Loading...
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
24#include "smpboot.h"
25
26#ifdef CONFIG_SMP
27/* Serializes the updates to cpu_online_mask, cpu_present_mask */
28static DEFINE_MUTEX(cpu_add_remove_lock);
29
30/*
31 * The following two APIs (cpu_maps_update_begin/done) must be used when
32 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
33 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
34 * hotplug callback (un)registration performed using __register_cpu_notifier()
35 * or __unregister_cpu_notifier().
36 */
37void cpu_maps_update_begin(void)
38{
39 mutex_lock(&cpu_add_remove_lock);
40}
41EXPORT_SYMBOL(cpu_notifier_register_begin);
42
43void cpu_maps_update_done(void)
44{
45 mutex_unlock(&cpu_add_remove_lock);
46}
47EXPORT_SYMBOL(cpu_notifier_register_done);
48
49static RAW_NOTIFIER_HEAD(cpu_chain);
50
51/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
52 * Should always be manipulated under cpu_add_remove_lock
53 */
54static int cpu_hotplug_disabled;
55
56#ifdef CONFIG_HOTPLUG_CPU
57
58static struct {
59 struct task_struct *active_writer;
60 struct mutex lock; /* Synchronizes accesses to refcount, */
61 /*
62 * Also blocks the new readers during
63 * an ongoing cpu hotplug operation.
64 */
65 int refcount;
66
67#ifdef CONFIG_DEBUG_LOCK_ALLOC
68 struct lockdep_map dep_map;
69#endif
70} cpu_hotplug = {
71 .active_writer = NULL,
72 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
73 .refcount = 0,
74#ifdef CONFIG_DEBUG_LOCK_ALLOC
75 .dep_map = {.name = "cpu_hotplug.lock" },
76#endif
77};
78
79/* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
80#define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
81#define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
82#define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
83
84void get_online_cpus(void)
85{
86 might_sleep();
87 if (cpu_hotplug.active_writer == current)
88 return;
89 cpuhp_lock_acquire_read();
90 mutex_lock(&cpu_hotplug.lock);
91 cpu_hotplug.refcount++;
92 mutex_unlock(&cpu_hotplug.lock);
93
94}
95EXPORT_SYMBOL_GPL(get_online_cpus);
96
97void put_online_cpus(void)
98{
99 if (cpu_hotplug.active_writer == current)
100 return;
101 mutex_lock(&cpu_hotplug.lock);
102
103 if (WARN_ON(!cpu_hotplug.refcount))
104 cpu_hotplug.refcount++; /* try to fix things up */
105
106 if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
107 wake_up_process(cpu_hotplug.active_writer);
108 mutex_unlock(&cpu_hotplug.lock);
109 cpuhp_lock_release();
110
111}
112EXPORT_SYMBOL_GPL(put_online_cpus);
113
114/*
115 * This ensures that the hotplug operation can begin only when the
116 * refcount goes to zero.
117 *
118 * Note that during a cpu-hotplug operation, the new readers, if any,
119 * will be blocked by the cpu_hotplug.lock
120 *
121 * Since cpu_hotplug_begin() is always called after invoking
122 * cpu_maps_update_begin(), we can be sure that only one writer is active.
123 *
124 * Note that theoretically, there is a possibility of a livelock:
125 * - Refcount goes to zero, last reader wakes up the sleeping
126 * writer.
127 * - Last reader unlocks the cpu_hotplug.lock.
128 * - A new reader arrives at this moment, bumps up the refcount.
129 * - The writer acquires the cpu_hotplug.lock finds the refcount
130 * non zero and goes to sleep again.
131 *
132 * However, this is very difficult to achieve in practice since
133 * get_online_cpus() not an api which is called all that often.
134 *
135 */
136void cpu_hotplug_begin(void)
137{
138 cpu_hotplug.active_writer = current;
139
140 cpuhp_lock_acquire();
141 for (;;) {
142 mutex_lock(&cpu_hotplug.lock);
143 if (likely(!cpu_hotplug.refcount))
144 break;
145 __set_current_state(TASK_UNINTERRUPTIBLE);
146 mutex_unlock(&cpu_hotplug.lock);
147 schedule();
148 }
149}
150
151void cpu_hotplug_done(void)
152{
153 cpu_hotplug.active_writer = NULL;
154 mutex_unlock(&cpu_hotplug.lock);
155 cpuhp_lock_release();
156}
157
158/*
159 * Wait for currently running CPU hotplug operations to complete (if any) and
160 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
161 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
162 * hotplug path before performing hotplug operations. So acquiring that lock
163 * guarantees mutual exclusion from any currently running hotplug operations.
164 */
165void cpu_hotplug_disable(void)
166{
167 cpu_maps_update_begin();
168 cpu_hotplug_disabled = 1;
169 cpu_maps_update_done();
170}
171
172void cpu_hotplug_enable(void)
173{
174 cpu_maps_update_begin();
175 cpu_hotplug_disabled = 0;
176 cpu_maps_update_done();
177}
178
179#endif /* CONFIG_HOTPLUG_CPU */
180
181/* Need to know about CPUs going up/down? */
182int __ref register_cpu_notifier(struct notifier_block *nb)
183{
184 int ret;
185 cpu_maps_update_begin();
186 ret = raw_notifier_chain_register(&cpu_chain, nb);
187 cpu_maps_update_done();
188 return ret;
189}
190
191int __ref __register_cpu_notifier(struct notifier_block *nb)
192{
193 return raw_notifier_chain_register(&cpu_chain, nb);
194}
195
196static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
197 int *nr_calls)
198{
199 int ret;
200
201 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
202 nr_calls);
203
204 return notifier_to_errno(ret);
205}
206
207static int cpu_notify(unsigned long val, void *v)
208{
209 return __cpu_notify(val, v, -1, NULL);
210}
211
212#ifdef CONFIG_HOTPLUG_CPU
213
214static void cpu_notify_nofail(unsigned long val, void *v)
215{
216 BUG_ON(cpu_notify(val, v));
217}
218EXPORT_SYMBOL(register_cpu_notifier);
219EXPORT_SYMBOL(__register_cpu_notifier);
220
221void __ref unregister_cpu_notifier(struct notifier_block *nb)
222{
223 cpu_maps_update_begin();
224 raw_notifier_chain_unregister(&cpu_chain, nb);
225 cpu_maps_update_done();
226}
227EXPORT_SYMBOL(unregister_cpu_notifier);
228
229void __ref __unregister_cpu_notifier(struct notifier_block *nb)
230{
231 raw_notifier_chain_unregister(&cpu_chain, nb);
232}
233EXPORT_SYMBOL(__unregister_cpu_notifier);
234
235/**
236 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
237 * @cpu: a CPU id
238 *
239 * This function walks all processes, finds a valid mm struct for each one and
240 * then clears a corresponding bit in mm's cpumask. While this all sounds
241 * trivial, there are various non-obvious corner cases, which this function
242 * tries to solve in a safe manner.
243 *
244 * Also note that the function uses a somewhat relaxed locking scheme, so it may
245 * be called only for an already offlined CPU.
246 */
247void clear_tasks_mm_cpumask(int cpu)
248{
249 struct task_struct *p;
250
251 /*
252 * This function is called after the cpu is taken down and marked
253 * offline, so its not like new tasks will ever get this cpu set in
254 * their mm mask. -- Peter Zijlstra
255 * Thus, we may use rcu_read_lock() here, instead of grabbing
256 * full-fledged tasklist_lock.
257 */
258 WARN_ON(cpu_online(cpu));
259 rcu_read_lock();
260 for_each_process(p) {
261 struct task_struct *t;
262
263 /*
264 * Main thread might exit, but other threads may still have
265 * a valid mm. Find one.
266 */
267 t = find_lock_task_mm(p);
268 if (!t)
269 continue;
270 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
271 task_unlock(t);
272 }
273 rcu_read_unlock();
274}
275
276static inline void check_for_tasks(int cpu)
277{
278 struct task_struct *p;
279 cputime_t utime, stime;
280
281 write_lock_irq(&tasklist_lock);
282 for_each_process(p) {
283 task_cputime(p, &utime, &stime);
284 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
285 (utime || stime))
286 printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
287 "(state = %ld, flags = %x)\n",
288 p->comm, task_pid_nr(p), cpu,
289 p->state, p->flags);
290 }
291 write_unlock_irq(&tasklist_lock);
292}
293
294struct take_cpu_down_param {
295 unsigned long mod;
296 void *hcpu;
297};
298
299/* Take this CPU down. */
300static int __ref take_cpu_down(void *_param)
301{
302 struct take_cpu_down_param *param = _param;
303 int err;
304
305 /* Ensure this CPU doesn't handle any more interrupts. */
306 err = __cpu_disable();
307 if (err < 0)
308 return err;
309
310 cpu_notify(CPU_DYING | param->mod, param->hcpu);
311 /* Park the stopper thread */
312 kthread_park(current);
313 return 0;
314}
315
316/* Requires cpu_add_remove_lock to be held */
317static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
318{
319 int err, nr_calls = 0;
320 void *hcpu = (void *)(long)cpu;
321 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
322 struct take_cpu_down_param tcd_param = {
323 .mod = mod,
324 .hcpu = hcpu,
325 };
326
327 if (num_online_cpus() == 1)
328 return -EBUSY;
329
330 if (!cpu_online(cpu))
331 return -EINVAL;
332
333 cpu_hotplug_begin();
334
335 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
336 if (err) {
337 nr_calls--;
338 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
339 printk("%s: attempt to take down CPU %u failed\n",
340 __func__, cpu);
341 goto out_release;
342 }
343
344 /*
345 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
346 * and RCU users of this state to go away such that all new such users
347 * will observe it.
348 *
349 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
350 * not imply sync_sched(), so explicitly call both.
351 *
352 * Do sync before park smpboot threads to take care the rcu boost case.
353 */
354#ifdef CONFIG_PREEMPT
355 synchronize_sched();
356#endif
357 synchronize_rcu();
358
359 smpboot_park_threads(cpu);
360
361 /*
362 * So now all preempt/rcu users must observe !cpu_active().
363 */
364
365 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
366 if (err) {
367 /* CPU didn't die: tell everyone. Can't complain. */
368 smpboot_unpark_threads(cpu);
369 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
370 goto out_release;
371 }
372 BUG_ON(cpu_online(cpu));
373
374 /*
375 * The migration_call() CPU_DYING callback will have removed all
376 * runnable tasks from the cpu, there's only the idle task left now
377 * that the migration thread is done doing the stop_machine thing.
378 *
379 * Wait for the stop thread to go away.
380 */
381 while (!idle_cpu(cpu))
382 cpu_relax();
383
384 /* This actually kills the CPU. */
385 __cpu_die(cpu);
386
387 /* CPU is completely dead: tell everyone. Too late to complain. */
388 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
389
390 check_for_tasks(cpu);
391
392out_release:
393 cpu_hotplug_done();
394 if (!err)
395 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
396 return err;
397}
398
399int __ref cpu_down(unsigned int cpu)
400{
401 int err;
402
403 cpu_maps_update_begin();
404
405 if (cpu_hotplug_disabled) {
406 err = -EBUSY;
407 goto out;
408 }
409
410 err = _cpu_down(cpu, 0);
411
412out:
413 cpu_maps_update_done();
414 return err;
415}
416EXPORT_SYMBOL(cpu_down);
417#endif /*CONFIG_HOTPLUG_CPU*/
418
419/* Requires cpu_add_remove_lock to be held */
420static int _cpu_up(unsigned int cpu, int tasks_frozen)
421{
422 int ret, nr_calls = 0;
423 void *hcpu = (void *)(long)cpu;
424 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
425 struct task_struct *idle;
426
427 cpu_hotplug_begin();
428
429 if (cpu_online(cpu) || !cpu_present(cpu)) {
430 ret = -EINVAL;
431 goto out;
432 }
433
434 idle = idle_thread_get(cpu);
435 if (IS_ERR(idle)) {
436 ret = PTR_ERR(idle);
437 goto out;
438 }
439
440 ret = smpboot_create_threads(cpu);
441 if (ret)
442 goto out;
443
444 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
445 if (ret) {
446 nr_calls--;
447 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
448 __func__, cpu);
449 goto out_notify;
450 }
451
452 /* Arch-specific enabling code. */
453 ret = __cpu_up(cpu, idle);
454 if (ret != 0)
455 goto out_notify;
456 BUG_ON(!cpu_online(cpu));
457
458 /* Wake the per cpu threads */
459 smpboot_unpark_threads(cpu);
460
461 /* Now call notifier in preparation. */
462 cpu_notify(CPU_ONLINE | mod, hcpu);
463
464out_notify:
465 if (ret != 0)
466 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
467out:
468 cpu_hotplug_done();
469
470 return ret;
471}
472
473int cpu_up(unsigned int cpu)
474{
475 int err = 0;
476
477 if (!cpu_possible(cpu)) {
478 printk(KERN_ERR "can't online cpu %d because it is not "
479 "configured as may-hotadd at boot time\n", cpu);
480#if defined(CONFIG_IA64)
481 printk(KERN_ERR "please check additional_cpus= boot "
482 "parameter\n");
483#endif
484 return -EINVAL;
485 }
486
487 err = try_online_node(cpu_to_node(cpu));
488 if (err)
489 return err;
490
491 cpu_maps_update_begin();
492
493 if (cpu_hotplug_disabled) {
494 err = -EBUSY;
495 goto out;
496 }
497
498 err = _cpu_up(cpu, 0);
499
500out:
501 cpu_maps_update_done();
502 return err;
503}
504EXPORT_SYMBOL_GPL(cpu_up);
505
506#ifdef CONFIG_PM_SLEEP_SMP
507static cpumask_var_t frozen_cpus;
508
509int disable_nonboot_cpus(void)
510{
511 int cpu, first_cpu, error = 0;
512
513 cpu_maps_update_begin();
514 first_cpu = cpumask_first(cpu_online_mask);
515 /*
516 * We take down all of the non-boot CPUs in one shot to avoid races
517 * with the userspace trying to use the CPU hotplug at the same time
518 */
519 cpumask_clear(frozen_cpus);
520
521 printk("Disabling non-boot CPUs ...\n");
522 for_each_online_cpu(cpu) {
523 if (cpu == first_cpu)
524 continue;
525 error = _cpu_down(cpu, 1);
526 if (!error)
527 cpumask_set_cpu(cpu, frozen_cpus);
528 else {
529 printk(KERN_ERR "Error taking CPU%d down: %d\n",
530 cpu, error);
531 break;
532 }
533 }
534
535 if (!error) {
536 BUG_ON(num_online_cpus() > 1);
537 /* Make sure the CPUs won't be enabled by someone else */
538 cpu_hotplug_disabled = 1;
539 } else {
540 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
541 }
542 cpu_maps_update_done();
543 return error;
544}
545
546void __weak arch_enable_nonboot_cpus_begin(void)
547{
548}
549
550void __weak arch_enable_nonboot_cpus_end(void)
551{
552}
553
554void __ref enable_nonboot_cpus(void)
555{
556 int cpu, error;
557
558 /* Allow everyone to use the CPU hotplug again */
559 cpu_maps_update_begin();
560 cpu_hotplug_disabled = 0;
561 if (cpumask_empty(frozen_cpus))
562 goto out;
563
564 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
565
566 arch_enable_nonboot_cpus_begin();
567
568 for_each_cpu(cpu, frozen_cpus) {
569 error = _cpu_up(cpu, 1);
570 if (!error) {
571 printk(KERN_INFO "CPU%d is up\n", cpu);
572 continue;
573 }
574 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
575 }
576
577 arch_enable_nonboot_cpus_end();
578
579 cpumask_clear(frozen_cpus);
580out:
581 cpu_maps_update_done();
582}
583
584static int __init alloc_frozen_cpus(void)
585{
586 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
587 return -ENOMEM;
588 return 0;
589}
590core_initcall(alloc_frozen_cpus);
591
592/*
593 * When callbacks for CPU hotplug notifications are being executed, we must
594 * ensure that the state of the system with respect to the tasks being frozen
595 * or not, as reported by the notification, remains unchanged *throughout the
596 * duration* of the execution of the callbacks.
597 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
598 *
599 * This synchronization is implemented by mutually excluding regular CPU
600 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
601 * Hibernate notifications.
602 */
603static int
604cpu_hotplug_pm_callback(struct notifier_block *nb,
605 unsigned long action, void *ptr)
606{
607 switch (action) {
608
609 case PM_SUSPEND_PREPARE:
610 case PM_HIBERNATION_PREPARE:
611 cpu_hotplug_disable();
612 break;
613
614 case PM_POST_SUSPEND:
615 case PM_POST_HIBERNATION:
616 cpu_hotplug_enable();
617 break;
618
619 default:
620 return NOTIFY_DONE;
621 }
622
623 return NOTIFY_OK;
624}
625
626
627static int __init cpu_hotplug_pm_sync_init(void)
628{
629 /*
630 * cpu_hotplug_pm_callback has higher priority than x86
631 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
632 * to disable cpu hotplug to avoid cpu hotplug race.
633 */
634 pm_notifier(cpu_hotplug_pm_callback, 0);
635 return 0;
636}
637core_initcall(cpu_hotplug_pm_sync_init);
638
639#endif /* CONFIG_PM_SLEEP_SMP */
640
641/**
642 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
643 * @cpu: cpu that just started
644 *
645 * This function calls the cpu_chain notifiers with CPU_STARTING.
646 * It must be called by the arch code on the new cpu, before the new cpu
647 * enables interrupts and before the "boot" cpu returns from __cpu_up().
648 */
649void notify_cpu_starting(unsigned int cpu)
650{
651 unsigned long val = CPU_STARTING;
652
653#ifdef CONFIG_PM_SLEEP_SMP
654 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
655 val = CPU_STARTING_FROZEN;
656#endif /* CONFIG_PM_SLEEP_SMP */
657 cpu_notify(val, (void *)(long)cpu);
658}
659
660#endif /* CONFIG_SMP */
661
662/*
663 * cpu_bit_bitmap[] is a special, "compressed" data structure that
664 * represents all NR_CPUS bits binary values of 1<<nr.
665 *
666 * It is used by cpumask_of() to get a constant address to a CPU
667 * mask value that has a single bit set only.
668 */
669
670/* cpu_bit_bitmap[0] is empty - so we can back into it */
671#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
672#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
673#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
674#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
675
676const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
677
678 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
679 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
680#if BITS_PER_LONG > 32
681 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
682 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
683#endif
684};
685EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
686
687const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
688EXPORT_SYMBOL(cpu_all_bits);
689
690#ifdef CONFIG_INIT_ALL_POSSIBLE
691static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
692 = CPU_BITS_ALL;
693#else
694static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
695#endif
696const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
697EXPORT_SYMBOL(cpu_possible_mask);
698
699static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
700const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
701EXPORT_SYMBOL(cpu_online_mask);
702
703static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
704const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
705EXPORT_SYMBOL(cpu_present_mask);
706
707static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
708const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
709EXPORT_SYMBOL(cpu_active_mask);
710
711void set_cpu_possible(unsigned int cpu, bool possible)
712{
713 if (possible)
714 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
715 else
716 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
717}
718
719void set_cpu_present(unsigned int cpu, bool present)
720{
721 if (present)
722 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
723 else
724 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
725}
726
727void set_cpu_online(unsigned int cpu, bool online)
728{
729 if (online) {
730 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
731 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
732 } else {
733 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
734 }
735}
736
737void set_cpu_active(unsigned int cpu, bool active)
738{
739 if (active)
740 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
741 else
742 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
743}
744
745void init_cpu_present(const struct cpumask *src)
746{
747 cpumask_copy(to_cpumask(cpu_present_bits), src);
748}
749
750void init_cpu_possible(const struct cpumask *src)
751{
752 cpumask_copy(to_cpumask(cpu_possible_bits), src);
753}
754
755void init_cpu_online(const struct cpumask *src)
756{
757 cpumask_copy(to_cpumask(cpu_online_bits), src);
758}
1/* CPU control.
2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
3 *
4 * This code is licenced under the GPL.
5 */
6#include <linux/sched/mm.h>
7#include <linux/proc_fs.h>
8#include <linux/smp.h>
9#include <linux/init.h>
10#include <linux/notifier.h>
11#include <linux/sched/signal.h>
12#include <linux/sched/hotplug.h>
13#include <linux/sched/isolation.h>
14#include <linux/sched/task.h>
15#include <linux/sched/smt.h>
16#include <linux/unistd.h>
17#include <linux/cpu.h>
18#include <linux/oom.h>
19#include <linux/rcupdate.h>
20#include <linux/delay.h>
21#include <linux/export.h>
22#include <linux/bug.h>
23#include <linux/kthread.h>
24#include <linux/stop_machine.h>
25#include <linux/mutex.h>
26#include <linux/gfp.h>
27#include <linux/suspend.h>
28#include <linux/lockdep.h>
29#include <linux/tick.h>
30#include <linux/irq.h>
31#include <linux/nmi.h>
32#include <linux/smpboot.h>
33#include <linux/relay.h>
34#include <linux/slab.h>
35#include <linux/scs.h>
36#include <linux/percpu-rwsem.h>
37#include <linux/cpuset.h>
38#include <linux/random.h>
39#include <linux/cc_platform.h>
40
41#include <trace/events/power.h>
42#define CREATE_TRACE_POINTS
43#include <trace/events/cpuhp.h>
44
45#include "smpboot.h"
46
47/**
48 * struct cpuhp_cpu_state - Per cpu hotplug state storage
49 * @state: The current cpu state
50 * @target: The target state
51 * @fail: Current CPU hotplug callback state
52 * @thread: Pointer to the hotplug thread
53 * @should_run: Thread should execute
54 * @rollback: Perform a rollback
55 * @single: Single callback invocation
56 * @bringup: Single callback bringup or teardown selector
57 * @node: Remote CPU node; for multi-instance, do a
58 * single entry callback for install/remove
59 * @last: For multi-instance rollback, remember how far we got
60 * @cb_state: The state for a single callback (install/uninstall)
61 * @result: Result of the operation
62 * @ap_sync_state: State for AP synchronization
63 * @done_up: Signal completion to the issuer of the task for cpu-up
64 * @done_down: Signal completion to the issuer of the task for cpu-down
65 */
66struct cpuhp_cpu_state {
67 enum cpuhp_state state;
68 enum cpuhp_state target;
69 enum cpuhp_state fail;
70#ifdef CONFIG_SMP
71 struct task_struct *thread;
72 bool should_run;
73 bool rollback;
74 bool single;
75 bool bringup;
76 struct hlist_node *node;
77 struct hlist_node *last;
78 enum cpuhp_state cb_state;
79 int result;
80 atomic_t ap_sync_state;
81 struct completion done_up;
82 struct completion done_down;
83#endif
84};
85
86static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
87 .fail = CPUHP_INVALID,
88};
89
90#ifdef CONFIG_SMP
91cpumask_t cpus_booted_once_mask;
92#endif
93
94#if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
95static struct lockdep_map cpuhp_state_up_map =
96 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
97static struct lockdep_map cpuhp_state_down_map =
98 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
99
100
101static inline void cpuhp_lock_acquire(bool bringup)
102{
103 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
104}
105
106static inline void cpuhp_lock_release(bool bringup)
107{
108 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
109}
110#else
111
112static inline void cpuhp_lock_acquire(bool bringup) { }
113static inline void cpuhp_lock_release(bool bringup) { }
114
115#endif
116
117/**
118 * struct cpuhp_step - Hotplug state machine step
119 * @name: Name of the step
120 * @startup: Startup function of the step
121 * @teardown: Teardown function of the step
122 * @cant_stop: Bringup/teardown can't be stopped at this step
123 * @multi_instance: State has multiple instances which get added afterwards
124 */
125struct cpuhp_step {
126 const char *name;
127 union {
128 int (*single)(unsigned int cpu);
129 int (*multi)(unsigned int cpu,
130 struct hlist_node *node);
131 } startup;
132 union {
133 int (*single)(unsigned int cpu);
134 int (*multi)(unsigned int cpu,
135 struct hlist_node *node);
136 } teardown;
137 /* private: */
138 struct hlist_head list;
139 /* public: */
140 bool cant_stop;
141 bool multi_instance;
142};
143
144static DEFINE_MUTEX(cpuhp_state_mutex);
145static struct cpuhp_step cpuhp_hp_states[];
146
147static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
148{
149 return cpuhp_hp_states + state;
150}
151
152static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
153{
154 return bringup ? !step->startup.single : !step->teardown.single;
155}
156
157/**
158 * cpuhp_invoke_callback - Invoke the callbacks for a given state
159 * @cpu: The cpu for which the callback should be invoked
160 * @state: The state to do callbacks for
161 * @bringup: True if the bringup callback should be invoked
162 * @node: For multi-instance, do a single entry callback for install/remove
163 * @lastp: For multi-instance rollback, remember how far we got
164 *
165 * Called from cpu hotplug and from the state register machinery.
166 *
167 * Return: %0 on success or a negative errno code
168 */
169static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
170 bool bringup, struct hlist_node *node,
171 struct hlist_node **lastp)
172{
173 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
174 struct cpuhp_step *step = cpuhp_get_step(state);
175 int (*cbm)(unsigned int cpu, struct hlist_node *node);
176 int (*cb)(unsigned int cpu);
177 int ret, cnt;
178
179 if (st->fail == state) {
180 st->fail = CPUHP_INVALID;
181 return -EAGAIN;
182 }
183
184 if (cpuhp_step_empty(bringup, step)) {
185 WARN_ON_ONCE(1);
186 return 0;
187 }
188
189 if (!step->multi_instance) {
190 WARN_ON_ONCE(lastp && *lastp);
191 cb = bringup ? step->startup.single : step->teardown.single;
192
193 trace_cpuhp_enter(cpu, st->target, state, cb);
194 ret = cb(cpu);
195 trace_cpuhp_exit(cpu, st->state, state, ret);
196 return ret;
197 }
198 cbm = bringup ? step->startup.multi : step->teardown.multi;
199
200 /* Single invocation for instance add/remove */
201 if (node) {
202 WARN_ON_ONCE(lastp && *lastp);
203 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
204 ret = cbm(cpu, node);
205 trace_cpuhp_exit(cpu, st->state, state, ret);
206 return ret;
207 }
208
209 /* State transition. Invoke on all instances */
210 cnt = 0;
211 hlist_for_each(node, &step->list) {
212 if (lastp && node == *lastp)
213 break;
214
215 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
216 ret = cbm(cpu, node);
217 trace_cpuhp_exit(cpu, st->state, state, ret);
218 if (ret) {
219 if (!lastp)
220 goto err;
221
222 *lastp = node;
223 return ret;
224 }
225 cnt++;
226 }
227 if (lastp)
228 *lastp = NULL;
229 return 0;
230err:
231 /* Rollback the instances if one failed */
232 cbm = !bringup ? step->startup.multi : step->teardown.multi;
233 if (!cbm)
234 return ret;
235
236 hlist_for_each(node, &step->list) {
237 if (!cnt--)
238 break;
239
240 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
241 ret = cbm(cpu, node);
242 trace_cpuhp_exit(cpu, st->state, state, ret);
243 /*
244 * Rollback must not fail,
245 */
246 WARN_ON_ONCE(ret);
247 }
248 return ret;
249}
250
251#ifdef CONFIG_SMP
252static bool cpuhp_is_ap_state(enum cpuhp_state state)
253{
254 /*
255 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
256 * purposes as that state is handled explicitly in cpu_down.
257 */
258 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
259}
260
261static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
262{
263 struct completion *done = bringup ? &st->done_up : &st->done_down;
264 wait_for_completion(done);
265}
266
267static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
268{
269 struct completion *done = bringup ? &st->done_up : &st->done_down;
270 complete(done);
271}
272
273/*
274 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
275 */
276static bool cpuhp_is_atomic_state(enum cpuhp_state state)
277{
278 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
279}
280
281/* Synchronization state management */
282enum cpuhp_sync_state {
283 SYNC_STATE_DEAD,
284 SYNC_STATE_KICKED,
285 SYNC_STATE_SHOULD_DIE,
286 SYNC_STATE_ALIVE,
287 SYNC_STATE_SHOULD_ONLINE,
288 SYNC_STATE_ONLINE,
289};
290
291#ifdef CONFIG_HOTPLUG_CORE_SYNC
292/**
293 * cpuhp_ap_update_sync_state - Update synchronization state during bringup/teardown
294 * @state: The synchronization state to set
295 *
296 * No synchronization point. Just update of the synchronization state, but implies
297 * a full barrier so that the AP changes are visible before the control CPU proceeds.
298 */
299static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state)
300{
301 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
302
303 (void)atomic_xchg(st, state);
304}
305
306void __weak arch_cpuhp_sync_state_poll(void) { cpu_relax(); }
307
308static bool cpuhp_wait_for_sync_state(unsigned int cpu, enum cpuhp_sync_state state,
309 enum cpuhp_sync_state next_state)
310{
311 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
312 ktime_t now, end, start = ktime_get();
313 int sync;
314
315 end = start + 10ULL * NSEC_PER_SEC;
316
317 sync = atomic_read(st);
318 while (1) {
319 if (sync == state) {
320 if (!atomic_try_cmpxchg(st, &sync, next_state))
321 continue;
322 return true;
323 }
324
325 now = ktime_get();
326 if (now > end) {
327 /* Timeout. Leave the state unchanged */
328 return false;
329 } else if (now - start < NSEC_PER_MSEC) {
330 /* Poll for one millisecond */
331 arch_cpuhp_sync_state_poll();
332 } else {
333 usleep_range(USEC_PER_MSEC, 2 * USEC_PER_MSEC);
334 }
335 sync = atomic_read(st);
336 }
337 return true;
338}
339#else /* CONFIG_HOTPLUG_CORE_SYNC */
340static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state) { }
341#endif /* !CONFIG_HOTPLUG_CORE_SYNC */
342
343#ifdef CONFIG_HOTPLUG_CORE_SYNC_DEAD
344/**
345 * cpuhp_ap_report_dead - Update synchronization state to DEAD
346 *
347 * No synchronization point. Just update of the synchronization state.
348 */
349void cpuhp_ap_report_dead(void)
350{
351 cpuhp_ap_update_sync_state(SYNC_STATE_DEAD);
352}
353
354void __weak arch_cpuhp_cleanup_dead_cpu(unsigned int cpu) { }
355
356/*
357 * Late CPU shutdown synchronization point. Cannot use cpuhp_state::done_down
358 * because the AP cannot issue complete() at this stage.
359 */
360static void cpuhp_bp_sync_dead(unsigned int cpu)
361{
362 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
363 int sync = atomic_read(st);
364
365 do {
366 /* CPU can have reported dead already. Don't overwrite that! */
367 if (sync == SYNC_STATE_DEAD)
368 break;
369 } while (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_SHOULD_DIE));
370
371 if (cpuhp_wait_for_sync_state(cpu, SYNC_STATE_DEAD, SYNC_STATE_DEAD)) {
372 /* CPU reached dead state. Invoke the cleanup function */
373 arch_cpuhp_cleanup_dead_cpu(cpu);
374 return;
375 }
376
377 /* No further action possible. Emit message and give up. */
378 pr_err("CPU%u failed to report dead state\n", cpu);
379}
380#else /* CONFIG_HOTPLUG_CORE_SYNC_DEAD */
381static inline void cpuhp_bp_sync_dead(unsigned int cpu) { }
382#endif /* !CONFIG_HOTPLUG_CORE_SYNC_DEAD */
383
384#ifdef CONFIG_HOTPLUG_CORE_SYNC_FULL
385/**
386 * cpuhp_ap_sync_alive - Synchronize AP with the control CPU once it is alive
387 *
388 * Updates the AP synchronization state to SYNC_STATE_ALIVE and waits
389 * for the BP to release it.
390 */
391void cpuhp_ap_sync_alive(void)
392{
393 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
394
395 cpuhp_ap_update_sync_state(SYNC_STATE_ALIVE);
396
397 /* Wait for the control CPU to release it. */
398 while (atomic_read(st) != SYNC_STATE_SHOULD_ONLINE)
399 cpu_relax();
400}
401
402static bool cpuhp_can_boot_ap(unsigned int cpu)
403{
404 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
405 int sync = atomic_read(st);
406
407again:
408 switch (sync) {
409 case SYNC_STATE_DEAD:
410 /* CPU is properly dead */
411 break;
412 case SYNC_STATE_KICKED:
413 /* CPU did not come up in previous attempt */
414 break;
415 case SYNC_STATE_ALIVE:
416 /* CPU is stuck cpuhp_ap_sync_alive(). */
417 break;
418 default:
419 /* CPU failed to report online or dead and is in limbo state. */
420 return false;
421 }
422
423 /* Prepare for booting */
424 if (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_KICKED))
425 goto again;
426
427 return true;
428}
429
430void __weak arch_cpuhp_cleanup_kick_cpu(unsigned int cpu) { }
431
432/*
433 * Early CPU bringup synchronization point. Cannot use cpuhp_state::done_up
434 * because the AP cannot issue complete() so early in the bringup.
435 */
436static int cpuhp_bp_sync_alive(unsigned int cpu)
437{
438 int ret = 0;
439
440 if (!IS_ENABLED(CONFIG_HOTPLUG_CORE_SYNC_FULL))
441 return 0;
442
443 if (!cpuhp_wait_for_sync_state(cpu, SYNC_STATE_ALIVE, SYNC_STATE_SHOULD_ONLINE)) {
444 pr_err("CPU%u failed to report alive state\n", cpu);
445 ret = -EIO;
446 }
447
448 /* Let the architecture cleanup the kick alive mechanics. */
449 arch_cpuhp_cleanup_kick_cpu(cpu);
450 return ret;
451}
452#else /* CONFIG_HOTPLUG_CORE_SYNC_FULL */
453static inline int cpuhp_bp_sync_alive(unsigned int cpu) { return 0; }
454static inline bool cpuhp_can_boot_ap(unsigned int cpu) { return true; }
455#endif /* !CONFIG_HOTPLUG_CORE_SYNC_FULL */
456
457/* Serializes the updates to cpu_online_mask, cpu_present_mask */
458static DEFINE_MUTEX(cpu_add_remove_lock);
459bool cpuhp_tasks_frozen;
460EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
461
462/*
463 * The following two APIs (cpu_maps_update_begin/done) must be used when
464 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
465 */
466void cpu_maps_update_begin(void)
467{
468 mutex_lock(&cpu_add_remove_lock);
469}
470
471void cpu_maps_update_done(void)
472{
473 mutex_unlock(&cpu_add_remove_lock);
474}
475
476/*
477 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
478 * Should always be manipulated under cpu_add_remove_lock
479 */
480static int cpu_hotplug_disabled;
481
482#ifdef CONFIG_HOTPLUG_CPU
483
484DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
485
486static bool cpu_hotplug_offline_disabled __ro_after_init;
487
488void cpus_read_lock(void)
489{
490 percpu_down_read(&cpu_hotplug_lock);
491}
492EXPORT_SYMBOL_GPL(cpus_read_lock);
493
494int cpus_read_trylock(void)
495{
496 return percpu_down_read_trylock(&cpu_hotplug_lock);
497}
498EXPORT_SYMBOL_GPL(cpus_read_trylock);
499
500void cpus_read_unlock(void)
501{
502 percpu_up_read(&cpu_hotplug_lock);
503}
504EXPORT_SYMBOL_GPL(cpus_read_unlock);
505
506void cpus_write_lock(void)
507{
508 percpu_down_write(&cpu_hotplug_lock);
509}
510
511void cpus_write_unlock(void)
512{
513 percpu_up_write(&cpu_hotplug_lock);
514}
515
516void lockdep_assert_cpus_held(void)
517{
518 /*
519 * We can't have hotplug operations before userspace starts running,
520 * and some init codepaths will knowingly not take the hotplug lock.
521 * This is all valid, so mute lockdep until it makes sense to report
522 * unheld locks.
523 */
524 if (system_state < SYSTEM_RUNNING)
525 return;
526
527 percpu_rwsem_assert_held(&cpu_hotplug_lock);
528}
529
530#ifdef CONFIG_LOCKDEP
531int lockdep_is_cpus_held(void)
532{
533 return percpu_rwsem_is_held(&cpu_hotplug_lock);
534}
535#endif
536
537static void lockdep_acquire_cpus_lock(void)
538{
539 rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
540}
541
542static void lockdep_release_cpus_lock(void)
543{
544 rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
545}
546
547/* Declare CPU offlining not supported */
548void cpu_hotplug_disable_offlining(void)
549{
550 cpu_maps_update_begin();
551 cpu_hotplug_offline_disabled = true;
552 cpu_maps_update_done();
553}
554
555/*
556 * Wait for currently running CPU hotplug operations to complete (if any) and
557 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
558 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
559 * hotplug path before performing hotplug operations. So acquiring that lock
560 * guarantees mutual exclusion from any currently running hotplug operations.
561 */
562void cpu_hotplug_disable(void)
563{
564 cpu_maps_update_begin();
565 cpu_hotplug_disabled++;
566 cpu_maps_update_done();
567}
568EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
569
570static void __cpu_hotplug_enable(void)
571{
572 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
573 return;
574 cpu_hotplug_disabled--;
575}
576
577void cpu_hotplug_enable(void)
578{
579 cpu_maps_update_begin();
580 __cpu_hotplug_enable();
581 cpu_maps_update_done();
582}
583EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
584
585#else
586
587static void lockdep_acquire_cpus_lock(void)
588{
589}
590
591static void lockdep_release_cpus_lock(void)
592{
593}
594
595#endif /* CONFIG_HOTPLUG_CPU */
596
597/*
598 * Architectures that need SMT-specific errata handling during SMT hotplug
599 * should override this.
600 */
601void __weak arch_smt_update(void) { }
602
603#ifdef CONFIG_HOTPLUG_SMT
604
605enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
606static unsigned int cpu_smt_max_threads __ro_after_init;
607unsigned int cpu_smt_num_threads __read_mostly = UINT_MAX;
608
609void __init cpu_smt_disable(bool force)
610{
611 if (!cpu_smt_possible())
612 return;
613
614 if (force) {
615 pr_info("SMT: Force disabled\n");
616 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
617 } else {
618 pr_info("SMT: disabled\n");
619 cpu_smt_control = CPU_SMT_DISABLED;
620 }
621 cpu_smt_num_threads = 1;
622}
623
624/*
625 * The decision whether SMT is supported can only be done after the full
626 * CPU identification. Called from architecture code.
627 */
628void __init cpu_smt_set_num_threads(unsigned int num_threads,
629 unsigned int max_threads)
630{
631 WARN_ON(!num_threads || (num_threads > max_threads));
632
633 if (max_threads == 1)
634 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
635
636 cpu_smt_max_threads = max_threads;
637
638 /*
639 * If SMT has been disabled via the kernel command line or SMT is
640 * not supported, set cpu_smt_num_threads to 1 for consistency.
641 * If enabled, take the architecture requested number of threads
642 * to bring up into account.
643 */
644 if (cpu_smt_control != CPU_SMT_ENABLED)
645 cpu_smt_num_threads = 1;
646 else if (num_threads < cpu_smt_num_threads)
647 cpu_smt_num_threads = num_threads;
648}
649
650static int __init smt_cmdline_disable(char *str)
651{
652 cpu_smt_disable(str && !strcmp(str, "force"));
653 return 0;
654}
655early_param("nosmt", smt_cmdline_disable);
656
657/*
658 * For Archicture supporting partial SMT states check if the thread is allowed.
659 * Otherwise this has already been checked through cpu_smt_max_threads when
660 * setting the SMT level.
661 */
662static inline bool cpu_smt_thread_allowed(unsigned int cpu)
663{
664#ifdef CONFIG_SMT_NUM_THREADS_DYNAMIC
665 return topology_smt_thread_allowed(cpu);
666#else
667 return true;
668#endif
669}
670
671static inline bool cpu_bootable(unsigned int cpu)
672{
673 if (cpu_smt_control == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
674 return true;
675
676 /* All CPUs are bootable if controls are not configured */
677 if (cpu_smt_control == CPU_SMT_NOT_IMPLEMENTED)
678 return true;
679
680 /* All CPUs are bootable if CPU is not SMT capable */
681 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
682 return true;
683
684 if (topology_is_primary_thread(cpu))
685 return true;
686
687 /*
688 * On x86 it's required to boot all logical CPUs at least once so
689 * that the init code can get a chance to set CR4.MCE on each
690 * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
691 * core will shutdown the machine.
692 */
693 return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
694}
695
696/* Returns true if SMT is supported and not forcefully (irreversibly) disabled */
697bool cpu_smt_possible(void)
698{
699 return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
700 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
701}
702EXPORT_SYMBOL_GPL(cpu_smt_possible);
703
704#else
705static inline bool cpu_bootable(unsigned int cpu) { return true; }
706#endif
707
708static inline enum cpuhp_state
709cpuhp_set_state(int cpu, struct cpuhp_cpu_state *st, enum cpuhp_state target)
710{
711 enum cpuhp_state prev_state = st->state;
712 bool bringup = st->state < target;
713
714 st->rollback = false;
715 st->last = NULL;
716
717 st->target = target;
718 st->single = false;
719 st->bringup = bringup;
720 if (cpu_dying(cpu) != !bringup)
721 set_cpu_dying(cpu, !bringup);
722
723 return prev_state;
724}
725
726static inline void
727cpuhp_reset_state(int cpu, struct cpuhp_cpu_state *st,
728 enum cpuhp_state prev_state)
729{
730 bool bringup = !st->bringup;
731
732 st->target = prev_state;
733
734 /*
735 * Already rolling back. No need invert the bringup value or to change
736 * the current state.
737 */
738 if (st->rollback)
739 return;
740
741 st->rollback = true;
742
743 /*
744 * If we have st->last we need to undo partial multi_instance of this
745 * state first. Otherwise start undo at the previous state.
746 */
747 if (!st->last) {
748 if (st->bringup)
749 st->state--;
750 else
751 st->state++;
752 }
753
754 st->bringup = bringup;
755 if (cpu_dying(cpu) != !bringup)
756 set_cpu_dying(cpu, !bringup);
757}
758
759/* Regular hotplug invocation of the AP hotplug thread */
760static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
761{
762 if (!st->single && st->state == st->target)
763 return;
764
765 st->result = 0;
766 /*
767 * Make sure the above stores are visible before should_run becomes
768 * true. Paired with the mb() above in cpuhp_thread_fun()
769 */
770 smp_mb();
771 st->should_run = true;
772 wake_up_process(st->thread);
773 wait_for_ap_thread(st, st->bringup);
774}
775
776static int cpuhp_kick_ap(int cpu, struct cpuhp_cpu_state *st,
777 enum cpuhp_state target)
778{
779 enum cpuhp_state prev_state;
780 int ret;
781
782 prev_state = cpuhp_set_state(cpu, st, target);
783 __cpuhp_kick_ap(st);
784 if ((ret = st->result)) {
785 cpuhp_reset_state(cpu, st, prev_state);
786 __cpuhp_kick_ap(st);
787 }
788
789 return ret;
790}
791
792static int bringup_wait_for_ap_online(unsigned int cpu)
793{
794 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
795
796 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
797 wait_for_ap_thread(st, true);
798 if (WARN_ON_ONCE((!cpu_online(cpu))))
799 return -ECANCELED;
800
801 /* Unpark the hotplug thread of the target cpu */
802 kthread_unpark(st->thread);
803
804 /*
805 * SMT soft disabling on X86 requires to bring the CPU out of the
806 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
807 * CPU marked itself as booted_once in notify_cpu_starting() so the
808 * cpu_bootable() check will now return false if this is not the
809 * primary sibling.
810 */
811 if (!cpu_bootable(cpu))
812 return -ECANCELED;
813 return 0;
814}
815
816#ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
817static int cpuhp_kick_ap_alive(unsigned int cpu)
818{
819 if (!cpuhp_can_boot_ap(cpu))
820 return -EAGAIN;
821
822 return arch_cpuhp_kick_ap_alive(cpu, idle_thread_get(cpu));
823}
824
825static int cpuhp_bringup_ap(unsigned int cpu)
826{
827 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
828 int ret;
829
830 /*
831 * Some architectures have to walk the irq descriptors to
832 * setup the vector space for the cpu which comes online.
833 * Prevent irq alloc/free across the bringup.
834 */
835 irq_lock_sparse();
836
837 ret = cpuhp_bp_sync_alive(cpu);
838 if (ret)
839 goto out_unlock;
840
841 ret = bringup_wait_for_ap_online(cpu);
842 if (ret)
843 goto out_unlock;
844
845 irq_unlock_sparse();
846
847 if (st->target <= CPUHP_AP_ONLINE_IDLE)
848 return 0;
849
850 return cpuhp_kick_ap(cpu, st, st->target);
851
852out_unlock:
853 irq_unlock_sparse();
854 return ret;
855}
856#else
857static int bringup_cpu(unsigned int cpu)
858{
859 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
860 struct task_struct *idle = idle_thread_get(cpu);
861 int ret;
862
863 if (!cpuhp_can_boot_ap(cpu))
864 return -EAGAIN;
865
866 /*
867 * Some architectures have to walk the irq descriptors to
868 * setup the vector space for the cpu which comes online.
869 *
870 * Prevent irq alloc/free across the bringup by acquiring the
871 * sparse irq lock. Hold it until the upcoming CPU completes the
872 * startup in cpuhp_online_idle() which allows to avoid
873 * intermediate synchronization points in the architecture code.
874 */
875 irq_lock_sparse();
876
877 ret = __cpu_up(cpu, idle);
878 if (ret)
879 goto out_unlock;
880
881 ret = cpuhp_bp_sync_alive(cpu);
882 if (ret)
883 goto out_unlock;
884
885 ret = bringup_wait_for_ap_online(cpu);
886 if (ret)
887 goto out_unlock;
888
889 irq_unlock_sparse();
890
891 if (st->target <= CPUHP_AP_ONLINE_IDLE)
892 return 0;
893
894 return cpuhp_kick_ap(cpu, st, st->target);
895
896out_unlock:
897 irq_unlock_sparse();
898 return ret;
899}
900#endif
901
902static int finish_cpu(unsigned int cpu)
903{
904 struct task_struct *idle = idle_thread_get(cpu);
905 struct mm_struct *mm = idle->active_mm;
906
907 /*
908 * idle_task_exit() will have switched to &init_mm, now
909 * clean up any remaining active_mm state.
910 */
911 if (mm != &init_mm)
912 idle->active_mm = &init_mm;
913 mmdrop_lazy_tlb(mm);
914 return 0;
915}
916
917/*
918 * Hotplug state machine related functions
919 */
920
921/*
922 * Get the next state to run. Empty ones will be skipped. Returns true if a
923 * state must be run.
924 *
925 * st->state will be modified ahead of time, to match state_to_run, as if it
926 * has already ran.
927 */
928static bool cpuhp_next_state(bool bringup,
929 enum cpuhp_state *state_to_run,
930 struct cpuhp_cpu_state *st,
931 enum cpuhp_state target)
932{
933 do {
934 if (bringup) {
935 if (st->state >= target)
936 return false;
937
938 *state_to_run = ++st->state;
939 } else {
940 if (st->state <= target)
941 return false;
942
943 *state_to_run = st->state--;
944 }
945
946 if (!cpuhp_step_empty(bringup, cpuhp_get_step(*state_to_run)))
947 break;
948 } while (true);
949
950 return true;
951}
952
953static int __cpuhp_invoke_callback_range(bool bringup,
954 unsigned int cpu,
955 struct cpuhp_cpu_state *st,
956 enum cpuhp_state target,
957 bool nofail)
958{
959 enum cpuhp_state state;
960 int ret = 0;
961
962 while (cpuhp_next_state(bringup, &state, st, target)) {
963 int err;
964
965 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
966 if (!err)
967 continue;
968
969 if (nofail) {
970 pr_warn("CPU %u %s state %s (%d) failed (%d)\n",
971 cpu, bringup ? "UP" : "DOWN",
972 cpuhp_get_step(st->state)->name,
973 st->state, err);
974 ret = -1;
975 } else {
976 ret = err;
977 break;
978 }
979 }
980
981 return ret;
982}
983
984static inline int cpuhp_invoke_callback_range(bool bringup,
985 unsigned int cpu,
986 struct cpuhp_cpu_state *st,
987 enum cpuhp_state target)
988{
989 return __cpuhp_invoke_callback_range(bringup, cpu, st, target, false);
990}
991
992static inline void cpuhp_invoke_callback_range_nofail(bool bringup,
993 unsigned int cpu,
994 struct cpuhp_cpu_state *st,
995 enum cpuhp_state target)
996{
997 __cpuhp_invoke_callback_range(bringup, cpu, st, target, true);
998}
999
1000static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
1001{
1002 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
1003 return true;
1004 /*
1005 * When CPU hotplug is disabled, then taking the CPU down is not
1006 * possible because takedown_cpu() and the architecture and
1007 * subsystem specific mechanisms are not available. So the CPU
1008 * which would be completely unplugged again needs to stay around
1009 * in the current state.
1010 */
1011 return st->state <= CPUHP_BRINGUP_CPU;
1012}
1013
1014static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1015 enum cpuhp_state target)
1016{
1017 enum cpuhp_state prev_state = st->state;
1018 int ret = 0;
1019
1020 ret = cpuhp_invoke_callback_range(true, cpu, st, target);
1021 if (ret) {
1022 pr_debug("CPU UP failed (%d) CPU %u state %s (%d)\n",
1023 ret, cpu, cpuhp_get_step(st->state)->name,
1024 st->state);
1025
1026 cpuhp_reset_state(cpu, st, prev_state);
1027 if (can_rollback_cpu(st))
1028 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
1029 prev_state));
1030 }
1031 return ret;
1032}
1033
1034/*
1035 * The cpu hotplug threads manage the bringup and teardown of the cpus
1036 */
1037static int cpuhp_should_run(unsigned int cpu)
1038{
1039 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1040
1041 return st->should_run;
1042}
1043
1044/*
1045 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
1046 * callbacks when a state gets [un]installed at runtime.
1047 *
1048 * Each invocation of this function by the smpboot thread does a single AP
1049 * state callback.
1050 *
1051 * It has 3 modes of operation:
1052 * - single: runs st->cb_state
1053 * - up: runs ++st->state, while st->state < st->target
1054 * - down: runs st->state--, while st->state > st->target
1055 *
1056 * When complete or on error, should_run is cleared and the completion is fired.
1057 */
1058static void cpuhp_thread_fun(unsigned int cpu)
1059{
1060 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1061 bool bringup = st->bringup;
1062 enum cpuhp_state state;
1063
1064 if (WARN_ON_ONCE(!st->should_run))
1065 return;
1066
1067 /*
1068 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
1069 * that if we see ->should_run we also see the rest of the state.
1070 */
1071 smp_mb();
1072
1073 /*
1074 * The BP holds the hotplug lock, but we're now running on the AP,
1075 * ensure that anybody asserting the lock is held, will actually find
1076 * it so.
1077 */
1078 lockdep_acquire_cpus_lock();
1079 cpuhp_lock_acquire(bringup);
1080
1081 if (st->single) {
1082 state = st->cb_state;
1083 st->should_run = false;
1084 } else {
1085 st->should_run = cpuhp_next_state(bringup, &state, st, st->target);
1086 if (!st->should_run)
1087 goto end;
1088 }
1089
1090 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
1091
1092 if (cpuhp_is_atomic_state(state)) {
1093 local_irq_disable();
1094 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1095 local_irq_enable();
1096
1097 /*
1098 * STARTING/DYING must not fail!
1099 */
1100 WARN_ON_ONCE(st->result);
1101 } else {
1102 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1103 }
1104
1105 if (st->result) {
1106 /*
1107 * If we fail on a rollback, we're up a creek without no
1108 * paddle, no way forward, no way back. We loose, thanks for
1109 * playing.
1110 */
1111 WARN_ON_ONCE(st->rollback);
1112 st->should_run = false;
1113 }
1114
1115end:
1116 cpuhp_lock_release(bringup);
1117 lockdep_release_cpus_lock();
1118
1119 if (!st->should_run)
1120 complete_ap_thread(st, bringup);
1121}
1122
1123/* Invoke a single callback on a remote cpu */
1124static int
1125cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
1126 struct hlist_node *node)
1127{
1128 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1129 int ret;
1130
1131 if (!cpu_online(cpu))
1132 return 0;
1133
1134 cpuhp_lock_acquire(false);
1135 cpuhp_lock_release(false);
1136
1137 cpuhp_lock_acquire(true);
1138 cpuhp_lock_release(true);
1139
1140 /*
1141 * If we are up and running, use the hotplug thread. For early calls
1142 * we invoke the thread function directly.
1143 */
1144 if (!st->thread)
1145 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1146
1147 st->rollback = false;
1148 st->last = NULL;
1149
1150 st->node = node;
1151 st->bringup = bringup;
1152 st->cb_state = state;
1153 st->single = true;
1154
1155 __cpuhp_kick_ap(st);
1156
1157 /*
1158 * If we failed and did a partial, do a rollback.
1159 */
1160 if ((ret = st->result) && st->last) {
1161 st->rollback = true;
1162 st->bringup = !bringup;
1163
1164 __cpuhp_kick_ap(st);
1165 }
1166
1167 /*
1168 * Clean up the leftovers so the next hotplug operation wont use stale
1169 * data.
1170 */
1171 st->node = st->last = NULL;
1172 return ret;
1173}
1174
1175static int cpuhp_kick_ap_work(unsigned int cpu)
1176{
1177 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1178 enum cpuhp_state prev_state = st->state;
1179 int ret;
1180
1181 cpuhp_lock_acquire(false);
1182 cpuhp_lock_release(false);
1183
1184 cpuhp_lock_acquire(true);
1185 cpuhp_lock_release(true);
1186
1187 trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
1188 ret = cpuhp_kick_ap(cpu, st, st->target);
1189 trace_cpuhp_exit(cpu, st->state, prev_state, ret);
1190
1191 return ret;
1192}
1193
1194static struct smp_hotplug_thread cpuhp_threads = {
1195 .store = &cpuhp_state.thread,
1196 .thread_should_run = cpuhp_should_run,
1197 .thread_fn = cpuhp_thread_fun,
1198 .thread_comm = "cpuhp/%u",
1199 .selfparking = true,
1200};
1201
1202static __init void cpuhp_init_state(void)
1203{
1204 struct cpuhp_cpu_state *st;
1205 int cpu;
1206
1207 for_each_possible_cpu(cpu) {
1208 st = per_cpu_ptr(&cpuhp_state, cpu);
1209 init_completion(&st->done_up);
1210 init_completion(&st->done_down);
1211 }
1212}
1213
1214void __init cpuhp_threads_init(void)
1215{
1216 cpuhp_init_state();
1217 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
1218 kthread_unpark(this_cpu_read(cpuhp_state.thread));
1219}
1220
1221#ifdef CONFIG_HOTPLUG_CPU
1222#ifndef arch_clear_mm_cpumask_cpu
1223#define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
1224#endif
1225
1226/**
1227 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
1228 * @cpu: a CPU id
1229 *
1230 * This function walks all processes, finds a valid mm struct for each one and
1231 * then clears a corresponding bit in mm's cpumask. While this all sounds
1232 * trivial, there are various non-obvious corner cases, which this function
1233 * tries to solve in a safe manner.
1234 *
1235 * Also note that the function uses a somewhat relaxed locking scheme, so it may
1236 * be called only for an already offlined CPU.
1237 */
1238void clear_tasks_mm_cpumask(int cpu)
1239{
1240 struct task_struct *p;
1241
1242 /*
1243 * This function is called after the cpu is taken down and marked
1244 * offline, so its not like new tasks will ever get this cpu set in
1245 * their mm mask. -- Peter Zijlstra
1246 * Thus, we may use rcu_read_lock() here, instead of grabbing
1247 * full-fledged tasklist_lock.
1248 */
1249 WARN_ON(cpu_online(cpu));
1250 rcu_read_lock();
1251 for_each_process(p) {
1252 struct task_struct *t;
1253
1254 /*
1255 * Main thread might exit, but other threads may still have
1256 * a valid mm. Find one.
1257 */
1258 t = find_lock_task_mm(p);
1259 if (!t)
1260 continue;
1261 arch_clear_mm_cpumask_cpu(cpu, t->mm);
1262 task_unlock(t);
1263 }
1264 rcu_read_unlock();
1265}
1266
1267/* Take this CPU down. */
1268static int take_cpu_down(void *_param)
1269{
1270 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1271 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
1272 int err, cpu = smp_processor_id();
1273
1274 /* Ensure this CPU doesn't handle any more interrupts. */
1275 err = __cpu_disable();
1276 if (err < 0)
1277 return err;
1278
1279 /*
1280 * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
1281 * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
1282 */
1283 WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
1284
1285 /*
1286 * Invoke the former CPU_DYING callbacks. DYING must not fail!
1287 */
1288 cpuhp_invoke_callback_range_nofail(false, cpu, st, target);
1289
1290 /* Park the stopper thread */
1291 stop_machine_park(cpu);
1292 return 0;
1293}
1294
1295static int takedown_cpu(unsigned int cpu)
1296{
1297 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1298 int err;
1299
1300 /* Park the smpboot threads */
1301 kthread_park(st->thread);
1302
1303 /*
1304 * Prevent irq alloc/free while the dying cpu reorganizes the
1305 * interrupt affinities.
1306 */
1307 irq_lock_sparse();
1308
1309 /*
1310 * So now all preempt/rcu users must observe !cpu_active().
1311 */
1312 err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
1313 if (err) {
1314 /* CPU refused to die */
1315 irq_unlock_sparse();
1316 /* Unpark the hotplug thread so we can rollback there */
1317 kthread_unpark(st->thread);
1318 return err;
1319 }
1320 BUG_ON(cpu_online(cpu));
1321
1322 /*
1323 * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1324 * all runnable tasks from the CPU, there's only the idle task left now
1325 * that the migration thread is done doing the stop_machine thing.
1326 *
1327 * Wait for the stop thread to go away.
1328 */
1329 wait_for_ap_thread(st, false);
1330 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1331
1332 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1333 irq_unlock_sparse();
1334
1335 hotplug_cpu__broadcast_tick_pull(cpu);
1336 /* This actually kills the CPU. */
1337 __cpu_die(cpu);
1338
1339 cpuhp_bp_sync_dead(cpu);
1340
1341 lockdep_cleanup_dead_cpu(cpu, idle_thread_get(cpu));
1342
1343 /*
1344 * Callbacks must be re-integrated right away to the RCU state machine.
1345 * Otherwise an RCU callback could block a further teardown function
1346 * waiting for its completion.
1347 */
1348 rcutree_migrate_callbacks(cpu);
1349
1350 return 0;
1351}
1352
1353static void cpuhp_complete_idle_dead(void *arg)
1354{
1355 struct cpuhp_cpu_state *st = arg;
1356
1357 complete_ap_thread(st, false);
1358}
1359
1360void cpuhp_report_idle_dead(void)
1361{
1362 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1363
1364 BUG_ON(st->state != CPUHP_AP_OFFLINE);
1365 tick_assert_timekeeping_handover();
1366 rcutree_report_cpu_dead();
1367 st->state = CPUHP_AP_IDLE_DEAD;
1368 /*
1369 * We cannot call complete after rcutree_report_cpu_dead() so we delegate it
1370 * to an online cpu.
1371 */
1372 smp_call_function_single(cpumask_first(cpu_online_mask),
1373 cpuhp_complete_idle_dead, st, 0);
1374}
1375
1376static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1377 enum cpuhp_state target)
1378{
1379 enum cpuhp_state prev_state = st->state;
1380 int ret = 0;
1381
1382 ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1383 if (ret) {
1384 pr_debug("CPU DOWN failed (%d) CPU %u state %s (%d)\n",
1385 ret, cpu, cpuhp_get_step(st->state)->name,
1386 st->state);
1387
1388 cpuhp_reset_state(cpu, st, prev_state);
1389
1390 if (st->state < prev_state)
1391 WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1392 prev_state));
1393 }
1394
1395 return ret;
1396}
1397
1398/* Requires cpu_add_remove_lock to be held */
1399static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1400 enum cpuhp_state target)
1401{
1402 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1403 int prev_state, ret = 0;
1404
1405 if (num_online_cpus() == 1)
1406 return -EBUSY;
1407
1408 if (!cpu_present(cpu))
1409 return -EINVAL;
1410
1411 cpus_write_lock();
1412
1413 cpuhp_tasks_frozen = tasks_frozen;
1414
1415 prev_state = cpuhp_set_state(cpu, st, target);
1416 /*
1417 * If the current CPU state is in the range of the AP hotplug thread,
1418 * then we need to kick the thread.
1419 */
1420 if (st->state > CPUHP_TEARDOWN_CPU) {
1421 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1422 ret = cpuhp_kick_ap_work(cpu);
1423 /*
1424 * The AP side has done the error rollback already. Just
1425 * return the error code..
1426 */
1427 if (ret)
1428 goto out;
1429
1430 /*
1431 * We might have stopped still in the range of the AP hotplug
1432 * thread. Nothing to do anymore.
1433 */
1434 if (st->state > CPUHP_TEARDOWN_CPU)
1435 goto out;
1436
1437 st->target = target;
1438 }
1439 /*
1440 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1441 * to do the further cleanups.
1442 */
1443 ret = cpuhp_down_callbacks(cpu, st, target);
1444 if (ret && st->state < prev_state) {
1445 if (st->state == CPUHP_TEARDOWN_CPU) {
1446 cpuhp_reset_state(cpu, st, prev_state);
1447 __cpuhp_kick_ap(st);
1448 } else {
1449 WARN(1, "DEAD callback error for CPU%d", cpu);
1450 }
1451 }
1452
1453out:
1454 cpus_write_unlock();
1455 /*
1456 * Do post unplug cleanup. This is still protected against
1457 * concurrent CPU hotplug via cpu_add_remove_lock.
1458 */
1459 lockup_detector_cleanup();
1460 arch_smt_update();
1461 return ret;
1462}
1463
1464struct cpu_down_work {
1465 unsigned int cpu;
1466 enum cpuhp_state target;
1467};
1468
1469static long __cpu_down_maps_locked(void *arg)
1470{
1471 struct cpu_down_work *work = arg;
1472
1473 return _cpu_down(work->cpu, 0, work->target);
1474}
1475
1476static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1477{
1478 struct cpu_down_work work = { .cpu = cpu, .target = target, };
1479
1480 /*
1481 * If the platform does not support hotplug, report it explicitly to
1482 * differentiate it from a transient offlining failure.
1483 */
1484 if (cpu_hotplug_offline_disabled)
1485 return -EOPNOTSUPP;
1486 if (cpu_hotplug_disabled)
1487 return -EBUSY;
1488
1489 /*
1490 * Ensure that the control task does not run on the to be offlined
1491 * CPU to prevent a deadlock against cfs_b->period_timer.
1492 * Also keep at least one housekeeping cpu onlined to avoid generating
1493 * an empty sched_domain span.
1494 */
1495 for_each_cpu_and(cpu, cpu_online_mask, housekeeping_cpumask(HK_TYPE_DOMAIN)) {
1496 if (cpu != work.cpu)
1497 return work_on_cpu(cpu, __cpu_down_maps_locked, &work);
1498 }
1499 return -EBUSY;
1500}
1501
1502static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1503{
1504 int err;
1505
1506 cpu_maps_update_begin();
1507 err = cpu_down_maps_locked(cpu, target);
1508 cpu_maps_update_done();
1509 return err;
1510}
1511
1512/**
1513 * cpu_device_down - Bring down a cpu device
1514 * @dev: Pointer to the cpu device to offline
1515 *
1516 * This function is meant to be used by device core cpu subsystem only.
1517 *
1518 * Other subsystems should use remove_cpu() instead.
1519 *
1520 * Return: %0 on success or a negative errno code
1521 */
1522int cpu_device_down(struct device *dev)
1523{
1524 return cpu_down(dev->id, CPUHP_OFFLINE);
1525}
1526
1527int remove_cpu(unsigned int cpu)
1528{
1529 int ret;
1530
1531 lock_device_hotplug();
1532 ret = device_offline(get_cpu_device(cpu));
1533 unlock_device_hotplug();
1534
1535 return ret;
1536}
1537EXPORT_SYMBOL_GPL(remove_cpu);
1538
1539void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1540{
1541 unsigned int cpu;
1542 int error;
1543
1544 cpu_maps_update_begin();
1545
1546 /*
1547 * Make certain the cpu I'm about to reboot on is online.
1548 *
1549 * This is inline to what migrate_to_reboot_cpu() already do.
1550 */
1551 if (!cpu_online(primary_cpu))
1552 primary_cpu = cpumask_first(cpu_online_mask);
1553
1554 for_each_online_cpu(cpu) {
1555 if (cpu == primary_cpu)
1556 continue;
1557
1558 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1559 if (error) {
1560 pr_err("Failed to offline CPU%d - error=%d",
1561 cpu, error);
1562 break;
1563 }
1564 }
1565
1566 /*
1567 * Ensure all but the reboot CPU are offline.
1568 */
1569 BUG_ON(num_online_cpus() > 1);
1570
1571 /*
1572 * Make sure the CPUs won't be enabled by someone else after this
1573 * point. Kexec will reboot to a new kernel shortly resetting
1574 * everything along the way.
1575 */
1576 cpu_hotplug_disabled++;
1577
1578 cpu_maps_update_done();
1579}
1580
1581#else
1582#define takedown_cpu NULL
1583#endif /*CONFIG_HOTPLUG_CPU*/
1584
1585/**
1586 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1587 * @cpu: cpu that just started
1588 *
1589 * It must be called by the arch code on the new cpu, before the new cpu
1590 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1591 */
1592void notify_cpu_starting(unsigned int cpu)
1593{
1594 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1595 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1596
1597 rcutree_report_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1598 cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1599
1600 /*
1601 * STARTING must not fail!
1602 */
1603 cpuhp_invoke_callback_range_nofail(true, cpu, st, target);
1604}
1605
1606/*
1607 * Called from the idle task. Wake up the controlling task which brings the
1608 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1609 * online bringup to the hotplug thread.
1610 */
1611void cpuhp_online_idle(enum cpuhp_state state)
1612{
1613 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1614
1615 /* Happens for the boot cpu */
1616 if (state != CPUHP_AP_ONLINE_IDLE)
1617 return;
1618
1619 cpuhp_ap_update_sync_state(SYNC_STATE_ONLINE);
1620
1621 /*
1622 * Unpark the stopper thread before we start the idle loop (and start
1623 * scheduling); this ensures the stopper task is always available.
1624 */
1625 stop_machine_unpark(smp_processor_id());
1626
1627 st->state = CPUHP_AP_ONLINE_IDLE;
1628 complete_ap_thread(st, true);
1629}
1630
1631/* Requires cpu_add_remove_lock to be held */
1632static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1633{
1634 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1635 struct task_struct *idle;
1636 int ret = 0;
1637
1638 cpus_write_lock();
1639
1640 if (!cpu_present(cpu)) {
1641 ret = -EINVAL;
1642 goto out;
1643 }
1644
1645 /*
1646 * The caller of cpu_up() might have raced with another
1647 * caller. Nothing to do.
1648 */
1649 if (st->state >= target)
1650 goto out;
1651
1652 if (st->state == CPUHP_OFFLINE) {
1653 /* Let it fail before we try to bring the cpu up */
1654 idle = idle_thread_get(cpu);
1655 if (IS_ERR(idle)) {
1656 ret = PTR_ERR(idle);
1657 goto out;
1658 }
1659
1660 /*
1661 * Reset stale stack state from the last time this CPU was online.
1662 */
1663 scs_task_reset(idle);
1664 kasan_unpoison_task_stack(idle);
1665 }
1666
1667 cpuhp_tasks_frozen = tasks_frozen;
1668
1669 cpuhp_set_state(cpu, st, target);
1670 /*
1671 * If the current CPU state is in the range of the AP hotplug thread,
1672 * then we need to kick the thread once more.
1673 */
1674 if (st->state > CPUHP_BRINGUP_CPU) {
1675 ret = cpuhp_kick_ap_work(cpu);
1676 /*
1677 * The AP side has done the error rollback already. Just
1678 * return the error code..
1679 */
1680 if (ret)
1681 goto out;
1682 }
1683
1684 /*
1685 * Try to reach the target state. We max out on the BP at
1686 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1687 * responsible for bringing it up to the target state.
1688 */
1689 target = min((int)target, CPUHP_BRINGUP_CPU);
1690 ret = cpuhp_up_callbacks(cpu, st, target);
1691out:
1692 cpus_write_unlock();
1693 arch_smt_update();
1694 return ret;
1695}
1696
1697static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1698{
1699 int err = 0;
1700
1701 if (!cpu_possible(cpu)) {
1702 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1703 cpu);
1704 return -EINVAL;
1705 }
1706
1707 err = try_online_node(cpu_to_node(cpu));
1708 if (err)
1709 return err;
1710
1711 cpu_maps_update_begin();
1712
1713 if (cpu_hotplug_disabled) {
1714 err = -EBUSY;
1715 goto out;
1716 }
1717 if (!cpu_bootable(cpu)) {
1718 err = -EPERM;
1719 goto out;
1720 }
1721
1722 err = _cpu_up(cpu, 0, target);
1723out:
1724 cpu_maps_update_done();
1725 return err;
1726}
1727
1728/**
1729 * cpu_device_up - Bring up a cpu device
1730 * @dev: Pointer to the cpu device to online
1731 *
1732 * This function is meant to be used by device core cpu subsystem only.
1733 *
1734 * Other subsystems should use add_cpu() instead.
1735 *
1736 * Return: %0 on success or a negative errno code
1737 */
1738int cpu_device_up(struct device *dev)
1739{
1740 return cpu_up(dev->id, CPUHP_ONLINE);
1741}
1742
1743int add_cpu(unsigned int cpu)
1744{
1745 int ret;
1746
1747 lock_device_hotplug();
1748 ret = device_online(get_cpu_device(cpu));
1749 unlock_device_hotplug();
1750
1751 return ret;
1752}
1753EXPORT_SYMBOL_GPL(add_cpu);
1754
1755/**
1756 * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1757 * @sleep_cpu: The cpu we hibernated on and should be brought up.
1758 *
1759 * On some architectures like arm64, we can hibernate on any CPU, but on
1760 * wake up the CPU we hibernated on might be offline as a side effect of
1761 * using maxcpus= for example.
1762 *
1763 * Return: %0 on success or a negative errno code
1764 */
1765int bringup_hibernate_cpu(unsigned int sleep_cpu)
1766{
1767 int ret;
1768
1769 if (!cpu_online(sleep_cpu)) {
1770 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1771 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1772 if (ret) {
1773 pr_err("Failed to bring hibernate-CPU up!\n");
1774 return ret;
1775 }
1776 }
1777 return 0;
1778}
1779
1780static void __init cpuhp_bringup_mask(const struct cpumask *mask, unsigned int ncpus,
1781 enum cpuhp_state target)
1782{
1783 unsigned int cpu;
1784
1785 for_each_cpu(cpu, mask) {
1786 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1787
1788 if (cpu_up(cpu, target) && can_rollback_cpu(st)) {
1789 /*
1790 * If this failed then cpu_up() might have only
1791 * rolled back to CPUHP_BP_KICK_AP for the final
1792 * online. Clean it up. NOOP if already rolled back.
1793 */
1794 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st, CPUHP_OFFLINE));
1795 }
1796
1797 if (!--ncpus)
1798 break;
1799 }
1800}
1801
1802#ifdef CONFIG_HOTPLUG_PARALLEL
1803static bool __cpuhp_parallel_bringup __ro_after_init = true;
1804
1805static int __init parallel_bringup_parse_param(char *arg)
1806{
1807 return kstrtobool(arg, &__cpuhp_parallel_bringup);
1808}
1809early_param("cpuhp.parallel", parallel_bringup_parse_param);
1810
1811#ifdef CONFIG_HOTPLUG_SMT
1812static inline bool cpuhp_smt_aware(void)
1813{
1814 return cpu_smt_max_threads > 1;
1815}
1816
1817static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1818{
1819 return cpu_primary_thread_mask;
1820}
1821#else
1822static inline bool cpuhp_smt_aware(void)
1823{
1824 return false;
1825}
1826static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1827{
1828 return cpu_none_mask;
1829}
1830#endif
1831
1832bool __weak arch_cpuhp_init_parallel_bringup(void)
1833{
1834 return true;
1835}
1836
1837/*
1838 * On architectures which have enabled parallel bringup this invokes all BP
1839 * prepare states for each of the to be onlined APs first. The last state
1840 * sends the startup IPI to the APs. The APs proceed through the low level
1841 * bringup code in parallel and then wait for the control CPU to release
1842 * them one by one for the final onlining procedure.
1843 *
1844 * This avoids waiting for each AP to respond to the startup IPI in
1845 * CPUHP_BRINGUP_CPU.
1846 */
1847static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
1848{
1849 const struct cpumask *mask = cpu_present_mask;
1850
1851 if (__cpuhp_parallel_bringup)
1852 __cpuhp_parallel_bringup = arch_cpuhp_init_parallel_bringup();
1853 if (!__cpuhp_parallel_bringup)
1854 return false;
1855
1856 if (cpuhp_smt_aware()) {
1857 const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
1858 static struct cpumask tmp_mask __initdata;
1859
1860 /*
1861 * X86 requires to prevent that SMT siblings stopped while
1862 * the primary thread does a microcode update for various
1863 * reasons. Bring the primary threads up first.
1864 */
1865 cpumask_and(&tmp_mask, mask, pmask);
1866 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_BP_KICK_AP);
1867 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_ONLINE);
1868 /* Account for the online CPUs */
1869 ncpus -= num_online_cpus();
1870 if (!ncpus)
1871 return true;
1872 /* Create the mask for secondary CPUs */
1873 cpumask_andnot(&tmp_mask, mask, pmask);
1874 mask = &tmp_mask;
1875 }
1876
1877 /* Bring the not-yet started CPUs up */
1878 cpuhp_bringup_mask(mask, ncpus, CPUHP_BP_KICK_AP);
1879 cpuhp_bringup_mask(mask, ncpus, CPUHP_ONLINE);
1880 return true;
1881}
1882#else
1883static inline bool cpuhp_bringup_cpus_parallel(unsigned int ncpus) { return false; }
1884#endif /* CONFIG_HOTPLUG_PARALLEL */
1885
1886void __init bringup_nonboot_cpus(unsigned int max_cpus)
1887{
1888 if (!max_cpus)
1889 return;
1890
1891 /* Try parallel bringup optimization if enabled */
1892 if (cpuhp_bringup_cpus_parallel(max_cpus))
1893 return;
1894
1895 /* Full per CPU serialized bringup */
1896 cpuhp_bringup_mask(cpu_present_mask, max_cpus, CPUHP_ONLINE);
1897}
1898
1899#ifdef CONFIG_PM_SLEEP_SMP
1900static cpumask_var_t frozen_cpus;
1901
1902int freeze_secondary_cpus(int primary)
1903{
1904 int cpu, error = 0;
1905
1906 cpu_maps_update_begin();
1907 if (primary == -1) {
1908 primary = cpumask_first(cpu_online_mask);
1909 if (!housekeeping_cpu(primary, HK_TYPE_TIMER))
1910 primary = housekeeping_any_cpu(HK_TYPE_TIMER);
1911 } else {
1912 if (!cpu_online(primary))
1913 primary = cpumask_first(cpu_online_mask);
1914 }
1915
1916 /*
1917 * We take down all of the non-boot CPUs in one shot to avoid races
1918 * with the userspace trying to use the CPU hotplug at the same time
1919 */
1920 cpumask_clear(frozen_cpus);
1921
1922 pr_info("Disabling non-boot CPUs ...\n");
1923 for (cpu = nr_cpu_ids - 1; cpu >= 0; cpu--) {
1924 if (!cpu_online(cpu) || cpu == primary)
1925 continue;
1926
1927 if (pm_wakeup_pending()) {
1928 pr_info("Wakeup pending. Abort CPU freeze\n");
1929 error = -EBUSY;
1930 break;
1931 }
1932
1933 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1934 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1935 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1936 if (!error)
1937 cpumask_set_cpu(cpu, frozen_cpus);
1938 else {
1939 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1940 break;
1941 }
1942 }
1943
1944 if (!error)
1945 BUG_ON(num_online_cpus() > 1);
1946 else
1947 pr_err("Non-boot CPUs are not disabled\n");
1948
1949 /*
1950 * Make sure the CPUs won't be enabled by someone else. We need to do
1951 * this even in case of failure as all freeze_secondary_cpus() users are
1952 * supposed to do thaw_secondary_cpus() on the failure path.
1953 */
1954 cpu_hotplug_disabled++;
1955
1956 cpu_maps_update_done();
1957 return error;
1958}
1959
1960void __weak arch_thaw_secondary_cpus_begin(void)
1961{
1962}
1963
1964void __weak arch_thaw_secondary_cpus_end(void)
1965{
1966}
1967
1968void thaw_secondary_cpus(void)
1969{
1970 int cpu, error;
1971
1972 /* Allow everyone to use the CPU hotplug again */
1973 cpu_maps_update_begin();
1974 __cpu_hotplug_enable();
1975 if (cpumask_empty(frozen_cpus))
1976 goto out;
1977
1978 pr_info("Enabling non-boot CPUs ...\n");
1979
1980 arch_thaw_secondary_cpus_begin();
1981
1982 for_each_cpu(cpu, frozen_cpus) {
1983 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1984 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1985 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1986 if (!error) {
1987 pr_info("CPU%d is up\n", cpu);
1988 continue;
1989 }
1990 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1991 }
1992
1993 arch_thaw_secondary_cpus_end();
1994
1995 cpumask_clear(frozen_cpus);
1996out:
1997 cpu_maps_update_done();
1998}
1999
2000static int __init alloc_frozen_cpus(void)
2001{
2002 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
2003 return -ENOMEM;
2004 return 0;
2005}
2006core_initcall(alloc_frozen_cpus);
2007
2008/*
2009 * When callbacks for CPU hotplug notifications are being executed, we must
2010 * ensure that the state of the system with respect to the tasks being frozen
2011 * or not, as reported by the notification, remains unchanged *throughout the
2012 * duration* of the execution of the callbacks.
2013 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
2014 *
2015 * This synchronization is implemented by mutually excluding regular CPU
2016 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
2017 * Hibernate notifications.
2018 */
2019static int
2020cpu_hotplug_pm_callback(struct notifier_block *nb,
2021 unsigned long action, void *ptr)
2022{
2023 switch (action) {
2024
2025 case PM_SUSPEND_PREPARE:
2026 case PM_HIBERNATION_PREPARE:
2027 cpu_hotplug_disable();
2028 break;
2029
2030 case PM_POST_SUSPEND:
2031 case PM_POST_HIBERNATION:
2032 cpu_hotplug_enable();
2033 break;
2034
2035 default:
2036 return NOTIFY_DONE;
2037 }
2038
2039 return NOTIFY_OK;
2040}
2041
2042
2043static int __init cpu_hotplug_pm_sync_init(void)
2044{
2045 /*
2046 * cpu_hotplug_pm_callback has higher priority than x86
2047 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
2048 * to disable cpu hotplug to avoid cpu hotplug race.
2049 */
2050 pm_notifier(cpu_hotplug_pm_callback, 0);
2051 return 0;
2052}
2053core_initcall(cpu_hotplug_pm_sync_init);
2054
2055#endif /* CONFIG_PM_SLEEP_SMP */
2056
2057int __boot_cpu_id;
2058
2059#endif /* CONFIG_SMP */
2060
2061/* Boot processor state steps */
2062static struct cpuhp_step cpuhp_hp_states[] = {
2063 [CPUHP_OFFLINE] = {
2064 .name = "offline",
2065 .startup.single = NULL,
2066 .teardown.single = NULL,
2067 },
2068#ifdef CONFIG_SMP
2069 [CPUHP_CREATE_THREADS]= {
2070 .name = "threads:prepare",
2071 .startup.single = smpboot_create_threads,
2072 .teardown.single = NULL,
2073 .cant_stop = true,
2074 },
2075 [CPUHP_PERF_PREPARE] = {
2076 .name = "perf:prepare",
2077 .startup.single = perf_event_init_cpu,
2078 .teardown.single = perf_event_exit_cpu,
2079 },
2080 [CPUHP_RANDOM_PREPARE] = {
2081 .name = "random:prepare",
2082 .startup.single = random_prepare_cpu,
2083 .teardown.single = NULL,
2084 },
2085 [CPUHP_WORKQUEUE_PREP] = {
2086 .name = "workqueue:prepare",
2087 .startup.single = workqueue_prepare_cpu,
2088 .teardown.single = NULL,
2089 },
2090 [CPUHP_HRTIMERS_PREPARE] = {
2091 .name = "hrtimers:prepare",
2092 .startup.single = hrtimers_prepare_cpu,
2093 .teardown.single = NULL,
2094 },
2095 [CPUHP_SMPCFD_PREPARE] = {
2096 .name = "smpcfd:prepare",
2097 .startup.single = smpcfd_prepare_cpu,
2098 .teardown.single = smpcfd_dead_cpu,
2099 },
2100 [CPUHP_RELAY_PREPARE] = {
2101 .name = "relay:prepare",
2102 .startup.single = relay_prepare_cpu,
2103 .teardown.single = NULL,
2104 },
2105 [CPUHP_RCUTREE_PREP] = {
2106 .name = "RCU/tree:prepare",
2107 .startup.single = rcutree_prepare_cpu,
2108 .teardown.single = rcutree_dead_cpu,
2109 },
2110 /*
2111 * On the tear-down path, timers_dead_cpu() must be invoked
2112 * before blk_mq_queue_reinit_notify() from notify_dead(),
2113 * otherwise a RCU stall occurs.
2114 */
2115 [CPUHP_TIMERS_PREPARE] = {
2116 .name = "timers:prepare",
2117 .startup.single = timers_prepare_cpu,
2118 .teardown.single = timers_dead_cpu,
2119 },
2120
2121#ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
2122 /*
2123 * Kicks the AP alive. AP will wait in cpuhp_ap_sync_alive() until
2124 * the next step will release it.
2125 */
2126 [CPUHP_BP_KICK_AP] = {
2127 .name = "cpu:kick_ap",
2128 .startup.single = cpuhp_kick_ap_alive,
2129 },
2130
2131 /*
2132 * Waits for the AP to reach cpuhp_ap_sync_alive() and then
2133 * releases it for the complete bringup.
2134 */
2135 [CPUHP_BRINGUP_CPU] = {
2136 .name = "cpu:bringup",
2137 .startup.single = cpuhp_bringup_ap,
2138 .teardown.single = finish_cpu,
2139 .cant_stop = true,
2140 },
2141#else
2142 /*
2143 * All-in-one CPU bringup state which includes the kick alive.
2144 */
2145 [CPUHP_BRINGUP_CPU] = {
2146 .name = "cpu:bringup",
2147 .startup.single = bringup_cpu,
2148 .teardown.single = finish_cpu,
2149 .cant_stop = true,
2150 },
2151#endif
2152 /* Final state before CPU kills itself */
2153 [CPUHP_AP_IDLE_DEAD] = {
2154 .name = "idle:dead",
2155 },
2156 /*
2157 * Last state before CPU enters the idle loop to die. Transient state
2158 * for synchronization.
2159 */
2160 [CPUHP_AP_OFFLINE] = {
2161 .name = "ap:offline",
2162 .cant_stop = true,
2163 },
2164 /* First state is scheduler control. Interrupts are disabled */
2165 [CPUHP_AP_SCHED_STARTING] = {
2166 .name = "sched:starting",
2167 .startup.single = sched_cpu_starting,
2168 .teardown.single = sched_cpu_dying,
2169 },
2170 [CPUHP_AP_RCUTREE_DYING] = {
2171 .name = "RCU/tree:dying",
2172 .startup.single = NULL,
2173 .teardown.single = rcutree_dying_cpu,
2174 },
2175 [CPUHP_AP_SMPCFD_DYING] = {
2176 .name = "smpcfd:dying",
2177 .startup.single = NULL,
2178 .teardown.single = smpcfd_dying_cpu,
2179 },
2180 [CPUHP_AP_HRTIMERS_DYING] = {
2181 .name = "hrtimers:dying",
2182 .startup.single = hrtimers_cpu_starting,
2183 .teardown.single = hrtimers_cpu_dying,
2184 },
2185 [CPUHP_AP_TICK_DYING] = {
2186 .name = "tick:dying",
2187 .startup.single = NULL,
2188 .teardown.single = tick_cpu_dying,
2189 },
2190 /* Entry state on starting. Interrupts enabled from here on. Transient
2191 * state for synchronsization */
2192 [CPUHP_AP_ONLINE] = {
2193 .name = "ap:online",
2194 },
2195 /*
2196 * Handled on control processor until the plugged processor manages
2197 * this itself.
2198 */
2199 [CPUHP_TEARDOWN_CPU] = {
2200 .name = "cpu:teardown",
2201 .startup.single = NULL,
2202 .teardown.single = takedown_cpu,
2203 .cant_stop = true,
2204 },
2205
2206 [CPUHP_AP_SCHED_WAIT_EMPTY] = {
2207 .name = "sched:waitempty",
2208 .startup.single = NULL,
2209 .teardown.single = sched_cpu_wait_empty,
2210 },
2211
2212 /* Handle smpboot threads park/unpark */
2213 [CPUHP_AP_SMPBOOT_THREADS] = {
2214 .name = "smpboot/threads:online",
2215 .startup.single = smpboot_unpark_threads,
2216 .teardown.single = smpboot_park_threads,
2217 },
2218 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
2219 .name = "irq/affinity:online",
2220 .startup.single = irq_affinity_online_cpu,
2221 .teardown.single = NULL,
2222 },
2223 [CPUHP_AP_PERF_ONLINE] = {
2224 .name = "perf:online",
2225 .startup.single = perf_event_init_cpu,
2226 .teardown.single = perf_event_exit_cpu,
2227 },
2228 [CPUHP_AP_WATCHDOG_ONLINE] = {
2229 .name = "lockup_detector:online",
2230 .startup.single = lockup_detector_online_cpu,
2231 .teardown.single = lockup_detector_offline_cpu,
2232 },
2233 [CPUHP_AP_WORKQUEUE_ONLINE] = {
2234 .name = "workqueue:online",
2235 .startup.single = workqueue_online_cpu,
2236 .teardown.single = workqueue_offline_cpu,
2237 },
2238 [CPUHP_AP_RANDOM_ONLINE] = {
2239 .name = "random:online",
2240 .startup.single = random_online_cpu,
2241 .teardown.single = NULL,
2242 },
2243 [CPUHP_AP_RCUTREE_ONLINE] = {
2244 .name = "RCU/tree:online",
2245 .startup.single = rcutree_online_cpu,
2246 .teardown.single = rcutree_offline_cpu,
2247 },
2248#endif
2249 /*
2250 * The dynamically registered state space is here
2251 */
2252
2253#ifdef CONFIG_SMP
2254 /* Last state is scheduler control setting the cpu active */
2255 [CPUHP_AP_ACTIVE] = {
2256 .name = "sched:active",
2257 .startup.single = sched_cpu_activate,
2258 .teardown.single = sched_cpu_deactivate,
2259 },
2260#endif
2261
2262 /* CPU is fully up and running. */
2263 [CPUHP_ONLINE] = {
2264 .name = "online",
2265 .startup.single = NULL,
2266 .teardown.single = NULL,
2267 },
2268};
2269
2270/* Sanity check for callbacks */
2271static int cpuhp_cb_check(enum cpuhp_state state)
2272{
2273 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
2274 return -EINVAL;
2275 return 0;
2276}
2277
2278/*
2279 * Returns a free for dynamic slot assignment of the Online state. The states
2280 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
2281 * by having no name assigned.
2282 */
2283static int cpuhp_reserve_state(enum cpuhp_state state)
2284{
2285 enum cpuhp_state i, end;
2286 struct cpuhp_step *step;
2287
2288 switch (state) {
2289 case CPUHP_AP_ONLINE_DYN:
2290 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
2291 end = CPUHP_AP_ONLINE_DYN_END;
2292 break;
2293 case CPUHP_BP_PREPARE_DYN:
2294 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
2295 end = CPUHP_BP_PREPARE_DYN_END;
2296 break;
2297 default:
2298 return -EINVAL;
2299 }
2300
2301 for (i = state; i <= end; i++, step++) {
2302 if (!step->name)
2303 return i;
2304 }
2305 WARN(1, "No more dynamic states available for CPU hotplug\n");
2306 return -ENOSPC;
2307}
2308
2309static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
2310 int (*startup)(unsigned int cpu),
2311 int (*teardown)(unsigned int cpu),
2312 bool multi_instance)
2313{
2314 /* (Un)Install the callbacks for further cpu hotplug operations */
2315 struct cpuhp_step *sp;
2316 int ret = 0;
2317
2318 /*
2319 * If name is NULL, then the state gets removed.
2320 *
2321 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
2322 * the first allocation from these dynamic ranges, so the removal
2323 * would trigger a new allocation and clear the wrong (already
2324 * empty) state, leaving the callbacks of the to be cleared state
2325 * dangling, which causes wreckage on the next hotplug operation.
2326 */
2327 if (name && (state == CPUHP_AP_ONLINE_DYN ||
2328 state == CPUHP_BP_PREPARE_DYN)) {
2329 ret = cpuhp_reserve_state(state);
2330 if (ret < 0)
2331 return ret;
2332 state = ret;
2333 }
2334 sp = cpuhp_get_step(state);
2335 if (name && sp->name)
2336 return -EBUSY;
2337
2338 sp->startup.single = startup;
2339 sp->teardown.single = teardown;
2340 sp->name = name;
2341 sp->multi_instance = multi_instance;
2342 INIT_HLIST_HEAD(&sp->list);
2343 return ret;
2344}
2345
2346static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
2347{
2348 return cpuhp_get_step(state)->teardown.single;
2349}
2350
2351/*
2352 * Call the startup/teardown function for a step either on the AP or
2353 * on the current CPU.
2354 */
2355static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
2356 struct hlist_node *node)
2357{
2358 struct cpuhp_step *sp = cpuhp_get_step(state);
2359 int ret;
2360
2361 /*
2362 * If there's nothing to do, we done.
2363 * Relies on the union for multi_instance.
2364 */
2365 if (cpuhp_step_empty(bringup, sp))
2366 return 0;
2367 /*
2368 * The non AP bound callbacks can fail on bringup. On teardown
2369 * e.g. module removal we crash for now.
2370 */
2371#ifdef CONFIG_SMP
2372 if (cpuhp_is_ap_state(state))
2373 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
2374 else
2375 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2376#else
2377 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2378#endif
2379 BUG_ON(ret && !bringup);
2380 return ret;
2381}
2382
2383/*
2384 * Called from __cpuhp_setup_state on a recoverable failure.
2385 *
2386 * Note: The teardown callbacks for rollback are not allowed to fail!
2387 */
2388static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
2389 struct hlist_node *node)
2390{
2391 int cpu;
2392
2393 /* Roll back the already executed steps on the other cpus */
2394 for_each_present_cpu(cpu) {
2395 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2396 int cpustate = st->state;
2397
2398 if (cpu >= failedcpu)
2399 break;
2400
2401 /* Did we invoke the startup call on that cpu ? */
2402 if (cpustate >= state)
2403 cpuhp_issue_call(cpu, state, false, node);
2404 }
2405}
2406
2407int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
2408 struct hlist_node *node,
2409 bool invoke)
2410{
2411 struct cpuhp_step *sp;
2412 int cpu;
2413 int ret;
2414
2415 lockdep_assert_cpus_held();
2416
2417 sp = cpuhp_get_step(state);
2418 if (sp->multi_instance == false)
2419 return -EINVAL;
2420
2421 mutex_lock(&cpuhp_state_mutex);
2422
2423 if (!invoke || !sp->startup.multi)
2424 goto add_node;
2425
2426 /*
2427 * Try to call the startup callback for each present cpu
2428 * depending on the hotplug state of the cpu.
2429 */
2430 for_each_present_cpu(cpu) {
2431 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2432 int cpustate = st->state;
2433
2434 if (cpustate < state)
2435 continue;
2436
2437 ret = cpuhp_issue_call(cpu, state, true, node);
2438 if (ret) {
2439 if (sp->teardown.multi)
2440 cpuhp_rollback_install(cpu, state, node);
2441 goto unlock;
2442 }
2443 }
2444add_node:
2445 ret = 0;
2446 hlist_add_head(node, &sp->list);
2447unlock:
2448 mutex_unlock(&cpuhp_state_mutex);
2449 return ret;
2450}
2451
2452int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
2453 bool invoke)
2454{
2455 int ret;
2456
2457 cpus_read_lock();
2458 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
2459 cpus_read_unlock();
2460 return ret;
2461}
2462EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
2463
2464/**
2465 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
2466 * @state: The state to setup
2467 * @name: Name of the step
2468 * @invoke: If true, the startup function is invoked for cpus where
2469 * cpu state >= @state
2470 * @startup: startup callback function
2471 * @teardown: teardown callback function
2472 * @multi_instance: State is set up for multiple instances which get
2473 * added afterwards.
2474 *
2475 * The caller needs to hold cpus read locked while calling this function.
2476 * Return:
2477 * On success:
2478 * Positive state number if @state is CPUHP_AP_ONLINE_DYN or CPUHP_BP_PREPARE_DYN;
2479 * 0 for all other states
2480 * On failure: proper (negative) error code
2481 */
2482int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
2483 const char *name, bool invoke,
2484 int (*startup)(unsigned int cpu),
2485 int (*teardown)(unsigned int cpu),
2486 bool multi_instance)
2487{
2488 int cpu, ret = 0;
2489 bool dynstate;
2490
2491 lockdep_assert_cpus_held();
2492
2493 if (cpuhp_cb_check(state) || !name)
2494 return -EINVAL;
2495
2496 mutex_lock(&cpuhp_state_mutex);
2497
2498 ret = cpuhp_store_callbacks(state, name, startup, teardown,
2499 multi_instance);
2500
2501 dynstate = state == CPUHP_AP_ONLINE_DYN || state == CPUHP_BP_PREPARE_DYN;
2502 if (ret > 0 && dynstate) {
2503 state = ret;
2504 ret = 0;
2505 }
2506
2507 if (ret || !invoke || !startup)
2508 goto out;
2509
2510 /*
2511 * Try to call the startup callback for each present cpu
2512 * depending on the hotplug state of the cpu.
2513 */
2514 for_each_present_cpu(cpu) {
2515 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2516 int cpustate = st->state;
2517
2518 if (cpustate < state)
2519 continue;
2520
2521 ret = cpuhp_issue_call(cpu, state, true, NULL);
2522 if (ret) {
2523 if (teardown)
2524 cpuhp_rollback_install(cpu, state, NULL);
2525 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2526 goto out;
2527 }
2528 }
2529out:
2530 mutex_unlock(&cpuhp_state_mutex);
2531 /*
2532 * If the requested state is CPUHP_AP_ONLINE_DYN or CPUHP_BP_PREPARE_DYN,
2533 * return the dynamically allocated state in case of success.
2534 */
2535 if (!ret && dynstate)
2536 return state;
2537 return ret;
2538}
2539EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2540
2541int __cpuhp_setup_state(enum cpuhp_state state,
2542 const char *name, bool invoke,
2543 int (*startup)(unsigned int cpu),
2544 int (*teardown)(unsigned int cpu),
2545 bool multi_instance)
2546{
2547 int ret;
2548
2549 cpus_read_lock();
2550 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2551 teardown, multi_instance);
2552 cpus_read_unlock();
2553 return ret;
2554}
2555EXPORT_SYMBOL(__cpuhp_setup_state);
2556
2557int __cpuhp_state_remove_instance(enum cpuhp_state state,
2558 struct hlist_node *node, bool invoke)
2559{
2560 struct cpuhp_step *sp = cpuhp_get_step(state);
2561 int cpu;
2562
2563 BUG_ON(cpuhp_cb_check(state));
2564
2565 if (!sp->multi_instance)
2566 return -EINVAL;
2567
2568 cpus_read_lock();
2569 mutex_lock(&cpuhp_state_mutex);
2570
2571 if (!invoke || !cpuhp_get_teardown_cb(state))
2572 goto remove;
2573 /*
2574 * Call the teardown callback for each present cpu depending
2575 * on the hotplug state of the cpu. This function is not
2576 * allowed to fail currently!
2577 */
2578 for_each_present_cpu(cpu) {
2579 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2580 int cpustate = st->state;
2581
2582 if (cpustate >= state)
2583 cpuhp_issue_call(cpu, state, false, node);
2584 }
2585
2586remove:
2587 hlist_del(node);
2588 mutex_unlock(&cpuhp_state_mutex);
2589 cpus_read_unlock();
2590
2591 return 0;
2592}
2593EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2594
2595/**
2596 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2597 * @state: The state to remove
2598 * @invoke: If true, the teardown function is invoked for cpus where
2599 * cpu state >= @state
2600 *
2601 * The caller needs to hold cpus read locked while calling this function.
2602 * The teardown callback is currently not allowed to fail. Think
2603 * about module removal!
2604 */
2605void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2606{
2607 struct cpuhp_step *sp = cpuhp_get_step(state);
2608 int cpu;
2609
2610 BUG_ON(cpuhp_cb_check(state));
2611
2612 lockdep_assert_cpus_held();
2613
2614 mutex_lock(&cpuhp_state_mutex);
2615 if (sp->multi_instance) {
2616 WARN(!hlist_empty(&sp->list),
2617 "Error: Removing state %d which has instances left.\n",
2618 state);
2619 goto remove;
2620 }
2621
2622 if (!invoke || !cpuhp_get_teardown_cb(state))
2623 goto remove;
2624
2625 /*
2626 * Call the teardown callback for each present cpu depending
2627 * on the hotplug state of the cpu. This function is not
2628 * allowed to fail currently!
2629 */
2630 for_each_present_cpu(cpu) {
2631 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2632 int cpustate = st->state;
2633
2634 if (cpustate >= state)
2635 cpuhp_issue_call(cpu, state, false, NULL);
2636 }
2637remove:
2638 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2639 mutex_unlock(&cpuhp_state_mutex);
2640}
2641EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2642
2643void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2644{
2645 cpus_read_lock();
2646 __cpuhp_remove_state_cpuslocked(state, invoke);
2647 cpus_read_unlock();
2648}
2649EXPORT_SYMBOL(__cpuhp_remove_state);
2650
2651#ifdef CONFIG_HOTPLUG_SMT
2652static void cpuhp_offline_cpu_device(unsigned int cpu)
2653{
2654 struct device *dev = get_cpu_device(cpu);
2655
2656 dev->offline = true;
2657 /* Tell user space about the state change */
2658 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2659}
2660
2661static void cpuhp_online_cpu_device(unsigned int cpu)
2662{
2663 struct device *dev = get_cpu_device(cpu);
2664
2665 dev->offline = false;
2666 /* Tell user space about the state change */
2667 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2668}
2669
2670int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2671{
2672 int cpu, ret = 0;
2673
2674 cpu_maps_update_begin();
2675 for_each_online_cpu(cpu) {
2676 if (topology_is_primary_thread(cpu))
2677 continue;
2678 /*
2679 * Disable can be called with CPU_SMT_ENABLED when changing
2680 * from a higher to lower number of SMT threads per core.
2681 */
2682 if (ctrlval == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
2683 continue;
2684 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2685 if (ret)
2686 break;
2687 /*
2688 * As this needs to hold the cpu maps lock it's impossible
2689 * to call device_offline() because that ends up calling
2690 * cpu_down() which takes cpu maps lock. cpu maps lock
2691 * needs to be held as this might race against in kernel
2692 * abusers of the hotplug machinery (thermal management).
2693 *
2694 * So nothing would update device:offline state. That would
2695 * leave the sysfs entry stale and prevent onlining after
2696 * smt control has been changed to 'off' again. This is
2697 * called under the sysfs hotplug lock, so it is properly
2698 * serialized against the regular offline usage.
2699 */
2700 cpuhp_offline_cpu_device(cpu);
2701 }
2702 if (!ret)
2703 cpu_smt_control = ctrlval;
2704 cpu_maps_update_done();
2705 return ret;
2706}
2707
2708/* Check if the core a CPU belongs to is online */
2709#if !defined(topology_is_core_online)
2710static inline bool topology_is_core_online(unsigned int cpu)
2711{
2712 return true;
2713}
2714#endif
2715
2716int cpuhp_smt_enable(void)
2717{
2718 int cpu, ret = 0;
2719
2720 cpu_maps_update_begin();
2721 cpu_smt_control = CPU_SMT_ENABLED;
2722 for_each_present_cpu(cpu) {
2723 /* Skip online CPUs and CPUs on offline nodes */
2724 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2725 continue;
2726 if (!cpu_smt_thread_allowed(cpu) || !topology_is_core_online(cpu))
2727 continue;
2728 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2729 if (ret)
2730 break;
2731 /* See comment in cpuhp_smt_disable() */
2732 cpuhp_online_cpu_device(cpu);
2733 }
2734 cpu_maps_update_done();
2735 return ret;
2736}
2737#endif
2738
2739#if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2740static ssize_t state_show(struct device *dev,
2741 struct device_attribute *attr, char *buf)
2742{
2743 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2744
2745 return sprintf(buf, "%d\n", st->state);
2746}
2747static DEVICE_ATTR_RO(state);
2748
2749static ssize_t target_store(struct device *dev, struct device_attribute *attr,
2750 const char *buf, size_t count)
2751{
2752 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2753 struct cpuhp_step *sp;
2754 int target, ret;
2755
2756 ret = kstrtoint(buf, 10, &target);
2757 if (ret)
2758 return ret;
2759
2760#ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2761 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2762 return -EINVAL;
2763#else
2764 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2765 return -EINVAL;
2766#endif
2767
2768 ret = lock_device_hotplug_sysfs();
2769 if (ret)
2770 return ret;
2771
2772 mutex_lock(&cpuhp_state_mutex);
2773 sp = cpuhp_get_step(target);
2774 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2775 mutex_unlock(&cpuhp_state_mutex);
2776 if (ret)
2777 goto out;
2778
2779 if (st->state < target)
2780 ret = cpu_up(dev->id, target);
2781 else if (st->state > target)
2782 ret = cpu_down(dev->id, target);
2783 else if (WARN_ON(st->target != target))
2784 st->target = target;
2785out:
2786 unlock_device_hotplug();
2787 return ret ? ret : count;
2788}
2789
2790static ssize_t target_show(struct device *dev,
2791 struct device_attribute *attr, char *buf)
2792{
2793 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2794
2795 return sprintf(buf, "%d\n", st->target);
2796}
2797static DEVICE_ATTR_RW(target);
2798
2799static ssize_t fail_store(struct device *dev, struct device_attribute *attr,
2800 const char *buf, size_t count)
2801{
2802 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2803 struct cpuhp_step *sp;
2804 int fail, ret;
2805
2806 ret = kstrtoint(buf, 10, &fail);
2807 if (ret)
2808 return ret;
2809
2810 if (fail == CPUHP_INVALID) {
2811 st->fail = fail;
2812 return count;
2813 }
2814
2815 if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2816 return -EINVAL;
2817
2818 /*
2819 * Cannot fail STARTING/DYING callbacks.
2820 */
2821 if (cpuhp_is_atomic_state(fail))
2822 return -EINVAL;
2823
2824 /*
2825 * DEAD callbacks cannot fail...
2826 * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2827 * triggering STARTING callbacks, a failure in this state would
2828 * hinder rollback.
2829 */
2830 if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2831 return -EINVAL;
2832
2833 /*
2834 * Cannot fail anything that doesn't have callbacks.
2835 */
2836 mutex_lock(&cpuhp_state_mutex);
2837 sp = cpuhp_get_step(fail);
2838 if (!sp->startup.single && !sp->teardown.single)
2839 ret = -EINVAL;
2840 mutex_unlock(&cpuhp_state_mutex);
2841 if (ret)
2842 return ret;
2843
2844 st->fail = fail;
2845
2846 return count;
2847}
2848
2849static ssize_t fail_show(struct device *dev,
2850 struct device_attribute *attr, char *buf)
2851{
2852 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2853
2854 return sprintf(buf, "%d\n", st->fail);
2855}
2856
2857static DEVICE_ATTR_RW(fail);
2858
2859static struct attribute *cpuhp_cpu_attrs[] = {
2860 &dev_attr_state.attr,
2861 &dev_attr_target.attr,
2862 &dev_attr_fail.attr,
2863 NULL
2864};
2865
2866static const struct attribute_group cpuhp_cpu_attr_group = {
2867 .attrs = cpuhp_cpu_attrs,
2868 .name = "hotplug",
2869};
2870
2871static ssize_t states_show(struct device *dev,
2872 struct device_attribute *attr, char *buf)
2873{
2874 ssize_t cur, res = 0;
2875 int i;
2876
2877 mutex_lock(&cpuhp_state_mutex);
2878 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2879 struct cpuhp_step *sp = cpuhp_get_step(i);
2880
2881 if (sp->name) {
2882 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2883 buf += cur;
2884 res += cur;
2885 }
2886 }
2887 mutex_unlock(&cpuhp_state_mutex);
2888 return res;
2889}
2890static DEVICE_ATTR_RO(states);
2891
2892static struct attribute *cpuhp_cpu_root_attrs[] = {
2893 &dev_attr_states.attr,
2894 NULL
2895};
2896
2897static const struct attribute_group cpuhp_cpu_root_attr_group = {
2898 .attrs = cpuhp_cpu_root_attrs,
2899 .name = "hotplug",
2900};
2901
2902#ifdef CONFIG_HOTPLUG_SMT
2903
2904static bool cpu_smt_num_threads_valid(unsigned int threads)
2905{
2906 if (IS_ENABLED(CONFIG_SMT_NUM_THREADS_DYNAMIC))
2907 return threads >= 1 && threads <= cpu_smt_max_threads;
2908 return threads == 1 || threads == cpu_smt_max_threads;
2909}
2910
2911static ssize_t
2912__store_smt_control(struct device *dev, struct device_attribute *attr,
2913 const char *buf, size_t count)
2914{
2915 int ctrlval, ret, num_threads, orig_threads;
2916 bool force_off;
2917
2918 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2919 return -EPERM;
2920
2921 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2922 return -ENODEV;
2923
2924 if (sysfs_streq(buf, "on")) {
2925 ctrlval = CPU_SMT_ENABLED;
2926 num_threads = cpu_smt_max_threads;
2927 } else if (sysfs_streq(buf, "off")) {
2928 ctrlval = CPU_SMT_DISABLED;
2929 num_threads = 1;
2930 } else if (sysfs_streq(buf, "forceoff")) {
2931 ctrlval = CPU_SMT_FORCE_DISABLED;
2932 num_threads = 1;
2933 } else if (kstrtoint(buf, 10, &num_threads) == 0) {
2934 if (num_threads == 1)
2935 ctrlval = CPU_SMT_DISABLED;
2936 else if (cpu_smt_num_threads_valid(num_threads))
2937 ctrlval = CPU_SMT_ENABLED;
2938 else
2939 return -EINVAL;
2940 } else {
2941 return -EINVAL;
2942 }
2943
2944 ret = lock_device_hotplug_sysfs();
2945 if (ret)
2946 return ret;
2947
2948 orig_threads = cpu_smt_num_threads;
2949 cpu_smt_num_threads = num_threads;
2950
2951 force_off = ctrlval != cpu_smt_control && ctrlval == CPU_SMT_FORCE_DISABLED;
2952
2953 if (num_threads > orig_threads)
2954 ret = cpuhp_smt_enable();
2955 else if (num_threads < orig_threads || force_off)
2956 ret = cpuhp_smt_disable(ctrlval);
2957
2958 unlock_device_hotplug();
2959 return ret ? ret : count;
2960}
2961
2962#else /* !CONFIG_HOTPLUG_SMT */
2963static ssize_t
2964__store_smt_control(struct device *dev, struct device_attribute *attr,
2965 const char *buf, size_t count)
2966{
2967 return -ENODEV;
2968}
2969#endif /* CONFIG_HOTPLUG_SMT */
2970
2971static const char *smt_states[] = {
2972 [CPU_SMT_ENABLED] = "on",
2973 [CPU_SMT_DISABLED] = "off",
2974 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2975 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2976 [CPU_SMT_NOT_IMPLEMENTED] = "notimplemented",
2977};
2978
2979static ssize_t control_show(struct device *dev,
2980 struct device_attribute *attr, char *buf)
2981{
2982 const char *state = smt_states[cpu_smt_control];
2983
2984#ifdef CONFIG_HOTPLUG_SMT
2985 /*
2986 * If SMT is enabled but not all threads are enabled then show the
2987 * number of threads. If all threads are enabled show "on". Otherwise
2988 * show the state name.
2989 */
2990 if (cpu_smt_control == CPU_SMT_ENABLED &&
2991 cpu_smt_num_threads != cpu_smt_max_threads)
2992 return sysfs_emit(buf, "%d\n", cpu_smt_num_threads);
2993#endif
2994
2995 return sysfs_emit(buf, "%s\n", state);
2996}
2997
2998static ssize_t control_store(struct device *dev, struct device_attribute *attr,
2999 const char *buf, size_t count)
3000{
3001 return __store_smt_control(dev, attr, buf, count);
3002}
3003static DEVICE_ATTR_RW(control);
3004
3005static ssize_t active_show(struct device *dev,
3006 struct device_attribute *attr, char *buf)
3007{
3008 return sysfs_emit(buf, "%d\n", sched_smt_active());
3009}
3010static DEVICE_ATTR_RO(active);
3011
3012static struct attribute *cpuhp_smt_attrs[] = {
3013 &dev_attr_control.attr,
3014 &dev_attr_active.attr,
3015 NULL
3016};
3017
3018static const struct attribute_group cpuhp_smt_attr_group = {
3019 .attrs = cpuhp_smt_attrs,
3020 .name = "smt",
3021};
3022
3023static int __init cpu_smt_sysfs_init(void)
3024{
3025 struct device *dev_root;
3026 int ret = -ENODEV;
3027
3028 dev_root = bus_get_dev_root(&cpu_subsys);
3029 if (dev_root) {
3030 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_smt_attr_group);
3031 put_device(dev_root);
3032 }
3033 return ret;
3034}
3035
3036static int __init cpuhp_sysfs_init(void)
3037{
3038 struct device *dev_root;
3039 int cpu, ret;
3040
3041 ret = cpu_smt_sysfs_init();
3042 if (ret)
3043 return ret;
3044
3045 dev_root = bus_get_dev_root(&cpu_subsys);
3046 if (dev_root) {
3047 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_cpu_root_attr_group);
3048 put_device(dev_root);
3049 if (ret)
3050 return ret;
3051 }
3052
3053 for_each_possible_cpu(cpu) {
3054 struct device *dev = get_cpu_device(cpu);
3055
3056 if (!dev)
3057 continue;
3058 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
3059 if (ret)
3060 return ret;
3061 }
3062 return 0;
3063}
3064device_initcall(cpuhp_sysfs_init);
3065#endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
3066
3067/*
3068 * cpu_bit_bitmap[] is a special, "compressed" data structure that
3069 * represents all NR_CPUS bits binary values of 1<<nr.
3070 *
3071 * It is used by cpumask_of() to get a constant address to a CPU
3072 * mask value that has a single bit set only.
3073 */
3074
3075/* cpu_bit_bitmap[0] is empty - so we can back into it */
3076#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
3077#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
3078#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
3079#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
3080
3081const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
3082
3083 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
3084 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
3085#if BITS_PER_LONG > 32
3086 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
3087 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
3088#endif
3089};
3090EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
3091
3092const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
3093EXPORT_SYMBOL(cpu_all_bits);
3094
3095#ifdef CONFIG_INIT_ALL_POSSIBLE
3096struct cpumask __cpu_possible_mask __ro_after_init
3097 = {CPU_BITS_ALL};
3098#else
3099struct cpumask __cpu_possible_mask __ro_after_init;
3100#endif
3101EXPORT_SYMBOL(__cpu_possible_mask);
3102
3103struct cpumask __cpu_online_mask __read_mostly;
3104EXPORT_SYMBOL(__cpu_online_mask);
3105
3106struct cpumask __cpu_enabled_mask __read_mostly;
3107EXPORT_SYMBOL(__cpu_enabled_mask);
3108
3109struct cpumask __cpu_present_mask __read_mostly;
3110EXPORT_SYMBOL(__cpu_present_mask);
3111
3112struct cpumask __cpu_active_mask __read_mostly;
3113EXPORT_SYMBOL(__cpu_active_mask);
3114
3115struct cpumask __cpu_dying_mask __read_mostly;
3116EXPORT_SYMBOL(__cpu_dying_mask);
3117
3118atomic_t __num_online_cpus __read_mostly;
3119EXPORT_SYMBOL(__num_online_cpus);
3120
3121void init_cpu_present(const struct cpumask *src)
3122{
3123 cpumask_copy(&__cpu_present_mask, src);
3124}
3125
3126void init_cpu_possible(const struct cpumask *src)
3127{
3128 cpumask_copy(&__cpu_possible_mask, src);
3129}
3130
3131void init_cpu_online(const struct cpumask *src)
3132{
3133 cpumask_copy(&__cpu_online_mask, src);
3134}
3135
3136void set_cpu_online(unsigned int cpu, bool online)
3137{
3138 /*
3139 * atomic_inc/dec() is required to handle the horrid abuse of this
3140 * function by the reboot and kexec code which invoke it from
3141 * IPI/NMI broadcasts when shutting down CPUs. Invocation from
3142 * regular CPU hotplug is properly serialized.
3143 *
3144 * Note, that the fact that __num_online_cpus is of type atomic_t
3145 * does not protect readers which are not serialized against
3146 * concurrent hotplug operations.
3147 */
3148 if (online) {
3149 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
3150 atomic_inc(&__num_online_cpus);
3151 } else {
3152 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
3153 atomic_dec(&__num_online_cpus);
3154 }
3155}
3156
3157/*
3158 * Activate the first processor.
3159 */
3160void __init boot_cpu_init(void)
3161{
3162 int cpu = smp_processor_id();
3163
3164 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
3165 set_cpu_online(cpu, true);
3166 set_cpu_active(cpu, true);
3167 set_cpu_present(cpu, true);
3168 set_cpu_possible(cpu, true);
3169
3170#ifdef CONFIG_SMP
3171 __boot_cpu_id = cpu;
3172#endif
3173}
3174
3175/*
3176 * Must be called _AFTER_ setting up the per_cpu areas
3177 */
3178void __init boot_cpu_hotplug_init(void)
3179{
3180#ifdef CONFIG_SMP
3181 cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
3182 atomic_set(this_cpu_ptr(&cpuhp_state.ap_sync_state), SYNC_STATE_ONLINE);
3183#endif
3184 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
3185 this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
3186}
3187
3188#ifdef CONFIG_CPU_MITIGATIONS
3189/*
3190 * These are used for a global "mitigations=" cmdline option for toggling
3191 * optional CPU mitigations.
3192 */
3193enum cpu_mitigations {
3194 CPU_MITIGATIONS_OFF,
3195 CPU_MITIGATIONS_AUTO,
3196 CPU_MITIGATIONS_AUTO_NOSMT,
3197};
3198
3199static enum cpu_mitigations cpu_mitigations __ro_after_init = CPU_MITIGATIONS_AUTO;
3200
3201static int __init mitigations_parse_cmdline(char *arg)
3202{
3203 if (!strcmp(arg, "off"))
3204 cpu_mitigations = CPU_MITIGATIONS_OFF;
3205 else if (!strcmp(arg, "auto"))
3206 cpu_mitigations = CPU_MITIGATIONS_AUTO;
3207 else if (!strcmp(arg, "auto,nosmt"))
3208 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
3209 else
3210 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
3211 arg);
3212
3213 return 0;
3214}
3215
3216/* mitigations=off */
3217bool cpu_mitigations_off(void)
3218{
3219 return cpu_mitigations == CPU_MITIGATIONS_OFF;
3220}
3221EXPORT_SYMBOL_GPL(cpu_mitigations_off);
3222
3223/* mitigations=auto,nosmt */
3224bool cpu_mitigations_auto_nosmt(void)
3225{
3226 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
3227}
3228EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);
3229#else
3230static int __init mitigations_parse_cmdline(char *arg)
3231{
3232 pr_crit("Kernel compiled without mitigations, ignoring 'mitigations'; system may still be vulnerable\n");
3233 return 0;
3234}
3235#endif
3236early_param("mitigations", mitigations_parse_cmdline);