 if (ret)
792 goto out;
793
794 /*
795 * We might have stopped still in the range of the AP hotplug
796 * thread. Nothing to do anymore.
797 */
798 if (st->state > CPUHP_TEARDOWN_CPU)
799 goto out;
800 }
801 /*
802 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
803 * to do the further cleanups.
804 */
805 ret = cpuhp_down_callbacks(cpu, st, target);
806 if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
807 st->target = prev_state;
808 st->rollback = true;
809 cpuhp_kick_ap_work(cpu);
810 }
811
812 hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
813out:
814 cpu_hotplug_done();
815 return ret;
816}
817
818static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
819{
820 int err;
821
822 cpu_maps_update_begin();
823
824 if (cpu_hotplug_disabled) {
825 err = -EBUSY;
826 goto out;
827 }
828
829 err = _cpu_down(cpu, 0, target);
830
831out:
832 cpu_maps_update_done();
833 return err;
834}
835int cpu_down(unsigned int cpu)
836{
837 return do_cpu_down(cpu, CPUHP_OFFLINE);
838}
839EXPORT_SYMBOL(cpu_down);
840#endif /*CONFIG_HOTPLUG_CPU*/
841
842/**
843 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
844 * @cpu: cpu that just started
845 *
846 * It must be called by the arch code on the new cpu, before the new cpu
847 * enables interrupts and before the "boot" cpu returns from __cpu_up().
848 */
849void notify_cpu_starting(unsigned int cpu)
850{
851 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
852 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
853
854 rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
855 while (st->state < target) {
856 st->state++;
857 cpuhp_invoke_callback(cpu, st->state, true, NULL);
858 }
859}
860
861/*
862 * Called from the idle task. We need to set active here, so we can kick off
863 * the stopper thread and unpark the smpboot threads. If the target state is
864 * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
865 * cpu further.
866 */
867void cpuhp_online_idle(enum cpuhp_state state)
868{
869 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
870 unsigned int cpu = smp_processor_id();
871
872 /* Happens for the boot cpu */
873 if (state != CPUHP_AP_ONLINE_IDLE)
874 return;
875
876 st->state = CPUHP_AP_ONLINE_IDLE;
877
878 /* Unpark the stopper thread and the hotplug thread of this cpu */
879 stop_machine_unpark(cpu);
880 kthread_unpark(st->thread);
881
882 /* Should we go further up ? */
883 if (st->target > CPUHP_AP_ONLINE_IDLE)
884 __cpuhp_kick_ap_work(st);
885 else
886 complete(&st->done);
887}
888
889/* Requires cpu_add_remove_lock to be held */
890static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
891{
892 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
893 struct task_struct *idle;
894 int ret = 0;
895
896 cpu_hotplug_begin();
897
898 if (!cpu_present(cpu)) {
899 ret = -EINVAL;
900 goto out;
901 }
902
903 /*
904 * The caller of do_cpu_up might have raced with another
905 * caller. Ignore it for now.
906 */
907 if (st->state >= target)
908 goto out;
909
910 if (st->state == CPUHP_OFFLINE) {
911 /* Let it fail before we try to bring the cpu up */
912 idle = idle_thread_get(cpu);
913 if (IS_ERR(idle)) {
914 ret = PTR_ERR(idle);
915 goto out;
916 }
917 }
918
919 cpuhp_tasks_frozen = tasks_frozen;
920
921 st->target = target;
922 /*
923 * If the current CPU state is in the range of the AP hotplug thread,
924 * then we need to kick the thread once more.
925 */
926 if (st->state > CPUHP_BRINGUP_CPU) {
927 ret = cpuhp_kick_ap_work(cpu);
928 /*
929 * The AP side has done the error rollback already. Just
930 * return the error code..
931 */
932 if (ret)
933 goto out;
934 }
935
936 /*
937 * Try to reach the target state. We max out on the BP at
938 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
939 * responsible for bringing it up to the target state.
940 */
941 target = min((int)target, CPUHP_BRINGUP_CPU);
942 ret = cpuhp_up_callbacks(cpu, st, target);
943out:
944 cpu_hotplug_done();
945 return ret;
946}
947
948static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
949{
950 int err = 0;
951
952 if (!cpu_possible(cpu)) {
953 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
954 cpu);
955#if defined(CONFIG_IA64)
956 pr_err("please check additional_cpus= boot parameter\n");
957#endif
958 return -EINVAL;
959 }
960
961 err = try_online_node(cpu_to_node(cpu));
962 if (err)
963 return err;
964
965 cpu_maps_update_begin();
966
967 if (cpu_hotplug_disabled) {
968 err = -EBUSY;
969 goto out;
970 }
971
972 err = _cpu_up(cpu, 0, target);
973out:
974 cpu_maps_update_done();
975 return err;
976}
977
978int cpu_up(unsigned int cpu)
979{
980 return do_cpu_up(cpu, CPUHP_ONLINE);
981}
982EXPORT_SYMBOL_GPL(cpu_up);
983
984#ifdef CONFIG_PM_SLEEP_SMP
985static cpumask_var_t frozen_cpus;
986
987int freeze_secondary_cpus(int primary)
988{
989 int cpu, error = 0;
990
991 cpu_maps_update_begin();
992 if (!cpu_online(primary))
993 primary = cpumask_first(cpu_online_mask);
994 /*
995 * We take down all of the non-boot CPUs in one shot to avoid races
996 * with the userspace trying to use the CPU hotplug at the same time
997 */
998 cpumask_clear(frozen_cpus);
999
1000 pr_info("Disabling non-boot CPUs ...\n");
1001 for_each_online_cpu(cpu) {
1002 if (cpu == primary)
1003 continue;
1004 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1005 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1006 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1007 if (!error)
1008 cpumask_set_cpu(cpu, frozen_cpus);
1009 else {
1010 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1011 break;
1012 }
1013 }
1014
1015 if (!error)
1016 BUG_ON(num_online_cpus() > 1);
1017 else
1018 pr_err("Non-boot CPUs are not disabled\n");
1019
1020 /*
1021 * Make sure the CPUs won't be enabled by someone else. We need to do
1022 * this even in case of failure as all disable_nonboot_cpus() users are
1023 * supposed to do enable_nonboot_cpus() on the failure path.
1024 */
1025 cpu_hotplug_disabled++;
1026
1027 cpu_maps_update_done();
1028 return error;
1029}
1030
1031void __weak arch_enable_nonboot_cpus_begin(void)
1032{
1033}
1034
1035void __weak arch_enable_nonboot_cpus_end(void)
1036{
1037}
1038
1039void enable_nonboot_cpus(void)
1040{
1041 int cpu, error;
1042
1043 /* Allow everyone to use the CPU hotplug again */
1044 cpu_maps_update_begin();
1045 __cpu_hotplug_enable();
1046 if (cpumask_empty(frozen_cpus))
1047 goto out;
1048
1049 pr_info("Enabling non-boot CPUs ...\n");
1050
1051 arch_enable_nonboot_cpus_begin();
1052
1053 for_each_cpu(cpu, frozen_cpus) {
1054 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1055 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1056 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1057 if (!error) {
1058 pr_info("CPU%d is up\n", cpu);
1059 continue;
1060 }
1061 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1062 }
1063
1064 arch_enable_nonboot_cpus_end();
1065
1066 cpumask_clear(frozen_cpus);
1067out:
1068 cpu_maps_update_done();
1069}
1070
1071static int __init alloc_frozen_cpus(void)
1072{
1073 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1074 return -ENOMEM;
1075 return 0;
1076}
1077core_initcall(alloc_frozen_cpus);
1078
1079/*
1080 * When callbacks for CPU hotplug notifications are being executed, we must
1081 * ensure that the state of the system with respect to the tasks being frozen
1082 * or not, as reported by the notification, remains unchanged *throughout the
1083 * duration* of the execution of the callbacks.
1084 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1085 *
1086 * This synchronization is implemented by mutually excluding regular CPU
1087 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1088 * Hibernate notifications.
1089 */
1090static int
1091cpu_hotplug_pm_callback(struct notifier_block *nb,
1092 unsigned long action, void *ptr)
1093{
1094 switch (action) {
1095
1096 case PM_SUSPEND_PREPARE:
1097 case PM_HIBERNATION_PREPARE:
1098 cpu_hotplug_disable();
1099 break;
1100
1101 case PM_POST_SUSPEND:
1102 case PM_POST_HIBERNATION:
1103 cpu_hotplug_enable();
1104 break;
1105
1106 default:
1107 return NOTIFY_DONE;
1108 }
1109
1110 return NOTIFY_OK;
1111}
1112
1113
1114static int __init cpu_hotplug_pm_sync_init(void)
1115{
1116 /*
1117 * cpu_hotplug_pm_callback has higher priority than x86
1118 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1119 * to disable cpu hotplug to avoid cpu hotplug race.
1120 */
1121 pm_notifier(cpu_hotplug_pm_callback, 0);
1122 return 0;
1123}
1124core_initcall(cpu_hotplug_pm_sync_init);
1125
1126#endif /* CONFIG_PM_SLEEP_SMP */
1127
1128#endif /* CONFIG_SMP */
1129
1130/* Boot processor state steps */
1131static struct cpuhp_step cpuhp_bp_states[] = {
1132 [CPUHP_OFFLINE] = {
1133 .name = "offline",
1134 .startup.single = NULL,
1135 .teardown.single = NULL,
1136 },
1137#ifdef CONFIG_SMP
1138 [CPUHP_CREATE_THREADS]= {
1139 .name = "threads:prepare",
1140 .startup.single = smpboot_create_threads,
1141 .teardown.single = NULL,
1142 .cant_stop = true,
1143 },
1144 [CPUHP_PERF_PREPARE] = {
1145 .name = "perf:prepare",
1146 .startup.single = perf_event_init_cpu,
1147 .teardown.single = perf_event_exit_cpu,
1148 },
1149 [CPUHP_WORKQUEUE_PREP] = {
1150 .name = "workqueue:prepare",
1151 .startup.single = workqueue_prepare_cpu,
1152 .teardown.single = NULL,
1153 },
1154 [CPUHP_HRTIMERS_PREPARE] = {
1155 .name = "hrtimers:prepare",
1156 .startup.single = hrtimers_prepare_cpu,
1157 .teardown.single = hrtimers_dead_cpu,
1158 },
1159 [CPUHP_SMPCFD_PREPARE] = {
1160 .name = "smpcfd:prepare",
1161 .startup.single = smpcfd_prepare_cpu,
1162 .teardown.single = smpcfd_dead_cpu,
1163 },
1164 [CPUHP_RELAY_PREPARE] = {
1165 .name = "relay:prepare",
1166 .startup.single = relay_prepare_cpu,
1167 .teardown.single = NULL,
1168 },
1169 [CPUHP_SLAB_PREPARE] = {
1170 .name = "slab:prepare",
1171 .startup.single = slab_prepare_cpu,
1172 .teardown.single = slab_dead_cpu,
1173 },
1174 [CPUHP_RCUTREE_PREP] = {
1175 .name = "RCU/tree:prepare",
1176 .startup.single = rcutree_prepare_cpu,
1177 .teardown.single = rcutree_dead_cpu,
1178 },
1179 /*
1180 * On the tear-down path, timers_dead_cpu() must be invoked
1181 * before blk_mq_queue_reinit_notify() from notify_dead(),
1182 * otherwise a RCU stall occurs.
1183 */
1184 [CPUHP_TIMERS_DEAD] = {
1185 .name = "timers:dead",
1186 .startup.single = NULL,
1187 .teardown.single = timers_dead_cpu,
1188 },
1189 /* Kicks the plugged cpu into life */
1190 [CPUHP_BRINGUP_CPU] = {
1191 .name = "cpu:bringup",
1192 .startup.single = bringup_cpu,
1193 .teardown.single = NULL,
1194 .cant_stop = true,
1195 },
1196 [CPUHP_AP_SMPCFD_DYING] = {
1197 .name = "smpcfd:dying",
1198 .startup.single = NULL,
1199 .teardown.single = smpcfd_dying_cpu,
1200 },
1201 /*
1202 * Handled on controll processor until the plugged processor manages
1203 * this itself.
1204 */
1205 [CPUHP_TEARDOWN_CPU] = {
1206 .name = "cpu:teardown",
1207 .startup.single = NULL,
1208 .teardown.single = takedown_cpu,
1209 .cant_stop = true,
1210 },
1211#else
1212 [CPUHP_BRINGUP_CPU] = { },
1213#endif
1214};
1215
1216/* Application processor state steps */
1217static struct cpuhp_step cpuhp_ap_states[] = {
1218#ifdef CONFIG_SMP
1219 /* Final state before CPU kills itself */
1220 [CPUHP_AP_IDLE_DEAD] = {
1221 .name = "idle:dead",
1222 },
1223 /*
1224 * Last state before CPU enters the idle loop to die. Transient state
1225 * for synchronization.
1226 */
1227 [CPUHP_AP_OFFLINE] = {
1228 .name = "ap:offline",
1229 .cant_stop = true,
1230 },
1231 /* First state is scheduler control. Interrupts are disabled */
1232 [CPUHP_AP_SCHED_STARTING] = {
1233 .name = "sched:starting",
1234 .startup.single = sched_cpu_starting,
1235 .teardown.single = sched_cpu_dying,
1236 },
1237 [CPUHP_AP_RCUTREE_DYING] = {
1238 .name = "RCU/tree:dying",
1239 .startup.single = NULL,
1240 .teardown.single = rcutree_dying_cpu,
1241 },
1242 /* Entry state on starting. Interrupts enabled from here on. Transient
1243 * state for synchronsization */
1244 [CPUHP_AP_ONLINE] = {
1245 .name = "ap:online",
1246 },
1247 /* Handle smpboot threads park/unpark */
1248 [CPUHP_AP_SMPBOOT_THREADS] = {
1249 .name = "smpboot/threads:online",
1250 .startup.single = smpboot_unpark_threads,
1251 .teardown.single = NULL,
1252 },
1253 [CPUHP_AP_PERF_ONLINE] = {
1254 .name = "perf:online",
1255 .startup.single = perf_event_init_cpu,
1256 .teardown.single = perf_event_exit_cpu,
1257 },
1258 [CPUHP_AP_WORKQUEUE_ONLINE] = {
1259 .name = "workqueue:online",
1260 .startup.single = workqueue_online_cpu,
1261 .teardown.single = workqueue_offline_cpu,
1262 },
1263 [CPUHP_AP_RCUTREE_ONLINE] = {
1264 .name = "RCU/tree:online",
1265 .startup.single = rcutree_online_cpu,
1266 .teardown.single = rcutree_offline_cpu,
1267 },
1268#endif
1269 /*
1270 * The dynamically registered state space is here
1271 */
1272
1273#ifdef CONFIG_SMP
1274 /* Last state is scheduler control setting the cpu active */
1275 [CPUHP_AP_ACTIVE] = {
1276 .name = "sched:active",
1277 .startup.single = sched_cpu_activate,
1278 .teardown.single = sched_cpu_deactivate,
1279 },
1280#endif
1281
1282 /* CPU is fully up and running. */
1283 [CPUHP_ONLINE] = {
1284 .name = "online",
1285 .startup.single = NULL,
1286 .teardown.single = NULL,
1287 },
1288};
1289
1290/* Sanity check for callbacks */
1291static int cpuhp_cb_check(enum cpuhp_state state)
1292{
1293 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1294 return -EINVAL;
1295 return 0;
1296}
1297
1298/*
1299 * Returns a free for dynamic slot assignment of the Online state. The states
1300 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1301 * by having no name assigned.
1302 */
1303static int cpuhp_reserve_state(enum cpuhp_state state)
1304{
1305 enum cpuhp_state i;
1306
1307 for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1308 if (!cpuhp_ap_states[i].name)
1309 return i;
1310 }
1311 WARN(1, "No more dynamic states available for CPU hotplug\n");
1312 return -ENOSPC;
1313}
1314
1315static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1316 int (*startup)(unsigned int cpu),
1317 int (*teardown)(unsigned int cpu),
1318 bool multi_instance)
1319{
1320 /* (Un)Install the callbacks for further cpu hotplug operations */
1321 struct cpuhp_step *sp;
1322 int ret = 0;
1323
1324 mutex_lock(&cpuhp_state_mutex);
1325
1326 if (state == CPUHP_AP_ONLINE_DYN) {
1327 ret = cpuhp_reserve_state(state);
1328 if (ret < 0)
1329 goto out;
1330 state = ret;
1331 }
1332 sp = cpuhp_get_step(state);
1333 if (name && sp->name) {
1334 ret = -EBUSY;
1335 goto out;
1336 }
1337 sp->startup.single = startup;
1338 sp->teardown.single = teardown;
1339 sp->name = name;
1340 sp->multi_instance = multi_instance;
1341 INIT_HLIST_HEAD(&sp->list);
1342out:
1343 mutex_unlock(&cpuhp_state_mutex);
1344 return ret;
1345}
1346
1347static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1348{
1349 return cpuhp_get_step(state)->teardown.single;
1350}
1351
1352/*
1353 * Call the startup/teardown function for a step either on the AP or
1354 * on the current CPU.
1355 */
1356static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1357 struct hlist_node *node)
1358{
1359 struct cpuhp_step *sp = cpuhp_get_step(state);
1360 int ret;
1361
1362 if ((bringup && !sp->startup.single) ||
1363 (!bringup && !sp->teardown.single))
1364 return 0;
1365 /*
1366 * The non AP bound callbacks can fail on bringup. On teardown
1367 * e.g. module removal we crash for now.
1368 */
1369#ifdef CONFIG_SMP
1370 if (cpuhp_is_ap_state(state))
1371 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1372 else
1373 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1374#else
1375 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1376#endif
1377 BUG_ON(ret && !bringup);
1378 return ret;
1379}
1380
1381/*
1382 * Called from __cpuhp_setup_state on a recoverable failure.
1383 *
1384 * Note: The teardown callbacks for rollback are not allowed to fail!
1385 */
1386static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1387 struct hlist_node *node)
1388{
1389 int cpu;
1390
1391 /* Roll back the already executed steps on the other cpus */
1392 for_each_present_cpu(cpu) {
1393 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1394 int cpustate = st->state;
1395
1396 if (cpu >= failedcpu)
1397 break;
1398
1399 /* Did we invoke the startup call on that cpu ? */
1400 if (cpustate >= state)
1401 cpuhp_issue_call(cpu, state, false, node);
1402 }
1403}
1404
1405int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1406 bool invoke)
1407{
1408 struct cpuhp_step *sp;
1409 int cpu;
1410 int ret;
1411
1412 sp = cpuhp_get_step(state);
1413 if (sp->multi_instance == false)
1414 return -EINVAL;
1415
1416 get_online_cpus();
1417
1418 if (!invoke || !sp->startup.multi)
1419 goto add_node;
1420
1421 /*
1422 * Try to call the startup callback for each present cpu
1423 * depending on the hotplug state of the cpu.
1424 */
1425 for_each_present_cpu(cpu) {
1426 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1427 int cpustate = st->state;
1428
1429 if (cpustate < state)
1430 continue;
1431
1432 ret = cpuhp_issue_call(cpu, state, true, node);
1433 if (ret) {
1434 if (sp->teardown.multi)
1435 cpuhp_rollback_install(cpu, state, node);
1436 goto err;
1437 }
1438 }
1439add_node:
1440 ret = 0;
1441 mutex_lock(&cpuhp_state_mutex);
1442 hlist_add_head(node, &sp->list);
1443 mutex_unlock(&cpuhp_state_mutex);
1444
1445err:
1446 put_online_cpus();
1447 return ret;
1448}
1449EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1450
1451/**
1452 * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1453 * @state: The state to setup
1454 * @invoke: If true, the startup function is invoked for cpus where
1455 * cpu state >= @state
1456 * @startup: startup callback function
1457 * @teardown: teardown callback function
1458 * @multi_instance: State is set up for multiple instances which get
1459 * added afterwards.
1460 *
1461 * Returns:
1462 * On success:
1463 * Positive state number if @state is CPUHP_AP_ONLINE_DYN
1464 * 0 for all other states
1465 * On failure: proper (negative) error code
1466 */
1467int __cpuhp_setup_state(enum cpuhp_state state,
1468 const char *name, bool invoke,
1469 int (*startup)(unsigned int cpu),
1470 int (*teardown)(unsigned int cpu),
1471 bool multi_instance)
1472{
1473 int cpu, ret = 0;
1474
1475 if (cpuhp_cb_check(state) || !name)
1476 return -EINVAL;
1477
1478 get_online_cpus();
1479
1480 ret = cpuhp_store_callbacks(state, name, startup, teardown,
1481 multi_instance);
1482
1483 if (ret || !invoke || !startup)
1484 goto out;
1485
1486 /*
1487 * Try to call the startup callback for each present cpu
1488 * depending on the hotplug state of the cpu.
1489 */
1490 for_each_present_cpu(cpu) {
1491 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1492 int cpustate = st->state;
1493
1494 if (cpustate < state)
1495 continue;
1496
1497 ret = cpuhp_issue_call(cpu, state, true, NULL);
1498 if (ret) {
1499 if (teardown)
1500 cpuhp_rollback_install(cpu, state, NULL);
1501 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1502 goto out;
1503 }
1504 }
1505out:
1506 put_online_cpus();
1507 /*
1508 * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1509 * dynamically allocated state in case of success.
1510 */
1511 if (!ret && state == CPUHP_AP_ONLINE_DYN)
1512 return state;
1513 return ret;
1514}
1515EXPORT_SYMBOL(__cpuhp_setup_state);
1516
1517int __cpuhp_state_remove_instance(enum cpuhp_state state,
1518 struct hlist_node *node, bool invoke)
1519{
1520 struct cpuhp_step *sp = cpuhp_get_step(state);
1521 int cpu;
1522
1523 BUG_ON(cpuhp_cb_check(state));
1524
1525 if (!sp->multi_instance)
1526 return -EINVAL;
1527
1528 get_online_cpus();
1529 if (!invoke || !cpuhp_get_teardown_cb(state))
1530 goto remove;
1531 /*
1532 * Call the teardown callback for each present cpu depending
1533 * on the hotplug state of the cpu. This function is not
1534 * allowed to fail currently!
1535 */
1536 for_each_present_cpu(cpu) {
1537 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1538 int cpustate = st->state;
1539
1540 if (cpustate >= state)
1541 cpuhp_issue_call(cpu, state, false, node);
1542 }
1543
1544remove:
1545 mutex_lock(&cpuhp_state_mutex);
1546 hlist_del(node);
1547 mutex_unlock(&cpuhp_state_mutex);
1548 put_online_cpus();
1549
1550 return 0;
1551}
1552EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1553/**
1554 * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1555 * @state: The state to remove
1556 * @invoke: If true, the teardown function is invoked for cpus where
1557 * cpu state >= @state
1558 *
1559 * The teardown callback is currently not allowed to fail. Think
1560 * about module removal!
1561 */
1562void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1563{
1564 struct cpuhp_step *sp = cpuhp_get_step(state);
1565 int cpu;
1566
1567 BUG_ON(cpuhp_cb_check(state));
1568
1569 get_online_cpus();
1570
1571 if (sp->multi_instance) {
1572 WARN(!hlist_empty(&sp->list),
1573 "Error: Removing state %d which has instances left.\n",
1574 state);
1575 goto remove;
1576 }
1577
1578 if (!invoke || !cpuhp_get_teardown_cb(state))
1579 goto remove;
1580
1581 /*
1582 * Call the teardown callback for each present cpu depending
1583 * on the hotplug state of the cpu. This function is not
1584 * allowed to fail currently!
1585 */
1586 for_each_present_cpu(cpu) {
1587 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1588 int cpustate = st->state;
1589
1590 if (cpustate >= state)
1591 cpuhp_issue_call(cpu, state, false, NULL);
1592 }
1593remove:
1594 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1595 put_online_cpus();
1596}
1597EXPORT_SYMBOL(__cpuhp_remove_state);
1598
1599#if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1600static ssize_t show_cpuhp_state(struct device *dev,
1601 struct device_attribute *attr, char *buf)
1602{
1603 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1604
1605 return sprintf(buf, "%d\n", st->state);
1606}
1607static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1608
1609static ssize_t write_cpuhp_target(struct device *dev,
1610 struct device_attribute *attr,
1611 const char *buf, size_t count)
1612{
1613 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1614 struct cpuhp_step *sp;
1615 int target, ret;
1616
1617 ret = kstrtoint(buf, 10, &target);
1618 if (ret)
1619 return ret;
1620
1621#ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1622 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1623 return -EINVAL;
1624#else
1625 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1626 return -EINVAL;
1627#endif
1628
1629 ret = lock_device_hotplug_sysfs();
1630 if (ret)
1631 return ret;
1632
1633 mutex_lock(&cpuhp_state_mutex);
1634 sp = cpuhp_get_step(target);
1635 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1636 mutex_unlock(&cpuhp_state_mutex);
1637 if (ret)
1638 return ret;
1639
1640 if (st->state < target)
1641 ret = do_cpu_up(dev->id, target);
1642 else
1643 ret = do_cpu_down(dev->id, target);
1644
1645 unlock_device_hotplug();
1646 return ret ? ret : count;
1647}
1648
1649static ssize_t show_cpuhp_target(struct device *dev,
1650 struct device_attribute *attr, char *buf)
1651{
1652 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1653
1654 return sprintf(buf, "%d\n", st->target);
1655}
1656static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1657
1658static struct attribute *cpuhp_cpu_attrs[] = {
1659 &dev_attr_state.attr,
1660 &dev_attr_target.attr,
1661 NULL
1662};
1663
1664static struct attribute_group cpuhp_cpu_attr_group = {
1665 .attrs = cpuhp_cpu_attrs,
1666 .name = "hotplug",
1667 NULL
1668};
1669
1670static ssize_t show_cpuhp_states(struct device *dev,
1671 struct device_attribute *attr, char *buf)
1672{
1673 ssize_t cur, res = 0;
1674 int i;
1675
1676 mutex_lock(&cpuhp_state_mutex);
1677 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1678 struct cpuhp_step *sp = cpuhp_get_step(i);
1679
1680 if (sp->name) {
1681 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1682 buf += cur;
1683 res += cur;
1684 }
1685 }
1686 mutex_unlock(&cpuhp_state_mutex);
1687 return res;
1688}
1689static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1690
1691static struct attribute *cpuhp_cpu_root_attrs[] = {
1692 &dev_attr_states.attr,
1693 NULL
1694};
1695
1696static struct attribute_group cpuhp_cpu_root_attr_group = {
1697 .attrs = cpuhp_cpu_root_attrs,
1698 .name = "hotplug",
1699 NULL
1700};
1701
1702static int __init cpuhp_sysfs_init(void)
1703{
1704 int cpu, ret;
1705
1706 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1707 &cpuhp_cpu_root_attr_group);
1708 if (ret)
1709 return ret;
1710
1711 for_each_possible_cpu(cpu) {
1712 struct device *dev = get_cpu_device(cpu);
1713
1714 if (!dev)
1715 continue;
1716 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1717 if (ret)
1718 return ret;
1719 }
1720 return 0;
1721}
1722device_initcall(cpuhp_sysfs_init);
1723#endif
1724
1725/*
1726 * cpu_bit_bitmap[] is a special, "compressed" data structure that
1727 * represents all NR_CPUS bits binary values of 1<<nr.
1728 *
1729 * It is used by cpumask_of() to get a constant address to a CPU
1730 * mask value that has a single bit set only.
1731 */
1732
1733/* cpu_bit_bitmap[0] is empty - so we can back into it */
1734#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
1735#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1736#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1737#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1738
1739const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1740
1741 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
1742 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
1743#if BITS_PER_LONG > 32
1744 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
1745 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
1746#endif
1747};
1748EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1749
1750const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1751EXPORT_SYMBOL(cpu_all_bits);
1752
1753#ifdef CONFIG_INIT_ALL_POSSIBLE
1754struct cpumask __cpu_possible_mask __read_mostly
1755 = {CPU_BITS_ALL};
1756#else
1757struct cpumask __cpu_possible_mask __read_mostly;
1758#endif
1759EXPORT_SYMBOL(__cpu_possible_mask);
1760
1761struct cpumask __cpu_online_mask __read_mostly;
1762EXPORT_SYMBOL(__cpu_online_mask);
1763
1764struct cpumask __cpu_present_mask __read_mostly;
1765EXPORT_SYMBOL(__cpu_present_mask);
1766
1767struct cpumask __cpu_active_mask __read_mostly;
1768EXPORT_SYMBOL(__cpu_active_mask);
1769
1770void init_cpu_present(const struct cpumask *src)
1771{
1772 cpumask_copy(&__cpu_present_mask, src);
1773}
1774
1775void init_cpu_possible(const struct cpumask *src)
1776{
1777 cpumask_copy(&__cpu_possible_mask, src);
1778}
1779
1780void init_cpu_online(const struct cpumask *src)
1781{
1782 cpumask_copy(&__cpu_online_mask, src);
1783}
1784
1785/*
1786 * Activate the first processor.
1787 */
1788void __init boot_cpu_init(void)
1789{
1790 int cpu = smp_processor_id();
1791
1792 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1793 set_cpu_online(cpu, true);
1794 set_cpu_active(cpu, true);
1795 set_cpu_present(cpu, true);
1796 set_cpu_possible(cpu, true);
1797}
1798
1799/*
1800 * Must be called _AFTER_ setting up the per_cpu areas
1801 */
1802void __init boot_cpu_state_init(void)
1803{
1804 per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1805}