2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
4 * This code is licenced under the GPL.
6 #include <linux/sched/mm.h>
7 #include <linux/proc_fs.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>
41 #include <trace/events/power.h>
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/cpuhp.h>
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
66 struct cpuhp_cpu_state {
67 enum cpuhp_state state;
68 enum cpuhp_state target;
69 enum cpuhp_state fail;
71 struct task_struct *thread;
76 struct hlist_node *node;
77 struct hlist_node *last;
78 enum cpuhp_state cb_state;
80 atomic_t ap_sync_state;
81 struct completion done_up;
82 struct completion done_down;
86 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
87 .fail = CPUHP_INVALID,
91 cpumask_t cpus_booted_once_mask;
94 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
95 static struct lockdep_map cpuhp_state_up_map =
96 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
97 static struct lockdep_map cpuhp_state_down_map =
98 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
101 static inline void cpuhp_lock_acquire(bool bringup)
103 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
106 static inline void cpuhp_lock_release(bool bringup)
108 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
112 static inline void cpuhp_lock_acquire(bool bringup) { }
113 static inline void cpuhp_lock_release(bool bringup) { }
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
128 int (*single)(unsigned int cpu);
129 int (*multi)(unsigned int cpu,
130 struct hlist_node *node);
133 int (*single)(unsigned int cpu);
134 int (*multi)(unsigned int cpu,
135 struct hlist_node *node);
138 struct hlist_head list;
144 static DEFINE_MUTEX(cpuhp_state_mutex);
145 static struct cpuhp_step cpuhp_hp_states[];
147 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
149 return cpuhp_hp_states + state;
152 static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
154 return bringup ? !step->startup.single : !step->teardown.single;
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
165 * Called from cpu hotplug and from the state register machinery.
167 * Return: %0 on success or a negative errno code
169 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
170 bool bringup, struct hlist_node *node,
171 struct hlist_node **lastp)
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);
179 if (st->fail == state) {
180 st->fail = CPUHP_INVALID;
184 if (cpuhp_step_empty(bringup, step)) {
189 if (!step->multi_instance) {
190 WARN_ON_ONCE(lastp && *lastp);
191 cb = bringup ? step->startup.single : step->teardown.single;
193 trace_cpuhp_enter(cpu, st->target, state, cb);
195 trace_cpuhp_exit(cpu, st->state, state, ret);
198 cbm = bringup ? step->startup.multi : step->teardown.multi;
200 /* Single invocation for instance add/remove */
202 WARN_ON_ONCE(lastp && *lastp);
203 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
204 ret = cbm(cpu, node);
205 trace_cpuhp_exit(cpu, st->state, state, ret);
209 /* State transition. Invoke on all instances */
211 hlist_for_each(node, &step->list) {
212 if (lastp && node == *lastp)
215 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
216 ret = cbm(cpu, node);
217 trace_cpuhp_exit(cpu, st->state, state, ret);
231 /* Rollback the instances if one failed */
232 cbm = !bringup ? step->startup.multi : step->teardown.multi;
236 hlist_for_each(node, &step->list) {
240 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
241 ret = cbm(cpu, node);
242 trace_cpuhp_exit(cpu, st->state, state, ret);
244 * Rollback must not fail,
252 static bool cpuhp_is_ap_state(enum cpuhp_state state)
255 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
256 * purposes as that state is handled explicitly in cpu_down.
258 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
261 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
263 struct completion *done = bringup ? &st->done_up : &st->done_down;
264 wait_for_completion(done);
267 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
269 struct completion *done = bringup ? &st->done_up : &st->done_down;
274 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
276 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
278 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
281 /* Synchronization state management */
282 enum cpuhp_sync_state {
285 SYNC_STATE_SHOULD_DIE,
287 SYNC_STATE_SHOULD_ONLINE,
291 #ifdef CONFIG_HOTPLUG_CORE_SYNC
293 * cpuhp_ap_update_sync_state - Update synchronization state during bringup/teardown
294 * @state: The synchronization state to set
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.
299 static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state)
301 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
303 (void)atomic_xchg(st, state);
306 void __weak arch_cpuhp_sync_state_poll(void) { cpu_relax(); }
308 static bool cpuhp_wait_for_sync_state(unsigned int cpu, enum cpuhp_sync_state state,
309 enum cpuhp_sync_state next_state)
311 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
312 ktime_t now, end, start = ktime_get();
315 end = start + 10ULL * NSEC_PER_SEC;
317 sync = atomic_read(st);
320 if (!atomic_try_cmpxchg(st, &sync, next_state))
327 /* Timeout. Leave the state unchanged */
329 } else if (now - start < NSEC_PER_MSEC) {
330 /* Poll for one millisecond */
331 arch_cpuhp_sync_state_poll();
333 usleep_range_state(USEC_PER_MSEC, 2 * USEC_PER_MSEC, TASK_UNINTERRUPTIBLE);
335 sync = atomic_read(st);
339 #else /* CONFIG_HOTPLUG_CORE_SYNC */
340 static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state) { }
341 #endif /* !CONFIG_HOTPLUG_CORE_SYNC */
343 #ifdef CONFIG_HOTPLUG_CORE_SYNC_DEAD
345 * cpuhp_ap_report_dead - Update synchronization state to DEAD
347 * No synchronization point. Just update of the synchronization state.
349 void cpuhp_ap_report_dead(void)
351 cpuhp_ap_update_sync_state(SYNC_STATE_DEAD);
354 void __weak arch_cpuhp_cleanup_dead_cpu(unsigned int cpu) { }
357 * Late CPU shutdown synchronization point. Cannot use cpuhp_state::done_down
358 * because the AP cannot issue complete() at this stage.
360 static void cpuhp_bp_sync_dead(unsigned int cpu)
362 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
363 int sync = atomic_read(st);
366 /* CPU can have reported dead already. Don't overwrite that! */
367 if (sync == SYNC_STATE_DEAD)
369 } while (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_SHOULD_DIE));
371 if (cpuhp_wait_for_sync_state(cpu, SYNC_STATE_DEAD, SYNC_STATE_DEAD)) {
372 /* CPU reached dead state. Invoke the cleanup function */
373 arch_cpuhp_cleanup_dead_cpu(cpu);
377 /* No further action possible. Emit message and give up. */
378 pr_err("CPU%u failed to report dead state\n", cpu);
380 #else /* CONFIG_HOTPLUG_CORE_SYNC_DEAD */
381 static inline void cpuhp_bp_sync_dead(unsigned int cpu) { }
382 #endif /* !CONFIG_HOTPLUG_CORE_SYNC_DEAD */
384 #ifdef CONFIG_HOTPLUG_CORE_SYNC_FULL
386 * cpuhp_ap_sync_alive - Synchronize AP with the control CPU once it is alive
388 * Updates the AP synchronization state to SYNC_STATE_ALIVE and waits
389 * for the BP to release it.
391 void cpuhp_ap_sync_alive(void)
393 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
395 cpuhp_ap_update_sync_state(SYNC_STATE_ALIVE);
397 /* Wait for the control CPU to release it. */
398 while (atomic_read(st) != SYNC_STATE_SHOULD_ONLINE)
402 static bool cpuhp_can_boot_ap(unsigned int cpu)
404 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
405 int sync = atomic_read(st);
409 case SYNC_STATE_DEAD:
410 /* CPU is properly dead */
412 case SYNC_STATE_KICKED:
413 /* CPU did not come up in previous attempt */
415 case SYNC_STATE_ALIVE:
416 /* CPU is stuck cpuhp_ap_sync_alive(). */
419 /* CPU failed to report online or dead and is in limbo state. */
423 /* Prepare for booting */
424 if (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_KICKED))
430 void __weak arch_cpuhp_cleanup_kick_cpu(unsigned int cpu) { }
433 * Early CPU bringup synchronization point. Cannot use cpuhp_state::done_up
434 * because the AP cannot issue complete() so early in the bringup.
436 static int cpuhp_bp_sync_alive(unsigned int cpu)
440 if (!IS_ENABLED(CONFIG_HOTPLUG_CORE_SYNC_FULL))
443 if (!cpuhp_wait_for_sync_state(cpu, SYNC_STATE_ALIVE, SYNC_STATE_SHOULD_ONLINE)) {
444 pr_err("CPU%u failed to report alive state\n", cpu);
448 /* Let the architecture cleanup the kick alive mechanics. */
449 arch_cpuhp_cleanup_kick_cpu(cpu);
452 #else /* CONFIG_HOTPLUG_CORE_SYNC_FULL */
453 static inline int cpuhp_bp_sync_alive(unsigned int cpu) { return 0; }
454 static inline bool cpuhp_can_boot_ap(unsigned int cpu) { return true; }
455 #endif /* !CONFIG_HOTPLUG_CORE_SYNC_FULL */
457 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
458 static DEFINE_MUTEX(cpu_add_remove_lock);
459 bool cpuhp_tasks_frozen;
460 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
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.
466 void cpu_maps_update_begin(void)
468 mutex_lock(&cpu_add_remove_lock);
471 void cpu_maps_update_done(void)
473 mutex_unlock(&cpu_add_remove_lock);
477 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
478 * Should always be manipulated under cpu_add_remove_lock
480 static int cpu_hotplug_disabled;
482 #ifdef CONFIG_HOTPLUG_CPU
484 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
486 void cpus_read_lock(void)
488 percpu_down_read(&cpu_hotplug_lock);
490 EXPORT_SYMBOL_GPL(cpus_read_lock);
492 int cpus_read_trylock(void)
494 return percpu_down_read_trylock(&cpu_hotplug_lock);
496 EXPORT_SYMBOL_GPL(cpus_read_trylock);
498 void cpus_read_unlock(void)
500 percpu_up_read(&cpu_hotplug_lock);
502 EXPORT_SYMBOL_GPL(cpus_read_unlock);
504 void cpus_write_lock(void)
506 percpu_down_write(&cpu_hotplug_lock);
509 void cpus_write_unlock(void)
511 percpu_up_write(&cpu_hotplug_lock);
514 void lockdep_assert_cpus_held(void)
517 * We can't have hotplug operations before userspace starts running,
518 * and some init codepaths will knowingly not take the hotplug lock.
519 * This is all valid, so mute lockdep until it makes sense to report
522 if (system_state < SYSTEM_RUNNING)
525 percpu_rwsem_assert_held(&cpu_hotplug_lock);
528 #ifdef CONFIG_LOCKDEP
529 int lockdep_is_cpus_held(void)
531 return percpu_rwsem_is_held(&cpu_hotplug_lock);
535 static void lockdep_acquire_cpus_lock(void)
537 rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
540 static void lockdep_release_cpus_lock(void)
542 rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
546 * Wait for currently running CPU hotplug operations to complete (if any) and
547 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
548 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
549 * hotplug path before performing hotplug operations. So acquiring that lock
550 * guarantees mutual exclusion from any currently running hotplug operations.
552 void cpu_hotplug_disable(void)
554 cpu_maps_update_begin();
555 cpu_hotplug_disabled++;
556 cpu_maps_update_done();
558 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
560 static void __cpu_hotplug_enable(void)
562 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
564 cpu_hotplug_disabled--;
567 void cpu_hotplug_enable(void)
569 cpu_maps_update_begin();
570 __cpu_hotplug_enable();
571 cpu_maps_update_done();
573 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
577 static void lockdep_acquire_cpus_lock(void)
581 static void lockdep_release_cpus_lock(void)
585 #endif /* CONFIG_HOTPLUG_CPU */
588 * Architectures that need SMT-specific errata handling during SMT hotplug
589 * should override this.
591 void __weak arch_smt_update(void) { }
593 #ifdef CONFIG_HOTPLUG_SMT
595 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
596 static unsigned int cpu_smt_max_threads __ro_after_init;
597 unsigned int cpu_smt_num_threads __read_mostly = UINT_MAX;
599 void __init cpu_smt_disable(bool force)
601 if (!cpu_smt_possible())
605 pr_info("SMT: Force disabled\n");
606 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
608 pr_info("SMT: disabled\n");
609 cpu_smt_control = CPU_SMT_DISABLED;
611 cpu_smt_num_threads = 1;
615 * The decision whether SMT is supported can only be done after the full
616 * CPU identification. Called from architecture code.
618 void __init cpu_smt_set_num_threads(unsigned int num_threads,
619 unsigned int max_threads)
621 WARN_ON(!num_threads || (num_threads > max_threads));
623 if (max_threads == 1)
624 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
626 cpu_smt_max_threads = max_threads;
629 * If SMT has been disabled via the kernel command line or SMT is
630 * not supported, set cpu_smt_num_threads to 1 for consistency.
631 * If enabled, take the architecture requested number of threads
632 * to bring up into account.
634 if (cpu_smt_control != CPU_SMT_ENABLED)
635 cpu_smt_num_threads = 1;
636 else if (num_threads < cpu_smt_num_threads)
637 cpu_smt_num_threads = num_threads;
640 static int __init smt_cmdline_disable(char *str)
642 cpu_smt_disable(str && !strcmp(str, "force"));
645 early_param("nosmt", smt_cmdline_disable);
648 * For Archicture supporting partial SMT states check if the thread is allowed.
649 * Otherwise this has already been checked through cpu_smt_max_threads when
650 * setting the SMT level.
652 static inline bool cpu_smt_thread_allowed(unsigned int cpu)
654 #ifdef CONFIG_SMT_NUM_THREADS_DYNAMIC
655 return topology_smt_thread_allowed(cpu);
661 static inline bool cpu_bootable(unsigned int cpu)
663 if (cpu_smt_control == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
666 /* All CPUs are bootable if controls are not configured */
667 if (cpu_smt_control == CPU_SMT_NOT_IMPLEMENTED)
670 /* All CPUs are bootable if CPU is not SMT capable */
671 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
674 if (topology_is_primary_thread(cpu))
678 * On x86 it's required to boot all logical CPUs at least once so
679 * that the init code can get a chance to set CR4.MCE on each
680 * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
681 * core will shutdown the machine.
683 return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
686 /* Returns true if SMT is supported and not forcefully (irreversibly) disabled */
687 bool cpu_smt_possible(void)
689 return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
690 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
692 EXPORT_SYMBOL_GPL(cpu_smt_possible);
695 static inline bool cpu_bootable(unsigned int cpu) { return true; }
698 static inline enum cpuhp_state
699 cpuhp_set_state(int cpu, struct cpuhp_cpu_state *st, enum cpuhp_state target)
701 enum cpuhp_state prev_state = st->state;
702 bool bringup = st->state < target;
704 st->rollback = false;
709 st->bringup = bringup;
710 if (cpu_dying(cpu) != !bringup)
711 set_cpu_dying(cpu, !bringup);
717 cpuhp_reset_state(int cpu, struct cpuhp_cpu_state *st,
718 enum cpuhp_state prev_state)
720 bool bringup = !st->bringup;
722 st->target = prev_state;
725 * Already rolling back. No need invert the bringup value or to change
734 * If we have st->last we need to undo partial multi_instance of this
735 * state first. Otherwise start undo at the previous state.
744 st->bringup = bringup;
745 if (cpu_dying(cpu) != !bringup)
746 set_cpu_dying(cpu, !bringup);
749 /* Regular hotplug invocation of the AP hotplug thread */
750 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
752 if (!st->single && st->state == st->target)
757 * Make sure the above stores are visible before should_run becomes
758 * true. Paired with the mb() above in cpuhp_thread_fun()
761 st->should_run = true;
762 wake_up_process(st->thread);
763 wait_for_ap_thread(st, st->bringup);
766 static int cpuhp_kick_ap(int cpu, struct cpuhp_cpu_state *st,
767 enum cpuhp_state target)
769 enum cpuhp_state prev_state;
772 prev_state = cpuhp_set_state(cpu, st, target);
774 if ((ret = st->result)) {
775 cpuhp_reset_state(cpu, st, prev_state);
782 static int bringup_wait_for_ap_online(unsigned int cpu)
784 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
786 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
787 wait_for_ap_thread(st, true);
788 if (WARN_ON_ONCE((!cpu_online(cpu))))
791 /* Unpark the hotplug thread of the target cpu */
792 kthread_unpark(st->thread);
795 * SMT soft disabling on X86 requires to bring the CPU out of the
796 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
797 * CPU marked itself as booted_once in notify_cpu_starting() so the
798 * cpu_bootable() check will now return false if this is not the
801 if (!cpu_bootable(cpu))
806 #ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
807 static int cpuhp_kick_ap_alive(unsigned int cpu)
809 if (!cpuhp_can_boot_ap(cpu))
812 return arch_cpuhp_kick_ap_alive(cpu, idle_thread_get(cpu));
815 static int cpuhp_bringup_ap(unsigned int cpu)
817 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
821 * Some architectures have to walk the irq descriptors to
822 * setup the vector space for the cpu which comes online.
823 * Prevent irq alloc/free across the bringup.
827 ret = cpuhp_bp_sync_alive(cpu);
831 ret = bringup_wait_for_ap_online(cpu);
837 if (st->target <= CPUHP_AP_ONLINE_IDLE)
840 return cpuhp_kick_ap(cpu, st, st->target);
847 static int bringup_cpu(unsigned int cpu)
849 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
850 struct task_struct *idle = idle_thread_get(cpu);
853 if (!cpuhp_can_boot_ap(cpu))
857 * Some architectures have to walk the irq descriptors to
858 * setup the vector space for the cpu which comes online.
860 * Prevent irq alloc/free across the bringup by acquiring the
861 * sparse irq lock. Hold it until the upcoming CPU completes the
862 * startup in cpuhp_online_idle() which allows to avoid
863 * intermediate synchronization points in the architecture code.
867 ret = __cpu_up(cpu, idle);
871 ret = cpuhp_bp_sync_alive(cpu);
875 ret = bringup_wait_for_ap_online(cpu);
881 if (st->target <= CPUHP_AP_ONLINE_IDLE)
884 return cpuhp_kick_ap(cpu, st, st->target);
892 static int finish_cpu(unsigned int cpu)
894 struct task_struct *idle = idle_thread_get(cpu);
895 struct mm_struct *mm = idle->active_mm;
898 * idle_task_exit() will have switched to &init_mm, now
899 * clean up any remaining active_mm state.
902 idle->active_mm = &init_mm;
908 * Hotplug state machine related functions
912 * Get the next state to run. Empty ones will be skipped. Returns true if a
915 * st->state will be modified ahead of time, to match state_to_run, as if it
918 static bool cpuhp_next_state(bool bringup,
919 enum cpuhp_state *state_to_run,
920 struct cpuhp_cpu_state *st,
921 enum cpuhp_state target)
925 if (st->state >= target)
928 *state_to_run = ++st->state;
930 if (st->state <= target)
933 *state_to_run = st->state--;
936 if (!cpuhp_step_empty(bringup, cpuhp_get_step(*state_to_run)))
943 static int __cpuhp_invoke_callback_range(bool bringup,
945 struct cpuhp_cpu_state *st,
946 enum cpuhp_state target,
949 enum cpuhp_state state;
952 while (cpuhp_next_state(bringup, &state, st, target)) {
955 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
960 pr_warn("CPU %u %s state %s (%d) failed (%d)\n",
961 cpu, bringup ? "UP" : "DOWN",
962 cpuhp_get_step(st->state)->name,
974 static inline int cpuhp_invoke_callback_range(bool bringup,
976 struct cpuhp_cpu_state *st,
977 enum cpuhp_state target)
979 return __cpuhp_invoke_callback_range(bringup, cpu, st, target, false);
982 static inline void cpuhp_invoke_callback_range_nofail(bool bringup,
984 struct cpuhp_cpu_state *st,
985 enum cpuhp_state target)
987 __cpuhp_invoke_callback_range(bringup, cpu, st, target, true);
990 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
992 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
995 * When CPU hotplug is disabled, then taking the CPU down is not
996 * possible because takedown_cpu() and the architecture and
997 * subsystem specific mechanisms are not available. So the CPU
998 * which would be completely unplugged again needs to stay around
999 * in the current state.
1001 return st->state <= CPUHP_BRINGUP_CPU;
1004 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1005 enum cpuhp_state target)
1007 enum cpuhp_state prev_state = st->state;
1010 ret = cpuhp_invoke_callback_range(true, cpu, st, target);
1012 pr_debug("CPU UP failed (%d) CPU %u state %s (%d)\n",
1013 ret, cpu, cpuhp_get_step(st->state)->name,
1016 cpuhp_reset_state(cpu, st, prev_state);
1017 if (can_rollback_cpu(st))
1018 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
1025 * The cpu hotplug threads manage the bringup and teardown of the cpus
1027 static int cpuhp_should_run(unsigned int cpu)
1029 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1031 return st->should_run;
1035 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
1036 * callbacks when a state gets [un]installed at runtime.
1038 * Each invocation of this function by the smpboot thread does a single AP
1041 * It has 3 modes of operation:
1042 * - single: runs st->cb_state
1043 * - up: runs ++st->state, while st->state < st->target
1044 * - down: runs st->state--, while st->state > st->target
1046 * When complete or on error, should_run is cleared and the completion is fired.
1048 static void cpuhp_thread_fun(unsigned int cpu)
1050 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1051 bool bringup = st->bringup;
1052 enum cpuhp_state state;
1054 if (WARN_ON_ONCE(!st->should_run))
1058 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
1059 * that if we see ->should_run we also see the rest of the state.
1064 * The BP holds the hotplug lock, but we're now running on the AP,
1065 * ensure that anybody asserting the lock is held, will actually find
1068 lockdep_acquire_cpus_lock();
1069 cpuhp_lock_acquire(bringup);
1072 state = st->cb_state;
1073 st->should_run = false;
1075 st->should_run = cpuhp_next_state(bringup, &state, st, st->target);
1076 if (!st->should_run)
1080 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
1082 if (cpuhp_is_atomic_state(state)) {
1083 local_irq_disable();
1084 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1088 * STARTING/DYING must not fail!
1090 WARN_ON_ONCE(st->result);
1092 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1097 * If we fail on a rollback, we're up a creek without no
1098 * paddle, no way forward, no way back. We loose, thanks for
1101 WARN_ON_ONCE(st->rollback);
1102 st->should_run = false;
1106 cpuhp_lock_release(bringup);
1107 lockdep_release_cpus_lock();
1109 if (!st->should_run)
1110 complete_ap_thread(st, bringup);
1113 /* Invoke a single callback on a remote cpu */
1115 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
1116 struct hlist_node *node)
1118 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1121 if (!cpu_online(cpu))
1124 cpuhp_lock_acquire(false);
1125 cpuhp_lock_release(false);
1127 cpuhp_lock_acquire(true);
1128 cpuhp_lock_release(true);
1131 * If we are up and running, use the hotplug thread. For early calls
1132 * we invoke the thread function directly.
1135 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1137 st->rollback = false;
1141 st->bringup = bringup;
1142 st->cb_state = state;
1145 __cpuhp_kick_ap(st);
1148 * If we failed and did a partial, do a rollback.
1150 if ((ret = st->result) && st->last) {
1151 st->rollback = true;
1152 st->bringup = !bringup;
1154 __cpuhp_kick_ap(st);
1158 * Clean up the leftovers so the next hotplug operation wont use stale
1161 st->node = st->last = NULL;
1165 static int cpuhp_kick_ap_work(unsigned int cpu)
1167 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1168 enum cpuhp_state prev_state = st->state;
1171 cpuhp_lock_acquire(false);
1172 cpuhp_lock_release(false);
1174 cpuhp_lock_acquire(true);
1175 cpuhp_lock_release(true);
1177 trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
1178 ret = cpuhp_kick_ap(cpu, st, st->target);
1179 trace_cpuhp_exit(cpu, st->state, prev_state, ret);
1184 static struct smp_hotplug_thread cpuhp_threads = {
1185 .store = &cpuhp_state.thread,
1186 .thread_should_run = cpuhp_should_run,
1187 .thread_fn = cpuhp_thread_fun,
1188 .thread_comm = "cpuhp/%u",
1189 .selfparking = true,
1192 static __init void cpuhp_init_state(void)
1194 struct cpuhp_cpu_state *st;
1197 for_each_possible_cpu(cpu) {
1198 st = per_cpu_ptr(&cpuhp_state, cpu);
1199 init_completion(&st->done_up);
1200 init_completion(&st->done_down);
1204 void __init cpuhp_threads_init(void)
1207 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
1208 kthread_unpark(this_cpu_read(cpuhp_state.thread));
1213 * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
1216 * The operation is still serialized against concurrent CPU hotplug via
1217 * cpu_add_remove_lock, i.e. CPU map protection. But it is _not_
1218 * serialized against other hotplug related activity like adding or
1219 * removing of state callbacks and state instances, which invoke either the
1220 * startup or the teardown callback of the affected state.
1222 * This is required for subsystems which are unfixable vs. CPU hotplug and
1223 * evade lock inversion problems by scheduling work which has to be
1224 * completed _before_ cpu_up()/_cpu_down() returns.
1226 * Don't even think about adding anything to this for any new code or even
1227 * drivers. It's only purpose is to keep existing lock order trainwrecks
1230 * For cpu_down() there might be valid reasons to finish cleanups which are
1231 * not required to be done under cpu_hotplug_lock, but that's a different
1232 * story and would be not invoked via this.
1234 static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
1237 * cpusets delegate hotplug operations to a worker to "solve" the
1238 * lock order problems. Wait for the worker, but only if tasks are
1239 * _not_ frozen (suspend, hibernate) as that would wait forever.
1241 * The wait is required because otherwise the hotplug operation
1242 * returns with inconsistent state, which could even be observed in
1243 * user space when a new CPU is brought up. The CPU plug uevent
1244 * would be delivered and user space reacting on it would fail to
1245 * move tasks to the newly plugged CPU up to the point where the
1246 * work has finished because up to that point the newly plugged CPU
1247 * is not assignable in cpusets/cgroups. On unplug that's not
1248 * necessarily a visible issue, but it is still inconsistent state,
1249 * which is the real problem which needs to be "fixed". This can't
1250 * prevent the transient state between scheduling the work and
1251 * returning from waiting for it.
1254 cpuset_wait_for_hotplug();
1257 #ifdef CONFIG_HOTPLUG_CPU
1258 #ifndef arch_clear_mm_cpumask_cpu
1259 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
1263 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
1266 * This function walks all processes, finds a valid mm struct for each one and
1267 * then clears a corresponding bit in mm's cpumask. While this all sounds
1268 * trivial, there are various non-obvious corner cases, which this function
1269 * tries to solve in a safe manner.
1271 * Also note that the function uses a somewhat relaxed locking scheme, so it may
1272 * be called only for an already offlined CPU.
1274 void clear_tasks_mm_cpumask(int cpu)
1276 struct task_struct *p;
1279 * This function is called after the cpu is taken down and marked
1280 * offline, so its not like new tasks will ever get this cpu set in
1281 * their mm mask. -- Peter Zijlstra
1282 * Thus, we may use rcu_read_lock() here, instead of grabbing
1283 * full-fledged tasklist_lock.
1285 WARN_ON(cpu_online(cpu));
1287 for_each_process(p) {
1288 struct task_struct *t;
1291 * Main thread might exit, but other threads may still have
1292 * a valid mm. Find one.
1294 t = find_lock_task_mm(p);
1297 arch_clear_mm_cpumask_cpu(cpu, t->mm);
1303 /* Take this CPU down. */
1304 static int take_cpu_down(void *_param)
1306 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1307 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
1308 int err, cpu = smp_processor_id();
1310 /* Ensure this CPU doesn't handle any more interrupts. */
1311 err = __cpu_disable();
1316 * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
1317 * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
1319 WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
1322 * Invoke the former CPU_DYING callbacks. DYING must not fail!
1324 cpuhp_invoke_callback_range_nofail(false, cpu, st, target);
1326 /* Park the stopper thread */
1327 stop_machine_park(cpu);
1331 static int takedown_cpu(unsigned int cpu)
1333 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1336 /* Park the smpboot threads */
1337 kthread_park(st->thread);
1340 * Prevent irq alloc/free while the dying cpu reorganizes the
1341 * interrupt affinities.
1346 * So now all preempt/rcu users must observe !cpu_active().
1348 err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
1350 /* CPU refused to die */
1351 irq_unlock_sparse();
1352 /* Unpark the hotplug thread so we can rollback there */
1353 kthread_unpark(st->thread);
1356 BUG_ON(cpu_online(cpu));
1359 * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1360 * all runnable tasks from the CPU, there's only the idle task left now
1361 * that the migration thread is done doing the stop_machine thing.
1363 * Wait for the stop thread to go away.
1365 wait_for_ap_thread(st, false);
1366 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1368 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1369 irq_unlock_sparse();
1371 hotplug_cpu__broadcast_tick_pull(cpu);
1372 /* This actually kills the CPU. */
1375 cpuhp_bp_sync_dead(cpu);
1377 tick_cleanup_dead_cpu(cpu);
1380 * Callbacks must be re-integrated right away to the RCU state machine.
1381 * Otherwise an RCU callback could block a further teardown function
1382 * waiting for its completion.
1384 rcutree_migrate_callbacks(cpu);
1389 static void cpuhp_complete_idle_dead(void *arg)
1391 struct cpuhp_cpu_state *st = arg;
1393 complete_ap_thread(st, false);
1396 void cpuhp_report_idle_dead(void)
1398 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1400 BUG_ON(st->state != CPUHP_AP_OFFLINE);
1401 tick_assert_timekeeping_handover();
1402 rcutree_report_cpu_dead();
1403 st->state = CPUHP_AP_IDLE_DEAD;
1405 * We cannot call complete after rcutree_report_cpu_dead() so we delegate it
1408 smp_call_function_single(cpumask_first(cpu_online_mask),
1409 cpuhp_complete_idle_dead, st, 0);
1412 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1413 enum cpuhp_state target)
1415 enum cpuhp_state prev_state = st->state;
1418 ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1420 pr_debug("CPU DOWN failed (%d) CPU %u state %s (%d)\n",
1421 ret, cpu, cpuhp_get_step(st->state)->name,
1424 cpuhp_reset_state(cpu, st, prev_state);
1426 if (st->state < prev_state)
1427 WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1434 /* Requires cpu_add_remove_lock to be held */
1435 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1436 enum cpuhp_state target)
1438 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1439 int prev_state, ret = 0;
1441 if (num_online_cpus() == 1)
1444 if (!cpu_present(cpu))
1449 cpuhp_tasks_frozen = tasks_frozen;
1451 prev_state = cpuhp_set_state(cpu, st, target);
1453 * If the current CPU state is in the range of the AP hotplug thread,
1454 * then we need to kick the thread.
1456 if (st->state > CPUHP_TEARDOWN_CPU) {
1457 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1458 ret = cpuhp_kick_ap_work(cpu);
1460 * The AP side has done the error rollback already. Just
1461 * return the error code..
1467 * We might have stopped still in the range of the AP hotplug
1468 * thread. Nothing to do anymore.
1470 if (st->state > CPUHP_TEARDOWN_CPU)
1473 st->target = target;
1476 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1477 * to do the further cleanups.
1479 ret = cpuhp_down_callbacks(cpu, st, target);
1480 if (ret && st->state < prev_state) {
1481 if (st->state == CPUHP_TEARDOWN_CPU) {
1482 cpuhp_reset_state(cpu, st, prev_state);
1483 __cpuhp_kick_ap(st);
1485 WARN(1, "DEAD callback error for CPU%d", cpu);
1490 cpus_write_unlock();
1492 * Do post unplug cleanup. This is still protected against
1493 * concurrent CPU hotplug via cpu_add_remove_lock.
1495 lockup_detector_cleanup();
1497 cpu_up_down_serialize_trainwrecks(tasks_frozen);
1501 struct cpu_down_work {
1503 enum cpuhp_state target;
1506 static long __cpu_down_maps_locked(void *arg)
1508 struct cpu_down_work *work = arg;
1510 return _cpu_down(work->cpu, 0, work->target);
1513 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1515 struct cpu_down_work work = { .cpu = cpu, .target = target, };
1518 * If the platform does not support hotplug, report it explicitly to
1519 * differentiate it from a transient offlining failure.
1521 if (cc_platform_has(CC_ATTR_HOTPLUG_DISABLED))
1523 if (cpu_hotplug_disabled)
1527 * Ensure that the control task does not run on the to be offlined
1528 * CPU to prevent a deadlock against cfs_b->period_timer.
1529 * Also keep at least one housekeeping cpu onlined to avoid generating
1530 * an empty sched_domain span.
1532 for_each_cpu_and(cpu, cpu_online_mask, housekeeping_cpumask(HK_TYPE_DOMAIN)) {
1533 if (cpu != work.cpu)
1534 return work_on_cpu(cpu, __cpu_down_maps_locked, &work);
1539 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1543 cpu_maps_update_begin();
1544 err = cpu_down_maps_locked(cpu, target);
1545 cpu_maps_update_done();
1550 * cpu_device_down - Bring down a cpu device
1551 * @dev: Pointer to the cpu device to offline
1553 * This function is meant to be used by device core cpu subsystem only.
1555 * Other subsystems should use remove_cpu() instead.
1557 * Return: %0 on success or a negative errno code
1559 int cpu_device_down(struct device *dev)
1561 return cpu_down(dev->id, CPUHP_OFFLINE);
1564 int remove_cpu(unsigned int cpu)
1568 lock_device_hotplug();
1569 ret = device_offline(get_cpu_device(cpu));
1570 unlock_device_hotplug();
1574 EXPORT_SYMBOL_GPL(remove_cpu);
1576 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1581 cpu_maps_update_begin();
1584 * Make certain the cpu I'm about to reboot on is online.
1586 * This is inline to what migrate_to_reboot_cpu() already do.
1588 if (!cpu_online(primary_cpu))
1589 primary_cpu = cpumask_first(cpu_online_mask);
1591 for_each_online_cpu(cpu) {
1592 if (cpu == primary_cpu)
1595 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1597 pr_err("Failed to offline CPU%d - error=%d",
1604 * Ensure all but the reboot CPU are offline.
1606 BUG_ON(num_online_cpus() > 1);
1609 * Make sure the CPUs won't be enabled by someone else after this
1610 * point. Kexec will reboot to a new kernel shortly resetting
1611 * everything along the way.
1613 cpu_hotplug_disabled++;
1615 cpu_maps_update_done();
1619 #define takedown_cpu NULL
1620 #endif /*CONFIG_HOTPLUG_CPU*/
1623 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1624 * @cpu: cpu that just started
1626 * It must be called by the arch code on the new cpu, before the new cpu
1627 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1629 void notify_cpu_starting(unsigned int cpu)
1631 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1632 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1634 rcutree_report_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1635 cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1638 * STARTING must not fail!
1640 cpuhp_invoke_callback_range_nofail(true, cpu, st, target);
1644 * Called from the idle task. Wake up the controlling task which brings the
1645 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1646 * online bringup to the hotplug thread.
1648 void cpuhp_online_idle(enum cpuhp_state state)
1650 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1652 /* Happens for the boot cpu */
1653 if (state != CPUHP_AP_ONLINE_IDLE)
1656 cpuhp_ap_update_sync_state(SYNC_STATE_ONLINE);
1659 * Unpark the stopper thread before we start the idle loop (and start
1660 * scheduling); this ensures the stopper task is always available.
1662 stop_machine_unpark(smp_processor_id());
1664 st->state = CPUHP_AP_ONLINE_IDLE;
1665 complete_ap_thread(st, true);
1668 /* Requires cpu_add_remove_lock to be held */
1669 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1671 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1672 struct task_struct *idle;
1677 if (!cpu_present(cpu)) {
1683 * The caller of cpu_up() might have raced with another
1684 * caller. Nothing to do.
1686 if (st->state >= target)
1689 if (st->state == CPUHP_OFFLINE) {
1690 /* Let it fail before we try to bring the cpu up */
1691 idle = idle_thread_get(cpu);
1693 ret = PTR_ERR(idle);
1698 * Reset stale stack state from the last time this CPU was online.
1700 scs_task_reset(idle);
1701 kasan_unpoison_task_stack(idle);
1704 cpuhp_tasks_frozen = tasks_frozen;
1706 cpuhp_set_state(cpu, st, target);
1708 * If the current CPU state is in the range of the AP hotplug thread,
1709 * then we need to kick the thread once more.
1711 if (st->state > CPUHP_BRINGUP_CPU) {
1712 ret = cpuhp_kick_ap_work(cpu);
1714 * The AP side has done the error rollback already. Just
1715 * return the error code..
1722 * Try to reach the target state. We max out on the BP at
1723 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1724 * responsible for bringing it up to the target state.
1726 target = min((int)target, CPUHP_BRINGUP_CPU);
1727 ret = cpuhp_up_callbacks(cpu, st, target);
1729 cpus_write_unlock();
1731 cpu_up_down_serialize_trainwrecks(tasks_frozen);
1735 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1739 if (!cpu_possible(cpu)) {
1740 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1745 err = try_online_node(cpu_to_node(cpu));
1749 cpu_maps_update_begin();
1751 if (cpu_hotplug_disabled) {
1755 if (!cpu_bootable(cpu)) {
1760 err = _cpu_up(cpu, 0, target);
1762 cpu_maps_update_done();
1767 * cpu_device_up - Bring up a cpu device
1768 * @dev: Pointer to the cpu device to online
1770 * This function is meant to be used by device core cpu subsystem only.
1772 * Other subsystems should use add_cpu() instead.
1774 * Return: %0 on success or a negative errno code
1776 int cpu_device_up(struct device *dev)
1778 return cpu_up(dev->id, CPUHP_ONLINE);
1781 int add_cpu(unsigned int cpu)
1785 lock_device_hotplug();
1786 ret = device_online(get_cpu_device(cpu));
1787 unlock_device_hotplug();
1791 EXPORT_SYMBOL_GPL(add_cpu);
1794 * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1795 * @sleep_cpu: The cpu we hibernated on and should be brought up.
1797 * On some architectures like arm64, we can hibernate on any CPU, but on
1798 * wake up the CPU we hibernated on might be offline as a side effect of
1799 * using maxcpus= for example.
1801 * Return: %0 on success or a negative errno code
1803 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1807 if (!cpu_online(sleep_cpu)) {
1808 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1809 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1811 pr_err("Failed to bring hibernate-CPU up!\n");
1818 static void __init cpuhp_bringup_mask(const struct cpumask *mask, unsigned int ncpus,
1819 enum cpuhp_state target)
1823 for_each_cpu(cpu, mask) {
1824 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1826 if (cpu_up(cpu, target) && can_rollback_cpu(st)) {
1828 * If this failed then cpu_up() might have only
1829 * rolled back to CPUHP_BP_KICK_AP for the final
1830 * online. Clean it up. NOOP if already rolled back.
1832 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st, CPUHP_OFFLINE));
1840 #ifdef CONFIG_HOTPLUG_PARALLEL
1841 static bool __cpuhp_parallel_bringup __ro_after_init = true;
1843 static int __init parallel_bringup_parse_param(char *arg)
1845 return kstrtobool(arg, &__cpuhp_parallel_bringup);
1847 early_param("cpuhp.parallel", parallel_bringup_parse_param);
1849 static inline bool cpuhp_smt_aware(void)
1851 return cpu_smt_max_threads > 1;
1854 static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1856 return cpu_primary_thread_mask;
1860 * On architectures which have enabled parallel bringup this invokes all BP
1861 * prepare states for each of the to be onlined APs first. The last state
1862 * sends the startup IPI to the APs. The APs proceed through the low level
1863 * bringup code in parallel and then wait for the control CPU to release
1864 * them one by one for the final onlining procedure.
1866 * This avoids waiting for each AP to respond to the startup IPI in
1867 * CPUHP_BRINGUP_CPU.
1869 static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
1871 const struct cpumask *mask = cpu_present_mask;
1873 if (__cpuhp_parallel_bringup)
1874 __cpuhp_parallel_bringup = arch_cpuhp_init_parallel_bringup();
1875 if (!__cpuhp_parallel_bringup)
1878 if (cpuhp_smt_aware()) {
1879 const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
1880 static struct cpumask tmp_mask __initdata;
1883 * X86 requires to prevent that SMT siblings stopped while
1884 * the primary thread does a microcode update for various
1885 * reasons. Bring the primary threads up first.
1887 cpumask_and(&tmp_mask, mask, pmask);
1888 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_BP_KICK_AP);
1889 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_ONLINE);
1890 /* Account for the online CPUs */
1891 ncpus -= num_online_cpus();
1894 /* Create the mask for secondary CPUs */
1895 cpumask_andnot(&tmp_mask, mask, pmask);
1899 /* Bring the not-yet started CPUs up */
1900 cpuhp_bringup_mask(mask, ncpus, CPUHP_BP_KICK_AP);
1901 cpuhp_bringup_mask(mask, ncpus, CPUHP_ONLINE);
1905 static inline bool cpuhp_bringup_cpus_parallel(unsigned int ncpus) { return false; }
1906 #endif /* CONFIG_HOTPLUG_PARALLEL */
1908 void __init bringup_nonboot_cpus(unsigned int max_cpus)
1910 /* Try parallel bringup optimization if enabled */
1911 if (cpuhp_bringup_cpus_parallel(max_cpus))
1914 /* Full per CPU serialized bringup */
1915 cpuhp_bringup_mask(cpu_present_mask, max_cpus, CPUHP_ONLINE);
1918 #ifdef CONFIG_PM_SLEEP_SMP
1919 static cpumask_var_t frozen_cpus;
1921 int freeze_secondary_cpus(int primary)
1925 cpu_maps_update_begin();
1926 if (primary == -1) {
1927 primary = cpumask_first(cpu_online_mask);
1928 if (!housekeeping_cpu(primary, HK_TYPE_TIMER))
1929 primary = housekeeping_any_cpu(HK_TYPE_TIMER);
1931 if (!cpu_online(primary))
1932 primary = cpumask_first(cpu_online_mask);
1936 * We take down all of the non-boot CPUs in one shot to avoid races
1937 * with the userspace trying to use the CPU hotplug at the same time
1939 cpumask_clear(frozen_cpus);
1941 pr_info("Disabling non-boot CPUs ...\n");
1942 for_each_online_cpu(cpu) {
1946 if (pm_wakeup_pending()) {
1947 pr_info("Wakeup pending. Abort CPU freeze\n");
1952 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1953 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1954 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1956 cpumask_set_cpu(cpu, frozen_cpus);
1958 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1964 BUG_ON(num_online_cpus() > 1);
1966 pr_err("Non-boot CPUs are not disabled\n");
1969 * Make sure the CPUs won't be enabled by someone else. We need to do
1970 * this even in case of failure as all freeze_secondary_cpus() users are
1971 * supposed to do thaw_secondary_cpus() on the failure path.
1973 cpu_hotplug_disabled++;
1975 cpu_maps_update_done();
1979 void __weak arch_thaw_secondary_cpus_begin(void)
1983 void __weak arch_thaw_secondary_cpus_end(void)
1987 void thaw_secondary_cpus(void)
1991 /* Allow everyone to use the CPU hotplug again */
1992 cpu_maps_update_begin();
1993 __cpu_hotplug_enable();
1994 if (cpumask_empty(frozen_cpus))
1997 pr_info("Enabling non-boot CPUs ...\n");
1999 arch_thaw_secondary_cpus_begin();
2001 for_each_cpu(cpu, frozen_cpus) {
2002 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
2003 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
2004 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
2006 pr_info("CPU%d is up\n", cpu);
2009 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
2012 arch_thaw_secondary_cpus_end();
2014 cpumask_clear(frozen_cpus);
2016 cpu_maps_update_done();
2019 static int __init alloc_frozen_cpus(void)
2021 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
2025 core_initcall(alloc_frozen_cpus);
2028 * When callbacks for CPU hotplug notifications are being executed, we must
2029 * ensure that the state of the system with respect to the tasks being frozen
2030 * or not, as reported by the notification, remains unchanged *throughout the
2031 * duration* of the execution of the callbacks.
2032 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
2034 * This synchronization is implemented by mutually excluding regular CPU
2035 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
2036 * Hibernate notifications.
2039 cpu_hotplug_pm_callback(struct notifier_block *nb,
2040 unsigned long action, void *ptr)
2044 case PM_SUSPEND_PREPARE:
2045 case PM_HIBERNATION_PREPARE:
2046 cpu_hotplug_disable();
2049 case PM_POST_SUSPEND:
2050 case PM_POST_HIBERNATION:
2051 cpu_hotplug_enable();
2062 static int __init cpu_hotplug_pm_sync_init(void)
2065 * cpu_hotplug_pm_callback has higher priority than x86
2066 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
2067 * to disable cpu hotplug to avoid cpu hotplug race.
2069 pm_notifier(cpu_hotplug_pm_callback, 0);
2072 core_initcall(cpu_hotplug_pm_sync_init);
2074 #endif /* CONFIG_PM_SLEEP_SMP */
2078 #endif /* CONFIG_SMP */
2080 /* Boot processor state steps */
2081 static struct cpuhp_step cpuhp_hp_states[] = {
2084 .startup.single = NULL,
2085 .teardown.single = NULL,
2088 [CPUHP_CREATE_THREADS]= {
2089 .name = "threads:prepare",
2090 .startup.single = smpboot_create_threads,
2091 .teardown.single = NULL,
2094 [CPUHP_PERF_PREPARE] = {
2095 .name = "perf:prepare",
2096 .startup.single = perf_event_init_cpu,
2097 .teardown.single = perf_event_exit_cpu,
2099 [CPUHP_RANDOM_PREPARE] = {
2100 .name = "random:prepare",
2101 .startup.single = random_prepare_cpu,
2102 .teardown.single = NULL,
2104 [CPUHP_WORKQUEUE_PREP] = {
2105 .name = "workqueue:prepare",
2106 .startup.single = workqueue_prepare_cpu,
2107 .teardown.single = NULL,
2109 [CPUHP_HRTIMERS_PREPARE] = {
2110 .name = "hrtimers:prepare",
2111 .startup.single = hrtimers_prepare_cpu,
2112 .teardown.single = NULL,
2114 [CPUHP_SMPCFD_PREPARE] = {
2115 .name = "smpcfd:prepare",
2116 .startup.single = smpcfd_prepare_cpu,
2117 .teardown.single = smpcfd_dead_cpu,
2119 [CPUHP_RELAY_PREPARE] = {
2120 .name = "relay:prepare",
2121 .startup.single = relay_prepare_cpu,
2122 .teardown.single = NULL,
2124 [CPUHP_RCUTREE_PREP] = {
2125 .name = "RCU/tree:prepare",
2126 .startup.single = rcutree_prepare_cpu,
2127 .teardown.single = rcutree_dead_cpu,
2130 * On the tear-down path, timers_dead_cpu() must be invoked
2131 * before blk_mq_queue_reinit_notify() from notify_dead(),
2132 * otherwise a RCU stall occurs.
2134 [CPUHP_TIMERS_PREPARE] = {
2135 .name = "timers:prepare",
2136 .startup.single = timers_prepare_cpu,
2137 .teardown.single = timers_dead_cpu,
2140 #ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
2142 * Kicks the AP alive. AP will wait in cpuhp_ap_sync_alive() until
2143 * the next step will release it.
2145 [CPUHP_BP_KICK_AP] = {
2146 .name = "cpu:kick_ap",
2147 .startup.single = cpuhp_kick_ap_alive,
2151 * Waits for the AP to reach cpuhp_ap_sync_alive() and then
2152 * releases it for the complete bringup.
2154 [CPUHP_BRINGUP_CPU] = {
2155 .name = "cpu:bringup",
2156 .startup.single = cpuhp_bringup_ap,
2157 .teardown.single = finish_cpu,
2162 * All-in-one CPU bringup state which includes the kick alive.
2164 [CPUHP_BRINGUP_CPU] = {
2165 .name = "cpu:bringup",
2166 .startup.single = bringup_cpu,
2167 .teardown.single = finish_cpu,
2171 /* Final state before CPU kills itself */
2172 [CPUHP_AP_IDLE_DEAD] = {
2173 .name = "idle:dead",
2176 * Last state before CPU enters the idle loop to die. Transient state
2177 * for synchronization.
2179 [CPUHP_AP_OFFLINE] = {
2180 .name = "ap:offline",
2183 /* First state is scheduler control. Interrupts are disabled */
2184 [CPUHP_AP_SCHED_STARTING] = {
2185 .name = "sched:starting",
2186 .startup.single = sched_cpu_starting,
2187 .teardown.single = sched_cpu_dying,
2189 [CPUHP_AP_RCUTREE_DYING] = {
2190 .name = "RCU/tree:dying",
2191 .startup.single = NULL,
2192 .teardown.single = rcutree_dying_cpu,
2194 [CPUHP_AP_SMPCFD_DYING] = {
2195 .name = "smpcfd:dying",
2196 .startup.single = NULL,
2197 .teardown.single = smpcfd_dying_cpu,
2199 [CPUHP_AP_HRTIMERS_DYING] = {
2200 .name = "hrtimers:dying",
2201 .startup.single = NULL,
2202 .teardown.single = hrtimers_cpu_dying,
2204 [CPUHP_AP_TICK_DYING] = {
2205 .name = "tick:dying",
2206 .startup.single = NULL,
2207 .teardown.single = tick_cpu_dying,
2209 /* Entry state on starting. Interrupts enabled from here on. Transient
2210 * state for synchronsization */
2211 [CPUHP_AP_ONLINE] = {
2212 .name = "ap:online",
2215 * Handled on control processor until the plugged processor manages
2218 [CPUHP_TEARDOWN_CPU] = {
2219 .name = "cpu:teardown",
2220 .startup.single = NULL,
2221 .teardown.single = takedown_cpu,
2225 [CPUHP_AP_SCHED_WAIT_EMPTY] = {
2226 .name = "sched:waitempty",
2227 .startup.single = NULL,
2228 .teardown.single = sched_cpu_wait_empty,
2231 /* Handle smpboot threads park/unpark */
2232 [CPUHP_AP_SMPBOOT_THREADS] = {
2233 .name = "smpboot/threads:online",
2234 .startup.single = smpboot_unpark_threads,
2235 .teardown.single = smpboot_park_threads,
2237 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
2238 .name = "irq/affinity:online",
2239 .startup.single = irq_affinity_online_cpu,
2240 .teardown.single = NULL,
2242 [CPUHP_AP_PERF_ONLINE] = {
2243 .name = "perf:online",
2244 .startup.single = perf_event_init_cpu,
2245 .teardown.single = perf_event_exit_cpu,
2247 [CPUHP_AP_WATCHDOG_ONLINE] = {
2248 .name = "lockup_detector:online",
2249 .startup.single = lockup_detector_online_cpu,
2250 .teardown.single = lockup_detector_offline_cpu,
2252 [CPUHP_AP_WORKQUEUE_ONLINE] = {
2253 .name = "workqueue:online",
2254 .startup.single = workqueue_online_cpu,
2255 .teardown.single = workqueue_offline_cpu,
2257 [CPUHP_AP_RANDOM_ONLINE] = {
2258 .name = "random:online",
2259 .startup.single = random_online_cpu,
2260 .teardown.single = NULL,
2262 [CPUHP_AP_RCUTREE_ONLINE] = {
2263 .name = "RCU/tree:online",
2264 .startup.single = rcutree_online_cpu,
2265 .teardown.single = rcutree_offline_cpu,
2269 * The dynamically registered state space is here
2273 /* Last state is scheduler control setting the cpu active */
2274 [CPUHP_AP_ACTIVE] = {
2275 .name = "sched:active",
2276 .startup.single = sched_cpu_activate,
2277 .teardown.single = sched_cpu_deactivate,
2281 /* CPU is fully up and running. */
2284 .startup.single = NULL,
2285 .teardown.single = NULL,
2289 /* Sanity check for callbacks */
2290 static int cpuhp_cb_check(enum cpuhp_state state)
2292 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
2298 * Returns a free for dynamic slot assignment of the Online state. The states
2299 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
2300 * by having no name assigned.
2302 static int cpuhp_reserve_state(enum cpuhp_state state)
2304 enum cpuhp_state i, end;
2305 struct cpuhp_step *step;
2308 case CPUHP_AP_ONLINE_DYN:
2309 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
2310 end = CPUHP_AP_ONLINE_DYN_END;
2312 case CPUHP_BP_PREPARE_DYN:
2313 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
2314 end = CPUHP_BP_PREPARE_DYN_END;
2320 for (i = state; i <= end; i++, step++) {
2324 WARN(1, "No more dynamic states available for CPU hotplug\n");
2328 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
2329 int (*startup)(unsigned int cpu),
2330 int (*teardown)(unsigned int cpu),
2331 bool multi_instance)
2333 /* (Un)Install the callbacks for further cpu hotplug operations */
2334 struct cpuhp_step *sp;
2338 * If name is NULL, then the state gets removed.
2340 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
2341 * the first allocation from these dynamic ranges, so the removal
2342 * would trigger a new allocation and clear the wrong (already
2343 * empty) state, leaving the callbacks of the to be cleared state
2344 * dangling, which causes wreckage on the next hotplug operation.
2346 if (name && (state == CPUHP_AP_ONLINE_DYN ||
2347 state == CPUHP_BP_PREPARE_DYN)) {
2348 ret = cpuhp_reserve_state(state);
2353 sp = cpuhp_get_step(state);
2354 if (name && sp->name)
2357 sp->startup.single = startup;
2358 sp->teardown.single = teardown;
2360 sp->multi_instance = multi_instance;
2361 INIT_HLIST_HEAD(&sp->list);
2365 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
2367 return cpuhp_get_step(state)->teardown.single;
2371 * Call the startup/teardown function for a step either on the AP or
2372 * on the current CPU.
2374 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
2375 struct hlist_node *node)
2377 struct cpuhp_step *sp = cpuhp_get_step(state);
2381 * If there's nothing to do, we done.
2382 * Relies on the union for multi_instance.
2384 if (cpuhp_step_empty(bringup, sp))
2387 * The non AP bound callbacks can fail on bringup. On teardown
2388 * e.g. module removal we crash for now.
2391 if (cpuhp_is_ap_state(state))
2392 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
2394 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2396 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2398 BUG_ON(ret && !bringup);
2403 * Called from __cpuhp_setup_state on a recoverable failure.
2405 * Note: The teardown callbacks for rollback are not allowed to fail!
2407 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
2408 struct hlist_node *node)
2412 /* Roll back the already executed steps on the other cpus */
2413 for_each_present_cpu(cpu) {
2414 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2415 int cpustate = st->state;
2417 if (cpu >= failedcpu)
2420 /* Did we invoke the startup call on that cpu ? */
2421 if (cpustate >= state)
2422 cpuhp_issue_call(cpu, state, false, node);
2426 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
2427 struct hlist_node *node,
2430 struct cpuhp_step *sp;
2434 lockdep_assert_cpus_held();
2436 sp = cpuhp_get_step(state);
2437 if (sp->multi_instance == false)
2440 mutex_lock(&cpuhp_state_mutex);
2442 if (!invoke || !sp->startup.multi)
2446 * Try to call the startup callback for each present cpu
2447 * depending on the hotplug state of the cpu.
2449 for_each_present_cpu(cpu) {
2450 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2451 int cpustate = st->state;
2453 if (cpustate < state)
2456 ret = cpuhp_issue_call(cpu, state, true, node);
2458 if (sp->teardown.multi)
2459 cpuhp_rollback_install(cpu, state, node);
2465 hlist_add_head(node, &sp->list);
2467 mutex_unlock(&cpuhp_state_mutex);
2471 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
2477 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
2481 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
2484 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
2485 * @state: The state to setup
2486 * @name: Name of the step
2487 * @invoke: If true, the startup function is invoked for cpus where
2488 * cpu state >= @state
2489 * @startup: startup callback function
2490 * @teardown: teardown callback function
2491 * @multi_instance: State is set up for multiple instances which get
2494 * The caller needs to hold cpus read locked while calling this function.
2497 * Positive state number if @state is CPUHP_AP_ONLINE_DYN;
2498 * 0 for all other states
2499 * On failure: proper (negative) error code
2501 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
2502 const char *name, bool invoke,
2503 int (*startup)(unsigned int cpu),
2504 int (*teardown)(unsigned int cpu),
2505 bool multi_instance)
2510 lockdep_assert_cpus_held();
2512 if (cpuhp_cb_check(state) || !name)
2515 mutex_lock(&cpuhp_state_mutex);
2517 ret = cpuhp_store_callbacks(state, name, startup, teardown,
2520 dynstate = state == CPUHP_AP_ONLINE_DYN;
2521 if (ret > 0 && dynstate) {
2526 if (ret || !invoke || !startup)
2530 * Try to call the startup callback for each present cpu
2531 * depending on the hotplug state of the cpu.
2533 for_each_present_cpu(cpu) {
2534 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2535 int cpustate = st->state;
2537 if (cpustate < state)
2540 ret = cpuhp_issue_call(cpu, state, true, NULL);
2543 cpuhp_rollback_install(cpu, state, NULL);
2544 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2549 mutex_unlock(&cpuhp_state_mutex);
2551 * If the requested state is CPUHP_AP_ONLINE_DYN, return the
2552 * dynamically allocated state in case of success.
2554 if (!ret && dynstate)
2558 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2560 int __cpuhp_setup_state(enum cpuhp_state state,
2561 const char *name, bool invoke,
2562 int (*startup)(unsigned int cpu),
2563 int (*teardown)(unsigned int cpu),
2564 bool multi_instance)
2569 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2570 teardown, multi_instance);
2574 EXPORT_SYMBOL(__cpuhp_setup_state);
2576 int __cpuhp_state_remove_instance(enum cpuhp_state state,
2577 struct hlist_node *node, bool invoke)
2579 struct cpuhp_step *sp = cpuhp_get_step(state);
2582 BUG_ON(cpuhp_cb_check(state));
2584 if (!sp->multi_instance)
2588 mutex_lock(&cpuhp_state_mutex);
2590 if (!invoke || !cpuhp_get_teardown_cb(state))
2593 * Call the teardown callback for each present cpu depending
2594 * on the hotplug state of the cpu. This function is not
2595 * allowed to fail currently!
2597 for_each_present_cpu(cpu) {
2598 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2599 int cpustate = st->state;
2601 if (cpustate >= state)
2602 cpuhp_issue_call(cpu, state, false, node);
2607 mutex_unlock(&cpuhp_state_mutex);
2612 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2615 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2616 * @state: The state to remove
2617 * @invoke: If true, the teardown function is invoked for cpus where
2618 * cpu state >= @state
2620 * The caller needs to hold cpus read locked while calling this function.
2621 * The teardown callback is currently not allowed to fail. Think
2622 * about module removal!
2624 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2626 struct cpuhp_step *sp = cpuhp_get_step(state);
2629 BUG_ON(cpuhp_cb_check(state));
2631 lockdep_assert_cpus_held();
2633 mutex_lock(&cpuhp_state_mutex);
2634 if (sp->multi_instance) {
2635 WARN(!hlist_empty(&sp->list),
2636 "Error: Removing state %d which has instances left.\n",
2641 if (!invoke || !cpuhp_get_teardown_cb(state))
2645 * Call the teardown callback for each present cpu depending
2646 * on the hotplug state of the cpu. This function is not
2647 * allowed to fail currently!
2649 for_each_present_cpu(cpu) {
2650 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2651 int cpustate = st->state;
2653 if (cpustate >= state)
2654 cpuhp_issue_call(cpu, state, false, NULL);
2657 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2658 mutex_unlock(&cpuhp_state_mutex);
2660 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2662 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2665 __cpuhp_remove_state_cpuslocked(state, invoke);
2668 EXPORT_SYMBOL(__cpuhp_remove_state);
2670 #ifdef CONFIG_HOTPLUG_SMT
2671 static void cpuhp_offline_cpu_device(unsigned int cpu)
2673 struct device *dev = get_cpu_device(cpu);
2675 dev->offline = true;
2676 /* Tell user space about the state change */
2677 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2680 static void cpuhp_online_cpu_device(unsigned int cpu)
2682 struct device *dev = get_cpu_device(cpu);
2684 dev->offline = false;
2685 /* Tell user space about the state change */
2686 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2689 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2693 cpu_maps_update_begin();
2694 for_each_online_cpu(cpu) {
2695 if (topology_is_primary_thread(cpu))
2698 * Disable can be called with CPU_SMT_ENABLED when changing
2699 * from a higher to lower number of SMT threads per core.
2701 if (ctrlval == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
2703 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2707 * As this needs to hold the cpu maps lock it's impossible
2708 * to call device_offline() because that ends up calling
2709 * cpu_down() which takes cpu maps lock. cpu maps lock
2710 * needs to be held as this might race against in kernel
2711 * abusers of the hotplug machinery (thermal management).
2713 * So nothing would update device:offline state. That would
2714 * leave the sysfs entry stale and prevent onlining after
2715 * smt control has been changed to 'off' again. This is
2716 * called under the sysfs hotplug lock, so it is properly
2717 * serialized against the regular offline usage.
2719 cpuhp_offline_cpu_device(cpu);
2722 cpu_smt_control = ctrlval;
2723 cpu_maps_update_done();
2727 int cpuhp_smt_enable(void)
2731 cpu_maps_update_begin();
2732 cpu_smt_control = CPU_SMT_ENABLED;
2733 for_each_present_cpu(cpu) {
2734 /* Skip online CPUs and CPUs on offline nodes */
2735 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2737 if (!cpu_smt_thread_allowed(cpu))
2739 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2742 /* See comment in cpuhp_smt_disable() */
2743 cpuhp_online_cpu_device(cpu);
2745 cpu_maps_update_done();
2750 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2751 static ssize_t state_show(struct device *dev,
2752 struct device_attribute *attr, char *buf)
2754 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2756 return sprintf(buf, "%d\n", st->state);
2758 static DEVICE_ATTR_RO(state);
2760 static ssize_t target_store(struct device *dev, struct device_attribute *attr,
2761 const char *buf, size_t count)
2763 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2764 struct cpuhp_step *sp;
2767 ret = kstrtoint(buf, 10, &target);
2771 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2772 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2775 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2779 ret = lock_device_hotplug_sysfs();
2783 mutex_lock(&cpuhp_state_mutex);
2784 sp = cpuhp_get_step(target);
2785 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2786 mutex_unlock(&cpuhp_state_mutex);
2790 if (st->state < target)
2791 ret = cpu_up(dev->id, target);
2792 else if (st->state > target)
2793 ret = cpu_down(dev->id, target);
2794 else if (WARN_ON(st->target != target))
2795 st->target = target;
2797 unlock_device_hotplug();
2798 return ret ? ret : count;
2801 static ssize_t target_show(struct device *dev,
2802 struct device_attribute *attr, char *buf)
2804 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2806 return sprintf(buf, "%d\n", st->target);
2808 static DEVICE_ATTR_RW(target);
2810 static ssize_t fail_store(struct device *dev, struct device_attribute *attr,
2811 const char *buf, size_t count)
2813 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2814 struct cpuhp_step *sp;
2817 ret = kstrtoint(buf, 10, &fail);
2821 if (fail == CPUHP_INVALID) {
2826 if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2830 * Cannot fail STARTING/DYING callbacks.
2832 if (cpuhp_is_atomic_state(fail))
2836 * DEAD callbacks cannot fail...
2837 * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2838 * triggering STARTING callbacks, a failure in this state would
2841 if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2845 * Cannot fail anything that doesn't have callbacks.
2847 mutex_lock(&cpuhp_state_mutex);
2848 sp = cpuhp_get_step(fail);
2849 if (!sp->startup.single && !sp->teardown.single)
2851 mutex_unlock(&cpuhp_state_mutex);
2860 static ssize_t fail_show(struct device *dev,
2861 struct device_attribute *attr, char *buf)
2863 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2865 return sprintf(buf, "%d\n", st->fail);
2868 static DEVICE_ATTR_RW(fail);
2870 static struct attribute *cpuhp_cpu_attrs[] = {
2871 &dev_attr_state.attr,
2872 &dev_attr_target.attr,
2873 &dev_attr_fail.attr,
2877 static const struct attribute_group cpuhp_cpu_attr_group = {
2878 .attrs = cpuhp_cpu_attrs,
2883 static ssize_t states_show(struct device *dev,
2884 struct device_attribute *attr, char *buf)
2886 ssize_t cur, res = 0;
2889 mutex_lock(&cpuhp_state_mutex);
2890 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2891 struct cpuhp_step *sp = cpuhp_get_step(i);
2894 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2899 mutex_unlock(&cpuhp_state_mutex);
2902 static DEVICE_ATTR_RO(states);
2904 static struct attribute *cpuhp_cpu_root_attrs[] = {
2905 &dev_attr_states.attr,
2909 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2910 .attrs = cpuhp_cpu_root_attrs,
2915 #ifdef CONFIG_HOTPLUG_SMT
2917 static bool cpu_smt_num_threads_valid(unsigned int threads)
2919 if (IS_ENABLED(CONFIG_SMT_NUM_THREADS_DYNAMIC))
2920 return threads >= 1 && threads <= cpu_smt_max_threads;
2921 return threads == 1 || threads == cpu_smt_max_threads;
2925 __store_smt_control(struct device *dev, struct device_attribute *attr,
2926 const char *buf, size_t count)
2928 int ctrlval, ret, num_threads, orig_threads;
2931 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2934 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2937 if (sysfs_streq(buf, "on")) {
2938 ctrlval = CPU_SMT_ENABLED;
2939 num_threads = cpu_smt_max_threads;
2940 } else if (sysfs_streq(buf, "off")) {
2941 ctrlval = CPU_SMT_DISABLED;
2943 } else if (sysfs_streq(buf, "forceoff")) {
2944 ctrlval = CPU_SMT_FORCE_DISABLED;
2946 } else if (kstrtoint(buf, 10, &num_threads) == 0) {
2947 if (num_threads == 1)
2948 ctrlval = CPU_SMT_DISABLED;
2949 else if (cpu_smt_num_threads_valid(num_threads))
2950 ctrlval = CPU_SMT_ENABLED;
2957 ret = lock_device_hotplug_sysfs();
2961 orig_threads = cpu_smt_num_threads;
2962 cpu_smt_num_threads = num_threads;
2964 force_off = ctrlval != cpu_smt_control && ctrlval == CPU_SMT_FORCE_DISABLED;
2966 if (num_threads > orig_threads)
2967 ret = cpuhp_smt_enable();
2968 else if (num_threads < orig_threads || force_off)
2969 ret = cpuhp_smt_disable(ctrlval);
2971 unlock_device_hotplug();
2972 return ret ? ret : count;
2975 #else /* !CONFIG_HOTPLUG_SMT */
2977 __store_smt_control(struct device *dev, struct device_attribute *attr,
2978 const char *buf, size_t count)
2982 #endif /* CONFIG_HOTPLUG_SMT */
2984 static const char *smt_states[] = {
2985 [CPU_SMT_ENABLED] = "on",
2986 [CPU_SMT_DISABLED] = "off",
2987 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2988 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2989 [CPU_SMT_NOT_IMPLEMENTED] = "notimplemented",
2992 static ssize_t control_show(struct device *dev,
2993 struct device_attribute *attr, char *buf)
2995 const char *state = smt_states[cpu_smt_control];
2997 #ifdef CONFIG_HOTPLUG_SMT
2999 * If SMT is enabled but not all threads are enabled then show the
3000 * number of threads. If all threads are enabled show "on". Otherwise
3001 * show the state name.
3003 if (cpu_smt_control == CPU_SMT_ENABLED &&
3004 cpu_smt_num_threads != cpu_smt_max_threads)
3005 return sysfs_emit(buf, "%d\n", cpu_smt_num_threads);
3008 return sysfs_emit(buf, "%s\n", state);
3011 static ssize_t control_store(struct device *dev, struct device_attribute *attr,
3012 const char *buf, size_t count)
3014 return __store_smt_control(dev, attr, buf, count);
3016 static DEVICE_ATTR_RW(control);
3018 static ssize_t active_show(struct device *dev,
3019 struct device_attribute *attr, char *buf)
3021 return sysfs_emit(buf, "%d\n", sched_smt_active());
3023 static DEVICE_ATTR_RO(active);
3025 static struct attribute *cpuhp_smt_attrs[] = {
3026 &dev_attr_control.attr,
3027 &dev_attr_active.attr,
3031 static const struct attribute_group cpuhp_smt_attr_group = {
3032 .attrs = cpuhp_smt_attrs,
3037 static int __init cpu_smt_sysfs_init(void)
3039 struct device *dev_root;
3042 dev_root = bus_get_dev_root(&cpu_subsys);
3044 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_smt_attr_group);
3045 put_device(dev_root);
3050 static int __init cpuhp_sysfs_init(void)
3052 struct device *dev_root;
3055 ret = cpu_smt_sysfs_init();
3059 dev_root = bus_get_dev_root(&cpu_subsys);
3061 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_cpu_root_attr_group);
3062 put_device(dev_root);
3067 for_each_possible_cpu(cpu) {
3068 struct device *dev = get_cpu_device(cpu);
3072 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
3078 device_initcall(cpuhp_sysfs_init);
3079 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
3082 * cpu_bit_bitmap[] is a special, "compressed" data structure that
3083 * represents all NR_CPUS bits binary values of 1<<nr.
3085 * It is used by cpumask_of() to get a constant address to a CPU
3086 * mask value that has a single bit set only.
3089 /* cpu_bit_bitmap[0] is empty - so we can back into it */
3090 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
3091 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
3092 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
3093 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
3095 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
3097 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
3098 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
3099 #if BITS_PER_LONG > 32
3100 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
3101 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
3104 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
3106 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
3107 EXPORT_SYMBOL(cpu_all_bits);
3109 #ifdef CONFIG_INIT_ALL_POSSIBLE
3110 struct cpumask __cpu_possible_mask __ro_after_init
3113 struct cpumask __cpu_possible_mask __ro_after_init;
3115 EXPORT_SYMBOL(__cpu_possible_mask);
3117 struct cpumask __cpu_online_mask __read_mostly;
3118 EXPORT_SYMBOL(__cpu_online_mask);
3120 struct cpumask __cpu_present_mask __read_mostly;
3121 EXPORT_SYMBOL(__cpu_present_mask);
3123 struct cpumask __cpu_active_mask __read_mostly;
3124 EXPORT_SYMBOL(__cpu_active_mask);
3126 struct cpumask __cpu_dying_mask __read_mostly;
3127 EXPORT_SYMBOL(__cpu_dying_mask);
3129 atomic_t __num_online_cpus __read_mostly;
3130 EXPORT_SYMBOL(__num_online_cpus);
3132 void init_cpu_present(const struct cpumask *src)
3134 cpumask_copy(&__cpu_present_mask, src);
3137 void init_cpu_possible(const struct cpumask *src)
3139 cpumask_copy(&__cpu_possible_mask, src);
3142 void init_cpu_online(const struct cpumask *src)
3144 cpumask_copy(&__cpu_online_mask, src);
3147 void set_cpu_online(unsigned int cpu, bool online)
3150 * atomic_inc/dec() is required to handle the horrid abuse of this
3151 * function by the reboot and kexec code which invoke it from
3152 * IPI/NMI broadcasts when shutting down CPUs. Invocation from
3153 * regular CPU hotplug is properly serialized.
3155 * Note, that the fact that __num_online_cpus is of type atomic_t
3156 * does not protect readers which are not serialized against
3157 * concurrent hotplug operations.
3160 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
3161 atomic_inc(&__num_online_cpus);
3163 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
3164 atomic_dec(&__num_online_cpus);
3169 * Activate the first processor.
3171 void __init boot_cpu_init(void)
3173 int cpu = smp_processor_id();
3175 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
3176 set_cpu_online(cpu, true);
3177 set_cpu_active(cpu, true);
3178 set_cpu_present(cpu, true);
3179 set_cpu_possible(cpu, true);
3182 __boot_cpu_id = cpu;
3187 * Must be called _AFTER_ setting up the per_cpu areas
3189 void __init boot_cpu_hotplug_init(void)
3192 cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
3193 atomic_set(this_cpu_ptr(&cpuhp_state.ap_sync_state), SYNC_STATE_ONLINE);
3195 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
3196 this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
3200 * These are used for a global "mitigations=" cmdline option for toggling
3201 * optional CPU mitigations.
3203 enum cpu_mitigations {
3204 CPU_MITIGATIONS_OFF,
3205 CPU_MITIGATIONS_AUTO,
3206 CPU_MITIGATIONS_AUTO_NOSMT,
3209 static enum cpu_mitigations cpu_mitigations __ro_after_init =
3210 IS_ENABLED(CONFIG_CPU_MITIGATIONS) ? CPU_MITIGATIONS_AUTO :
3211 CPU_MITIGATIONS_OFF;
3213 static int __init mitigations_parse_cmdline(char *arg)
3215 if (!strcmp(arg, "off"))
3216 cpu_mitigations = CPU_MITIGATIONS_OFF;
3217 else if (!strcmp(arg, "auto"))
3218 cpu_mitigations = CPU_MITIGATIONS_AUTO;
3219 else if (!strcmp(arg, "auto,nosmt"))
3220 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
3222 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
3227 early_param("mitigations", mitigations_parse_cmdline);
3229 /* mitigations=off */
3230 bool cpu_mitigations_off(void)
3232 return cpu_mitigations == CPU_MITIGATIONS_OFF;
3234 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
3236 /* mitigations=auto,nosmt */
3237 bool cpu_mitigations_auto_nosmt(void)
3239 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
3241 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);