projects/linux·blob v6.7-rc5 · refs

3245 lines · 79.6 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/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 * @cpu: CPU number
58 * @node: Remote CPU node; for multi-instance, do a
59 * single entry callback for install/remove
60 * @last: For multi-instance rollback, remember how far we got
61 * @cb_state: The state for a single callback (install/uninstall)
62 * @result: Result of the operation
63 * @ap_sync_state: State for AP synchronization
64 * @done_up: Signal completion to the issuer of the task for cpu-up
65 * @done_down: Signal completion to the issuer of the task for cpu-down
66 */
67struct cpuhp_cpu_state {
68 enum cpuhp_state state;
69 enum cpuhp_state target;
70 enum cpuhp_state fail;
71#ifdef CONFIG_SMP
72 struct task_struct *thread;
73 bool should_run;
74 bool rollback;
75 bool single;
76 bool bringup;
77 struct hlist_node *node;
78 struct hlist_node *last;
79 enum cpuhp_state cb_state;
80 int result;
81 atomic_t ap_sync_state;
82 struct completion done_up;
83 struct completion done_down;
84#endif
85};
86
87static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
88 .fail = CPUHP_INVALID,
89};
90
91#ifdef CONFIG_SMP
92cpumask_t cpus_booted_once_mask;
93#endif
94
95#if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
96static struct lockdep_map cpuhp_state_up_map =
97 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
98static struct lockdep_map cpuhp_state_down_map =
99 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
100
101
102static inline void cpuhp_lock_acquire(bool bringup)
103{
104 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
105}
106
107static inline void cpuhp_lock_release(bool bringup)
108{
109 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
110}
111#else
112
113static inline void cpuhp_lock_acquire(bool bringup) { }
114static inline void cpuhp_lock_release(bool bringup) { }
115
116#endif
117
118/**
119 * struct cpuhp_step - Hotplug state machine step
120 * @name: Name of the step
121 * @startup: Startup function of the step
122 * @teardown: Teardown function of the step
123 * @cant_stop: Bringup/teardown can't be stopped at this step
124 * @multi_instance: State has multiple instances which get added afterwards
125 */
126struct cpuhp_step {
127 const char *name;
128 union {
129 int (*single)(unsigned int cpu);
130 int (*multi)(unsigned int cpu,
131 struct hlist_node *node);
132 } startup;
133 union {
134 int (*single)(unsigned int cpu);
135 int (*multi)(unsigned int cpu,
136 struct hlist_node *node);
137 } teardown;
138 /* private: */
139 struct hlist_head list;
140 /* public: */
141 bool cant_stop;
142 bool multi_instance;
143};
144
145static DEFINE_MUTEX(cpuhp_state_mutex);
146static struct cpuhp_step cpuhp_hp_states[];
147
148static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
149{
150 return cpuhp_hp_states + state;
151}
152
153static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
154{
155 return bringup ? !step->startup.single : !step->teardown.single;
156}
157
158/**
159 * cpuhp_invoke_callback - Invoke the callbacks for a given state
160 * @cpu: The cpu for which the callback should be invoked
161 * @state: The state to do callbacks for
162 * @bringup: True if the bringup callback should be invoked
163 * @node: For multi-instance, do a single entry callback for install/remove
164 * @lastp: For multi-instance rollback, remember how far we got
165 *
166 * Called from cpu hotplug and from the state register machinery.
167 *
168 * Return: %0 on success or a negative errno code
169 */
170static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
171 bool bringup, struct hlist_node *node,
172 struct hlist_node **lastp)
173{
174 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
175 struct cpuhp_step *step = cpuhp_get_step(state);
176 int (*cbm)(unsigned int cpu, struct hlist_node *node);
177 int (*cb)(unsigned int cpu);
178 int ret, cnt;
179
180 if (st->fail == state) {
181 st->fail = CPUHP_INVALID;
182 return -EAGAIN;
183 }
184
185 if (cpuhp_step_empty(bringup, step)) {
186 WARN_ON_ONCE(1);
187 return 0;
188 }
189
190 if (!step->multi_instance) {
191 WARN_ON_ONCE(lastp && *lastp);
192 cb = bringup ? step->startup.single : step->teardown.single;
193
194 trace_cpuhp_enter(cpu, st->target, state, cb);
195 ret = cb(cpu);
196 trace_cpuhp_exit(cpu, st->state, state, ret);
197 return ret;
198 }
199 cbm = bringup ? step->startup.multi : step->teardown.multi;
200
201 /* Single invocation for instance add/remove */
202 if (node) {
203 WARN_ON_ONCE(lastp && *lastp);
204 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
205 ret = cbm(cpu, node);
206 trace_cpuhp_exit(cpu, st->state, state, ret);
207 return ret;
208 }
209
210 /* State transition. Invoke on all instances */
211 cnt = 0;
212 hlist_for_each(node, &step->list) {
213 if (lastp && node == *lastp)
214 break;
215
216 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
217 ret = cbm(cpu, node);
218 trace_cpuhp_exit(cpu, st->state, state, ret);
219 if (ret) {
220 if (!lastp)
221 goto err;
222
223 *lastp = node;
224 return ret;
225 }
226 cnt++;
227 }
228 if (lastp)
229 *lastp = NULL;
230 return 0;
231err:
232 /* Rollback the instances if one failed */
233 cbm = !bringup ? step->startup.multi : step->teardown.multi;
234 if (!cbm)
235 return ret;
236
237 hlist_for_each(node, &step->list) {
238 if (!cnt--)
239 break;
240
241 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
242 ret = cbm(cpu, node);
243 trace_cpuhp_exit(cpu, st->state, state, ret);
244 /*
245 * Rollback must not fail,
246 */
247 WARN_ON_ONCE(ret);
248 }
249 return ret;
250}
251
252#ifdef CONFIG_SMP
253static bool cpuhp_is_ap_state(enum cpuhp_state state)
254{
255 /*
256 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
257 * purposes as that state is handled explicitly in cpu_down.
258 */
259 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
260}
261
262static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
263{
264 struct completion *done = bringup ? &st->done_up : &st->done_down;
265 wait_for_completion(done);
266}
267
268static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
269{
270 struct completion *done = bringup ? &st->done_up : &st->done_down;
271 complete(done);
272}
273
274/*
275 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
276 */
277static bool cpuhp_is_atomic_state(enum cpuhp_state state)
278{
279 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
280}
281
282/* Synchronization state management */
283enum cpuhp_sync_state {
284 SYNC_STATE_DEAD,
285 SYNC_STATE_KICKED,
286 SYNC_STATE_SHOULD_DIE,
287 SYNC_STATE_ALIVE,
288 SYNC_STATE_SHOULD_ONLINE,
289 SYNC_STATE_ONLINE,
290};
291
292#ifdef CONFIG_HOTPLUG_CORE_SYNC
293/**
294 * cpuhp_ap_update_sync_state - Update synchronization state during bringup/teardown
295 * @state: The synchronization state to set
296 *
297 * No synchronization point. Just update of the synchronization state, but implies
298 * a full barrier so that the AP changes are visible before the control CPU proceeds.
299 */
300static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state)
301{
302 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
303
304 (void)atomic_xchg(st, state);
305}
306
307void __weak arch_cpuhp_sync_state_poll(void) { cpu_relax(); }
308
309static bool cpuhp_wait_for_sync_state(unsigned int cpu, enum cpuhp_sync_state state,
310 enum cpuhp_sync_state next_state)
311{
312 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
313 ktime_t now, end, start = ktime_get();
314 int sync;
315
316 end = start + 10ULL * NSEC_PER_SEC;
317
318 sync = atomic_read(st);
319 while (1) {
320 if (sync == state) {
321 if (!atomic_try_cmpxchg(st, &sync, next_state))
322 continue;
323 return true;
324 }
325
326 now = ktime_get();
327 if (now > end) {
328 /* Timeout. Leave the state unchanged */
329 return false;
330 } else if (now - start < NSEC_PER_MSEC) {
331 /* Poll for one millisecond */
332 arch_cpuhp_sync_state_poll();
333 } else {
334 usleep_range_state(USEC_PER_MSEC, 2 * USEC_PER_MSEC, TASK_UNINTERRUPTIBLE);
335 }
336 sync = atomic_read(st);
337 }
338 return true;
339}
340#else /* CONFIG_HOTPLUG_CORE_SYNC */
341static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state) { }
342#endif /* !CONFIG_HOTPLUG_CORE_SYNC */
343
344#ifdef CONFIG_HOTPLUG_CORE_SYNC_DEAD
345/**
346 * cpuhp_ap_report_dead - Update synchronization state to DEAD
347 *
348 * No synchronization point. Just update of the synchronization state.
349 */
350void cpuhp_ap_report_dead(void)
351{
352 cpuhp_ap_update_sync_state(SYNC_STATE_DEAD);
353}
354
355void __weak arch_cpuhp_cleanup_dead_cpu(unsigned int cpu) { }
356
357/*
358 * Late CPU shutdown synchronization point. Cannot use cpuhp_state::done_down
359 * because the AP cannot issue complete() at this stage.
360 */
361static void cpuhp_bp_sync_dead(unsigned int cpu)
362{
363 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
364 int sync = atomic_read(st);
365
366 do {
367 /* CPU can have reported dead already. Don't overwrite that! */
368 if (sync == SYNC_STATE_DEAD)
369 break;
370 } while (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_SHOULD_DIE));
371
372 if (cpuhp_wait_for_sync_state(cpu, SYNC_STATE_DEAD, SYNC_STATE_DEAD)) {
373 /* CPU reached dead state. Invoke the cleanup function */
374 arch_cpuhp_cleanup_dead_cpu(cpu);
375 return;
376 }
377
378 /* No further action possible. Emit message and give up. */
379 pr_err("CPU%u failed to report dead state\n", cpu);
380}
381#else /* CONFIG_HOTPLUG_CORE_SYNC_DEAD */
382static inline void cpuhp_bp_sync_dead(unsigned int cpu) { }
383#endif /* !CONFIG_HOTPLUG_CORE_SYNC_DEAD */
384
385#ifdef CONFIG_HOTPLUG_CORE_SYNC_FULL
386/**
387 * cpuhp_ap_sync_alive - Synchronize AP with the control CPU once it is alive
388 *
389 * Updates the AP synchronization state to SYNC_STATE_ALIVE and waits
390 * for the BP to release it.
391 */
392void cpuhp_ap_sync_alive(void)
393{
394 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
395
396 cpuhp_ap_update_sync_state(SYNC_STATE_ALIVE);
397
398 /* Wait for the control CPU to release it. */
399 while (atomic_read(st) != SYNC_STATE_SHOULD_ONLINE)
400 cpu_relax();
401}
402
403static bool cpuhp_can_boot_ap(unsigned int cpu)
404{
405 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
406 int sync = atomic_read(st);
407
408again:
409 switch (sync) {
410 case SYNC_STATE_DEAD:
411 /* CPU is properly dead */
412 break;
413 case SYNC_STATE_KICKED:
414 /* CPU did not come up in previous attempt */
415 break;
416 case SYNC_STATE_ALIVE:
417 /* CPU is stuck cpuhp_ap_sync_alive(). */
418 break;
419 default:
420 /* CPU failed to report online or dead and is in limbo state. */
421 return false;
422 }
423
424 /* Prepare for booting */
425 if (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_KICKED))
426 goto again;
427
428 return true;
429}
430
431void __weak arch_cpuhp_cleanup_kick_cpu(unsigned int cpu) { }
432
433/*
434 * Early CPU bringup synchronization point. Cannot use cpuhp_state::done_up
435 * because the AP cannot issue complete() so early in the bringup.
436 */
437static int cpuhp_bp_sync_alive(unsigned int cpu)
438{
439 int ret = 0;
440
441 if (!IS_ENABLED(CONFIG_HOTPLUG_CORE_SYNC_FULL))
442 return 0;
443
444 if (!cpuhp_wait_for_sync_state(cpu, SYNC_STATE_ALIVE, SYNC_STATE_SHOULD_ONLINE)) {
445 pr_err("CPU%u failed to report alive state\n", cpu);
446 ret = -EIO;
447 }
448
449 /* Let the architecture cleanup the kick alive mechanics. */
450 arch_cpuhp_cleanup_kick_cpu(cpu);
451 return ret;
452}
453#else /* CONFIG_HOTPLUG_CORE_SYNC_FULL */
454static inline int cpuhp_bp_sync_alive(unsigned int cpu) { return 0; }
455static inline bool cpuhp_can_boot_ap(unsigned int cpu) { return true; }
456#endif /* !CONFIG_HOTPLUG_CORE_SYNC_FULL */
457
458/* Serializes the updates to cpu_online_mask, cpu_present_mask */
459static DEFINE_MUTEX(cpu_add_remove_lock);
460bool cpuhp_tasks_frozen;
461EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
462
463/*
464 * The following two APIs (cpu_maps_update_begin/done) must be used when
465 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
466 */
467void cpu_maps_update_begin(void)
468{
469 mutex_lock(&cpu_add_remove_lock);
470}
471
472void cpu_maps_update_done(void)
473{
474 mutex_unlock(&cpu_add_remove_lock);
475}
476
477/*
478 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
479 * Should always be manipulated under cpu_add_remove_lock
480 */
481static int cpu_hotplug_disabled;
482
483#ifdef CONFIG_HOTPLUG_CPU
484
485DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
486
487void cpus_read_lock(void)
488{
489 percpu_down_read(&cpu_hotplug_lock);
490}
491EXPORT_SYMBOL_GPL(cpus_read_lock);
492
493int cpus_read_trylock(void)
494{
495 return percpu_down_read_trylock(&cpu_hotplug_lock);
496}
497EXPORT_SYMBOL_GPL(cpus_read_trylock);
498
499void cpus_read_unlock(void)
500{
501 percpu_up_read(&cpu_hotplug_lock);
502}
503EXPORT_SYMBOL_GPL(cpus_read_unlock);
504
505void cpus_write_lock(void)
506{
507 percpu_down_write(&cpu_hotplug_lock);
508}
509
510void cpus_write_unlock(void)
511{
512 percpu_up_write(&cpu_hotplug_lock);
513}
514
515void lockdep_assert_cpus_held(void)
516{
517 /*
518 * We can't have hotplug operations before userspace starts running,
519 * and some init codepaths will knowingly not take the hotplug lock.
520 * This is all valid, so mute lockdep until it makes sense to report
521 * unheld locks.
522 */
523 if (system_state < SYSTEM_RUNNING)
524 return;
525
526 percpu_rwsem_assert_held(&cpu_hotplug_lock);
527}
528
529#ifdef CONFIG_LOCKDEP
530int lockdep_is_cpus_held(void)
531{
532 return percpu_rwsem_is_held(&cpu_hotplug_lock);
533}
534#endif
535
536static void lockdep_acquire_cpus_lock(void)
537{
538 rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
539}
540
541static void lockdep_release_cpus_lock(void)
542{
543 rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
544}
545
546/*
547 * Wait for currently running CPU hotplug operations to complete (if any) and
548 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
549 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
550 * hotplug path before performing hotplug operations. So acquiring that lock
551 * guarantees mutual exclusion from any currently running hotplug operations.
552 */
553void cpu_hotplug_disable(void)
554{
555 cpu_maps_update_begin();
556 cpu_hotplug_disabled++;
557 cpu_maps_update_done();
558}
559EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
560
561static void __cpu_hotplug_enable(void)
562{
563 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
564 return;
565 cpu_hotplug_disabled--;
566}
567
568void cpu_hotplug_enable(void)
569{
570 cpu_maps_update_begin();
571 __cpu_hotplug_enable();
572 cpu_maps_update_done();
573}
574EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
575
576#else
577
578static void lockdep_acquire_cpus_lock(void)
579{
580}
581
582static void lockdep_release_cpus_lock(void)
583{
584}
585
586#endif /* CONFIG_HOTPLUG_CPU */
587
588/*
589 * Architectures that need SMT-specific errata handling during SMT hotplug
590 * should override this.
591 */
592void __weak arch_smt_update(void) { }
593
594#ifdef CONFIG_HOTPLUG_SMT
595
596enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
597static unsigned int cpu_smt_max_threads __ro_after_init;
598unsigned int cpu_smt_num_threads __read_mostly = UINT_MAX;
599
600void __init cpu_smt_disable(bool force)
601{
602 if (!cpu_smt_possible())
603 return;
604
605 if (force) {
606 pr_info("SMT: Force disabled\n");
607 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
608 } else {
609 pr_info("SMT: disabled\n");
610 cpu_smt_control = CPU_SMT_DISABLED;
611 }
612 cpu_smt_num_threads = 1;
613}
614
615/*
616 * The decision whether SMT is supported can only be done after the full
617 * CPU identification. Called from architecture code.
618 */
619void __init cpu_smt_set_num_threads(unsigned int num_threads,
620 unsigned int max_threads)
621{
622 WARN_ON(!num_threads || (num_threads > max_threads));
623
624 if (max_threads == 1)
625 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
626
627 cpu_smt_max_threads = max_threads;
628
629 /*
630 * If SMT has been disabled via the kernel command line or SMT is
631 * not supported, set cpu_smt_num_threads to 1 for consistency.
632 * If enabled, take the architecture requested number of threads
633 * to bring up into account.
634 */
635 if (cpu_smt_control != CPU_SMT_ENABLED)
636 cpu_smt_num_threads = 1;
637 else if (num_threads < cpu_smt_num_threads)
638 cpu_smt_num_threads = num_threads;
639}
640
641static int __init smt_cmdline_disable(char *str)
642{
643 cpu_smt_disable(str && !strcmp(str, "force"));
644 return 0;
645}
646early_param("nosmt", smt_cmdline_disable);
647
648/*
649 * For Archicture supporting partial SMT states check if the thread is allowed.
650 * Otherwise this has already been checked through cpu_smt_max_threads when
651 * setting the SMT level.
652 */
653static inline bool cpu_smt_thread_allowed(unsigned int cpu)
654{
655#ifdef CONFIG_SMT_NUM_THREADS_DYNAMIC
656 return topology_smt_thread_allowed(cpu);
657#else
658 return true;
659#endif
660}
661
662static inline bool cpu_bootable(unsigned int cpu)
663{
664 if (cpu_smt_control == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
665 return true;
666
667 /* All CPUs are bootable if controls are not configured */
668 if (cpu_smt_control == CPU_SMT_NOT_IMPLEMENTED)
669 return true;
670
671 /* All CPUs are bootable if CPU is not SMT capable */
672 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
673 return true;
674
675 if (topology_is_primary_thread(cpu))
676 return true;
677
678 /*
679 * On x86 it's required to boot all logical CPUs at least once so
680 * that the init code can get a chance to set CR4.MCE on each
681 * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
682 * core will shutdown the machine.
683 */
684 return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
685}
686
687/* Returns true if SMT is supported and not forcefully (irreversibly) disabled */
688bool cpu_smt_possible(void)
689{
690 return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
691 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
692}
693EXPORT_SYMBOL_GPL(cpu_smt_possible);
694
695#else
696static inline bool cpu_bootable(unsigned int cpu) { return true; }
697#endif
698
699static inline enum cpuhp_state
700cpuhp_set_state(int cpu, struct cpuhp_cpu_state *st, enum cpuhp_state target)
701{
702 enum cpuhp_state prev_state = st->state;
703 bool bringup = st->state < target;
704
705 st->rollback = false;
706 st->last = NULL;
707
708 st->target = target;
709 st->single = false;
710 st->bringup = bringup;
711 if (cpu_dying(cpu) != !bringup)
712 set_cpu_dying(cpu, !bringup);
713
714 return prev_state;
715}
716
717static inline void
718cpuhp_reset_state(int cpu, struct cpuhp_cpu_state *st,
719 enum cpuhp_state prev_state)
720{
721 bool bringup = !st->bringup;
722
723 st->target = prev_state;
724
725 /*
726 * Already rolling back. No need invert the bringup value or to change
727 * the current state.
728 */
729 if (st->rollback)
730 return;
731
732 st->rollback = true;
733
734 /*
735 * If we have st->last we need to undo partial multi_instance of this
736 * state first. Otherwise start undo at the previous state.
737 */
738 if (!st->last) {
739 if (st->bringup)
740 st->state--;
741 else
742 st->state++;
743 }
744
745 st->bringup = bringup;
746 if (cpu_dying(cpu) != !bringup)
747 set_cpu_dying(cpu, !bringup);
748}
749
750/* Regular hotplug invocation of the AP hotplug thread */
751static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
752{
753 if (!st->single && st->state == st->target)
754 return;
755
756 st->result = 0;
757 /*
758 * Make sure the above stores are visible before should_run becomes
759 * true. Paired with the mb() above in cpuhp_thread_fun()
760 */
761 smp_mb();
762 st->should_run = true;
763 wake_up_process(st->thread);
764 wait_for_ap_thread(st, st->bringup);
765}
766
767static int cpuhp_kick_ap(int cpu, struct cpuhp_cpu_state *st,
768 enum cpuhp_state target)
769{
770 enum cpuhp_state prev_state;
771 int ret;
772
773 prev_state = cpuhp_set_state(cpu, st, target);
774 __cpuhp_kick_ap(st);
775 if ((ret = st->result)) {
776 cpuhp_reset_state(cpu, st, prev_state);
777 __cpuhp_kick_ap(st);
778 }
779
780 return ret;
781}
782
783static int bringup_wait_for_ap_online(unsigned int cpu)
784{
785 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
786
787 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
788 wait_for_ap_thread(st, true);
789 if (WARN_ON_ONCE((!cpu_online(cpu))))
790 return -ECANCELED;
791
792 /* Unpark the hotplug thread of the target cpu */
793 kthread_unpark(st->thread);
794
795 /*
796 * SMT soft disabling on X86 requires to bring the CPU out of the
797 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
798 * CPU marked itself as booted_once in notify_cpu_starting() so the
799 * cpu_bootable() check will now return false if this is not the
800 * primary sibling.
801 */
802 if (!cpu_bootable(cpu))
803 return -ECANCELED;
804 return 0;
805}
806
807#ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
808static int cpuhp_kick_ap_alive(unsigned int cpu)
809{
810 if (!cpuhp_can_boot_ap(cpu))
811 return -EAGAIN;
812
813 return arch_cpuhp_kick_ap_alive(cpu, idle_thread_get(cpu));
814}
815
816static int cpuhp_bringup_ap(unsigned int cpu)
817{
818 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
819 int ret;
820
821 /*
822 * Some architectures have to walk the irq descriptors to
823 * setup the vector space for the cpu which comes online.
824 * Prevent irq alloc/free across the bringup.
825 */
826 irq_lock_sparse();
827
828 ret = cpuhp_bp_sync_alive(cpu);
829 if (ret)
830 goto out_unlock;
831
832 ret = bringup_wait_for_ap_online(cpu);
833 if (ret)
834 goto out_unlock;
835
836 irq_unlock_sparse();
837
838 if (st->target <= CPUHP_AP_ONLINE_IDLE)
839 return 0;
840
841 return cpuhp_kick_ap(cpu, st, st->target);
842
843out_unlock:
844 irq_unlock_sparse();
845 return ret;
846}
847#else
848static int bringup_cpu(unsigned int cpu)
849{
850 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
851 struct task_struct *idle = idle_thread_get(cpu);
852 int ret;
853
854 if (!cpuhp_can_boot_ap(cpu))
855 return -EAGAIN;
856
857 /*
858 * Some architectures have to walk the irq descriptors to
859 * setup the vector space for the cpu which comes online.
860 *
861 * Prevent irq alloc/free across the bringup by acquiring the
862 * sparse irq lock. Hold it until the upcoming CPU completes the
863 * startup in cpuhp_online_idle() which allows to avoid
864 * intermediate synchronization points in the architecture code.
865 */
866 irq_lock_sparse();
867
868 ret = __cpu_up(cpu, idle);
869 if (ret)
870 goto out_unlock;
871
872 ret = cpuhp_bp_sync_alive(cpu);
873 if (ret)
874 goto out_unlock;
875
876 ret = bringup_wait_for_ap_online(cpu);
877 if (ret)
878 goto out_unlock;
879
880 irq_unlock_sparse();
881
882 if (st->target <= CPUHP_AP_ONLINE_IDLE)
883 return 0;
884
885 return cpuhp_kick_ap(cpu, st, st->target);
886
887out_unlock:
888 irq_unlock_sparse();
889 return ret;
890}
891#endif
892
893static int finish_cpu(unsigned int cpu)
894{
895 struct task_struct *idle = idle_thread_get(cpu);
896 struct mm_struct *mm = idle->active_mm;
897
898 /*
899 * idle_task_exit() will have switched to &init_mm, now
900 * clean up any remaining active_mm state.
901 */
902 if (mm != &init_mm)
903 idle->active_mm = &init_mm;
904 mmdrop_lazy_tlb(mm);
905 return 0;
906}
907
908/*
909 * Hotplug state machine related functions
910 */
911
912/*
913 * Get the next state to run. Empty ones will be skipped. Returns true if a
914 * state must be run.
915 *
916 * st->state will be modified ahead of time, to match state_to_run, as if it
917 * has already ran.
918 */
919static bool cpuhp_next_state(bool bringup,
920 enum cpuhp_state *state_to_run,
921 struct cpuhp_cpu_state *st,
922 enum cpuhp_state target)
923{
924 do {
925 if (bringup) {
926 if (st->state >= target)
927 return false;
928
929 *state_to_run = ++st->state;
930 } else {
931 if (st->state <= target)
932 return false;
933
934 *state_to_run = st->state--;
935 }
936
937 if (!cpuhp_step_empty(bringup, cpuhp_get_step(*state_to_run)))
938 break;
939 } while (true);
940
941 return true;
942}
943
944static int __cpuhp_invoke_callback_range(bool bringup,
945 unsigned int cpu,
946 struct cpuhp_cpu_state *st,
947 enum cpuhp_state target,
948 bool nofail)
949{
950 enum cpuhp_state state;
951 int ret = 0;
952
953 while (cpuhp_next_state(bringup, &state, st, target)) {
954 int err;
955
956 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
957 if (!err)
958 continue;
959
960 if (nofail) {
961 pr_warn("CPU %u %s state %s (%d) failed (%d)\n",
962 cpu, bringup ? "UP" : "DOWN",
963 cpuhp_get_step(st->state)->name,
964 st->state, err);
965 ret = -1;
966 } else {
967 ret = err;
968 break;
969 }
970 }
971
972 return ret;
973}
974
975static inline int cpuhp_invoke_callback_range(bool bringup,
976 unsigned int cpu,
977 struct cpuhp_cpu_state *st,
978 enum cpuhp_state target)
979{
980 return __cpuhp_invoke_callback_range(bringup, cpu, st, target, false);
981}
982
983static inline void cpuhp_invoke_callback_range_nofail(bool bringup,
984 unsigned int cpu,
985 struct cpuhp_cpu_state *st,
986 enum cpuhp_state target)
987{
988 __cpuhp_invoke_callback_range(bringup, cpu, st, target, true);
989}
990
991static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
992{
993 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
994 return true;
995 /*
996 * When CPU hotplug is disabled, then taking the CPU down is not
997 * possible because takedown_cpu() and the architecture and
998 * subsystem specific mechanisms are not available. So the CPU
999 * which would be completely unplugged again needs to stay around
1000 * in the current state.
1001 */
1002 return st->state <= CPUHP_BRINGUP_CPU;
1003}
1004
1005static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1006 enum cpuhp_state target)
1007{
1008 enum cpuhp_state prev_state = st->state;
1009 int ret = 0;
1010
1011 ret = cpuhp_invoke_callback_range(true, cpu, st, target);
1012 if (ret) {
1013 pr_debug("CPU UP failed (%d) CPU %u state %s (%d)\n",
1014 ret, cpu, cpuhp_get_step(st->state)->name,
1015 st->state);
1016
1017 cpuhp_reset_state(cpu, st, prev_state);
1018 if (can_rollback_cpu(st))
1019 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
1020 prev_state));
1021 }
1022 return ret;
1023}
1024
1025/*
1026 * The cpu hotplug threads manage the bringup and teardown of the cpus
1027 */
1028static int cpuhp_should_run(unsigned int cpu)
1029{
1030 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1031
1032 return st->should_run;
1033}
1034
1035/*
1036 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
1037 * callbacks when a state gets [un]installed at runtime.
1038 *
1039 * Each invocation of this function by the smpboot thread does a single AP
1040 * state callback.
1041 *
1042 * It has 3 modes of operation:
1043 * - single: runs st->cb_state
1044 * - up: runs ++st->state, while st->state < st->target
1045 * - down: runs st->state--, while st->state > st->target
1046 *
1047 * When complete or on error, should_run is cleared and the completion is fired.
1048 */
1049static void cpuhp_thread_fun(unsigned int cpu)
1050{
1051 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1052 bool bringup = st->bringup;
1053 enum cpuhp_state state;
1054
1055 if (WARN_ON_ONCE(!st->should_run))
1056 return;
1057
1058 /*
1059 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
1060 * that if we see ->should_run we also see the rest of the state.
1061 */
1062 smp_mb();
1063
1064 /*
1065 * The BP holds the hotplug lock, but we're now running on the AP,
1066 * ensure that anybody asserting the lock is held, will actually find
1067 * it so.
1068 */
1069 lockdep_acquire_cpus_lock();
1070 cpuhp_lock_acquire(bringup);
1071
1072 if (st->single) {
1073 state = st->cb_state;
1074 st->should_run = false;
1075 } else {
1076 st->should_run = cpuhp_next_state(bringup, &state, st, st->target);
1077 if (!st->should_run)
1078 goto end;
1079 }
1080
1081 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
1082
1083 if (cpuhp_is_atomic_state(state)) {
1084 local_irq_disable();
1085 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1086 local_irq_enable();
1087
1088 /*
1089 * STARTING/DYING must not fail!
1090 */
1091 WARN_ON_ONCE(st->result);
1092 } else {
1093 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1094 }
1095
1096 if (st->result) {
1097 /*
1098 * If we fail on a rollback, we're up a creek without no
1099 * paddle, no way forward, no way back. We loose, thanks for
1100 * playing.
1101 */
1102 WARN_ON_ONCE(st->rollback);
1103 st->should_run = false;
1104 }
1105
1106end:
1107 cpuhp_lock_release(bringup);
1108 lockdep_release_cpus_lock();
1109
1110 if (!st->should_run)
1111 complete_ap_thread(st, bringup);
1112}
1113
1114/* Invoke a single callback on a remote cpu */
1115static int
1116cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
1117 struct hlist_node *node)
1118{
1119 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1120 int ret;
1121
1122 if (!cpu_online(cpu))
1123 return 0;
1124
1125 cpuhp_lock_acquire(false);
1126 cpuhp_lock_release(false);
1127
1128 cpuhp_lock_acquire(true);
1129 cpuhp_lock_release(true);
1130
1131 /*
1132 * If we are up and running, use the hotplug thread. For early calls
1133 * we invoke the thread function directly.
1134 */
1135 if (!st->thread)
1136 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1137
1138 st->rollback = false;
1139 st->last = NULL;
1140
1141 st->node = node;
1142 st->bringup = bringup;
1143 st->cb_state = state;
1144 st->single = true;
1145
1146 __cpuhp_kick_ap(st);
1147
1148 /*
1149 * If we failed and did a partial, do a rollback.
1150 */
1151 if ((ret = st->result) && st->last) {
1152 st->rollback = true;
1153 st->bringup = !bringup;
1154
1155 __cpuhp_kick_ap(st);
1156 }
1157
1158 /*
1159 * Clean up the leftovers so the next hotplug operation wont use stale
1160 * data.
1161 */
1162 st->node = st->last = NULL;
1163 return ret;
1164}
1165
1166static int cpuhp_kick_ap_work(unsigned int cpu)
1167{
1168 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1169 enum cpuhp_state prev_state = st->state;
1170 int ret;
1171
1172 cpuhp_lock_acquire(false);
1173 cpuhp_lock_release(false);
1174
1175 cpuhp_lock_acquire(true);
1176 cpuhp_lock_release(true);
1177
1178 trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
1179 ret = cpuhp_kick_ap(cpu, st, st->target);
1180 trace_cpuhp_exit(cpu, st->state, prev_state, ret);
1181
1182 return ret;
1183}
1184
1185static struct smp_hotplug_thread cpuhp_threads = {
1186 .store = &cpuhp_state.thread,
1187 .thread_should_run = cpuhp_should_run,
1188 .thread_fn = cpuhp_thread_fun,
1189 .thread_comm = "cpuhp/%u",
1190 .selfparking = true,
1191};
1192
1193static __init void cpuhp_init_state(void)
1194{
1195 struct cpuhp_cpu_state *st;
1196 int cpu;
1197
1198 for_each_possible_cpu(cpu) {
1199 st = per_cpu_ptr(&cpuhp_state, cpu);
1200 init_completion(&st->done_up);
1201 init_completion(&st->done_down);
1202 }
1203}
1204
1205void __init cpuhp_threads_init(void)
1206{
1207 cpuhp_init_state();
1208 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
1209 kthread_unpark(this_cpu_read(cpuhp_state.thread));
1210}
1211
1212/*
1213 *
1214 * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
1215 * protected region.
1216 *
1217 * The operation is still serialized against concurrent CPU hotplug via
1218 * cpu_add_remove_lock, i.e. CPU map protection. But it is _not_
1219 * serialized against other hotplug related activity like adding or
1220 * removing of state callbacks and state instances, which invoke either the
1221 * startup or the teardown callback of the affected state.
1222 *
1223 * This is required for subsystems which are unfixable vs. CPU hotplug and
1224 * evade lock inversion problems by scheduling work which has to be
1225 * completed _before_ cpu_up()/_cpu_down() returns.
1226 *
1227 * Don't even think about adding anything to this for any new code or even
1228 * drivers. It's only purpose is to keep existing lock order trainwrecks
1229 * working.
1230 *
1231 * For cpu_down() there might be valid reasons to finish cleanups which are
1232 * not required to be done under cpu_hotplug_lock, but that's a different
1233 * story and would be not invoked via this.
1234 */
1235static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
1236{
1237 /*
1238 * cpusets delegate hotplug operations to a worker to "solve" the
1239 * lock order problems. Wait for the worker, but only if tasks are
1240 * _not_ frozen (suspend, hibernate) as that would wait forever.
1241 *
1242 * The wait is required because otherwise the hotplug operation
1243 * returns with inconsistent state, which could even be observed in
1244 * user space when a new CPU is brought up. The CPU plug uevent
1245 * would be delivered and user space reacting on it would fail to
1246 * move tasks to the newly plugged CPU up to the point where the
1247 * work has finished because up to that point the newly plugged CPU
1248 * is not assignable in cpusets/cgroups. On unplug that's not
1249 * necessarily a visible issue, but it is still inconsistent state,
1250 * which is the real problem which needs to be "fixed". This can't
1251 * prevent the transient state between scheduling the work and
1252 * returning from waiting for it.
1253 */
1254 if (!tasks_frozen)
1255 cpuset_wait_for_hotplug();
1256}
1257
1258#ifdef CONFIG_HOTPLUG_CPU
1259#ifndef arch_clear_mm_cpumask_cpu
1260#define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
1261#endif
1262
1263/**
1264 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
1265 * @cpu: a CPU id
1266 *
1267 * This function walks all processes, finds a valid mm struct for each one and
1268 * then clears a corresponding bit in mm's cpumask. While this all sounds
1269 * trivial, there are various non-obvious corner cases, which this function
1270 * tries to solve in a safe manner.
1271 *
1272 * Also note that the function uses a somewhat relaxed locking scheme, so it may
1273 * be called only for an already offlined CPU.
1274 */
1275void clear_tasks_mm_cpumask(int cpu)
1276{
1277 struct task_struct *p;
1278
1279 /*
1280 * This function is called after the cpu is taken down and marked
1281 * offline, so its not like new tasks will ever get this cpu set in
1282 * their mm mask. -- Peter Zijlstra
1283 * Thus, we may use rcu_read_lock() here, instead of grabbing
1284 * full-fledged tasklist_lock.
1285 */
1286 WARN_ON(cpu_online(cpu));
1287 rcu_read_lock();
1288 for_each_process(p) {
1289 struct task_struct *t;
1290
1291 /*
1292 * Main thread might exit, but other threads may still have
1293 * a valid mm. Find one.
1294 */
1295 t = find_lock_task_mm(p);
1296 if (!t)
1297 continue;
1298 arch_clear_mm_cpumask_cpu(cpu, t->mm);
1299 task_unlock(t);
1300 }
1301 rcu_read_unlock();
1302}
1303
1304/* Take this CPU down. */
1305static int take_cpu_down(void *_param)
1306{
1307 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1308 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
1309 int err, cpu = smp_processor_id();
1310
1311 /* Ensure this CPU doesn't handle any more interrupts. */
1312 err = __cpu_disable();
1313 if (err < 0)
1314 return err;
1315
1316 /*
1317 * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
1318 * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
1319 */
1320 WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
1321
1322 /*
1323 * Invoke the former CPU_DYING callbacks. DYING must not fail!
1324 */
1325 cpuhp_invoke_callback_range_nofail(false, cpu, st, target);
1326
1327 /* Give up timekeeping duties */
1328 tick_handover_do_timer();
1329 /* Remove CPU from timer broadcasting */
1330 tick_offline_cpu(cpu);
1331 /* Park the stopper thread */
1332 stop_machine_park(cpu);
1333 return 0;
1334}
1335
1336static int takedown_cpu(unsigned int cpu)
1337{
1338 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1339 int err;
1340
1341 /* Park the smpboot threads */
1342 kthread_park(st->thread);
1343
1344 /*
1345 * Prevent irq alloc/free while the dying cpu reorganizes the
1346 * interrupt affinities.
1347 */
1348 irq_lock_sparse();
1349
1350 /*
1351 * So now all preempt/rcu users must observe !cpu_active().
1352 */
1353 err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
1354 if (err) {
1355 /* CPU refused to die */
1356 irq_unlock_sparse();
1357 /* Unpark the hotplug thread so we can rollback there */
1358 kthread_unpark(st->thread);
1359 return err;
1360 }
1361 BUG_ON(cpu_online(cpu));
1362
1363 /*
1364 * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1365 * all runnable tasks from the CPU, there's only the idle task left now
1366 * that the migration thread is done doing the stop_machine thing.
1367 *
1368 * Wait for the stop thread to go away.
1369 */
1370 wait_for_ap_thread(st, false);
1371 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1372
1373 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1374 irq_unlock_sparse();
1375
1376 hotplug_cpu__broadcast_tick_pull(cpu);
1377 /* This actually kills the CPU. */
1378 __cpu_die(cpu);
1379
1380 cpuhp_bp_sync_dead(cpu);
1381
1382 tick_cleanup_dead_cpu(cpu);
1383
1384 /*
1385 * Callbacks must be re-integrated right away to the RCU state machine.
1386 * Otherwise an RCU callback could block a further teardown function
1387 * waiting for its completion.
1388 */
1389 rcutree_migrate_callbacks(cpu);
1390
1391 return 0;
1392}
1393
1394static void cpuhp_complete_idle_dead(void *arg)
1395{
1396 struct cpuhp_cpu_state *st = arg;
1397
1398 complete_ap_thread(st, false);
1399}
1400
1401void cpuhp_report_idle_dead(void)
1402{
1403 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1404
1405 BUG_ON(st->state != CPUHP_AP_OFFLINE);
1406 rcutree_report_cpu_dead();
1407 st->state = CPUHP_AP_IDLE_DEAD;
1408 /*
1409 * We cannot call complete after rcutree_report_cpu_dead() so we delegate it
1410 * to an online cpu.
1411 */
1412 smp_call_function_single(cpumask_first(cpu_online_mask),
1413 cpuhp_complete_idle_dead, st, 0);
1414}
1415
1416static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1417 enum cpuhp_state target)
1418{
1419 enum cpuhp_state prev_state = st->state;
1420 int ret = 0;
1421
1422 ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1423 if (ret) {
1424 pr_debug("CPU DOWN failed (%d) CPU %u state %s (%d)\n",
1425 ret, cpu, cpuhp_get_step(st->state)->name,
1426 st->state);
1427
1428 cpuhp_reset_state(cpu, st, prev_state);
1429
1430 if (st->state < prev_state)
1431 WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1432 prev_state));
1433 }
1434
1435 return ret;
1436}
1437
1438/* Requires cpu_add_remove_lock to be held */
1439static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1440 enum cpuhp_state target)
1441{
1442 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1443 int prev_state, ret = 0;
1444
1445 if (num_online_cpus() == 1)
1446 return -EBUSY;
1447
1448 if (!cpu_present(cpu))
1449 return -EINVAL;
1450
1451 cpus_write_lock();
1452
1453 cpuhp_tasks_frozen = tasks_frozen;
1454
1455 prev_state = cpuhp_set_state(cpu, st, target);
1456 /*
1457 * If the current CPU state is in the range of the AP hotplug thread,
1458 * then we need to kick the thread.
1459 */
1460 if (st->state > CPUHP_TEARDOWN_CPU) {
1461 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1462 ret = cpuhp_kick_ap_work(cpu);
1463 /*
1464 * The AP side has done the error rollback already. Just
1465 * return the error code..
1466 */
1467 if (ret)
1468 goto out;
1469
1470 /*
1471 * We might have stopped still in the range of the AP hotplug
1472 * thread. Nothing to do anymore.
1473 */
1474 if (st->state > CPUHP_TEARDOWN_CPU)
1475 goto out;
1476
1477 st->target = target;
1478 }
1479 /*
1480 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1481 * to do the further cleanups.
1482 */
1483 ret = cpuhp_down_callbacks(cpu, st, target);
1484 if (ret && st->state < prev_state) {
1485 if (st->state == CPUHP_TEARDOWN_CPU) {
1486 cpuhp_reset_state(cpu, st, prev_state);
1487 __cpuhp_kick_ap(st);
1488 } else {
1489 WARN(1, "DEAD callback error for CPU%d", cpu);
1490 }
1491 }
1492
1493out:
1494 cpus_write_unlock();
1495 /*
1496 * Do post unplug cleanup. This is still protected against
1497 * concurrent CPU hotplug via cpu_add_remove_lock.
1498 */
1499 lockup_detector_cleanup();
1500 arch_smt_update();
1501 cpu_up_down_serialize_trainwrecks(tasks_frozen);
1502 return ret;
1503}
1504
1505struct cpu_down_work {
1506 unsigned int cpu;
1507 enum cpuhp_state target;
1508};
1509
1510static long __cpu_down_maps_locked(void *arg)
1511{
1512 struct cpu_down_work *work = arg;
1513
1514 return _cpu_down(work->cpu, 0, work->target);
1515}
1516
1517static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1518{
1519 struct cpu_down_work work = { .cpu = cpu, .target = target, };
1520
1521 /*
1522 * If the platform does not support hotplug, report it explicitly to
1523 * differentiate it from a transient offlining failure.
1524 */
1525 if (cc_platform_has(CC_ATTR_HOTPLUG_DISABLED))
1526 return -EOPNOTSUPP;
1527 if (cpu_hotplug_disabled)
1528 return -EBUSY;
1529
1530 /*
1531 * Ensure that the control task does not run on the to be offlined
1532 * CPU to prevent a deadlock against cfs_b->period_timer.
1533 * Also keep at least one housekeeping cpu onlined to avoid generating
1534 * an empty sched_domain span.
1535 */
1536 for_each_cpu_and(cpu, cpu_online_mask, housekeeping_cpumask(HK_TYPE_DOMAIN)) {
1537 if (cpu != work.cpu)
1538 return work_on_cpu(cpu, __cpu_down_maps_locked, &work);
1539 }
1540 return -EBUSY;
1541}
1542
1543static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1544{
1545 int err;
1546
1547 cpu_maps_update_begin();
1548 err = cpu_down_maps_locked(cpu, target);
1549 cpu_maps_update_done();
1550 return err;
1551}
1552
1553/**
1554 * cpu_device_down - Bring down a cpu device
1555 * @dev: Pointer to the cpu device to offline
1556 *
1557 * This function is meant to be used by device core cpu subsystem only.
1558 *
1559 * Other subsystems should use remove_cpu() instead.
1560 *
1561 * Return: %0 on success or a negative errno code
1562 */
1563int cpu_device_down(struct device *dev)
1564{
1565 return cpu_down(dev->id, CPUHP_OFFLINE);
1566}
1567
1568int remove_cpu(unsigned int cpu)
1569{
1570 int ret;
1571
1572 lock_device_hotplug();
1573 ret = device_offline(get_cpu_device(cpu));
1574 unlock_device_hotplug();
1575
1576 return ret;
1577}
1578EXPORT_SYMBOL_GPL(remove_cpu);
1579
1580void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1581{
1582 unsigned int cpu;
1583 int error;
1584
1585 cpu_maps_update_begin();
1586
1587 /*
1588 * Make certain the cpu I'm about to reboot on is online.
1589 *
1590 * This is inline to what migrate_to_reboot_cpu() already do.
1591 */
1592 if (!cpu_online(primary_cpu))
1593 primary_cpu = cpumask_first(cpu_online_mask);
1594
1595 for_each_online_cpu(cpu) {
1596 if (cpu == primary_cpu)
1597 continue;
1598
1599 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1600 if (error) {
1601 pr_err("Failed to offline CPU%d - error=%d",
1602 cpu, error);
1603 break;
1604 }
1605 }
1606
1607 /*
1608 * Ensure all but the reboot CPU are offline.
1609 */
1610 BUG_ON(num_online_cpus() > 1);
1611
1612 /*
1613 * Make sure the CPUs won't be enabled by someone else after this
1614 * point. Kexec will reboot to a new kernel shortly resetting
1615 * everything along the way.
1616 */
1617 cpu_hotplug_disabled++;
1618
1619 cpu_maps_update_done();
1620}
1621
1622#else
1623#define takedown_cpu NULL
1624#endif /*CONFIG_HOTPLUG_CPU*/
1625
1626/**
1627 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1628 * @cpu: cpu that just started
1629 *
1630 * It must be called by the arch code on the new cpu, before the new cpu
1631 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1632 */
1633void notify_cpu_starting(unsigned int cpu)
1634{
1635 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1636 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1637
1638 rcutree_report_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1639 cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1640
1641 /*
1642 * STARTING must not fail!
1643 */
1644 cpuhp_invoke_callback_range_nofail(true, cpu, st, target);
1645}
1646
1647/*
1648 * Called from the idle task. Wake up the controlling task which brings the
1649 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1650 * online bringup to the hotplug thread.
1651 */
1652void cpuhp_online_idle(enum cpuhp_state state)
1653{
1654 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1655
1656 /* Happens for the boot cpu */
1657 if (state != CPUHP_AP_ONLINE_IDLE)
1658 return;
1659
1660 cpuhp_ap_update_sync_state(SYNC_STATE_ONLINE);
1661
1662 /*
1663 * Unpark the stopper thread before we start the idle loop (and start
1664 * scheduling); this ensures the stopper task is always available.
1665 */
1666 stop_machine_unpark(smp_processor_id());
1667
1668 st->state = CPUHP_AP_ONLINE_IDLE;
1669 complete_ap_thread(st, true);
1670}
1671
1672/* Requires cpu_add_remove_lock to be held */
1673static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1674{
1675 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1676 struct task_struct *idle;
1677 int ret = 0;
1678
1679 cpus_write_lock();
1680
1681 if (!cpu_present(cpu)) {
1682 ret = -EINVAL;
1683 goto out;
1684 }
1685
1686 /*
1687 * The caller of cpu_up() might have raced with another
1688 * caller. Nothing to do.
1689 */
1690 if (st->state >= target)
1691 goto out;
1692
1693 if (st->state == CPUHP_OFFLINE) {
1694 /* Let it fail before we try to bring the cpu up */
1695 idle = idle_thread_get(cpu);
1696 if (IS_ERR(idle)) {
1697 ret = PTR_ERR(idle);
1698 goto out;
1699 }
1700
1701 /*
1702 * Reset stale stack state from the last time this CPU was online.
1703 */
1704 scs_task_reset(idle);
1705 kasan_unpoison_task_stack(idle);
1706 }
1707
1708 cpuhp_tasks_frozen = tasks_frozen;
1709
1710 cpuhp_set_state(cpu, st, target);
1711 /*
1712 * If the current CPU state is in the range of the AP hotplug thread,
1713 * then we need to kick the thread once more.
1714 */
1715 if (st->state > CPUHP_BRINGUP_CPU) {
1716 ret = cpuhp_kick_ap_work(cpu);
1717 /*
1718 * The AP side has done the error rollback already. Just
1719 * return the error code..
1720 */
1721 if (ret)
1722 goto out;
1723 }
1724
1725 /*
1726 * Try to reach the target state. We max out on the BP at
1727 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1728 * responsible for bringing it up to the target state.
1729 */
1730 target = min((int)target, CPUHP_BRINGUP_CPU);
1731 ret = cpuhp_up_callbacks(cpu, st, target);
1732out:
1733 cpus_write_unlock();
1734 arch_smt_update();
1735 cpu_up_down_serialize_trainwrecks(tasks_frozen);
1736 return ret;
1737}
1738
1739static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1740{
1741 int err = 0;
1742
1743 if (!cpu_possible(cpu)) {
1744 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1745 cpu);
1746 return -EINVAL;
1747 }
1748
1749 err = try_online_node(cpu_to_node(cpu));
1750 if (err)
1751 return err;
1752
1753 cpu_maps_update_begin();
1754
1755 if (cpu_hotplug_disabled) {
1756 err = -EBUSY;
1757 goto out;
1758 }
1759 if (!cpu_bootable(cpu)) {
1760 err = -EPERM;
1761 goto out;
1762 }
1763
1764 err = _cpu_up(cpu, 0, target);
1765out:
1766 cpu_maps_update_done();
1767 return err;
1768}
1769
1770/**
1771 * cpu_device_up - Bring up a cpu device
1772 * @dev: Pointer to the cpu device to online
1773 *
1774 * This function is meant to be used by device core cpu subsystem only.
1775 *
1776 * Other subsystems should use add_cpu() instead.
1777 *
1778 * Return: %0 on success or a negative errno code
1779 */
1780int cpu_device_up(struct device *dev)
1781{
1782 return cpu_up(dev->id, CPUHP_ONLINE);
1783}
1784
1785int add_cpu(unsigned int cpu)
1786{
1787 int ret;
1788
1789 lock_device_hotplug();
1790 ret = device_online(get_cpu_device(cpu));
1791 unlock_device_hotplug();
1792
1793 return ret;
1794}
1795EXPORT_SYMBOL_GPL(add_cpu);
1796
1797/**
1798 * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1799 * @sleep_cpu: The cpu we hibernated on and should be brought up.
1800 *
1801 * On some architectures like arm64, we can hibernate on any CPU, but on
1802 * wake up the CPU we hibernated on might be offline as a side effect of
1803 * using maxcpus= for example.
1804 *
1805 * Return: %0 on success or a negative errno code
1806 */
1807int bringup_hibernate_cpu(unsigned int sleep_cpu)
1808{
1809 int ret;
1810
1811 if (!cpu_online(sleep_cpu)) {
1812 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1813 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1814 if (ret) {
1815 pr_err("Failed to bring hibernate-CPU up!\n");
1816 return ret;
1817 }
1818 }
1819 return 0;
1820}
1821
1822static void __init cpuhp_bringup_mask(const struct cpumask *mask, unsigned int ncpus,
1823 enum cpuhp_state target)
1824{
1825 unsigned int cpu;
1826
1827 for_each_cpu(cpu, mask) {
1828 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1829
1830 if (cpu_up(cpu, target) && can_rollback_cpu(st)) {
1831 /*
1832 * If this failed then cpu_up() might have only
1833 * rolled back to CPUHP_BP_KICK_AP for the final
1834 * online. Clean it up. NOOP if already rolled back.
1835 */
1836 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st, CPUHP_OFFLINE));
1837 }
1838
1839 if (!--ncpus)
1840 break;
1841 }
1842}
1843
1844#ifdef CONFIG_HOTPLUG_PARALLEL
1845static bool __cpuhp_parallel_bringup __ro_after_init = true;
1846
1847static int __init parallel_bringup_parse_param(char *arg)
1848{
1849 return kstrtobool(arg, &__cpuhp_parallel_bringup);
1850}
1851early_param("cpuhp.parallel", parallel_bringup_parse_param);
1852
1853static inline bool cpuhp_smt_aware(void)
1854{
1855 return cpu_smt_max_threads > 1;
1856}
1857
1858static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1859{
1860 return cpu_primary_thread_mask;
1861}
1862
1863/*
1864 * On architectures which have enabled parallel bringup this invokes all BP
1865 * prepare states for each of the to be onlined APs first. The last state
1866 * sends the startup IPI to the APs. The APs proceed through the low level
1867 * bringup code in parallel and then wait for the control CPU to release
1868 * them one by one for the final onlining procedure.
1869 *
1870 * This avoids waiting for each AP to respond to the startup IPI in
1871 * CPUHP_BRINGUP_CPU.
1872 */
1873static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
1874{
1875 const struct cpumask *mask = cpu_present_mask;
1876
1877 if (__cpuhp_parallel_bringup)
1878 __cpuhp_parallel_bringup = arch_cpuhp_init_parallel_bringup();
1879 if (!__cpuhp_parallel_bringup)
1880 return false;
1881
1882 if (cpuhp_smt_aware()) {
1883 const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
1884 static struct cpumask tmp_mask __initdata;
1885
1886 /*
1887 * X86 requires to prevent that SMT siblings stopped while
1888 * the primary thread does a microcode update for various
1889 * reasons. Bring the primary threads up first.
1890 */
1891 cpumask_and(&tmp_mask, mask, pmask);
1892 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_BP_KICK_AP);
1893 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_ONLINE);
1894 /* Account for the online CPUs */
1895 ncpus -= num_online_cpus();
1896 if (!ncpus)
1897 return true;
1898 /* Create the mask for secondary CPUs */
1899 cpumask_andnot(&tmp_mask, mask, pmask);
1900 mask = &tmp_mask;
1901 }
1902
1903 /* Bring the not-yet started CPUs up */
1904 cpuhp_bringup_mask(mask, ncpus, CPUHP_BP_KICK_AP);
1905 cpuhp_bringup_mask(mask, ncpus, CPUHP_ONLINE);
1906 return true;
1907}
1908#else
1909static inline bool cpuhp_bringup_cpus_parallel(unsigned int ncpus) { return false; }
1910#endif /* CONFIG_HOTPLUG_PARALLEL */
1911
1912void __init bringup_nonboot_cpus(unsigned int setup_max_cpus)
1913{
1914 /* Try parallel bringup optimization if enabled */
1915 if (cpuhp_bringup_cpus_parallel(setup_max_cpus))
1916 return;
1917
1918 /* Full per CPU serialized bringup */
1919 cpuhp_bringup_mask(cpu_present_mask, setup_max_cpus, CPUHP_ONLINE);
1920}
1921
1922#ifdef CONFIG_PM_SLEEP_SMP
1923static cpumask_var_t frozen_cpus;
1924
1925int freeze_secondary_cpus(int primary)
1926{
1927 int cpu, error = 0;
1928
1929 cpu_maps_update_begin();
1930 if (primary == -1) {
1931 primary = cpumask_first(cpu_online_mask);
1932 if (!housekeeping_cpu(primary, HK_TYPE_TIMER))
1933 primary = housekeeping_any_cpu(HK_TYPE_TIMER);
1934 } else {
1935 if (!cpu_online(primary))
1936 primary = cpumask_first(cpu_online_mask);
1937 }
1938
1939 /*
1940 * We take down all of the non-boot CPUs in one shot to avoid races
1941 * with the userspace trying to use the CPU hotplug at the same time
1942 */
1943 cpumask_clear(frozen_cpus);
1944
1945 pr_info("Disabling non-boot CPUs ...\n");
1946 for_each_online_cpu(cpu) {
1947 if (cpu == primary)
1948 continue;
1949
1950 if (pm_wakeup_pending()) {
1951 pr_info("Wakeup pending. Abort CPU freeze\n");
1952 error = -EBUSY;
1953 break;
1954 }
1955
1956 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1957 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1958 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1959 if (!error)
1960 cpumask_set_cpu(cpu, frozen_cpus);
1961 else {
1962 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1963 break;
1964 }
1965 }
1966
1967 if (!error)
1968 BUG_ON(num_online_cpus() > 1);
1969 else
1970 pr_err("Non-boot CPUs are not disabled\n");
1971
1972 /*
1973 * Make sure the CPUs won't be enabled by someone else. We need to do
1974 * this even in case of failure as all freeze_secondary_cpus() users are
1975 * supposed to do thaw_secondary_cpus() on the failure path.
1976 */
1977 cpu_hotplug_disabled++;
1978
1979 cpu_maps_update_done();
1980 return error;
1981}
1982
1983void __weak arch_thaw_secondary_cpus_begin(void)
1984{
1985}
1986
1987void __weak arch_thaw_secondary_cpus_end(void)
1988{
1989}
1990
1991void thaw_secondary_cpus(void)
1992{
1993 int cpu, error;
1994
1995 /* Allow everyone to use the CPU hotplug again */
1996 cpu_maps_update_begin();
1997 __cpu_hotplug_enable();
1998 if (cpumask_empty(frozen_cpus))
1999 goto out;
2000
2001 pr_info("Enabling non-boot CPUs ...\n");
2002
2003 arch_thaw_secondary_cpus_begin();
2004
2005 for_each_cpu(cpu, frozen_cpus) {
2006 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
2007 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
2008 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
2009 if (!error) {
2010 pr_info("CPU%d is up\n", cpu);
2011 continue;
2012 }
2013 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
2014 }
2015
2016 arch_thaw_secondary_cpus_end();
2017
2018 cpumask_clear(frozen_cpus);
2019out:
2020 cpu_maps_update_done();
2021}
2022
2023static int __init alloc_frozen_cpus(void)
2024{
2025 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
2026 return -ENOMEM;
2027 return 0;
2028}
2029core_initcall(alloc_frozen_cpus);
2030
2031/*
2032 * When callbacks for CPU hotplug notifications are being executed, we must
2033 * ensure that the state of the system with respect to the tasks being frozen
2034 * or not, as reported by the notification, remains unchanged *throughout the
2035 * duration* of the execution of the callbacks.
2036 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
2037 *
2038 * This synchronization is implemented by mutually excluding regular CPU
2039 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
2040 * Hibernate notifications.
2041 */
2042static int
2043cpu_hotplug_pm_callback(struct notifier_block *nb,
2044 unsigned long action, void *ptr)
2045{
2046 switch (action) {
2047
2048 case PM_SUSPEND_PREPARE:
2049 case PM_HIBERNATION_PREPARE:
2050 cpu_hotplug_disable();
2051 break;
2052
2053 case PM_POST_SUSPEND:
2054 case PM_POST_HIBERNATION:
2055 cpu_hotplug_enable();
2056 break;
2057
2058 default:
2059 return NOTIFY_DONE;
2060 }
2061
2062 return NOTIFY_OK;
2063}
2064
2065
2066static int __init cpu_hotplug_pm_sync_init(void)
2067{
2068 /*
2069 * cpu_hotplug_pm_callback has higher priority than x86
2070 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
2071 * to disable cpu hotplug to avoid cpu hotplug race.
2072 */
2073 pm_notifier(cpu_hotplug_pm_callback, 0);
2074 return 0;
2075}
2076core_initcall(cpu_hotplug_pm_sync_init);
2077
2078#endif /* CONFIG_PM_SLEEP_SMP */
2079
2080int __boot_cpu_id;
2081
2082#endif /* CONFIG_SMP */
2083
2084/* Boot processor state steps */
2085static struct cpuhp_step cpuhp_hp_states[] = {
2086 [CPUHP_OFFLINE] = {
2087 .name = "offline",
2088 .startup.single = NULL,
2089 .teardown.single = NULL,
2090 },
2091#ifdef CONFIG_SMP
2092 [CPUHP_CREATE_THREADS]= {
2093 .name = "threads:prepare",
2094 .startup.single = smpboot_create_threads,
2095 .teardown.single = NULL,
2096 .cant_stop = true,
2097 },
2098 [CPUHP_PERF_PREPARE] = {
2099 .name = "perf:prepare",
2100 .startup.single = perf_event_init_cpu,
2101 .teardown.single = perf_event_exit_cpu,
2102 },
2103 [CPUHP_RANDOM_PREPARE] = {
2104 .name = "random:prepare",
2105 .startup.single = random_prepare_cpu,
2106 .teardown.single = NULL,
2107 },
2108 [CPUHP_WORKQUEUE_PREP] = {
2109 .name = "workqueue:prepare",
2110 .startup.single = workqueue_prepare_cpu,
2111 .teardown.single = NULL,
2112 },
2113 [CPUHP_HRTIMERS_PREPARE] = {
2114 .name = "hrtimers:prepare",
2115 .startup.single = hrtimers_prepare_cpu,
2116 .teardown.single = NULL,
2117 },
2118 [CPUHP_SMPCFD_PREPARE] = {
2119 .name = "smpcfd:prepare",
2120 .startup.single = smpcfd_prepare_cpu,
2121 .teardown.single = smpcfd_dead_cpu,
2122 },
2123 [CPUHP_RELAY_PREPARE] = {
2124 .name = "relay:prepare",
2125 .startup.single = relay_prepare_cpu,
2126 .teardown.single = NULL,
2127 },
2128 [CPUHP_SLAB_PREPARE] = {
2129 .name = "slab:prepare",
2130 .startup.single = slab_prepare_cpu,
2131 .teardown.single = slab_dead_cpu,
2132 },
2133 [CPUHP_RCUTREE_PREP] = {
2134 .name = "RCU/tree:prepare",
2135 .startup.single = rcutree_prepare_cpu,
2136 .teardown.single = rcutree_dead_cpu,
2137 },
2138 /*
2139 * On the tear-down path, timers_dead_cpu() must be invoked
2140 * before blk_mq_queue_reinit_notify() from notify_dead(),
2141 * otherwise a RCU stall occurs.
2142 */
2143 [CPUHP_TIMERS_PREPARE] = {
2144 .name = "timers:prepare",
2145 .startup.single = timers_prepare_cpu,
2146 .teardown.single = timers_dead_cpu,
2147 },
2148
2149#ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
2150 /*
2151 * Kicks the AP alive. AP will wait in cpuhp_ap_sync_alive() until
2152 * the next step will release it.
2153 */
2154 [CPUHP_BP_KICK_AP] = {
2155 .name = "cpu:kick_ap",
2156 .startup.single = cpuhp_kick_ap_alive,
2157 },
2158
2159 /*
2160 * Waits for the AP to reach cpuhp_ap_sync_alive() and then
2161 * releases it for the complete bringup.
2162 */
2163 [CPUHP_BRINGUP_CPU] = {
2164 .name = "cpu:bringup",
2165 .startup.single = cpuhp_bringup_ap,
2166 .teardown.single = finish_cpu,
2167 .cant_stop = true,
2168 },
2169#else
2170 /*
2171 * All-in-one CPU bringup state which includes the kick alive.
2172 */
2173 [CPUHP_BRINGUP_CPU] = {
2174 .name = "cpu:bringup",
2175 .startup.single = bringup_cpu,
2176 .teardown.single = finish_cpu,
2177 .cant_stop = true,
2178 },
2179#endif
2180 /* Final state before CPU kills itself */
2181 [CPUHP_AP_IDLE_DEAD] = {
2182 .name = "idle:dead",
2183 },
2184 /*
2185 * Last state before CPU enters the idle loop to die. Transient state
2186 * for synchronization.
2187 */
2188 [CPUHP_AP_OFFLINE] = {
2189 .name = "ap:offline",
2190 .cant_stop = true,
2191 },
2192 /* First state is scheduler control. Interrupts are disabled */
2193 [CPUHP_AP_SCHED_STARTING] = {
2194 .name = "sched:starting",
2195 .startup.single = sched_cpu_starting,
2196 .teardown.single = sched_cpu_dying,
2197 },
2198 [CPUHP_AP_RCUTREE_DYING] = {
2199 .name = "RCU/tree:dying",
2200 .startup.single = NULL,
2201 .teardown.single = rcutree_dying_cpu,
2202 },
2203 [CPUHP_AP_SMPCFD_DYING] = {
2204 .name = "smpcfd:dying",
2205 .startup.single = NULL,
2206 .teardown.single = smpcfd_dying_cpu,
2207 },
2208 [CPUHP_AP_HRTIMERS_DYING] = {
2209 .name = "hrtimers:dying",
2210 .startup.single = NULL,
2211 .teardown.single = hrtimers_cpu_dying,
2212 },
2213
2214 /* Entry state on starting. Interrupts enabled from here on. Transient
2215 * state for synchronsization */
2216 [CPUHP_AP_ONLINE] = {
2217 .name = "ap:online",
2218 },
2219 /*
2220 * Handled on control processor until the plugged processor manages
2221 * this itself.
2222 */
2223 [CPUHP_TEARDOWN_CPU] = {
2224 .name = "cpu:teardown",
2225 .startup.single = NULL,
2226 .teardown.single = takedown_cpu,
2227 .cant_stop = true,
2228 },
2229
2230 [CPUHP_AP_SCHED_WAIT_EMPTY] = {
2231 .name = "sched:waitempty",
2232 .startup.single = NULL,
2233 .teardown.single = sched_cpu_wait_empty,
2234 },
2235
2236 /* Handle smpboot threads park/unpark */
2237 [CPUHP_AP_SMPBOOT_THREADS] = {
2238 .name = "smpboot/threads:online",
2239 .startup.single = smpboot_unpark_threads,
2240 .teardown.single = smpboot_park_threads,
2241 },
2242 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
2243 .name = "irq/affinity:online",
2244 .startup.single = irq_affinity_online_cpu,
2245 .teardown.single = NULL,
2246 },
2247 [CPUHP_AP_PERF_ONLINE] = {
2248 .name = "perf:online",
2249 .startup.single = perf_event_init_cpu,
2250 .teardown.single = perf_event_exit_cpu,
2251 },
2252 [CPUHP_AP_WATCHDOG_ONLINE] = {
2253 .name = "lockup_detector:online",
2254 .startup.single = lockup_detector_online_cpu,
2255 .teardown.single = lockup_detector_offline_cpu,
2256 },
2257 [CPUHP_AP_WORKQUEUE_ONLINE] = {
2258 .name = "workqueue:online",
2259 .startup.single = workqueue_online_cpu,
2260 .teardown.single = workqueue_offline_cpu,
2261 },
2262 [CPUHP_AP_RANDOM_ONLINE] = {
2263 .name = "random:online",
2264 .startup.single = random_online_cpu,
2265 .teardown.single = NULL,
2266 },
2267 [CPUHP_AP_RCUTREE_ONLINE] = {
2268 .name = "RCU/tree:online",
2269 .startup.single = rcutree_online_cpu,
2270 .teardown.single = rcutree_offline_cpu,
2271 },
2272#endif
2273 /*
2274 * The dynamically registered state space is here
2275 */
2276
2277#ifdef CONFIG_SMP
2278 /* Last state is scheduler control setting the cpu active */
2279 [CPUHP_AP_ACTIVE] = {
2280 .name = "sched:active",
2281 .startup.single = sched_cpu_activate,
2282 .teardown.single = sched_cpu_deactivate,
2283 },
2284#endif
2285
2286 /* CPU is fully up and running. */
2287 [CPUHP_ONLINE] = {
2288 .name = "online",
2289 .startup.single = NULL,
2290 .teardown.single = NULL,
2291 },
2292};
2293
2294/* Sanity check for callbacks */
2295static int cpuhp_cb_check(enum cpuhp_state state)
2296{
2297 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
2298 return -EINVAL;
2299 return 0;
2300}
2301
2302/*
2303 * Returns a free for dynamic slot assignment of the Online state. The states
2304 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
2305 * by having no name assigned.
2306 */
2307static int cpuhp_reserve_state(enum cpuhp_state state)
2308{
2309 enum cpuhp_state i, end;
2310 struct cpuhp_step *step;
2311
2312 switch (state) {
2313 case CPUHP_AP_ONLINE_DYN:
2314 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
2315 end = CPUHP_AP_ONLINE_DYN_END;
2316 break;
2317 case CPUHP_BP_PREPARE_DYN:
2318 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
2319 end = CPUHP_BP_PREPARE_DYN_END;
2320 break;
2321 default:
2322 return -EINVAL;
2323 }
2324
2325 for (i = state; i <= end; i++, step++) {
2326 if (!step->name)
2327 return i;
2328 }
2329 WARN(1, "No more dynamic states available for CPU hotplug\n");
2330 return -ENOSPC;
2331}
2332
2333static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
2334 int (*startup)(unsigned int cpu),
2335 int (*teardown)(unsigned int cpu),
2336 bool multi_instance)
2337{
2338 /* (Un)Install the callbacks for further cpu hotplug operations */
2339 struct cpuhp_step *sp;
2340 int ret = 0;
2341
2342 /*
2343 * If name is NULL, then the state gets removed.
2344 *
2345 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
2346 * the first allocation from these dynamic ranges, so the removal
2347 * would trigger a new allocation and clear the wrong (already
2348 * empty) state, leaving the callbacks of the to be cleared state
2349 * dangling, which causes wreckage on the next hotplug operation.
2350 */
2351 if (name && (state == CPUHP_AP_ONLINE_DYN ||
2352 state == CPUHP_BP_PREPARE_DYN)) {
2353 ret = cpuhp_reserve_state(state);
2354 if (ret < 0)
2355 return ret;
2356 state = ret;
2357 }
2358 sp = cpuhp_get_step(state);
2359 if (name && sp->name)
2360 return -EBUSY;
2361
2362 sp->startup.single = startup;
2363 sp->teardown.single = teardown;
2364 sp->name = name;
2365 sp->multi_instance = multi_instance;
2366 INIT_HLIST_HEAD(&sp->list);
2367 return ret;
2368}
2369
2370static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
2371{
2372 return cpuhp_get_step(state)->teardown.single;
2373}
2374
2375/*
2376 * Call the startup/teardown function for a step either on the AP or
2377 * on the current CPU.
2378 */
2379static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
2380 struct hlist_node *node)
2381{
2382 struct cpuhp_step *sp = cpuhp_get_step(state);
2383 int ret;
2384
2385 /*
2386 * If there's nothing to do, we done.
2387 * Relies on the union for multi_instance.
2388 */
2389 if (cpuhp_step_empty(bringup, sp))
2390 return 0;
2391 /*
2392 * The non AP bound callbacks can fail on bringup. On teardown
2393 * e.g. module removal we crash for now.
2394 */
2395#ifdef CONFIG_SMP
2396 if (cpuhp_is_ap_state(state))
2397 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
2398 else
2399 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2400#else
2401 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2402#endif
2403 BUG_ON(ret && !bringup);
2404 return ret;
2405}
2406
2407/*
2408 * Called from __cpuhp_setup_state on a recoverable failure.
2409 *
2410 * Note: The teardown callbacks for rollback are not allowed to fail!
2411 */
2412static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
2413 struct hlist_node *node)
2414{
2415 int cpu;
2416
2417 /* Roll back the already executed steps on the other cpus */
2418 for_each_present_cpu(cpu) {
2419 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2420 int cpustate = st->state;
2421
2422 if (cpu >= failedcpu)
2423 break;
2424
2425 /* Did we invoke the startup call on that cpu ? */
2426 if (cpustate >= state)
2427 cpuhp_issue_call(cpu, state, false, node);
2428 }
2429}
2430
2431int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
2432 struct hlist_node *node,
2433 bool invoke)
2434{
2435 struct cpuhp_step *sp;
2436 int cpu;
2437 int ret;
2438
2439 lockdep_assert_cpus_held();
2440
2441 sp = cpuhp_get_step(state);
2442 if (sp->multi_instance == false)
2443 return -EINVAL;
2444
2445 mutex_lock(&cpuhp_state_mutex);
2446
2447 if (!invoke || !sp->startup.multi)
2448 goto add_node;
2449
2450 /*
2451 * Try to call the startup callback for each present cpu
2452 * depending on the hotplug state of the cpu.
2453 */
2454 for_each_present_cpu(cpu) {
2455 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2456 int cpustate = st->state;
2457
2458 if (cpustate < state)
2459 continue;
2460
2461 ret = cpuhp_issue_call(cpu, state, true, node);
2462 if (ret) {
2463 if (sp->teardown.multi)
2464 cpuhp_rollback_install(cpu, state, node);
2465 goto unlock;
2466 }
2467 }
2468add_node:
2469 ret = 0;
2470 hlist_add_head(node, &sp->list);
2471unlock:
2472 mutex_unlock(&cpuhp_state_mutex);
2473 return ret;
2474}
2475
2476int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
2477 bool invoke)
2478{
2479 int ret;
2480
2481 cpus_read_lock();
2482 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
2483 cpus_read_unlock();
2484 return ret;
2485}
2486EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
2487
2488/**
2489 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
2490 * @state: The state to setup
2491 * @name: Name of the step
2492 * @invoke: If true, the startup function is invoked for cpus where
2493 * cpu state >= @state
2494 * @startup: startup callback function
2495 * @teardown: teardown callback function
2496 * @multi_instance: State is set up for multiple instances which get
2497 * added afterwards.
2498 *
2499 * The caller needs to hold cpus read locked while calling this function.
2500 * Return:
2501 * On success:
2502 * Positive state number if @state is CPUHP_AP_ONLINE_DYN;
2503 * 0 for all other states
2504 * On failure: proper (negative) error code
2505 */
2506int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
2507 const char *name, bool invoke,
2508 int (*startup)(unsigned int cpu),
2509 int (*teardown)(unsigned int cpu),
2510 bool multi_instance)
2511{
2512 int cpu, ret = 0;
2513 bool dynstate;
2514
2515 lockdep_assert_cpus_held();
2516
2517 if (cpuhp_cb_check(state) || !name)
2518 return -EINVAL;
2519
2520 mutex_lock(&cpuhp_state_mutex);
2521
2522 ret = cpuhp_store_callbacks(state, name, startup, teardown,
2523 multi_instance);
2524
2525 dynstate = state == CPUHP_AP_ONLINE_DYN;
2526 if (ret > 0 && dynstate) {
2527 state = ret;
2528 ret = 0;
2529 }
2530
2531 if (ret || !invoke || !startup)
2532 goto out;
2533
2534 /*
2535 * Try to call the startup callback for each present cpu
2536 * depending on the hotplug state of the cpu.
2537 */
2538 for_each_present_cpu(cpu) {
2539 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2540 int cpustate = st->state;
2541
2542 if (cpustate < state)
2543 continue;
2544
2545 ret = cpuhp_issue_call(cpu, state, true, NULL);
2546 if (ret) {
2547 if (teardown)
2548 cpuhp_rollback_install(cpu, state, NULL);
2549 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2550 goto out;
2551 }
2552 }
2553out:
2554 mutex_unlock(&cpuhp_state_mutex);
2555 /*
2556 * If the requested state is CPUHP_AP_ONLINE_DYN, return the
2557 * dynamically allocated state in case of success.
2558 */
2559 if (!ret && dynstate)
2560 return state;
2561 return ret;
2562}
2563EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2564
2565int __cpuhp_setup_state(enum cpuhp_state state,
2566 const char *name, bool invoke,
2567 int (*startup)(unsigned int cpu),
2568 int (*teardown)(unsigned int cpu),
2569 bool multi_instance)
2570{
2571 int ret;
2572
2573 cpus_read_lock();
2574 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2575 teardown, multi_instance);
2576 cpus_read_unlock();
2577 return ret;
2578}
2579EXPORT_SYMBOL(__cpuhp_setup_state);
2580
2581int __cpuhp_state_remove_instance(enum cpuhp_state state,
2582 struct hlist_node *node, bool invoke)
2583{
2584 struct cpuhp_step *sp = cpuhp_get_step(state);
2585 int cpu;
2586
2587 BUG_ON(cpuhp_cb_check(state));
2588
2589 if (!sp->multi_instance)
2590 return -EINVAL;
2591
2592 cpus_read_lock();
2593 mutex_lock(&cpuhp_state_mutex);
2594
2595 if (!invoke || !cpuhp_get_teardown_cb(state))
2596 goto remove;
2597 /*
2598 * Call the teardown callback for each present cpu depending
2599 * on the hotplug state of the cpu. This function is not
2600 * allowed to fail currently!
2601 */
2602 for_each_present_cpu(cpu) {
2603 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2604 int cpustate = st->state;
2605
2606 if (cpustate >= state)
2607 cpuhp_issue_call(cpu, state, false, node);
2608 }
2609
2610remove:
2611 hlist_del(node);
2612 mutex_unlock(&cpuhp_state_mutex);
2613 cpus_read_unlock();
2614
2615 return 0;
2616}
2617EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2618
2619/**
2620 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2621 * @state: The state to remove
2622 * @invoke: If true, the teardown function is invoked for cpus where
2623 * cpu state >= @state
2624 *
2625 * The caller needs to hold cpus read locked while calling this function.
2626 * The teardown callback is currently not allowed to fail. Think
2627 * about module removal!
2628 */
2629void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2630{
2631 struct cpuhp_step *sp = cpuhp_get_step(state);
2632 int cpu;
2633
2634 BUG_ON(cpuhp_cb_check(state));
2635
2636 lockdep_assert_cpus_held();
2637
2638 mutex_lock(&cpuhp_state_mutex);
2639 if (sp->multi_instance) {
2640 WARN(!hlist_empty(&sp->list),
2641 "Error: Removing state %d which has instances left.\n",
2642 state);
2643 goto remove;
2644 }
2645
2646 if (!invoke || !cpuhp_get_teardown_cb(state))
2647 goto remove;
2648
2649 /*
2650 * Call the teardown callback for each present cpu depending
2651 * on the hotplug state of the cpu. This function is not
2652 * allowed to fail currently!
2653 */
2654 for_each_present_cpu(cpu) {
2655 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2656 int cpustate = st->state;
2657
2658 if (cpustate >= state)
2659 cpuhp_issue_call(cpu, state, false, NULL);
2660 }
2661remove:
2662 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2663 mutex_unlock(&cpuhp_state_mutex);
2664}
2665EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2666
2667void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2668{
2669 cpus_read_lock();
2670 __cpuhp_remove_state_cpuslocked(state, invoke);
2671 cpus_read_unlock();
2672}
2673EXPORT_SYMBOL(__cpuhp_remove_state);
2674
2675#ifdef CONFIG_HOTPLUG_SMT
2676static void cpuhp_offline_cpu_device(unsigned int cpu)
2677{
2678 struct device *dev = get_cpu_device(cpu);
2679
2680 dev->offline = true;
2681 /* Tell user space about the state change */
2682 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2683}
2684
2685static void cpuhp_online_cpu_device(unsigned int cpu)
2686{
2687 struct device *dev = get_cpu_device(cpu);
2688
2689 dev->offline = false;
2690 /* Tell user space about the state change */
2691 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2692}
2693
2694int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2695{
2696 int cpu, ret = 0;
2697
2698 cpu_maps_update_begin();
2699 for_each_online_cpu(cpu) {
2700 if (topology_is_primary_thread(cpu))
2701 continue;
2702 /*
2703 * Disable can be called with CPU_SMT_ENABLED when changing
2704 * from a higher to lower number of SMT threads per core.
2705 */
2706 if (ctrlval == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
2707 continue;
2708 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2709 if (ret)
2710 break;
2711 /*
2712 * As this needs to hold the cpu maps lock it's impossible
2713 * to call device_offline() because that ends up calling
2714 * cpu_down() which takes cpu maps lock. cpu maps lock
2715 * needs to be held as this might race against in kernel
2716 * abusers of the hotplug machinery (thermal management).
2717 *
2718 * So nothing would update device:offline state. That would
2719 * leave the sysfs entry stale and prevent onlining after
2720 * smt control has been changed to 'off' again. This is
2721 * called under the sysfs hotplug lock, so it is properly
2722 * serialized against the regular offline usage.
2723 */
2724 cpuhp_offline_cpu_device(cpu);
2725 }
2726 if (!ret)
2727 cpu_smt_control = ctrlval;
2728 cpu_maps_update_done();
2729 return ret;
2730}
2731
2732int cpuhp_smt_enable(void)
2733{
2734 int cpu, ret = 0;
2735
2736 cpu_maps_update_begin();
2737 cpu_smt_control = CPU_SMT_ENABLED;
2738 for_each_present_cpu(cpu) {
2739 /* Skip online CPUs and CPUs on offline nodes */
2740 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2741 continue;
2742 if (!cpu_smt_thread_allowed(cpu))
2743 continue;
2744 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2745 if (ret)
2746 break;
2747 /* See comment in cpuhp_smt_disable() */
2748 cpuhp_online_cpu_device(cpu);
2749 }
2750 cpu_maps_update_done();
2751 return ret;
2752}
2753#endif
2754
2755#if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2756static ssize_t state_show(struct device *dev,
2757 struct device_attribute *attr, char *buf)
2758{
2759 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2760
2761 return sprintf(buf, "%d\n", st->state);
2762}
2763static DEVICE_ATTR_RO(state);
2764
2765static ssize_t target_store(struct device *dev, struct device_attribute *attr,
2766 const char *buf, size_t count)
2767{
2768 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2769 struct cpuhp_step *sp;
2770 int target, ret;
2771
2772 ret = kstrtoint(buf, 10, &target);
2773 if (ret)
2774 return ret;
2775
2776#ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2777 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2778 return -EINVAL;
2779#else
2780 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2781 return -EINVAL;
2782#endif
2783
2784 ret = lock_device_hotplug_sysfs();
2785 if (ret)
2786 return ret;
2787
2788 mutex_lock(&cpuhp_state_mutex);
2789 sp = cpuhp_get_step(target);
2790 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2791 mutex_unlock(&cpuhp_state_mutex);
2792 if (ret)
2793 goto out;
2794
2795 if (st->state < target)
2796 ret = cpu_up(dev->id, target);
2797 else if (st->state > target)
2798 ret = cpu_down(dev->id, target);
2799 else if (WARN_ON(st->target != target))
2800 st->target = target;
2801out:
2802 unlock_device_hotplug();
2803 return ret ? ret : count;
2804}
2805
2806static ssize_t target_show(struct device *dev,
2807 struct device_attribute *attr, char *buf)
2808{
2809 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2810
2811 return sprintf(buf, "%d\n", st->target);
2812}
2813static DEVICE_ATTR_RW(target);
2814
2815static ssize_t fail_store(struct device *dev, struct device_attribute *attr,
2816 const char *buf, size_t count)
2817{
2818 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2819 struct cpuhp_step *sp;
2820 int fail, ret;
2821
2822 ret = kstrtoint(buf, 10, &fail);
2823 if (ret)
2824 return ret;
2825
2826 if (fail == CPUHP_INVALID) {
2827 st->fail = fail;
2828 return count;
2829 }
2830
2831 if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2832 return -EINVAL;
2833
2834 /*
2835 * Cannot fail STARTING/DYING callbacks.
2836 */
2837 if (cpuhp_is_atomic_state(fail))
2838 return -EINVAL;
2839
2840 /*
2841 * DEAD callbacks cannot fail...
2842 * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2843 * triggering STARTING callbacks, a failure in this state would
2844 * hinder rollback.
2845 */
2846 if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2847 return -EINVAL;
2848
2849 /*
2850 * Cannot fail anything that doesn't have callbacks.
2851 */
2852 mutex_lock(&cpuhp_state_mutex);
2853 sp = cpuhp_get_step(fail);
2854 if (!sp->startup.single && !sp->teardown.single)
2855 ret = -EINVAL;
2856 mutex_unlock(&cpuhp_state_mutex);
2857 if (ret)
2858 return ret;
2859
2860 st->fail = fail;
2861
2862 return count;
2863}
2864
2865static ssize_t fail_show(struct device *dev,
2866 struct device_attribute *attr, char *buf)
2867{
2868 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2869
2870 return sprintf(buf, "%d\n", st->fail);
2871}
2872
2873static DEVICE_ATTR_RW(fail);
2874
2875static struct attribute *cpuhp_cpu_attrs[] = {
2876 &dev_attr_state.attr,
2877 &dev_attr_target.attr,
2878 &dev_attr_fail.attr,
2879 NULL
2880};
2881
2882static const struct attribute_group cpuhp_cpu_attr_group = {
2883 .attrs = cpuhp_cpu_attrs,
2884 .name = "hotplug",
2885 NULL
2886};
2887
2888static ssize_t states_show(struct device *dev,
2889 struct device_attribute *attr, char *buf)
2890{
2891 ssize_t cur, res = 0;
2892 int i;
2893
2894 mutex_lock(&cpuhp_state_mutex);
2895 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2896 struct cpuhp_step *sp = cpuhp_get_step(i);
2897
2898 if (sp->name) {
2899 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2900 buf += cur;
2901 res += cur;
2902 }
2903 }
2904 mutex_unlock(&cpuhp_state_mutex);
2905 return res;
2906}
2907static DEVICE_ATTR_RO(states);
2908
2909static struct attribute *cpuhp_cpu_root_attrs[] = {
2910 &dev_attr_states.attr,
2911 NULL
2912};
2913
2914static const struct attribute_group cpuhp_cpu_root_attr_group = {
2915 .attrs = cpuhp_cpu_root_attrs,
2916 .name = "hotplug",
2917 NULL
2918};
2919
2920#ifdef CONFIG_HOTPLUG_SMT
2921
2922static bool cpu_smt_num_threads_valid(unsigned int threads)
2923{
2924 if (IS_ENABLED(CONFIG_SMT_NUM_THREADS_DYNAMIC))
2925 return threads >= 1 && threads <= cpu_smt_max_threads;
2926 return threads == 1 || threads == cpu_smt_max_threads;
2927}
2928
2929static ssize_t
2930__store_smt_control(struct device *dev, struct device_attribute *attr,
2931 const char *buf, size_t count)
2932{
2933 int ctrlval, ret, num_threads, orig_threads;
2934 bool force_off;
2935
2936 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2937 return -EPERM;
2938
2939 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2940 return -ENODEV;
2941
2942 if (sysfs_streq(buf, "on")) {
2943 ctrlval = CPU_SMT_ENABLED;
2944 num_threads = cpu_smt_max_threads;
2945 } else if (sysfs_streq(buf, "off")) {
2946 ctrlval = CPU_SMT_DISABLED;
2947 num_threads = 1;
2948 } else if (sysfs_streq(buf, "forceoff")) {
2949 ctrlval = CPU_SMT_FORCE_DISABLED;
2950 num_threads = 1;
2951 } else if (kstrtoint(buf, 10, &num_threads) == 0) {
2952 if (num_threads == 1)
2953 ctrlval = CPU_SMT_DISABLED;
2954 else if (cpu_smt_num_threads_valid(num_threads))
2955 ctrlval = CPU_SMT_ENABLED;
2956 else
2957 return -EINVAL;
2958 } else {
2959 return -EINVAL;
2960 }
2961
2962 ret = lock_device_hotplug_sysfs();
2963 if (ret)
2964 return ret;
2965
2966 orig_threads = cpu_smt_num_threads;
2967 cpu_smt_num_threads = num_threads;
2968
2969 force_off = ctrlval != cpu_smt_control && ctrlval == CPU_SMT_FORCE_DISABLED;
2970
2971 if (num_threads > orig_threads)
2972 ret = cpuhp_smt_enable();
2973 else if (num_threads < orig_threads || force_off)
2974 ret = cpuhp_smt_disable(ctrlval);
2975
2976 unlock_device_hotplug();
2977 return ret ? ret : count;
2978}
2979
2980#else /* !CONFIG_HOTPLUG_SMT */
2981static ssize_t
2982__store_smt_control(struct device *dev, struct device_attribute *attr,
2983 const char *buf, size_t count)
2984{
2985 return -ENODEV;
2986}
2987#endif /* CONFIG_HOTPLUG_SMT */
2988
2989static const char *smt_states[] = {
2990 [CPU_SMT_ENABLED] = "on",
2991 [CPU_SMT_DISABLED] = "off",
2992 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2993 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2994 [CPU_SMT_NOT_IMPLEMENTED] = "notimplemented",
2995};
2996
2997static ssize_t control_show(struct device *dev,
2998 struct device_attribute *attr, char *buf)
2999{
3000 const char *state = smt_states[cpu_smt_control];
3001
3002#ifdef CONFIG_HOTPLUG_SMT
3003 /*
3004 * If SMT is enabled but not all threads are enabled then show the
3005 * number of threads. If all threads are enabled show "on". Otherwise
3006 * show the state name.
3007 */
3008 if (cpu_smt_control == CPU_SMT_ENABLED &&
3009 cpu_smt_num_threads != cpu_smt_max_threads)
3010 return sysfs_emit(buf, "%d\n", cpu_smt_num_threads);
3011#endif
3012
3013 return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
3014}
3015
3016static ssize_t control_store(struct device *dev, struct device_attribute *attr,
3017 const char *buf, size_t count)
3018{
3019 return __store_smt_control(dev, attr, buf, count);
3020}
3021static DEVICE_ATTR_RW(control);
3022
3023static ssize_t active_show(struct device *dev,
3024 struct device_attribute *attr, char *buf)
3025{
3026 return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
3027}
3028static DEVICE_ATTR_RO(active);
3029
3030static struct attribute *cpuhp_smt_attrs[] = {
3031 &dev_attr_control.attr,
3032 &dev_attr_active.attr,
3033 NULL
3034};
3035
3036static const struct attribute_group cpuhp_smt_attr_group = {
3037 .attrs = cpuhp_smt_attrs,
3038 .name = "smt",
3039 NULL
3040};
3041
3042static int __init cpu_smt_sysfs_init(void)
3043{
3044 struct device *dev_root;
3045 int ret = -ENODEV;
3046
3047 dev_root = bus_get_dev_root(&cpu_subsys);
3048 if (dev_root) {
3049 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_smt_attr_group);
3050 put_device(dev_root);
3051 }
3052 return ret;
3053}
3054
3055static int __init cpuhp_sysfs_init(void)
3056{
3057 struct device *dev_root;
3058 int cpu, ret;
3059
3060 ret = cpu_smt_sysfs_init();
3061 if (ret)
3062 return ret;
3063
3064 dev_root = bus_get_dev_root(&cpu_subsys);
3065 if (dev_root) {
3066 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_cpu_root_attr_group);
3067 put_device(dev_root);
3068 if (ret)
3069 return ret;
3070 }
3071
3072 for_each_possible_cpu(cpu) {
3073 struct device *dev = get_cpu_device(cpu);
3074
3075 if (!dev)
3076 continue;
3077 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
3078 if (ret)
3079 return ret;
3080 }
3081 return 0;
3082}
3083device_initcall(cpuhp_sysfs_init);
3084#endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
3085
3086/*
3087 * cpu_bit_bitmap[] is a special, "compressed" data structure that
3088 * represents all NR_CPUS bits binary values of 1<<nr.
3089 *
3090 * It is used by cpumask_of() to get a constant address to a CPU
3091 * mask value that has a single bit set only.
3092 */
3093
3094/* cpu_bit_bitmap[0] is empty - so we can back into it */
3095#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
3096#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
3097#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
3098#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
3099
3100const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
3101
3102 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
3103 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
3104#if BITS_PER_LONG > 32
3105 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
3106 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
3107#endif
3108};
3109EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
3110
3111const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
3112EXPORT_SYMBOL(cpu_all_bits);
3113
3114#ifdef CONFIG_INIT_ALL_POSSIBLE
3115struct cpumask __cpu_possible_mask __read_mostly
3116 = {CPU_BITS_ALL};
3117#else
3118struct cpumask __cpu_possible_mask __read_mostly;
3119#endif
3120EXPORT_SYMBOL(__cpu_possible_mask);
3121
3122struct cpumask __cpu_online_mask __read_mostly;
3123EXPORT_SYMBOL(__cpu_online_mask);
3124
3125struct cpumask __cpu_present_mask __read_mostly;
3126EXPORT_SYMBOL(__cpu_present_mask);
3127
3128struct cpumask __cpu_active_mask __read_mostly;
3129EXPORT_SYMBOL(__cpu_active_mask);
3130
3131struct cpumask __cpu_dying_mask __read_mostly;
3132EXPORT_SYMBOL(__cpu_dying_mask);
3133
3134atomic_t __num_online_cpus __read_mostly;
3135EXPORT_SYMBOL(__num_online_cpus);
3136
3137void init_cpu_present(const struct cpumask *src)
3138{
3139 cpumask_copy(&__cpu_present_mask, src);
3140}
3141
3142void init_cpu_possible(const struct cpumask *src)
3143{
3144 cpumask_copy(&__cpu_possible_mask, src);
3145}
3146
3147void init_cpu_online(const struct cpumask *src)
3148{
3149 cpumask_copy(&__cpu_online_mask, src);
3150}
3151
3152void set_cpu_online(unsigned int cpu, bool online)
3153{
3154 /*
3155 * atomic_inc/dec() is required to handle the horrid abuse of this
3156 * function by the reboot and kexec code which invoke it from
3157 * IPI/NMI broadcasts when shutting down CPUs. Invocation from
3158 * regular CPU hotplug is properly serialized.
3159 *
3160 * Note, that the fact that __num_online_cpus is of type atomic_t
3161 * does not protect readers which are not serialized against
3162 * concurrent hotplug operations.
3163 */
3164 if (online) {
3165 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
3166 atomic_inc(&__num_online_cpus);
3167 } else {
3168 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
3169 atomic_dec(&__num_online_cpus);
3170 }
3171}
3172
3173/*
3174 * Activate the first processor.
3175 */
3176void __init boot_cpu_init(void)
3177{
3178 int cpu = smp_processor_id();
3179
3180 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
3181 set_cpu_online(cpu, true);
3182 set_cpu_active(cpu, true);
3183 set_cpu_present(cpu, true);
3184 set_cpu_possible(cpu, true);
3185
3186#ifdef CONFIG_SMP
3187 __boot_cpu_id = cpu;
3188#endif
3189}
3190
3191/*
3192 * Must be called _AFTER_ setting up the per_cpu areas
3193 */
3194void __init boot_cpu_hotplug_init(void)
3195{
3196#ifdef CONFIG_SMP
3197 cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
3198 atomic_set(this_cpu_ptr(&cpuhp_state.ap_sync_state), SYNC_STATE_ONLINE);
3199#endif
3200 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
3201 this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
3202}
3203
3204/*
3205 * These are used for a global "mitigations=" cmdline option for toggling
3206 * optional CPU mitigations.
3207 */
3208enum cpu_mitigations {
3209 CPU_MITIGATIONS_OFF,
3210 CPU_MITIGATIONS_AUTO,
3211 CPU_MITIGATIONS_AUTO_NOSMT,
3212};
3213
3214static enum cpu_mitigations cpu_mitigations __ro_after_init =
3215 CPU_MITIGATIONS_AUTO;
3216
3217static int __init mitigations_parse_cmdline(char *arg)
3218{
3219 if (!strcmp(arg, "off"))
3220 cpu_mitigations = CPU_MITIGATIONS_OFF;
3221 else if (!strcmp(arg, "auto"))
3222 cpu_mitigations = CPU_MITIGATIONS_AUTO;
3223 else if (!strcmp(arg, "auto,nosmt"))
3224 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
3225 else
3226 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
3227 arg);
3228
3229 return 0;
3230}
3231early_param("mitigations", mitigations_parse_cmdline);
3232
3233/* mitigations=off */
3234bool cpu_mitigations_off(void)
3235{
3236 return cpu_mitigations == CPU_MITIGATIONS_OFF;
3237}
3238EXPORT_SYMBOL_GPL(cpu_mitigations_off);
3239
3240/* mitigations=auto,nosmt */
3241bool cpu_mitigations_auto_nosmt(void)
3242{
3243 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
3244}
3245EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);