projects/linux·blob v5.18-rc7 · refs

2723 lines · 65.5 KiB

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