1/* CPU control.
2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
3 *
4 * This code is licenced under the GPL.
5 */
6#include <linux/sched/mm.h>
7#include <linux/proc_fs.h>
8#include <linux/smp.h>
9#include <linux/init.h>
10#include <linux/notifier.h>
11#include <linux/sched/signal.h>
12#include <linux/sched/hotplug.h>
13#include <linux/sched/isolation.h>
14#include <linux/sched/task.h>
15#include <linux/sched/smt.h>
16#include <linux/unistd.h>
17#include <linux/cpu.h>
18#include <linux/oom.h>
19#include <linux/rcupdate.h>
20#include <linux/delay.h>
21#include <linux/export.h>
22#include <linux/bug.h>
23#include <linux/kthread.h>
24#include <linux/stop_machine.h>
25#include <linux/mutex.h>
26#include <linux/gfp.h>
27#include <linux/suspend.h>
28#include <linux/lockdep.h>
29#include <linux/tick.h>
30#include <linux/irq.h>
31#include <linux/nmi.h>
32#include <linux/smpboot.h>
33#include <linux/relay.h>
34#include <linux/slab.h>
35#include <linux/scs.h>
36#include <linux/percpu-rwsem.h>
37#include <linux/cpuset.h>
38#include <linux/random.h>
39#include <linux/cc_platform.h>
40
41#include <trace/events/power.h>
42#define CREATE_TRACE_POINTS
43#include <trace/events/cpuhp.h>
44
45#include "smpboot.h"
46
47/**
48 * struct cpuhp_cpu_state - Per cpu hotplug state storage
49 * @state: The current cpu state
50 * @target: The target state
51 * @fail: Current CPU hotplug callback state
52 * @thread: Pointer to the hotplug thread
53 * @should_run: Thread should execute
54 * @rollback: Perform a rollback
55 * @single: Single callback invocation
56 * @bringup: Single callback bringup or teardown selector
57 * @node: Remote CPU node; for multi-instance, do a
58 * single entry callback for install/remove
59 * @last: For multi-instance rollback, remember how far we got
60 * @cb_state: The state for a single callback (install/uninstall)
61 * @result: Result of the operation
62 * @ap_sync_state: State for AP synchronization
63 * @done_up: Signal completion to the issuer of the task for cpu-up
64 * @done_down: Signal completion to the issuer of the task for cpu-down
65 */
66struct cpuhp_cpu_state {
67 enum cpuhp_state state;
68 enum cpuhp_state target;
69 enum cpuhp_state fail;
70#ifdef CONFIG_SMP
71 struct task_struct *thread;
72 bool should_run;
73 bool rollback;
74 bool single;
75 bool bringup;
76 struct hlist_node *node;
77 struct hlist_node *last;
78 enum cpuhp_state cb_state;
79 int result;
80 atomic_t ap_sync_state;
81 struct completion done_up;
82 struct completion done_down;
83#endif
84};
85
86static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
87 .fail = CPUHP_INVALID,
88};
89
90#ifdef CONFIG_SMP
91cpumask_t cpus_booted_once_mask;
92#endif
93
94#if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
95static struct lockdep_map cpuhp_state_up_map =
96 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
97static struct lockdep_map cpuhp_state_down_map =
98 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
99
100
101static inline void cpuhp_lock_acquire(bool bringup)
102{
103 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
104}
105
106static inline void cpuhp_lock_release(bool bringup)
107{
108 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
109}
110#else
111
112static inline void cpuhp_lock_acquire(bool bringup) { }
113static inline void cpuhp_lock_release(bool bringup) { }
114
115#endif
116
117/**
118 * struct cpuhp_step - Hotplug state machine step
119 * @name: Name of the step
120 * @startup: Startup function of the step
121 * @teardown: Teardown function of the step
122 * @cant_stop: Bringup/teardown can't be stopped at this step
123 * @multi_instance: State has multiple instances which get added afterwards
124 */
125struct cpuhp_step {
126 const char *name;
127 union {
128 int (*single)(unsigned int cpu);
129 int (*multi)(unsigned int cpu,
130 struct hlist_node *node);
131 } startup;
132 union {
133 int (*single)(unsigned int cpu);
134 int (*multi)(unsigned int cpu,
135 struct hlist_node *node);
136 } teardown;
137 /* private: */
138 struct hlist_head list;
139 /* public: */
140 bool cant_stop;
141 bool multi_instance;
142};
143
144static DEFINE_MUTEX(cpuhp_state_mutex);
145static struct cpuhp_step cpuhp_hp_states[];
146
147static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
148{
149 return cpuhp_hp_states + state;
150}
151
152static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
153{
154 return bringup ? !step->startup.single : !step->teardown.single;
155}
156
157/**
158 * cpuhp_invoke_callback - Invoke the callbacks for a given state
159 * @cpu: The cpu for which the callback should be invoked
160 * @state: The state to do callbacks for
161 * @bringup: True if the bringup callback should be invoked
162 * @node: For multi-instance, do a single entry callback for install/remove
163 * @lastp: For multi-instance rollback, remember how far we got
164 *
165 * Called from cpu hotplug and from the state register machinery.
166 *
167 * Return: %0 on success or a negative errno code
168 */
169static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
170 bool bringup, struct hlist_node *node,
171 struct hlist_node **lastp)
172{
173 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
174 struct cpuhp_step *step = cpuhp_get_step(state);
175 int (*cbm)(unsigned int cpu, struct hlist_node *node);
176 int (*cb)(unsigned int cpu);
177 int ret, cnt;
178
179 if (st->fail == state) {
180 st->fail = CPUHP_INVALID;
181 return -EAGAIN;
182 }
183
184 if (cpuhp_step_empty(bringup, step)) {
185 WARN_ON_ONCE(1);
186 return 0;
187 }
188
189 if (!step->multi_instance) {
190 WARN_ON_ONCE(lastp && *lastp);
191 cb = bringup ? step->startup.single : step->teardown.single;
192
193 trace_cpuhp_enter(cpu, target: st->target, idx: state, fun: cb);
194 ret = cb(cpu);
195 trace_cpuhp_exit(cpu, state: st->state, idx: state, ret);
196 return ret;
197 }
198 cbm = bringup ? step->startup.multi : step->teardown.multi;
199
200 /* Single invocation for instance add/remove */
201 if (node) {
202 WARN_ON_ONCE(lastp && *lastp);
203 trace_cpuhp_multi_enter(cpu, target: st->target, idx: state, fun: cbm, node);
204 ret = cbm(cpu, node);
205 trace_cpuhp_exit(cpu, state: st->state, idx: state, ret);
206 return ret;
207 }
208
209 /* State transition. Invoke on all instances */
210 cnt = 0;
211 hlist_for_each(node, &step->list) {
212 if (lastp && node == *lastp)
213 break;
214
215 trace_cpuhp_multi_enter(cpu, target: st->target, idx: state, fun: cbm, node);
216 ret = cbm(cpu, node);
217 trace_cpuhp_exit(cpu, state: st->state, idx: state, ret);
218 if (ret) {
219 if (!lastp)
220 goto err;
221
222 *lastp = node;
223 return ret;
224 }
225 cnt++;
226 }
227 if (lastp)
228 *lastp = NULL;
229 return 0;
230err:
231 /* Rollback the instances if one failed */
232 cbm = !bringup ? step->startup.multi : step->teardown.multi;
233 if (!cbm)
234 return ret;
235
236 hlist_for_each(node, &step->list) {
237 if (!cnt--)
238 break;
239
240 trace_cpuhp_multi_enter(cpu, target: st->target, idx: state, fun: cbm, node);
241 ret = cbm(cpu, node);
242 trace_cpuhp_exit(cpu, state: st->state, idx: state, ret);
243 /*
244 * Rollback must not fail,
245 */
246 WARN_ON_ONCE(ret);
247 }
248 return ret;
249}
250
251#ifdef CONFIG_SMP
252static bool cpuhp_is_ap_state(enum cpuhp_state state)
253{
254 /*
255 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
256 * purposes as that state is handled explicitly in cpu_down.
257 */
258 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
259}
260
261static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
262{
263 struct completion *done = bringup ? &st->done_up : &st->done_down;
264 wait_for_completion(done);
265}
266
267static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
268{
269 struct completion *done = bringup ? &st->done_up : &st->done_down;
270 complete(done);
271}
272
273/*
274 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
275 */
276static bool cpuhp_is_atomic_state(enum cpuhp_state state)
277{
278 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
279}
280
281/* Synchronization state management */
282enum cpuhp_sync_state {
283 SYNC_STATE_DEAD,
284 SYNC_STATE_KICKED,
285 SYNC_STATE_SHOULD_DIE,
286 SYNC_STATE_ALIVE,
287 SYNC_STATE_SHOULD_ONLINE,
288 SYNC_STATE_ONLINE,
289};
290
291#ifdef CONFIG_HOTPLUG_CORE_SYNC
292/**
293 * cpuhp_ap_update_sync_state - Update synchronization state during bringup/teardown
294 * @state: The synchronization state to set
295 *
296 * No synchronization point. Just update of the synchronization state, but implies
297 * a full barrier so that the AP changes are visible before the control CPU proceeds.
298 */
299static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state)
300{
301 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
302
303 (void)atomic_xchg(v: st, new: state);
304}
305
306void __weak arch_cpuhp_sync_state_poll(void) { cpu_relax(); }
307
308static bool cpuhp_wait_for_sync_state(unsigned int cpu, enum cpuhp_sync_state state,
309 enum cpuhp_sync_state next_state)
310{
311 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
312 ktime_t now, end, start = ktime_get();
313 int sync;
314
315 end = start + 10ULL * NSEC_PER_SEC;
316
317 sync = atomic_read(v: st);
318 while (1) {
319 if (sync == state) {
320 if (!atomic_try_cmpxchg(v: st, old: &sync, new: next_state))
321 continue;
322 return true;
323 }
324
325 now = ktime_get();
326 if (now > end) {
327 /* Timeout. Leave the state unchanged */
328 return false;
329 } else if (now - start < NSEC_PER_MSEC) {
330 /* Poll for one millisecond */
331 arch_cpuhp_sync_state_poll();
332 } else {
333 usleep_range(USEC_PER_MSEC, max: 2 * USEC_PER_MSEC);
334 }
335 sync = atomic_read(v: st);
336 }
337 return true;
338}
339#else /* CONFIG_HOTPLUG_CORE_SYNC */
340static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state) { }
341#endif /* !CONFIG_HOTPLUG_CORE_SYNC */
342
343#ifdef CONFIG_HOTPLUG_CORE_SYNC_DEAD
344/**
345 * cpuhp_ap_report_dead - Update synchronization state to DEAD
346 *
347 * No synchronization point. Just update of the synchronization state.
348 */
349void cpuhp_ap_report_dead(void)
350{
351 cpuhp_ap_update_sync_state(state: SYNC_STATE_DEAD);
352}
353
354void __weak arch_cpuhp_cleanup_dead_cpu(unsigned int cpu) { }
355
356/*
357 * Late CPU shutdown synchronization point. Cannot use cpuhp_state::done_down
358 * because the AP cannot issue complete() at this stage.
359 */
360static void cpuhp_bp_sync_dead(unsigned int cpu)
361{
362 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
363 int sync = atomic_read(v: st);
364
365 do {
366 /* CPU can have reported dead already. Don't overwrite that! */
367 if (sync == SYNC_STATE_DEAD)
368 break;
369 } while (!atomic_try_cmpxchg(v: st, old: &sync, new: SYNC_STATE_SHOULD_DIE));
370
371 if (cpuhp_wait_for_sync_state(cpu, state: SYNC_STATE_DEAD, next_state: SYNC_STATE_DEAD)) {
372 /* CPU reached dead state. Invoke the cleanup function */
373 arch_cpuhp_cleanup_dead_cpu(cpu);
374 return;
375 }
376
377 /* No further action possible. Emit message and give up. */
378 pr_err("CPU%u failed to report dead state\n", cpu);
379}
380#else /* CONFIG_HOTPLUG_CORE_SYNC_DEAD */
381static inline void cpuhp_bp_sync_dead(unsigned int cpu) { }
382#endif /* !CONFIG_HOTPLUG_CORE_SYNC_DEAD */
383
384#ifdef CONFIG_HOTPLUG_CORE_SYNC_FULL
385/**
386 * cpuhp_ap_sync_alive - Synchronize AP with the control CPU once it is alive
387 *
388 * Updates the AP synchronization state to SYNC_STATE_ALIVE and waits
389 * for the BP to release it.
390 */
391void cpuhp_ap_sync_alive(void)
392{
393 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
394
395 cpuhp_ap_update_sync_state(state: SYNC_STATE_ALIVE);
396
397 /* Wait for the control CPU to release it. */
398 while (atomic_read(v: st) != SYNC_STATE_SHOULD_ONLINE)
399 cpu_relax();
400}
401
402static bool cpuhp_can_boot_ap(unsigned int cpu)
403{
404 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
405 int sync = atomic_read(v: st);
406
407again:
408 switch (sync) {
409 case SYNC_STATE_DEAD:
410 /* CPU is properly dead */
411 break;
412 case SYNC_STATE_KICKED:
413 /* CPU did not come up in previous attempt */
414 break;
415 case SYNC_STATE_ALIVE:
416 /* CPU is stuck cpuhp_ap_sync_alive(). */
417 break;
418 default:
419 /* CPU failed to report online or dead and is in limbo state. */
420 return false;
421 }
422
423 /* Prepare for booting */
424 if (!atomic_try_cmpxchg(v: st, old: &sync, new: SYNC_STATE_KICKED))
425 goto again;
426
427 return true;
428}
429
430void __weak arch_cpuhp_cleanup_kick_cpu(unsigned int cpu) { }
431
432/*
433 * Early CPU bringup synchronization point. Cannot use cpuhp_state::done_up
434 * because the AP cannot issue complete() so early in the bringup.
435 */
436static int cpuhp_bp_sync_alive(unsigned int cpu)
437{
438 int ret = 0;
439
440 if (!IS_ENABLED(CONFIG_HOTPLUG_CORE_SYNC_FULL))
441 return 0;
442
443 if (!cpuhp_wait_for_sync_state(cpu, state: SYNC_STATE_ALIVE, next_state: SYNC_STATE_SHOULD_ONLINE)) {
444 pr_err("CPU%u failed to report alive state\n", cpu);
445 ret = -EIO;
446 }
447
448 /* Let the architecture cleanup the kick alive mechanics. */
449 arch_cpuhp_cleanup_kick_cpu(cpu);
450 return ret;
451}
452#else /* CONFIG_HOTPLUG_CORE_SYNC_FULL */
453static inline int cpuhp_bp_sync_alive(unsigned int cpu) { return 0; }
454static inline bool cpuhp_can_boot_ap(unsigned int cpu) { return true; }
455#endif /* !CONFIG_HOTPLUG_CORE_SYNC_FULL */
456
457/* Serializes the updates to cpu_online_mask, cpu_present_mask */
458static DEFINE_MUTEX(cpu_add_remove_lock);
459bool cpuhp_tasks_frozen;
460EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
461
462/*
463 * The following two APIs (cpu_maps_update_begin/done) must be used when
464 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
465 */
466void cpu_maps_update_begin(void)
467{
468 mutex_lock(&cpu_add_remove_lock);
469}
470
471void cpu_maps_update_done(void)
472{
473 mutex_unlock(lock: &cpu_add_remove_lock);
474}
475
476/*
477 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
478 * Should always be manipulated under cpu_add_remove_lock
479 */
480static int cpu_hotplug_disabled;
481
482#ifdef CONFIG_HOTPLUG_CPU
483
484DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
485
486static bool cpu_hotplug_offline_disabled __ro_after_init;
487
488void cpus_read_lock(void)
489{
490 percpu_down_read(sem: &cpu_hotplug_lock);
491}
492EXPORT_SYMBOL_GPL(cpus_read_lock);
493
494int cpus_read_trylock(void)
495{
496 return percpu_down_read_trylock(sem: &cpu_hotplug_lock);
497}
498EXPORT_SYMBOL_GPL(cpus_read_trylock);
499
500void cpus_read_unlock(void)
501{
502 percpu_up_read(sem: &cpu_hotplug_lock);
503}
504EXPORT_SYMBOL_GPL(cpus_read_unlock);
505
506void cpus_write_lock(void)
507{
508 percpu_down_write(&cpu_hotplug_lock);
509}
510
511void cpus_write_unlock(void)
512{
513 percpu_up_write(&cpu_hotplug_lock);
514}
515
516void lockdep_assert_cpus_held(void)
517{
518 /*
519 * We can't have hotplug operations before userspace starts running,
520 * and some init codepaths will knowingly not take the hotplug lock.
521 * This is all valid, so mute lockdep until it makes sense to report
522 * unheld locks.
523 */
524 if (system_state < SYSTEM_RUNNING)
525 return;
526
527 percpu_rwsem_assert_held(&cpu_hotplug_lock);
528}
529EXPORT_SYMBOL_GPL(lockdep_assert_cpus_held);
530
531#ifdef CONFIG_LOCKDEP
532int lockdep_is_cpus_held(void)
533{
534 return percpu_rwsem_is_held(&cpu_hotplug_lock);
535}
536#endif
537
538static void lockdep_acquire_cpus_lock(void)
539{
540 rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
541}
542
543static void lockdep_release_cpus_lock(void)
544{
545 rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
546}
547
548/* Declare CPU offlining not supported */
549void cpu_hotplug_disable_offlining(void)
550{
551 cpu_maps_update_begin();
552 cpu_hotplug_offline_disabled = true;
553 cpu_maps_update_done();
554}
555
556/*
557 * Wait for currently running CPU hotplug operations to complete (if any) and
558 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
559 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
560 * hotplug path before performing hotplug operations. So acquiring that lock
561 * guarantees mutual exclusion from any currently running hotplug operations.
562 */
563void cpu_hotplug_disable(void)
564{
565 cpu_maps_update_begin();
566 cpu_hotplug_disabled++;
567 cpu_maps_update_done();
568}
569EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
570
571static void __cpu_hotplug_enable(void)
572{
573 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
574 return;
575 cpu_hotplug_disabled--;
576}
577
578void cpu_hotplug_enable(void)
579{
580 cpu_maps_update_begin();
581 __cpu_hotplug_enable();
582 cpu_maps_update_done();
583}
584EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
585
586#else
587
588static void lockdep_acquire_cpus_lock(void)
589{
590}
591
592static void lockdep_release_cpus_lock(void)
593{
594}
595
596#endif /* CONFIG_HOTPLUG_CPU */
597
598/*
599 * Architectures that need SMT-specific errata handling during SMT hotplug
600 * should override this.
601 */
602void __weak arch_smt_update(void) { }
603
604#ifdef CONFIG_HOTPLUG_SMT
605
606enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
607static unsigned int cpu_smt_max_threads __ro_after_init;
608unsigned int cpu_smt_num_threads __read_mostly = UINT_MAX;
609
610void __init cpu_smt_disable(bool force)
611{
612 if (!cpu_smt_possible())
613 return;
614
615 if (force) {
616 pr_info("SMT: Force disabled\n");
617 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
618 } else {
619 pr_info("SMT: disabled\n");
620 cpu_smt_control = CPU_SMT_DISABLED;
621 }
622 cpu_smt_num_threads = 1;
623}
624
625/*
626 * The decision whether SMT is supported can only be done after the full
627 * CPU identification. Called from architecture code.
628 */
629void __init cpu_smt_set_num_threads(unsigned int num_threads,
630 unsigned int max_threads)
631{
632 WARN_ON(!num_threads || (num_threads > max_threads));
633
634 if (max_threads == 1)
635 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
636
637 cpu_smt_max_threads = max_threads;
638
639 /*
640 * If SMT has been disabled via the kernel command line or SMT is
641 * not supported, set cpu_smt_num_threads to 1 for consistency.
642 * If enabled, take the architecture requested number of threads
643 * to bring up into account.
644 */
645 if (cpu_smt_control != CPU_SMT_ENABLED)
646 cpu_smt_num_threads = 1;
647 else if (num_threads < cpu_smt_num_threads)
648 cpu_smt_num_threads = num_threads;
649}
650
651static int __init smt_cmdline_disable(char *str)
652{
653 cpu_smt_disable(force: str && !strcmp(str, "force"));
654 return 0;
655}
656early_param("nosmt", smt_cmdline_disable);
657
658/*
659 * For Archicture supporting partial SMT states check if the thread is allowed.
660 * Otherwise this has already been checked through cpu_smt_max_threads when
661 * setting the SMT level.
662 */
663static inline bool cpu_smt_thread_allowed(unsigned int cpu)
664{
665#ifdef CONFIG_SMT_NUM_THREADS_DYNAMIC
666 return topology_smt_thread_allowed(cpu);
667#else
668 return true;
669#endif
670}
671
672static inline bool cpu_bootable(unsigned int cpu)
673{
674 if (cpu_smt_control == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
675 return true;
676
677 /* All CPUs are bootable if controls are not configured */
678 if (cpu_smt_control == CPU_SMT_NOT_IMPLEMENTED)
679 return true;
680
681 /* All CPUs are bootable if CPU is not SMT capable */
682 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
683 return true;
684
685 if (topology_is_primary_thread(cpu))
686 return true;
687
688 /*
689 * On x86 it's required to boot all logical CPUs at least once so
690 * that the init code can get a chance to set CR4.MCE on each
691 * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
692 * core will shutdown the machine.
693 */
694 return !cpumask_test_cpu(cpu, cpumask: &cpus_booted_once_mask);
695}
696
697/* Returns true if SMT is supported and not forcefully (irreversibly) disabled */
698bool cpu_smt_possible(void)
699{
700 return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
701 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
702}
703EXPORT_SYMBOL_GPL(cpu_smt_possible);
704
705#else
706static inline bool cpu_bootable(unsigned int cpu) { return true; }
707#endif
708
709static inline enum cpuhp_state
710cpuhp_set_state(int cpu, struct cpuhp_cpu_state *st, enum cpuhp_state target)
711{
712 enum cpuhp_state prev_state = st->state;
713 bool bringup = st->state < target;
714
715 st->rollback = false;
716 st->last = NULL;
717
718 st->target = target;
719 st->single = false;
720 st->bringup = bringup;
721 if (cpu_dying(cpu) != !bringup)
722 set_cpu_dying(cpu, !bringup);
723
724 return prev_state;
725}
726
727static inline void
728cpuhp_reset_state(int cpu, struct cpuhp_cpu_state *st,
729 enum cpuhp_state prev_state)
730{
731 bool bringup = !st->bringup;
732
733 st->target = prev_state;
734
735 /*
736 * Already rolling back. No need invert the bringup value or to change
737 * the current state.
738 */
739 if (st->rollback)
740 return;
741
742 st->rollback = true;
743
744 /*
745 * If we have st->last we need to undo partial multi_instance of this
746 * state first. Otherwise start undo at the previous state.
747 */
748 if (!st->last) {
749 if (st->bringup)
750 st->state--;
751 else
752 st->state++;
753 }
754
755 st->bringup = bringup;
756 if (cpu_dying(cpu) != !bringup)
757 set_cpu_dying(cpu, !bringup);
758}
759
760/* Regular hotplug invocation of the AP hotplug thread */
761static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
762{
763 if (!st->single && st->state == st->target)
764 return;
765
766 st->result = 0;
767 /*
768 * Make sure the above stores are visible before should_run becomes
769 * true. Paired with the mb() above in cpuhp_thread_fun()
770 */
771 smp_mb();
772 st->should_run = true;
773 wake_up_process(tsk: st->thread);
774 wait_for_ap_thread(st, bringup: st->bringup);
775}
776
777static int cpuhp_kick_ap(int cpu, struct cpuhp_cpu_state *st,
778 enum cpuhp_state target)
779{
780 enum cpuhp_state prev_state;
781 int ret;
782
783 prev_state = cpuhp_set_state(cpu, st, target);
784 __cpuhp_kick_ap(st);
785 if ((ret = st->result)) {
786 cpuhp_reset_state(cpu, st, prev_state);
787 __cpuhp_kick_ap(st);
788 }
789
790 return ret;
791}
792
793static int bringup_wait_for_ap_online(unsigned int cpu)
794{
795 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
796
797 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
798 wait_for_ap_thread(st, bringup: true);
799 if (WARN_ON_ONCE((!cpu_online(cpu))))
800 return -ECANCELED;
801
802 /* Unpark the hotplug thread of the target cpu */
803 kthread_unpark(k: st->thread);
804
805 /*
806 * SMT soft disabling on X86 requires to bring the CPU out of the
807 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
808 * CPU marked itself as booted_once in notify_cpu_starting() so the
809 * cpu_bootable() check will now return false if this is not the
810 * primary sibling.
811 */
812 if (!cpu_bootable(cpu))
813 return -ECANCELED;
814 return 0;
815}
816
817#ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
818static int cpuhp_kick_ap_alive(unsigned int cpu)
819{
820 if (!cpuhp_can_boot_ap(cpu))
821 return -EAGAIN;
822
823 return arch_cpuhp_kick_ap_alive(cpu, tidle: idle_thread_get(cpu));
824}
825
826static int cpuhp_bringup_ap(unsigned int cpu)
827{
828 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
829 int ret;
830
831 /*
832 * Some architectures have to walk the irq descriptors to
833 * setup the vector space for the cpu which comes online.
834 * Prevent irq alloc/free across the bringup.
835 */
836 irq_lock_sparse();
837
838 ret = cpuhp_bp_sync_alive(cpu);
839 if (ret)
840 goto out_unlock;
841
842 ret = bringup_wait_for_ap_online(cpu);
843 if (ret)
844 goto out_unlock;
845
846 irq_unlock_sparse();
847
848 if (st->target <= CPUHP_AP_ONLINE_IDLE)
849 return 0;
850
851 return cpuhp_kick_ap(cpu, st, target: st->target);
852
853out_unlock:
854 irq_unlock_sparse();
855 return ret;
856}
857#else
858static int bringup_cpu(unsigned int cpu)
859{
860 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
861 struct task_struct *idle = idle_thread_get(cpu);
862 int ret;
863
864 if (!cpuhp_can_boot_ap(cpu))
865 return -EAGAIN;
866
867 /*
868 * Some architectures have to walk the irq descriptors to
869 * setup the vector space for the cpu which comes online.
870 *
871 * Prevent irq alloc/free across the bringup by acquiring the
872 * sparse irq lock. Hold it until the upcoming CPU completes the
873 * startup in cpuhp_online_idle() which allows to avoid
874 * intermediate synchronization points in the architecture code.
875 */
876 irq_lock_sparse();
877
878 ret = __cpu_up(cpu, idle);
879 if (ret)
880 goto out_unlock;
881
882 ret = cpuhp_bp_sync_alive(cpu);
883 if (ret)
884 goto out_unlock;
885
886 ret = bringup_wait_for_ap_online(cpu);
887 if (ret)
888 goto out_unlock;
889
890 irq_unlock_sparse();
891
892 if (st->target <= CPUHP_AP_ONLINE_IDLE)
893 return 0;
894
895 return cpuhp_kick_ap(cpu, st, st->target);
896
897out_unlock:
898 irq_unlock_sparse();
899 return ret;
900}
901#endif
902
903static int finish_cpu(unsigned int cpu)
904{
905 struct task_struct *idle = idle_thread_get(cpu);
906 struct mm_struct *mm = idle->active_mm;
907
908 /*
909 * sched_force_init_mm() ensured the use of &init_mm,
910 * drop that refcount now that the CPU has stopped.
911 */
912 WARN_ON(mm != &init_mm);
913 idle->active_mm = NULL;
914 mmdrop_lazy_tlb(mm);
915
916 return 0;
917}
918
919/*
920 * Hotplug state machine related functions
921 */
922
923/*
924 * Get the next state to run. Empty ones will be skipped. Returns true if a
925 * state must be run.
926 *
927 * st->state will be modified ahead of time, to match state_to_run, as if it
928 * has already ran.
929 */
930static bool cpuhp_next_state(bool bringup,
931 enum cpuhp_state *state_to_run,
932 struct cpuhp_cpu_state *st,
933 enum cpuhp_state target)
934{
935 do {
936 if (bringup) {
937 if (st->state >= target)
938 return false;
939
940 *state_to_run = ++st->state;
941 } else {
942 if (st->state <= target)
943 return false;
944
945 *state_to_run = st->state--;
946 }
947
948 if (!cpuhp_step_empty(bringup, step: cpuhp_get_step(state: *state_to_run)))
949 break;
950 } while (true);
951
952 return true;
953}
954
955static int __cpuhp_invoke_callback_range(bool bringup,
956 unsigned int cpu,
957 struct cpuhp_cpu_state *st,
958 enum cpuhp_state target,
959 bool nofail)
960{
961 enum cpuhp_state state;
962 int ret = 0;
963
964 while (cpuhp_next_state(bringup, state_to_run: &state, st, target)) {
965 int err;
966
967 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
968 if (!err)
969 continue;
970
971 if (nofail) {
972 pr_warn("CPU %u %s state %s (%d) failed (%d)\n",
973 cpu, bringup ? "UP" : "DOWN",
974 cpuhp_get_step(st->state)->name,
975 st->state, err);
976 ret = -1;
977 } else {
978 ret = err;
979 break;
980 }
981 }
982
983 return ret;
984}
985
986static inline int cpuhp_invoke_callback_range(bool bringup,
987 unsigned int cpu,
988 struct cpuhp_cpu_state *st,
989 enum cpuhp_state target)
990{
991 return __cpuhp_invoke_callback_range(bringup, cpu, st, target, nofail: false);
992}
993
994static inline void cpuhp_invoke_callback_range_nofail(bool bringup,
995 unsigned int cpu,
996 struct cpuhp_cpu_state *st,
997 enum cpuhp_state target)
998{
999 __cpuhp_invoke_callback_range(bringup, cpu, st, target, nofail: true);
1000}
1001
1002static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
1003{
1004 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
1005 return true;
1006 /*
1007 * When CPU hotplug is disabled, then taking the CPU down is not
1008 * possible because takedown_cpu() and the architecture and
1009 * subsystem specific mechanisms are not available. So the CPU
1010 * which would be completely unplugged again needs to stay around
1011 * in the current state.
1012 */
1013 return st->state <= CPUHP_BRINGUP_CPU;
1014}
1015
1016static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1017 enum cpuhp_state target)
1018{
1019 enum cpuhp_state prev_state = st->state;
1020 int ret = 0;
1021
1022 ret = cpuhp_invoke_callback_range(bringup: true, cpu, st, target);
1023 if (ret) {
1024 pr_debug("CPU UP failed (%d) CPU %u state %s (%d)\n",
1025 ret, cpu, cpuhp_get_step(st->state)->name,
1026 st->state);
1027
1028 cpuhp_reset_state(cpu, st, prev_state);
1029 if (can_rollback_cpu(st))
1030 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
1031 prev_state));
1032 }
1033 return ret;
1034}
1035
1036/*
1037 * The cpu hotplug threads manage the bringup and teardown of the cpus
1038 */
1039static int cpuhp_should_run(unsigned int cpu)
1040{
1041 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1042
1043 return st->should_run;
1044}
1045
1046/*
1047 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
1048 * callbacks when a state gets [un]installed at runtime.
1049 *
1050 * Each invocation of this function by the smpboot thread does a single AP
1051 * state callback.
1052 *
1053 * It has 3 modes of operation:
1054 * - single: runs st->cb_state
1055 * - up: runs ++st->state, while st->state < st->target
1056 * - down: runs st->state--, while st->state > st->target
1057 *
1058 * When complete or on error, should_run is cleared and the completion is fired.
1059 */
1060static void cpuhp_thread_fun(unsigned int cpu)
1061{
1062 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1063 bool bringup = st->bringup;
1064 enum cpuhp_state state;
1065
1066 if (WARN_ON_ONCE(!st->should_run))
1067 return;
1068
1069 /*
1070 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
1071 * that if we see ->should_run we also see the rest of the state.
1072 */
1073 smp_mb();
1074
1075 /*
1076 * The BP holds the hotplug lock, but we're now running on the AP,
1077 * ensure that anybody asserting the lock is held, will actually find
1078 * it so.
1079 */
1080 lockdep_acquire_cpus_lock();
1081 cpuhp_lock_acquire(bringup);
1082
1083 if (st->single) {
1084 state = st->cb_state;
1085 st->should_run = false;
1086 } else {
1087 st->should_run = cpuhp_next_state(bringup, state_to_run: &state, st, target: st->target);
1088 if (!st->should_run)
1089 goto end;
1090 }
1091
1092 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
1093
1094 if (cpuhp_is_atomic_state(state)) {
1095 local_irq_disable();
1096 st->result = cpuhp_invoke_callback(cpu, state, bringup, node: st->node, lastp: &st->last);
1097 local_irq_enable();
1098
1099 /*
1100 * STARTING/DYING must not fail!
1101 */
1102 WARN_ON_ONCE(st->result);
1103 } else {
1104 st->result = cpuhp_invoke_callback(cpu, state, bringup, node: st->node, lastp: &st->last);
1105 }
1106
1107 if (st->result) {
1108 /*
1109 * If we fail on a rollback, we're up a creek without no
1110 * paddle, no way forward, no way back. We loose, thanks for
1111 * playing.
1112 */
1113 WARN_ON_ONCE(st->rollback);
1114 st->should_run = false;
1115 }
1116
1117end:
1118 cpuhp_lock_release(bringup);
1119 lockdep_release_cpus_lock();
1120
1121 if (!st->should_run)
1122 complete_ap_thread(st, bringup);
1123}
1124
1125/* Invoke a single callback on a remote cpu */
1126static int
1127cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
1128 struct hlist_node *node)
1129{
1130 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1131 int ret;
1132
1133 if (!cpu_online(cpu))
1134 return 0;
1135
1136 cpuhp_lock_acquire(bringup: false);
1137 cpuhp_lock_release(bringup: false);
1138
1139 cpuhp_lock_acquire(bringup: true);
1140 cpuhp_lock_release(bringup: true);
1141
1142 /*
1143 * If we are up and running, use the hotplug thread. For early calls
1144 * we invoke the thread function directly.
1145 */
1146 if (!st->thread)
1147 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1148
1149 st->rollback = false;
1150 st->last = NULL;
1151
1152 st->node = node;
1153 st->bringup = bringup;
1154 st->cb_state = state;
1155 st->single = true;
1156
1157 __cpuhp_kick_ap(st);
1158
1159 /*
1160 * If we failed and did a partial, do a rollback.
1161 */
1162 if ((ret = st->result) && st->last) {
1163 st->rollback = true;
1164 st->bringup = !bringup;
1165
1166 __cpuhp_kick_ap(st);
1167 }
1168
1169 /*
1170 * Clean up the leftovers so the next hotplug operation wont use stale
1171 * data.
1172 */
1173 st->node = st->last = NULL;
1174 return ret;
1175}
1176
1177static int cpuhp_kick_ap_work(unsigned int cpu)
1178{
1179 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1180 enum cpuhp_state prev_state = st->state;
1181 int ret;
1182
1183 cpuhp_lock_acquire(bringup: false);
1184 cpuhp_lock_release(bringup: false);
1185
1186 cpuhp_lock_acquire(bringup: true);
1187 cpuhp_lock_release(bringup: true);
1188
1189 trace_cpuhp_enter(cpu, target: st->target, idx: prev_state, fun: cpuhp_kick_ap_work);
1190 ret = cpuhp_kick_ap(cpu, st, target: st->target);
1191 trace_cpuhp_exit(cpu, state: st->state, idx: prev_state, ret);
1192
1193 return ret;
1194}
1195
1196static struct smp_hotplug_thread cpuhp_threads = {
1197 .store = &cpuhp_state.thread,
1198 .thread_should_run = cpuhp_should_run,
1199 .thread_fn = cpuhp_thread_fun,
1200 .thread_comm = "cpuhp/%u",
1201 .selfparking = true,
1202};
1203
1204static __init void cpuhp_init_state(void)
1205{
1206 struct cpuhp_cpu_state *st;
1207 int cpu;
1208
1209 for_each_possible_cpu(cpu) {
1210 st = per_cpu_ptr(&cpuhp_state, cpu);
1211 init_completion(x: &st->done_up);
1212 init_completion(x: &st->done_down);
1213 }
1214}
1215
1216void __init cpuhp_threads_init(void)
1217{
1218 cpuhp_init_state();
1219 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
1220 kthread_unpark(this_cpu_read(cpuhp_state.thread));
1221}
1222
1223#ifdef CONFIG_HOTPLUG_CPU
1224#ifndef arch_clear_mm_cpumask_cpu
1225#define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
1226#endif
1227
1228/**
1229 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
1230 * @cpu: a CPU id
1231 *
1232 * This function walks all processes, finds a valid mm struct for each one and
1233 * then clears a corresponding bit in mm's cpumask. While this all sounds
1234 * trivial, there are various non-obvious corner cases, which this function
1235 * tries to solve in a safe manner.
1236 *
1237 * Also note that the function uses a somewhat relaxed locking scheme, so it may
1238 * be called only for an already offlined CPU.
1239 */
1240void clear_tasks_mm_cpumask(int cpu)
1241{
1242 struct task_struct *p;
1243
1244 /*
1245 * This function is called after the cpu is taken down and marked
1246 * offline, so its not like new tasks will ever get this cpu set in
1247 * their mm mask. -- Peter Zijlstra
1248 * Thus, we may use rcu_read_lock() here, instead of grabbing
1249 * full-fledged tasklist_lock.
1250 */
1251 WARN_ON(cpu_online(cpu));
1252 rcu_read_lock();
1253 for_each_process(p) {
1254 struct task_struct *t;
1255
1256 /*
1257 * Main thread might exit, but other threads may still have
1258 * a valid mm. Find one.
1259 */
1260 t = find_lock_task_mm(p);
1261 if (!t)
1262 continue;
1263 arch_clear_mm_cpumask_cpu(cpu, t->mm);
1264 task_unlock(p: t);
1265 }
1266 rcu_read_unlock();
1267}
1268
1269/* Take this CPU down. */
1270static int take_cpu_down(void *_param)
1271{
1272 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1273 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
1274 int err, cpu = smp_processor_id();
1275
1276 /* Ensure this CPU doesn't handle any more interrupts. */
1277 err = __cpu_disable();
1278 if (err < 0)
1279 return err;
1280
1281 /*
1282 * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
1283 * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
1284 */
1285 WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
1286
1287 /*
1288 * Invoke the former CPU_DYING callbacks. DYING must not fail!
1289 */
1290 cpuhp_invoke_callback_range_nofail(bringup: false, cpu, st, target);
1291
1292 /* Park the stopper thread */
1293 stop_machine_park(cpu);
1294 return 0;
1295}
1296
1297static int takedown_cpu(unsigned int cpu)
1298{
1299 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1300 int err;
1301
1302 /* Park the smpboot threads */
1303 kthread_park(k: st->thread);
1304
1305 /*
1306 * Prevent irq alloc/free while the dying cpu reorganizes the
1307 * interrupt affinities.
1308 */
1309 irq_lock_sparse();
1310
1311 /*
1312 * So now all preempt/rcu users must observe !cpu_active().
1313 */
1314 err = stop_machine_cpuslocked(fn: take_cpu_down, NULL, cpumask_of(cpu));
1315 if (err) {
1316 /* CPU refused to die */
1317 irq_unlock_sparse();
1318 /* Unpark the hotplug thread so we can rollback there */
1319 kthread_unpark(k: st->thread);
1320 return err;
1321 }
1322 BUG_ON(cpu_online(cpu));
1323
1324 /*
1325 * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1326 * all runnable tasks from the CPU, there's only the idle task left now
1327 * that the migration thread is done doing the stop_machine thing.
1328 *
1329 * Wait for the stop thread to go away.
1330 */
1331 wait_for_ap_thread(st, bringup: false);
1332 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1333
1334 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1335 irq_unlock_sparse();
1336
1337 hotplug_cpu__broadcast_tick_pull(dead_cpu: cpu);
1338 /* This actually kills the CPU. */
1339 __cpu_die(cpu);
1340
1341 cpuhp_bp_sync_dead(cpu);
1342
1343 lockdep_cleanup_dead_cpu(cpu, idle: idle_thread_get(cpu));
1344
1345 /*
1346 * Callbacks must be re-integrated right away to the RCU state machine.
1347 * Otherwise an RCU callback could block a further teardown function
1348 * waiting for its completion.
1349 */
1350 rcutree_migrate_callbacks(cpu);
1351
1352 return 0;
1353}
1354
1355static void cpuhp_complete_idle_dead(void *arg)
1356{
1357 struct cpuhp_cpu_state *st = arg;
1358
1359 complete_ap_thread(st, bringup: false);
1360}
1361
1362void cpuhp_report_idle_dead(void)
1363{
1364 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1365
1366 BUG_ON(st->state != CPUHP_AP_OFFLINE);
1367 tick_assert_timekeeping_handover();
1368 rcutree_report_cpu_dead();
1369 st->state = CPUHP_AP_IDLE_DEAD;
1370 /*
1371 * We cannot call complete after rcutree_report_cpu_dead() so we delegate it
1372 * to an online cpu.
1373 */
1374 smp_call_function_single(cpuid: cpumask_first(cpu_online_mask),
1375 func: cpuhp_complete_idle_dead, info: st, wait: 0);
1376}
1377
1378static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1379 enum cpuhp_state target)
1380{
1381 enum cpuhp_state prev_state = st->state;
1382 int ret = 0;
1383
1384 ret = cpuhp_invoke_callback_range(bringup: false, cpu, st, target);
1385 if (ret) {
1386 pr_debug("CPU DOWN failed (%d) CPU %u state %s (%d)\n",
1387 ret, cpu, cpuhp_get_step(st->state)->name,
1388 st->state);
1389
1390 cpuhp_reset_state(cpu, st, prev_state);
1391
1392 if (st->state < prev_state)
1393 WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1394 prev_state));
1395 }
1396
1397 return ret;
1398}
1399
1400/* Requires cpu_add_remove_lock to be held */
1401static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1402 enum cpuhp_state target)
1403{
1404 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1405 int prev_state, ret = 0;
1406
1407 if (num_online_cpus() == 1)
1408 return -EBUSY;
1409
1410 if (!cpu_present(cpu))
1411 return -EINVAL;
1412
1413 cpus_write_lock();
1414
1415 cpuhp_tasks_frozen = tasks_frozen;
1416
1417 prev_state = cpuhp_set_state(cpu, st, target);
1418 /*
1419 * If the current CPU state is in the range of the AP hotplug thread,
1420 * then we need to kick the thread.
1421 */
1422 if (st->state > CPUHP_TEARDOWN_CPU) {
1423 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1424 ret = cpuhp_kick_ap_work(cpu);
1425 /*
1426 * The AP side has done the error rollback already. Just
1427 * return the error code..
1428 */
1429 if (ret)
1430 goto out;
1431
1432 /*
1433 * We might have stopped still in the range of the AP hotplug
1434 * thread. Nothing to do anymore.
1435 */
1436 if (st->state > CPUHP_TEARDOWN_CPU)
1437 goto out;
1438
1439 st->target = target;
1440 }
1441 /*
1442 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1443 * to do the further cleanups.
1444 */
1445 ret = cpuhp_down_callbacks(cpu, st, target);
1446 if (ret && st->state < prev_state) {
1447 if (st->state == CPUHP_TEARDOWN_CPU) {
1448 cpuhp_reset_state(cpu, st, prev_state);
1449 __cpuhp_kick_ap(st);
1450 } else {
1451 WARN(1, "DEAD callback error for CPU%d", cpu);
1452 }
1453 }
1454
1455out:
1456 cpus_write_unlock();
1457 arch_smt_update();
1458 return ret;
1459}
1460
1461struct cpu_down_work {
1462 unsigned int cpu;
1463 enum cpuhp_state target;
1464};
1465
1466static long __cpu_down_maps_locked(void *arg)
1467{
1468 struct cpu_down_work *work = arg;
1469
1470 return _cpu_down(cpu: work->cpu, tasks_frozen: 0, target: work->target);
1471}
1472
1473static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1474{
1475 struct cpu_down_work work = { .cpu = cpu, .target = target, };
1476
1477 /*
1478 * If the platform does not support hotplug, report it explicitly to
1479 * differentiate it from a transient offlining failure.
1480 */
1481 if (cpu_hotplug_offline_disabled)
1482 return -EOPNOTSUPP;
1483 if (cpu_hotplug_disabled)
1484 return -EBUSY;
1485
1486 /*
1487 * Ensure that the control task does not run on the to be offlined
1488 * CPU to prevent a deadlock against cfs_b->period_timer.
1489 * Also keep at least one housekeeping cpu onlined to avoid generating
1490 * an empty sched_domain span.
1491 */
1492 for_each_cpu_and(cpu, cpu_online_mask, housekeeping_cpumask(HK_TYPE_DOMAIN)) {
1493 if (cpu != work.cpu)
1494 return work_on_cpu(cpu, __cpu_down_maps_locked, &work);
1495 }
1496 return -EBUSY;
1497}
1498
1499static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1500{
1501 int err;
1502
1503 cpu_maps_update_begin();
1504 err = cpu_down_maps_locked(cpu, target);
1505 cpu_maps_update_done();
1506 return err;
1507}
1508
1509/**
1510 * cpu_device_down - Bring down a cpu device
1511 * @dev: Pointer to the cpu device to offline
1512 *
1513 * This function is meant to be used by device core cpu subsystem only.
1514 *
1515 * Other subsystems should use remove_cpu() instead.
1516 *
1517 * Return: %0 on success or a negative errno code
1518 */
1519int cpu_device_down(struct device *dev)
1520{
1521 return cpu_down(cpu: dev->id, target: CPUHP_OFFLINE);
1522}
1523
1524int remove_cpu(unsigned int cpu)
1525{
1526 int ret;
1527
1528 lock_device_hotplug();
1529 ret = device_offline(dev: get_cpu_device(cpu));
1530 unlock_device_hotplug();
1531
1532 return ret;
1533}
1534EXPORT_SYMBOL_GPL(remove_cpu);
1535
1536void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1537{
1538 unsigned int cpu;
1539 int error;
1540
1541 cpu_maps_update_begin();
1542
1543 /*
1544 * Make certain the cpu I'm about to reboot on is online.
1545 *
1546 * This is inline to what migrate_to_reboot_cpu() already do.
1547 */
1548 if (!cpu_online(cpu: primary_cpu))
1549 primary_cpu = cpumask_first(cpu_online_mask);
1550
1551 for_each_online_cpu(cpu) {
1552 if (cpu == primary_cpu)
1553 continue;
1554
1555 error = cpu_down_maps_locked(cpu, target: CPUHP_OFFLINE);
1556 if (error) {
1557 pr_err("Failed to offline CPU%d - error=%d",
1558 cpu, error);
1559 break;
1560 }
1561 }
1562
1563 /*
1564 * Ensure all but the reboot CPU are offline.
1565 */
1566 BUG_ON(num_online_cpus() > 1);
1567
1568 /*
1569 * Make sure the CPUs won't be enabled by someone else after this
1570 * point. Kexec will reboot to a new kernel shortly resetting
1571 * everything along the way.
1572 */
1573 cpu_hotplug_disabled++;
1574
1575 cpu_maps_update_done();
1576}
1577
1578#else
1579#define takedown_cpu NULL
1580#endif /*CONFIG_HOTPLUG_CPU*/
1581
1582/**
1583 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1584 * @cpu: cpu that just started
1585 *
1586 * It must be called by the arch code on the new cpu, before the new cpu
1587 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1588 */
1589void notify_cpu_starting(unsigned int cpu)
1590{
1591 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1592 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1593
1594 rcutree_report_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1595 cpumask_set_cpu(cpu, dstp: &cpus_booted_once_mask);
1596
1597 /*
1598 * STARTING must not fail!
1599 */
1600 cpuhp_invoke_callback_range_nofail(bringup: true, cpu, st, target);
1601}
1602
1603/*
1604 * Called from the idle task. Wake up the controlling task which brings the
1605 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1606 * online bringup to the hotplug thread.
1607 */
1608void cpuhp_online_idle(enum cpuhp_state state)
1609{
1610 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1611
1612 /* Happens for the boot cpu */
1613 if (state != CPUHP_AP_ONLINE_IDLE)
1614 return;
1615
1616 cpuhp_ap_update_sync_state(state: SYNC_STATE_ONLINE);
1617
1618 /*
1619 * Unpark the stopper thread before we start the idle loop (and start
1620 * scheduling); this ensures the stopper task is always available.
1621 */
1622 stop_machine_unpark(smp_processor_id());
1623
1624 st->state = CPUHP_AP_ONLINE_IDLE;
1625 complete_ap_thread(st, bringup: true);
1626}
1627
1628/* Requires cpu_add_remove_lock to be held */
1629static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1630{
1631 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1632 struct task_struct *idle;
1633 int ret = 0;
1634
1635 cpus_write_lock();
1636
1637 if (!cpu_present(cpu)) {
1638 ret = -EINVAL;
1639 goto out;
1640 }
1641
1642 /*
1643 * The caller of cpu_up() might have raced with another
1644 * caller. Nothing to do.
1645 */
1646 if (st->state >= target)
1647 goto out;
1648
1649 if (st->state == CPUHP_OFFLINE) {
1650 /* Let it fail before we try to bring the cpu up */
1651 idle = idle_thread_get(cpu);
1652 if (IS_ERR(ptr: idle)) {
1653 ret = PTR_ERR(ptr: idle);
1654 goto out;
1655 }
1656
1657 /*
1658 * Reset stale stack state from the last time this CPU was online.
1659 */
1660 scs_task_reset(tsk: idle);
1661 kasan_unpoison_task_stack(task: idle);
1662 }
1663
1664 cpuhp_tasks_frozen = tasks_frozen;
1665
1666 cpuhp_set_state(cpu, st, target);
1667 /*
1668 * If the current CPU state is in the range of the AP hotplug thread,
1669 * then we need to kick the thread once more.
1670 */
1671 if (st->state > CPUHP_BRINGUP_CPU) {
1672 ret = cpuhp_kick_ap_work(cpu);
1673 /*
1674 * The AP side has done the error rollback already. Just
1675 * return the error code..
1676 */
1677 if (ret)
1678 goto out;
1679 }
1680
1681 /*
1682 * Try to reach the target state. We max out on the BP at
1683 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1684 * responsible for bringing it up to the target state.
1685 */
1686 target = min((int)target, CPUHP_BRINGUP_CPU);
1687 ret = cpuhp_up_callbacks(cpu, st, target);
1688out:
1689 cpus_write_unlock();
1690 arch_smt_update();
1691 return ret;
1692}
1693
1694static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1695{
1696 int err = 0;
1697
1698 if (!cpu_possible(cpu)) {
1699 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1700 cpu);
1701 return -EINVAL;
1702 }
1703
1704 err = try_online_node(cpu_to_node(cpu));
1705 if (err)
1706 return err;
1707
1708 cpu_maps_update_begin();
1709
1710 if (cpu_hotplug_disabled) {
1711 err = -EBUSY;
1712 goto out;
1713 }
1714 if (!cpu_bootable(cpu)) {
1715 err = -EPERM;
1716 goto out;
1717 }
1718
1719 err = _cpu_up(cpu, tasks_frozen: 0, target);
1720out:
1721 cpu_maps_update_done();
1722 return err;
1723}
1724
1725/**
1726 * cpu_device_up - Bring up a cpu device
1727 * @dev: Pointer to the cpu device to online
1728 *
1729 * This function is meant to be used by device core cpu subsystem only.
1730 *
1731 * Other subsystems should use add_cpu() instead.
1732 *
1733 * Return: %0 on success or a negative errno code
1734 */
1735int cpu_device_up(struct device *dev)
1736{
1737 return cpu_up(cpu: dev->id, target: CPUHP_ONLINE);
1738}
1739
1740int add_cpu(unsigned int cpu)
1741{
1742 int ret;
1743
1744 lock_device_hotplug();
1745 ret = device_online(dev: get_cpu_device(cpu));
1746 unlock_device_hotplug();
1747
1748 return ret;
1749}
1750EXPORT_SYMBOL_GPL(add_cpu);
1751
1752/**
1753 * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1754 * @sleep_cpu: The cpu we hibernated on and should be brought up.
1755 *
1756 * On some architectures like arm64, we can hibernate on any CPU, but on
1757 * wake up the CPU we hibernated on might be offline as a side effect of
1758 * using maxcpus= for example.
1759 *
1760 * Return: %0 on success or a negative errno code
1761 */
1762int bringup_hibernate_cpu(unsigned int sleep_cpu)
1763{
1764 int ret;
1765
1766 if (!cpu_online(cpu: sleep_cpu)) {
1767 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1768 ret = cpu_up(cpu: sleep_cpu, target: CPUHP_ONLINE);
1769 if (ret) {
1770 pr_err("Failed to bring hibernate-CPU up!\n");
1771 return ret;
1772 }
1773 }
1774 return 0;
1775}
1776
1777static void __init cpuhp_bringup_mask(const struct cpumask *mask, unsigned int ncpus,
1778 enum cpuhp_state target)
1779{
1780 unsigned int cpu;
1781
1782 for_each_cpu(cpu, mask) {
1783 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1784
1785 if (cpu_up(cpu, target) && can_rollback_cpu(st)) {
1786 /*
1787 * If this failed then cpu_up() might have only
1788 * rolled back to CPUHP_BP_KICK_AP for the final
1789 * online. Clean it up. NOOP if already rolled back.
1790 */
1791 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st, CPUHP_OFFLINE));
1792 }
1793
1794 if (!--ncpus)
1795 break;
1796 }
1797}
1798
1799#ifdef CONFIG_HOTPLUG_PARALLEL
1800static bool __cpuhp_parallel_bringup __ro_after_init = true;
1801
1802static int __init parallel_bringup_parse_param(char *arg)
1803{
1804 return kstrtobool(s: arg, res: &__cpuhp_parallel_bringup);
1805}
1806early_param("cpuhp.parallel", parallel_bringup_parse_param);
1807
1808#ifdef CONFIG_HOTPLUG_SMT
1809static inline bool cpuhp_smt_aware(void)
1810{
1811 return cpu_smt_max_threads > 1;
1812}
1813
1814static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1815{
1816 return cpu_primary_thread_mask;
1817}
1818#else
1819static inline bool cpuhp_smt_aware(void)
1820{
1821 return false;
1822}
1823static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1824{
1825 return cpu_none_mask;
1826}
1827#endif
1828
1829bool __weak arch_cpuhp_init_parallel_bringup(void)
1830{
1831 return true;
1832}
1833
1834/*
1835 * On architectures which have enabled parallel bringup this invokes all BP
1836 * prepare states for each of the to be onlined APs first. The last state
1837 * sends the startup IPI to the APs. The APs proceed through the low level
1838 * bringup code in parallel and then wait for the control CPU to release
1839 * them one by one for the final onlining procedure.
1840 *
1841 * This avoids waiting for each AP to respond to the startup IPI in
1842 * CPUHP_BRINGUP_CPU.
1843 */
1844static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
1845{
1846 const struct cpumask *mask = cpu_present_mask;
1847
1848 if (__cpuhp_parallel_bringup)
1849 __cpuhp_parallel_bringup = arch_cpuhp_init_parallel_bringup();
1850 if (!__cpuhp_parallel_bringup)
1851 return false;
1852
1853 if (cpuhp_smt_aware()) {
1854 const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
1855 static struct cpumask tmp_mask __initdata;
1856
1857 /*
1858 * X86 requires to prevent that SMT siblings stopped while
1859 * the primary thread does a microcode update for various
1860 * reasons. Bring the primary threads up first.
1861 */
1862 cpumask_and(dstp: &tmp_mask, src1p: mask, src2p: pmask);
1863 cpuhp_bringup_mask(mask: &tmp_mask, ncpus, target: CPUHP_BP_KICK_AP);
1864 cpuhp_bringup_mask(mask: &tmp_mask, ncpus, target: CPUHP_ONLINE);
1865 /* Account for the online CPUs */
1866 ncpus -= num_online_cpus();
1867 if (!ncpus)
1868 return true;
1869 /* Create the mask for secondary CPUs */
1870 cpumask_andnot(dstp: &tmp_mask, src1p: mask, src2p: pmask);
1871 mask = &tmp_mask;
1872 }
1873
1874 /* Bring the not-yet started CPUs up */
1875 cpuhp_bringup_mask(mask, ncpus, target: CPUHP_BP_KICK_AP);
1876 cpuhp_bringup_mask(mask, ncpus, target: CPUHP_ONLINE);
1877 return true;
1878}
1879#else
1880static inline bool cpuhp_bringup_cpus_parallel(unsigned int ncpus) { return false; }
1881#endif /* CONFIG_HOTPLUG_PARALLEL */
1882
1883void __init bringup_nonboot_cpus(unsigned int max_cpus)
1884{
1885 if (!max_cpus)
1886 return;
1887
1888 /* Try parallel bringup optimization if enabled */
1889 if (cpuhp_bringup_cpus_parallel(ncpus: max_cpus))
1890 return;
1891
1892 /* Full per CPU serialized bringup */
1893 cpuhp_bringup_mask(cpu_present_mask, ncpus: max_cpus, target: CPUHP_ONLINE);
1894}
1895
1896#ifdef CONFIG_PM_SLEEP_SMP
1897static cpumask_var_t frozen_cpus;
1898
1899int freeze_secondary_cpus(int primary)
1900{
1901 int cpu, error = 0;
1902
1903 cpu_maps_update_begin();
1904 if (primary == -1) {
1905 primary = cpumask_first(cpu_online_mask);
1906 if (!housekeeping_cpu(cpu: primary, type: HK_TYPE_TIMER))
1907 primary = housekeeping_any_cpu(type: HK_TYPE_TIMER);
1908 } else {
1909 if (!cpu_online(cpu: primary))
1910 primary = cpumask_first(cpu_online_mask);
1911 }
1912
1913 /*
1914 * We take down all of the non-boot CPUs in one shot to avoid races
1915 * with the userspace trying to use the CPU hotplug at the same time
1916 */
1917 cpumask_clear(dstp: frozen_cpus);
1918
1919 pr_info("Disabling non-boot CPUs ...\n");
1920 for (cpu = nr_cpu_ids - 1; cpu >= 0; cpu--) {
1921 if (!cpu_online(cpu) || cpu == primary)
1922 continue;
1923
1924 if (pm_wakeup_pending()) {
1925 pr_info("Wakeup pending. Abort CPU freeze\n");
1926 error = -EBUSY;
1927 break;
1928 }
1929
1930 trace_suspend_resume(TPS("CPU_OFF"), val: cpu, start: true);
1931 error = _cpu_down(cpu, tasks_frozen: 1, target: CPUHP_OFFLINE);
1932 trace_suspend_resume(TPS("CPU_OFF"), val: cpu, start: false);
1933 if (!error)
1934 cpumask_set_cpu(cpu, dstp: frozen_cpus);
1935 else {
1936 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1937 break;
1938 }
1939 }
1940
1941 if (!error)
1942 BUG_ON(num_online_cpus() > 1);
1943 else
1944 pr_err("Non-boot CPUs are not disabled\n");
1945
1946 /*
1947 * Make sure the CPUs won't be enabled by someone else. We need to do
1948 * this even in case of failure as all freeze_secondary_cpus() users are
1949 * supposed to do thaw_secondary_cpus() on the failure path.
1950 */
1951 cpu_hotplug_disabled++;
1952
1953 cpu_maps_update_done();
1954 return error;
1955}
1956
1957void __weak arch_thaw_secondary_cpus_begin(void)
1958{
1959}
1960
1961void __weak arch_thaw_secondary_cpus_end(void)
1962{
1963}
1964
1965void thaw_secondary_cpus(void)
1966{
1967 int cpu, error;
1968
1969 /* Allow everyone to use the CPU hotplug again */
1970 cpu_maps_update_begin();
1971 __cpu_hotplug_enable();
1972 if (cpumask_empty(srcp: frozen_cpus))
1973 goto out;
1974
1975 pr_info("Enabling non-boot CPUs ...\n");
1976
1977 arch_thaw_secondary_cpus_begin();
1978
1979 for_each_cpu(cpu, frozen_cpus) {
1980 trace_suspend_resume(TPS("CPU_ON"), val: cpu, start: true);
1981 error = _cpu_up(cpu, tasks_frozen: 1, target: CPUHP_ONLINE);
1982 trace_suspend_resume(TPS("CPU_ON"), val: cpu, start: false);
1983 if (!error) {
1984 pr_info("CPU%d is up\n", cpu);
1985 continue;
1986 }
1987 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1988 }
1989
1990 arch_thaw_secondary_cpus_end();
1991
1992 cpumask_clear(dstp: frozen_cpus);
1993out:
1994 cpu_maps_update_done();
1995}
1996
1997static int __init alloc_frozen_cpus(void)
1998{
1999 if (!alloc_cpumask_var(mask: &frozen_cpus, GFP_KERNEL|__GFP_ZERO))
2000 return -ENOMEM;
2001 return 0;
2002}
2003core_initcall(alloc_frozen_cpus);
2004
2005/*
2006 * When callbacks for CPU hotplug notifications are being executed, we must
2007 * ensure that the state of the system with respect to the tasks being frozen
2008 * or not, as reported by the notification, remains unchanged *throughout the
2009 * duration* of the execution of the callbacks.
2010 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
2011 *
2012 * This synchronization is implemented by mutually excluding regular CPU
2013 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
2014 * Hibernate notifications.
2015 */
2016static int
2017cpu_hotplug_pm_callback(struct notifier_block *nb,
2018 unsigned long action, void *ptr)
2019{
2020 switch (action) {
2021
2022 case PM_SUSPEND_PREPARE:
2023 case PM_HIBERNATION_PREPARE:
2024 cpu_hotplug_disable();
2025 break;
2026
2027 case PM_POST_SUSPEND:
2028 case PM_POST_HIBERNATION:
2029 cpu_hotplug_enable();
2030 break;
2031
2032 default:
2033 return NOTIFY_DONE;
2034 }
2035
2036 return NOTIFY_OK;
2037}
2038
2039
2040static int __init cpu_hotplug_pm_sync_init(void)
2041{
2042 /*
2043 * cpu_hotplug_pm_callback has higher priority than x86
2044 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
2045 * to disable cpu hotplug to avoid cpu hotplug race.
2046 */
2047 pm_notifier(cpu_hotplug_pm_callback, 0);
2048 return 0;
2049}
2050core_initcall(cpu_hotplug_pm_sync_init);
2051
2052#endif /* CONFIG_PM_SLEEP_SMP */
2053
2054int __boot_cpu_id;
2055
2056#endif /* CONFIG_SMP */
2057
2058/* Boot processor state steps */
2059static struct cpuhp_step cpuhp_hp_states[] = {
2060 [CPUHP_OFFLINE] = {
2061 .name = "offline",
2062 .startup.single = NULL,
2063 .teardown.single = NULL,
2064 },
2065#ifdef CONFIG_SMP
2066 [CPUHP_CREATE_THREADS]= {
2067 .name = "threads:prepare",
2068 .startup.single = smpboot_create_threads,
2069 .teardown.single = NULL,
2070 .cant_stop = true,
2071 },
2072 [CPUHP_RANDOM_PREPARE] = {
2073 .name = "random:prepare",
2074 .startup.single = random_prepare_cpu,
2075 .teardown.single = NULL,
2076 },
2077 [CPUHP_WORKQUEUE_PREP] = {
2078 .name = "workqueue:prepare",
2079 .startup.single = workqueue_prepare_cpu,
2080 .teardown.single = NULL,
2081 },
2082 [CPUHP_HRTIMERS_PREPARE] = {
2083 .name = "hrtimers:prepare",
2084 .startup.single = hrtimers_prepare_cpu,
2085 .teardown.single = NULL,
2086 },
2087 [CPUHP_SMPCFD_PREPARE] = {
2088 .name = "smpcfd:prepare",
2089 .startup.single = smpcfd_prepare_cpu,
2090 .teardown.single = smpcfd_dead_cpu,
2091 },
2092 [CPUHP_RELAY_PREPARE] = {
2093 .name = "relay:prepare",
2094 .startup.single = relay_prepare_cpu,
2095 .teardown.single = NULL,
2096 },
2097 [CPUHP_RCUTREE_PREP] = {
2098 .name = "RCU/tree:prepare",
2099 .startup.single = rcutree_prepare_cpu,
2100 .teardown.single = rcutree_dead_cpu,
2101 },
2102 /*
2103 * On the tear-down path, timers_dead_cpu() must be invoked
2104 * before blk_mq_queue_reinit_notify() from notify_dead(),
2105 * otherwise a RCU stall occurs.
2106 */
2107 [CPUHP_TIMERS_PREPARE] = {
2108 .name = "timers:prepare",
2109 .startup.single = timers_prepare_cpu,
2110 .teardown.single = timers_dead_cpu,
2111 },
2112
2113#ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
2114 /*
2115 * Kicks the AP alive. AP will wait in cpuhp_ap_sync_alive() until
2116 * the next step will release it.
2117 */
2118 [CPUHP_BP_KICK_AP] = {
2119 .name = "cpu:kick_ap",
2120 .startup.single = cpuhp_kick_ap_alive,
2121 },
2122
2123 /*
2124 * Waits for the AP to reach cpuhp_ap_sync_alive() and then
2125 * releases it for the complete bringup.
2126 */
2127 [CPUHP_BRINGUP_CPU] = {
2128 .name = "cpu:bringup",
2129 .startup.single = cpuhp_bringup_ap,
2130 .teardown.single = finish_cpu,
2131 .cant_stop = true,
2132 },
2133#else
2134 /*
2135 * All-in-one CPU bringup state which includes the kick alive.
2136 */
2137 [CPUHP_BRINGUP_CPU] = {
2138 .name = "cpu:bringup",
2139 .startup.single = bringup_cpu,
2140 .teardown.single = finish_cpu,
2141 .cant_stop = true,
2142 },
2143#endif
2144 /* Final state before CPU kills itself */
2145 [CPUHP_AP_IDLE_DEAD] = {
2146 .name = "idle:dead",
2147 },
2148 /*
2149 * Last state before CPU enters the idle loop to die. Transient state
2150 * for synchronization.
2151 */
2152 [CPUHP_AP_OFFLINE] = {
2153 .name = "ap:offline",
2154 .cant_stop = true,
2155 },
2156 /* First state is scheduler control. Interrupts are disabled */
2157 [CPUHP_AP_SCHED_STARTING] = {
2158 .name = "sched:starting",
2159 .startup.single = sched_cpu_starting,
2160 .teardown.single = sched_cpu_dying,
2161 },
2162 [CPUHP_AP_RCUTREE_DYING] = {
2163 .name = "RCU/tree:dying",
2164 .startup.single = NULL,
2165 .teardown.single = rcutree_dying_cpu,
2166 },
2167 [CPUHP_AP_SMPCFD_DYING] = {
2168 .name = "smpcfd:dying",
2169 .startup.single = NULL,
2170 .teardown.single = smpcfd_dying_cpu,
2171 },
2172 [CPUHP_AP_HRTIMERS_DYING] = {
2173 .name = "hrtimers:dying",
2174 .startup.single = hrtimers_cpu_starting,
2175 .teardown.single = hrtimers_cpu_dying,
2176 },
2177 [CPUHP_AP_TICK_DYING] = {
2178 .name = "tick:dying",
2179 .startup.single = NULL,
2180 .teardown.single = tick_cpu_dying,
2181 },
2182 /* Entry state on starting. Interrupts enabled from here on. Transient
2183 * state for synchronsization */
2184 [CPUHP_AP_ONLINE] = {
2185 .name = "ap:online",
2186 },
2187 /*
2188 * Handled on control processor until the plugged processor manages
2189 * this itself.
2190 */
2191 [CPUHP_TEARDOWN_CPU] = {
2192 .name = "cpu:teardown",
2193 .startup.single = NULL,
2194 .teardown.single = takedown_cpu,
2195 .cant_stop = true,
2196 },
2197
2198 [CPUHP_AP_SCHED_WAIT_EMPTY] = {
2199 .name = "sched:waitempty",
2200 .startup.single = NULL,
2201 .teardown.single = sched_cpu_wait_empty,
2202 },
2203
2204 /* Handle smpboot threads park/unpark */
2205 [CPUHP_AP_SMPBOOT_THREADS] = {
2206 .name = "smpboot/threads:online",
2207 .startup.single = smpboot_unpark_threads,
2208 .teardown.single = smpboot_park_threads,
2209 },
2210 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
2211 .name = "irq/affinity:online",
2212 .startup.single = irq_affinity_online_cpu,
2213 .teardown.single = NULL,
2214 },
2215 [CPUHP_AP_PERF_ONLINE] = {
2216 .name = "perf:online",
2217 .startup.single = perf_event_init_cpu,
2218 .teardown.single = perf_event_exit_cpu,
2219 },
2220 [CPUHP_AP_WATCHDOG_ONLINE] = {
2221 .name = "lockup_detector:online",
2222 .startup.single = lockup_detector_online_cpu,
2223 .teardown.single = lockup_detector_offline_cpu,
2224 },
2225 [CPUHP_AP_WORKQUEUE_ONLINE] = {
2226 .name = "workqueue:online",
2227 .startup.single = workqueue_online_cpu,
2228 .teardown.single = workqueue_offline_cpu,
2229 },
2230 [CPUHP_AP_RANDOM_ONLINE] = {
2231 .name = "random:online",
2232 .startup.single = random_online_cpu,
2233 .teardown.single = NULL,
2234 },
2235 [CPUHP_AP_RCUTREE_ONLINE] = {
2236 .name = "RCU/tree:online",
2237 .startup.single = rcutree_online_cpu,
2238 .teardown.single = rcutree_offline_cpu,
2239 },
2240#endif
2241 /*
2242 * The dynamically registered state space is here
2243 */
2244
2245#ifdef CONFIG_SMP
2246 /* Last state is scheduler control setting the cpu active */
2247 [CPUHP_AP_ACTIVE] = {
2248 .name = "sched:active",
2249 .startup.single = sched_cpu_activate,
2250 .teardown.single = sched_cpu_deactivate,
2251 },
2252#endif
2253
2254 /* CPU is fully up and running. */
2255 [CPUHP_ONLINE] = {
2256 .name = "online",
2257 .startup.single = NULL,
2258 .teardown.single = NULL,
2259 },
2260};
2261
2262/* Sanity check for callbacks */
2263static int cpuhp_cb_check(enum cpuhp_state state)
2264{
2265 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
2266 return -EINVAL;
2267 return 0;
2268}
2269
2270/*
2271 * Returns a free for dynamic slot assignment of the Online state. The states
2272 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
2273 * by having no name assigned.
2274 */
2275static int cpuhp_reserve_state(enum cpuhp_state state)
2276{
2277 enum cpuhp_state i, end;
2278 struct cpuhp_step *step;
2279
2280 switch (state) {
2281 case CPUHP_AP_ONLINE_DYN:
2282 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
2283 end = CPUHP_AP_ONLINE_DYN_END;
2284 break;
2285 case CPUHP_BP_PREPARE_DYN:
2286 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
2287 end = CPUHP_BP_PREPARE_DYN_END;
2288 break;
2289 default:
2290 return -EINVAL;
2291 }
2292
2293 for (i = state; i <= end; i++, step++) {
2294 if (!step->name)
2295 return i;
2296 }
2297 WARN(1, "No more dynamic states available for CPU hotplug\n");
2298 return -ENOSPC;
2299}
2300
2301static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
2302 int (*startup)(unsigned int cpu),
2303 int (*teardown)(unsigned int cpu),
2304 bool multi_instance)
2305{
2306 /* (Un)Install the callbacks for further cpu hotplug operations */
2307 struct cpuhp_step *sp;
2308 int ret = 0;
2309
2310 /*
2311 * If name is NULL, then the state gets removed.
2312 *
2313 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
2314 * the first allocation from these dynamic ranges, so the removal
2315 * would trigger a new allocation and clear the wrong (already
2316 * empty) state, leaving the callbacks of the to be cleared state
2317 * dangling, which causes wreckage on the next hotplug operation.
2318 */
2319 if (name && (state == CPUHP_AP_ONLINE_DYN ||
2320 state == CPUHP_BP_PREPARE_DYN)) {
2321 ret = cpuhp_reserve_state(state);
2322 if (ret < 0)
2323 return ret;
2324 state = ret;
2325 }
2326 sp = cpuhp_get_step(state);
2327 if (name && sp->name)
2328 return -EBUSY;
2329
2330 sp->startup.single = startup;
2331 sp->teardown.single = teardown;
2332 sp->name = name;
2333 sp->multi_instance = multi_instance;
2334 INIT_HLIST_HEAD(&sp->list);
2335 return ret;
2336}
2337
2338static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
2339{
2340 return cpuhp_get_step(state)->teardown.single;
2341}
2342
2343/*
2344 * Call the startup/teardown function for a step either on the AP or
2345 * on the current CPU.
2346 */
2347static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
2348 struct hlist_node *node)
2349{
2350 struct cpuhp_step *sp = cpuhp_get_step(state);
2351 int ret;
2352
2353 /*
2354 * If there's nothing to do, we done.
2355 * Relies on the union for multi_instance.
2356 */
2357 if (cpuhp_step_empty(bringup, step: sp))
2358 return 0;
2359 /*
2360 * The non AP bound callbacks can fail on bringup. On teardown
2361 * e.g. module removal we crash for now.
2362 */
2363#ifdef CONFIG_SMP
2364 if (cpuhp_is_ap_state(state))
2365 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
2366 else
2367 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2368#else
2369 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2370#endif
2371 BUG_ON(ret && !bringup);
2372 return ret;
2373}
2374
2375/*
2376 * Called from __cpuhp_setup_state on a recoverable failure.
2377 *
2378 * Note: The teardown callbacks for rollback are not allowed to fail!
2379 */
2380static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
2381 struct hlist_node *node)
2382{
2383 int cpu;
2384
2385 /* Roll back the already executed steps on the other cpus */
2386 for_each_present_cpu(cpu) {
2387 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2388 int cpustate = st->state;
2389
2390 if (cpu >= failedcpu)
2391 break;
2392
2393 /* Did we invoke the startup call on that cpu ? */
2394 if (cpustate >= state)
2395 cpuhp_issue_call(cpu, state, bringup: false, node);
2396 }
2397}
2398
2399int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
2400 struct hlist_node *node,
2401 bool invoke)
2402{
2403 struct cpuhp_step *sp;
2404 int cpu;
2405 int ret;
2406
2407 lockdep_assert_cpus_held();
2408
2409 sp = cpuhp_get_step(state);
2410 if (sp->multi_instance == false)
2411 return -EINVAL;
2412
2413 mutex_lock(&cpuhp_state_mutex);
2414
2415 if (!invoke || !sp->startup.multi)
2416 goto add_node;
2417
2418 /*
2419 * Try to call the startup callback for each present cpu
2420 * depending on the hotplug state of the cpu.
2421 */
2422 for_each_present_cpu(cpu) {
2423 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2424 int cpustate = st->state;
2425
2426 if (cpustate < state)
2427 continue;
2428
2429 ret = cpuhp_issue_call(cpu, state, bringup: true, node);
2430 if (ret) {
2431 if (sp->teardown.multi)
2432 cpuhp_rollback_install(failedcpu: cpu, state, node);
2433 goto unlock;
2434 }
2435 }
2436add_node:
2437 ret = 0;
2438 hlist_add_head(n: node, h: &sp->list);
2439unlock:
2440 mutex_unlock(lock: &cpuhp_state_mutex);
2441 return ret;
2442}
2443
2444int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
2445 bool invoke)
2446{
2447 int ret;
2448
2449 cpus_read_lock();
2450 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
2451 cpus_read_unlock();
2452 return ret;
2453}
2454EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
2455
2456/**
2457 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
2458 * @state: The state to setup
2459 * @name: Name of the step
2460 * @invoke: If true, the startup function is invoked for cpus where
2461 * cpu state >= @state
2462 * @startup: startup callback function
2463 * @teardown: teardown callback function
2464 * @multi_instance: State is set up for multiple instances which get
2465 * added afterwards.
2466 *
2467 * The caller needs to hold cpus read locked while calling this function.
2468 * Return:
2469 * On success:
2470 * Positive state number if @state is CPUHP_AP_ONLINE_DYN or CPUHP_BP_PREPARE_DYN;
2471 * 0 for all other states
2472 * On failure: proper (negative) error code
2473 */
2474int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
2475 const char *name, bool invoke,
2476 int (*startup)(unsigned int cpu),
2477 int (*teardown)(unsigned int cpu),
2478 bool multi_instance)
2479{
2480 int cpu, ret = 0;
2481 bool dynstate;
2482
2483 lockdep_assert_cpus_held();
2484
2485 if (cpuhp_cb_check(state) || !name)
2486 return -EINVAL;
2487
2488 mutex_lock(&cpuhp_state_mutex);
2489
2490 ret = cpuhp_store_callbacks(state, name, startup, teardown,
2491 multi_instance);
2492
2493 dynstate = state == CPUHP_AP_ONLINE_DYN || state == CPUHP_BP_PREPARE_DYN;
2494 if (ret > 0 && dynstate) {
2495 state = ret;
2496 ret = 0;
2497 }
2498
2499 if (ret || !invoke || !startup)
2500 goto out;
2501
2502 /*
2503 * Try to call the startup callback for each present cpu
2504 * depending on the hotplug state of the cpu.
2505 */
2506 for_each_present_cpu(cpu) {
2507 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2508 int cpustate = st->state;
2509
2510 if (cpustate < state)
2511 continue;
2512
2513 ret = cpuhp_issue_call(cpu, state, bringup: true, NULL);
2514 if (ret) {
2515 if (teardown)
2516 cpuhp_rollback_install(failedcpu: cpu, state, NULL);
2517 cpuhp_store_callbacks(state, NULL, NULL, NULL, multi_instance: false);
2518 goto out;
2519 }
2520 }
2521out:
2522 mutex_unlock(lock: &cpuhp_state_mutex);
2523 /*
2524 * If the requested state is CPUHP_AP_ONLINE_DYN or CPUHP_BP_PREPARE_DYN,
2525 * return the dynamically allocated state in case of success.
2526 */
2527 if (!ret && dynstate)
2528 return state;
2529 return ret;
2530}
2531EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2532
2533int __cpuhp_setup_state(enum cpuhp_state state,
2534 const char *name, bool invoke,
2535 int (*startup)(unsigned int cpu),
2536 int (*teardown)(unsigned int cpu),
2537 bool multi_instance)
2538{
2539 int ret;
2540
2541 cpus_read_lock();
2542 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2543 teardown, multi_instance);
2544 cpus_read_unlock();
2545 return ret;
2546}
2547EXPORT_SYMBOL(__cpuhp_setup_state);
2548
2549int __cpuhp_state_remove_instance(enum cpuhp_state state,
2550 struct hlist_node *node, bool invoke)
2551{
2552 struct cpuhp_step *sp = cpuhp_get_step(state);
2553 int cpu;
2554
2555 BUG_ON(cpuhp_cb_check(state));
2556
2557 if (!sp->multi_instance)
2558 return -EINVAL;
2559
2560 cpus_read_lock();
2561 mutex_lock(&cpuhp_state_mutex);
2562
2563 if (!invoke || !cpuhp_get_teardown_cb(state))
2564 goto remove;
2565 /*
2566 * Call the teardown callback for each present cpu depending
2567 * on the hotplug state of the cpu. This function is not
2568 * allowed to fail currently!
2569 */
2570 for_each_present_cpu(cpu) {
2571 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2572 int cpustate = st->state;
2573
2574 if (cpustate >= state)
2575 cpuhp_issue_call(cpu, state, bringup: false, node);
2576 }
2577
2578remove:
2579 hlist_del(n: node);
2580 mutex_unlock(lock: &cpuhp_state_mutex);
2581 cpus_read_unlock();
2582
2583 return 0;
2584}
2585EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2586
2587/**
2588 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2589 * @state: The state to remove
2590 * @invoke: If true, the teardown function is invoked for cpus where
2591 * cpu state >= @state
2592 *
2593 * The caller needs to hold cpus read locked while calling this function.
2594 * The teardown callback is currently not allowed to fail. Think
2595 * about module removal!
2596 */
2597void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2598{
2599 struct cpuhp_step *sp = cpuhp_get_step(state);
2600 int cpu;
2601
2602 BUG_ON(cpuhp_cb_check(state));
2603
2604 lockdep_assert_cpus_held();
2605
2606 mutex_lock(&cpuhp_state_mutex);
2607 if (sp->multi_instance) {
2608 WARN(!hlist_empty(&sp->list),
2609 "Error: Removing state %d which has instances left.\n",
2610 state);
2611 goto remove;
2612 }
2613
2614 if (!invoke || !cpuhp_get_teardown_cb(state))
2615 goto remove;
2616
2617 /*
2618 * Call the teardown callback for each present cpu depending
2619 * on the hotplug state of the cpu. This function is not
2620 * allowed to fail currently!
2621 */
2622 for_each_present_cpu(cpu) {
2623 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2624 int cpustate = st->state;
2625
2626 if (cpustate >= state)
2627 cpuhp_issue_call(cpu, state, bringup: false, NULL);
2628 }
2629remove:
2630 cpuhp_store_callbacks(state, NULL, NULL, NULL, multi_instance: false);
2631 mutex_unlock(lock: &cpuhp_state_mutex);
2632}
2633EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2634
2635void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2636{
2637 cpus_read_lock();
2638 __cpuhp_remove_state_cpuslocked(state, invoke);
2639 cpus_read_unlock();
2640}
2641EXPORT_SYMBOL(__cpuhp_remove_state);
2642
2643#ifdef CONFIG_HOTPLUG_SMT
2644static void cpuhp_offline_cpu_device(unsigned int cpu)
2645{
2646 struct device *dev = get_cpu_device(cpu);
2647
2648 dev->offline = true;
2649 /* Tell user space about the state change */
2650 kobject_uevent(kobj: &dev->kobj, action: KOBJ_OFFLINE);
2651}
2652
2653static void cpuhp_online_cpu_device(unsigned int cpu)
2654{
2655 struct device *dev = get_cpu_device(cpu);
2656
2657 dev->offline = false;
2658 /* Tell user space about the state change */
2659 kobject_uevent(kobj: &dev->kobj, action: KOBJ_ONLINE);
2660}
2661
2662int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2663{
2664 int cpu, ret = 0;
2665
2666 cpu_maps_update_begin();
2667 for_each_online_cpu(cpu) {
2668 if (topology_is_primary_thread(cpu))
2669 continue;
2670 /*
2671 * Disable can be called with CPU_SMT_ENABLED when changing
2672 * from a higher to lower number of SMT threads per core.
2673 */
2674 if (ctrlval == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
2675 continue;
2676 ret = cpu_down_maps_locked(cpu, target: CPUHP_OFFLINE);
2677 if (ret)
2678 break;
2679 /*
2680 * As this needs to hold the cpu maps lock it's impossible
2681 * to call device_offline() because that ends up calling
2682 * cpu_down() which takes cpu maps lock. cpu maps lock
2683 * needs to be held as this might race against in kernel
2684 * abusers of the hotplug machinery (thermal management).
2685 *
2686 * So nothing would update device:offline state. That would
2687 * leave the sysfs entry stale and prevent onlining after
2688 * smt control has been changed to 'off' again. This is
2689 * called under the sysfs hotplug lock, so it is properly
2690 * serialized against the regular offline usage.
2691 */
2692 cpuhp_offline_cpu_device(cpu);
2693 }
2694 if (!ret)
2695 cpu_smt_control = ctrlval;
2696 cpu_maps_update_done();
2697 return ret;
2698}
2699
2700/* Check if the core a CPU belongs to is online */
2701#if !defined(topology_is_core_online)
2702static inline bool topology_is_core_online(unsigned int cpu)
2703{
2704 return true;
2705}
2706#endif
2707
2708int cpuhp_smt_enable(void)
2709{
2710 int cpu, ret = 0;
2711
2712 cpu_maps_update_begin();
2713 cpu_smt_control = CPU_SMT_ENABLED;
2714 for_each_present_cpu(cpu) {
2715 /* Skip online CPUs and CPUs on offline nodes */
2716 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2717 continue;
2718 if (!cpu_smt_thread_allowed(cpu) || !topology_is_core_online(cpu))
2719 continue;
2720 ret = _cpu_up(cpu, tasks_frozen: 0, target: CPUHP_ONLINE);
2721 if (ret)
2722 break;
2723 /* See comment in cpuhp_smt_disable() */
2724 cpuhp_online_cpu_device(cpu);
2725 }
2726 cpu_maps_update_done();
2727 return ret;
2728}
2729#endif
2730
2731#if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2732static ssize_t state_show(struct device *dev,
2733 struct device_attribute *attr, char *buf)
2734{
2735 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2736
2737 return sprintf(buf, fmt: "%d\n", st->state);
2738}
2739static DEVICE_ATTR_RO(state);
2740
2741static ssize_t target_store(struct device *dev, struct device_attribute *attr,
2742 const char *buf, size_t count)
2743{
2744 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2745 struct cpuhp_step *sp;
2746 int target, ret;
2747
2748 ret = kstrtoint(s: buf, base: 10, res: &target);
2749 if (ret)
2750 return ret;
2751
2752#ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2753 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2754 return -EINVAL;
2755#else
2756 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2757 return -EINVAL;
2758#endif
2759
2760 ret = lock_device_hotplug_sysfs();
2761 if (ret)
2762 return ret;
2763
2764 mutex_lock(&cpuhp_state_mutex);
2765 sp = cpuhp_get_step(state: target);
2766 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2767 mutex_unlock(lock: &cpuhp_state_mutex);
2768 if (ret)
2769 goto out;
2770
2771 if (st->state < target)
2772 ret = cpu_up(cpu: dev->id, target);
2773 else if (st->state > target)
2774 ret = cpu_down(cpu: dev->id, target);
2775 else if (WARN_ON(st->target != target))
2776 st->target = target;
2777out:
2778 unlock_device_hotplug();
2779 return ret ? ret : count;
2780}
2781
2782static ssize_t target_show(struct device *dev,
2783 struct device_attribute *attr, char *buf)
2784{
2785 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2786
2787 return sprintf(buf, fmt: "%d\n", st->target);
2788}
2789static DEVICE_ATTR_RW(target);
2790
2791static ssize_t fail_store(struct device *dev, struct device_attribute *attr,
2792 const char *buf, size_t count)
2793{
2794 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2795 struct cpuhp_step *sp;
2796 int fail, ret;
2797
2798 ret = kstrtoint(s: buf, base: 10, res: &fail);
2799 if (ret)
2800 return ret;
2801
2802 if (fail == CPUHP_INVALID) {
2803 st->fail = fail;
2804 return count;
2805 }
2806
2807 if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2808 return -EINVAL;
2809
2810 /*
2811 * Cannot fail STARTING/DYING callbacks.
2812 */
2813 if (cpuhp_is_atomic_state(state: fail))
2814 return -EINVAL;
2815
2816 /*
2817 * DEAD callbacks cannot fail...
2818 * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2819 * triggering STARTING callbacks, a failure in this state would
2820 * hinder rollback.
2821 */
2822 if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2823 return -EINVAL;
2824
2825 /*
2826 * Cannot fail anything that doesn't have callbacks.
2827 */
2828 mutex_lock(&cpuhp_state_mutex);
2829 sp = cpuhp_get_step(state: fail);
2830 if (!sp->startup.single && !sp->teardown.single)
2831 ret = -EINVAL;
2832 mutex_unlock(lock: &cpuhp_state_mutex);
2833 if (ret)
2834 return ret;
2835
2836 st->fail = fail;
2837
2838 return count;
2839}
2840
2841static ssize_t fail_show(struct device *dev,
2842 struct device_attribute *attr, char *buf)
2843{
2844 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2845
2846 return sprintf(buf, fmt: "%d\n", st->fail);
2847}
2848
2849static DEVICE_ATTR_RW(fail);
2850
2851static struct attribute *cpuhp_cpu_attrs[] = {
2852 &dev_attr_state.attr,
2853 &dev_attr_target.attr,
2854 &dev_attr_fail.attr,
2855 NULL
2856};
2857
2858static const struct attribute_group cpuhp_cpu_attr_group = {
2859 .attrs = cpuhp_cpu_attrs,
2860 .name = "hotplug",
2861};
2862
2863static ssize_t states_show(struct device *dev,
2864 struct device_attribute *attr, char *buf)
2865{
2866 ssize_t cur, res = 0;
2867 int i;
2868
2869 mutex_lock(&cpuhp_state_mutex);
2870 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2871 struct cpuhp_step *sp = cpuhp_get_step(state: i);
2872
2873 if (sp->name) {
2874 cur = sprintf(buf, fmt: "%3d: %s\n", i, sp->name);
2875 buf += cur;
2876 res += cur;
2877 }
2878 }
2879 mutex_unlock(lock: &cpuhp_state_mutex);
2880 return res;
2881}
2882static DEVICE_ATTR_RO(states);
2883
2884static struct attribute *cpuhp_cpu_root_attrs[] = {
2885 &dev_attr_states.attr,
2886 NULL
2887};
2888
2889static const struct attribute_group cpuhp_cpu_root_attr_group = {
2890 .attrs = cpuhp_cpu_root_attrs,
2891 .name = "hotplug",
2892};
2893
2894#ifdef CONFIG_HOTPLUG_SMT
2895
2896static bool cpu_smt_num_threads_valid(unsigned int threads)
2897{
2898 if (IS_ENABLED(CONFIG_SMT_NUM_THREADS_DYNAMIC))
2899 return threads >= 1 && threads <= cpu_smt_max_threads;
2900 return threads == 1 || threads == cpu_smt_max_threads;
2901}
2902
2903static ssize_t
2904__store_smt_control(struct device *dev, struct device_attribute *attr,
2905 const char *buf, size_t count)
2906{
2907 int ctrlval, ret, num_threads, orig_threads;
2908 bool force_off;
2909
2910 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2911 return -EPERM;
2912
2913 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2914 return -ENODEV;
2915
2916 if (sysfs_streq(s1: buf, s2: "on")) {
2917 ctrlval = CPU_SMT_ENABLED;
2918 num_threads = cpu_smt_max_threads;
2919 } else if (sysfs_streq(s1: buf, s2: "off")) {
2920 ctrlval = CPU_SMT_DISABLED;
2921 num_threads = 1;
2922 } else if (sysfs_streq(s1: buf, s2: "forceoff")) {
2923 ctrlval = CPU_SMT_FORCE_DISABLED;
2924 num_threads = 1;
2925 } else if (kstrtoint(s: buf, base: 10, res: &num_threads) == 0) {
2926 if (num_threads == 1)
2927 ctrlval = CPU_SMT_DISABLED;
2928 else if (cpu_smt_num_threads_valid(threads: num_threads))
2929 ctrlval = CPU_SMT_ENABLED;
2930 else
2931 return -EINVAL;
2932 } else {
2933 return -EINVAL;
2934 }
2935
2936 ret = lock_device_hotplug_sysfs();
2937 if (ret)
2938 return ret;
2939
2940 orig_threads = cpu_smt_num_threads;
2941 cpu_smt_num_threads = num_threads;
2942
2943 force_off = ctrlval != cpu_smt_control && ctrlval == CPU_SMT_FORCE_DISABLED;
2944
2945 if (num_threads > orig_threads)
2946 ret = cpuhp_smt_enable();
2947 else if (num_threads < orig_threads || force_off)
2948 ret = cpuhp_smt_disable(ctrlval);
2949
2950 unlock_device_hotplug();
2951 return ret ? ret : count;
2952}
2953
2954#else /* !CONFIG_HOTPLUG_SMT */
2955static ssize_t
2956__store_smt_control(struct device *dev, struct device_attribute *attr,
2957 const char *buf, size_t count)
2958{
2959 return -ENODEV;
2960}
2961#endif /* CONFIG_HOTPLUG_SMT */
2962
2963static const char *smt_states[] = {
2964 [CPU_SMT_ENABLED] = "on",
2965 [CPU_SMT_DISABLED] = "off",
2966 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2967 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2968 [CPU_SMT_NOT_IMPLEMENTED] = "notimplemented",
2969};
2970
2971static ssize_t control_show(struct device *dev,
2972 struct device_attribute *attr, char *buf)
2973{
2974 const char *state = smt_states[cpu_smt_control];
2975
2976#ifdef CONFIG_HOTPLUG_SMT
2977 /*
2978 * If SMT is enabled but not all threads are enabled then show the
2979 * number of threads. If all threads are enabled show "on". Otherwise
2980 * show the state name.
2981 */
2982 if (cpu_smt_control == CPU_SMT_ENABLED &&
2983 cpu_smt_num_threads != cpu_smt_max_threads)
2984 return sysfs_emit(buf, fmt: "%d\n", cpu_smt_num_threads);
2985#endif
2986
2987 return sysfs_emit(buf, fmt: "%s\n", state);
2988}
2989
2990static ssize_t control_store(struct device *dev, struct device_attribute *attr,
2991 const char *buf, size_t count)
2992{
2993 return __store_smt_control(dev, attr, buf, count);
2994}
2995static DEVICE_ATTR_RW(control);
2996
2997static ssize_t active_show(struct device *dev,
2998 struct device_attribute *attr, char *buf)
2999{
3000 return sysfs_emit(buf, fmt: "%d\n", sched_smt_active());
3001}
3002static DEVICE_ATTR_RO(active);
3003
3004static struct attribute *cpuhp_smt_attrs[] = {
3005 &dev_attr_control.attr,
3006 &dev_attr_active.attr,
3007 NULL
3008};
3009
3010static const struct attribute_group cpuhp_smt_attr_group = {
3011 .attrs = cpuhp_smt_attrs,
3012 .name = "smt",
3013};
3014
3015static int __init cpu_smt_sysfs_init(void)
3016{
3017 struct device *dev_root;
3018 int ret = -ENODEV;
3019
3020 dev_root = bus_get_dev_root(bus: &cpu_subsys);
3021 if (dev_root) {
3022 ret = sysfs_create_group(kobj: &dev_root->kobj, grp: &cpuhp_smt_attr_group);
3023 put_device(dev: dev_root);
3024 }
3025 return ret;
3026}
3027
3028static int __init cpuhp_sysfs_init(void)
3029{
3030 struct device *dev_root;
3031 int cpu, ret;
3032
3033 ret = cpu_smt_sysfs_init();
3034 if (ret)
3035 return ret;
3036
3037 dev_root = bus_get_dev_root(bus: &cpu_subsys);
3038 if (dev_root) {
3039 ret = sysfs_create_group(kobj: &dev_root->kobj, grp: &cpuhp_cpu_root_attr_group);
3040 put_device(dev: dev_root);
3041 if (ret)
3042 return ret;
3043 }
3044
3045 for_each_possible_cpu(cpu) {
3046 struct device *dev = get_cpu_device(cpu);
3047
3048 if (!dev)
3049 continue;
3050 ret = sysfs_create_group(kobj: &dev->kobj, grp: &cpuhp_cpu_attr_group);
3051 if (ret)
3052 return ret;
3053 }
3054 return 0;
3055}
3056device_initcall(cpuhp_sysfs_init);
3057#endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
3058
3059/*
3060 * cpu_bit_bitmap[] is a special, "compressed" data structure that
3061 * represents all NR_CPUS bits binary values of 1<<nr.
3062 *
3063 * It is used by cpumask_of() to get a constant address to a CPU
3064 * mask value that has a single bit set only.
3065 */
3066
3067/* cpu_bit_bitmap[0] is empty - so we can back into it */
3068#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
3069#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
3070#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
3071#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
3072
3073const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
3074
3075 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
3076 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
3077#if BITS_PER_LONG > 32
3078 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
3079 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
3080#endif
3081};
3082EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
3083
3084const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
3085EXPORT_SYMBOL(cpu_all_bits);
3086
3087#ifdef CONFIG_INIT_ALL_POSSIBLE
3088struct cpumask __cpu_possible_mask __ro_after_init
3089 = {CPU_BITS_ALL};
3090#else
3091struct cpumask __cpu_possible_mask __ro_after_init;
3092#endif
3093EXPORT_SYMBOL(__cpu_possible_mask);
3094
3095struct cpumask __cpu_online_mask __read_mostly;
3096EXPORT_SYMBOL(__cpu_online_mask);
3097
3098struct cpumask __cpu_enabled_mask __read_mostly;
3099EXPORT_SYMBOL(__cpu_enabled_mask);
3100
3101struct cpumask __cpu_present_mask __read_mostly;
3102EXPORT_SYMBOL(__cpu_present_mask);
3103
3104struct cpumask __cpu_active_mask __read_mostly;
3105EXPORT_SYMBOL(__cpu_active_mask);
3106
3107struct cpumask __cpu_dying_mask __read_mostly;
3108EXPORT_SYMBOL(__cpu_dying_mask);
3109
3110atomic_t __num_online_cpus __read_mostly;
3111EXPORT_SYMBOL(__num_online_cpus);
3112
3113void init_cpu_present(const struct cpumask *src)
3114{
3115 cpumask_copy(dstp: &__cpu_present_mask, srcp: src);
3116}
3117
3118void init_cpu_possible(const struct cpumask *src)
3119{
3120 cpumask_copy(dstp: &__cpu_possible_mask, srcp: src);
3121}
3122
3123void set_cpu_online(unsigned int cpu, bool online)
3124{
3125 /*
3126 * atomic_inc/dec() is required to handle the horrid abuse of this
3127 * function by the reboot and kexec code which invoke it from
3128 * IPI/NMI broadcasts when shutting down CPUs. Invocation from
3129 * regular CPU hotplug is properly serialized.
3130 *
3131 * Note, that the fact that __num_online_cpus is of type atomic_t
3132 * does not protect readers which are not serialized against
3133 * concurrent hotplug operations.
3134 */
3135 if (online) {
3136 if (!cpumask_test_and_set_cpu(cpu, cpumask: &__cpu_online_mask))
3137 atomic_inc(v: &__num_online_cpus);
3138 } else {
3139 if (cpumask_test_and_clear_cpu(cpu, cpumask: &__cpu_online_mask))
3140 atomic_dec(v: &__num_online_cpus);
3141 }
3142}
3143
3144/*
3145 * Activate the first processor.
3146 */
3147void __init boot_cpu_init(void)
3148{
3149 int cpu = smp_processor_id();
3150
3151 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
3152 set_cpu_online(cpu, online: true);
3153 set_cpu_active(cpu, true);
3154 set_cpu_present(cpu, true);
3155 set_cpu_possible(cpu, true);
3156
3157#ifdef CONFIG_SMP
3158 __boot_cpu_id = cpu;
3159#endif
3160}
3161
3162/*
3163 * Must be called _AFTER_ setting up the per_cpu areas
3164 */
3165void __init boot_cpu_hotplug_init(void)
3166{
3167#ifdef CONFIG_SMP
3168 cpumask_set_cpu(smp_processor_id(), dstp: &cpus_booted_once_mask);
3169 atomic_set(this_cpu_ptr(&cpuhp_state.ap_sync_state), i: SYNC_STATE_ONLINE);
3170#endif
3171 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
3172 this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
3173}
3174
3175#ifdef CONFIG_CPU_MITIGATIONS
3176/*
3177 * These are used for a global "mitigations=" cmdline option for toggling
3178 * optional CPU mitigations.
3179 */
3180enum cpu_mitigations {
3181 CPU_MITIGATIONS_OFF,
3182 CPU_MITIGATIONS_AUTO,
3183 CPU_MITIGATIONS_AUTO_NOSMT,
3184};
3185
3186static enum cpu_mitigations cpu_mitigations __ro_after_init = CPU_MITIGATIONS_AUTO;
3187
3188static int __init mitigations_parse_cmdline(char *arg)
3189{
3190 if (!strcmp(arg, "off"))
3191 cpu_mitigations = CPU_MITIGATIONS_OFF;
3192 else if (!strcmp(arg, "auto"))
3193 cpu_mitigations = CPU_MITIGATIONS_AUTO;
3194 else if (!strcmp(arg, "auto,nosmt"))
3195 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
3196 else
3197 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
3198 arg);
3199
3200 return 0;
3201}
3202
3203/* mitigations=off */
3204bool cpu_mitigations_off(void)
3205{
3206 return cpu_mitigations == CPU_MITIGATIONS_OFF;
3207}
3208EXPORT_SYMBOL_GPL(cpu_mitigations_off);
3209
3210/* mitigations=auto,nosmt */
3211bool cpu_mitigations_auto_nosmt(void)
3212{
3213 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
3214}
3215EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);
3216#else
3217static int __init mitigations_parse_cmdline(char *arg)
3218{
3219 pr_crit("Kernel compiled without mitigations, ignoring 'mitigations'; system may still be vulnerable\n");
3220 return 0;
3221}
3222#endif
3223early_param("mitigations", mitigations_parse_cmdline);
3224

Provided by KDAB

Privacy Policy
Improve your Profiling and Debugging skills
Find out more

source code of linux/kernel/cpu.c