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
58 * @node: Remote CPU node; for multi-instance, do a
59 * single entry callback for install/remove
60 * @last: For multi-instance rollback, remember how far we got
61 * @cb_state: The state for a single callback (install/uninstall)
62 * @result: Result of the operation
63 * @ap_sync_state: State for AP synchronization
64 * @done_up: Signal completion to the issuer of the task for cpu-up
65 * @done_down: Signal completion to the issuer of the task for cpu-down
67 struct cpuhp_cpu_state {
68 enum cpuhp_state state;
69 enum cpuhp_state target;
70 enum cpuhp_state fail;
72 struct task_struct *thread;
77 struct hlist_node *node;
78 struct hlist_node *last;
79 enum cpuhp_state cb_state;
81 atomic_t ap_sync_state;
82 struct completion done_up;
83 struct completion done_down;
87 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
88 .fail = CPUHP_INVALID,
92 cpumask_t cpus_booted_once_mask;
95 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
96 static struct lockdep_map cpuhp_state_up_map =
97 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
98 static struct lockdep_map cpuhp_state_down_map =
99 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
102 static inline void cpuhp_lock_acquire(bool bringup)
104 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
107 static inline void cpuhp_lock_release(bool bringup)
109 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
113 static inline void cpuhp_lock_acquire(bool bringup) { }
114 static inline void cpuhp_lock_release(bool bringup) { }
119 * struct cpuhp_step - Hotplug state machine step
120 * @name: Name of the step
121 * @startup: Startup function of the step
122 * @teardown: Teardown function of the step
123 * @cant_stop: Bringup/teardown can't be stopped at this step
124 * @multi_instance: State has multiple instances which get added afterwards
129 int (*single)(unsigned int cpu);
130 int (*multi)(unsigned int cpu,
131 struct hlist_node *node);
134 int (*single)(unsigned int cpu);
135 int (*multi)(unsigned int cpu,
136 struct hlist_node *node);
139 struct hlist_head list;
145 static DEFINE_MUTEX(cpuhp_state_mutex);
146 static struct cpuhp_step cpuhp_hp_states[];
148 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
150 return cpuhp_hp_states + state;
153 static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
155 return bringup ? !step->startup.single : !step->teardown.single;
159 * cpuhp_invoke_callback - Invoke the callbacks for a given state
160 * @cpu: The cpu for which the callback should be invoked
161 * @state: The state to do callbacks for
162 * @bringup: True if the bringup callback should be invoked
163 * @node: For multi-instance, do a single entry callback for install/remove
164 * @lastp: For multi-instance rollback, remember how far we got
166 * Called from cpu hotplug and from the state register machinery.
168 * Return: %0 on success or a negative errno code
170 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
171 bool bringup, struct hlist_node *node,
172 struct hlist_node **lastp)
174 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
175 struct cpuhp_step *step = cpuhp_get_step(state);
176 int (*cbm)(unsigned int cpu, struct hlist_node *node);
177 int (*cb)(unsigned int cpu);
180 if (st->fail == state) {
181 st->fail = CPUHP_INVALID;
185 if (cpuhp_step_empty(bringup, step)) {
190 if (!step->multi_instance) {
191 WARN_ON_ONCE(lastp && *lastp);
192 cb = bringup ? step->startup.single : step->teardown.single;
194 trace_cpuhp_enter(cpu, st->target, state, cb);
196 trace_cpuhp_exit(cpu, st->state, state, ret);
199 cbm = bringup ? step->startup.multi : step->teardown.multi;
201 /* Single invocation for instance add/remove */
203 WARN_ON_ONCE(lastp && *lastp);
204 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
205 ret = cbm(cpu, node);
206 trace_cpuhp_exit(cpu, st->state, state, ret);
210 /* State transition. Invoke on all instances */
212 hlist_for_each(node, &step->list) {
213 if (lastp && node == *lastp)
216 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
217 ret = cbm(cpu, node);
218 trace_cpuhp_exit(cpu, st->state, state, ret);
232 /* Rollback the instances if one failed */
233 cbm = !bringup ? step->startup.multi : step->teardown.multi;
237 hlist_for_each(node, &step->list) {
241 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
242 ret = cbm(cpu, node);
243 trace_cpuhp_exit(cpu, st->state, state, ret);
245 * Rollback must not fail,
253 static bool cpuhp_is_ap_state(enum cpuhp_state state)
256 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
257 * purposes as that state is handled explicitly in cpu_down.
259 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
262 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
264 struct completion *done = bringup ? &st->done_up : &st->done_down;
265 wait_for_completion(done);
268 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
270 struct completion *done = bringup ? &st->done_up : &st->done_down;
275 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
277 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
279 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
282 /* Synchronization state management */
283 enum cpuhp_sync_state {
286 SYNC_STATE_SHOULD_DIE,
288 SYNC_STATE_SHOULD_ONLINE,
292 #ifdef CONFIG_HOTPLUG_CORE_SYNC
294 * cpuhp_ap_update_sync_state - Update synchronization state during bringup/teardown
295 * @state: The synchronization state to set
297 * No synchronization point. Just update of the synchronization state, but implies
298 * a full barrier so that the AP changes are visible before the control CPU proceeds.
300 static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state)
302 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
304 (void)atomic_xchg(st, state);
307 void __weak arch_cpuhp_sync_state_poll(void) { cpu_relax(); }
309 static bool cpuhp_wait_for_sync_state(unsigned int cpu, enum cpuhp_sync_state state,
310 enum cpuhp_sync_state next_state)
312 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
313 ktime_t now, end, start = ktime_get();
316 end = start + 10ULL * NSEC_PER_SEC;
318 sync = atomic_read(st);
321 if (!atomic_try_cmpxchg(st, &sync, next_state))
328 /* Timeout. Leave the state unchanged */
330 } else if (now - start < NSEC_PER_MSEC) {
331 /* Poll for one millisecond */
332 arch_cpuhp_sync_state_poll();
334 usleep_range_state(USEC_PER_MSEC, 2 * USEC_PER_MSEC, TASK_UNINTERRUPTIBLE);
336 sync = atomic_read(st);
340 #else /* CONFIG_HOTPLUG_CORE_SYNC */
341 static inline void cpuhp_ap_update_sync_state(enum cpuhp_sync_state state) { }
342 #endif /* !CONFIG_HOTPLUG_CORE_SYNC */
344 #ifdef CONFIG_HOTPLUG_CORE_SYNC_DEAD
346 * cpuhp_ap_report_dead - Update synchronization state to DEAD
348 * No synchronization point. Just update of the synchronization state.
350 void cpuhp_ap_report_dead(void)
352 cpuhp_ap_update_sync_state(SYNC_STATE_DEAD);
355 void __weak arch_cpuhp_cleanup_dead_cpu(unsigned int cpu) { }
358 * Late CPU shutdown synchronization point. Cannot use cpuhp_state::done_down
359 * because the AP cannot issue complete() at this stage.
361 static void cpuhp_bp_sync_dead(unsigned int cpu)
363 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
364 int sync = atomic_read(st);
367 /* CPU can have reported dead already. Don't overwrite that! */
368 if (sync == SYNC_STATE_DEAD)
370 } while (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_SHOULD_DIE));
372 if (cpuhp_wait_for_sync_state(cpu, SYNC_STATE_DEAD, SYNC_STATE_DEAD)) {
373 /* CPU reached dead state. Invoke the cleanup function */
374 arch_cpuhp_cleanup_dead_cpu(cpu);
378 /* No further action possible. Emit message and give up. */
379 pr_err("CPU%u failed to report dead state\n", cpu);
381 #else /* CONFIG_HOTPLUG_CORE_SYNC_DEAD */
382 static inline void cpuhp_bp_sync_dead(unsigned int cpu) { }
383 #endif /* !CONFIG_HOTPLUG_CORE_SYNC_DEAD */
385 #ifdef CONFIG_HOTPLUG_CORE_SYNC_FULL
387 * cpuhp_ap_sync_alive - Synchronize AP with the control CPU once it is alive
389 * Updates the AP synchronization state to SYNC_STATE_ALIVE and waits
390 * for the BP to release it.
392 void cpuhp_ap_sync_alive(void)
394 atomic_t *st = this_cpu_ptr(&cpuhp_state.ap_sync_state);
396 cpuhp_ap_update_sync_state(SYNC_STATE_ALIVE);
398 /* Wait for the control CPU to release it. */
399 while (atomic_read(st) != SYNC_STATE_SHOULD_ONLINE)
403 static bool cpuhp_can_boot_ap(unsigned int cpu)
405 atomic_t *st = per_cpu_ptr(&cpuhp_state.ap_sync_state, cpu);
406 int sync = atomic_read(st);
410 case SYNC_STATE_DEAD:
411 /* CPU is properly dead */
413 case SYNC_STATE_KICKED:
414 /* CPU did not come up in previous attempt */
416 case SYNC_STATE_ALIVE:
417 /* CPU is stuck cpuhp_ap_sync_alive(). */
420 /* CPU failed to report online or dead and is in limbo state. */
424 /* Prepare for booting */
425 if (!atomic_try_cmpxchg(st, &sync, SYNC_STATE_KICKED))
431 void __weak arch_cpuhp_cleanup_kick_cpu(unsigned int cpu) { }
434 * Early CPU bringup synchronization point. Cannot use cpuhp_state::done_up
435 * because the AP cannot issue complete() so early in the bringup.
437 static int cpuhp_bp_sync_alive(unsigned int cpu)
441 if (!IS_ENABLED(CONFIG_HOTPLUG_CORE_SYNC_FULL))
444 if (!cpuhp_wait_for_sync_state(cpu, SYNC_STATE_ALIVE, SYNC_STATE_SHOULD_ONLINE)) {
445 pr_err("CPU%u failed to report alive state\n", cpu);
449 /* Let the architecture cleanup the kick alive mechanics. */
450 arch_cpuhp_cleanup_kick_cpu(cpu);
453 #else /* CONFIG_HOTPLUG_CORE_SYNC_FULL */
454 static inline int cpuhp_bp_sync_alive(unsigned int cpu) { return 0; }
455 static inline bool cpuhp_can_boot_ap(unsigned int cpu) { return true; }
456 #endif /* !CONFIG_HOTPLUG_CORE_SYNC_FULL */
458 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
459 static DEFINE_MUTEX(cpu_add_remove_lock);
460 bool cpuhp_tasks_frozen;
461 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
464 * The following two APIs (cpu_maps_update_begin/done) must be used when
465 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
467 void cpu_maps_update_begin(void)
469 mutex_lock(&cpu_add_remove_lock);
472 void cpu_maps_update_done(void)
474 mutex_unlock(&cpu_add_remove_lock);
478 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
479 * Should always be manipulated under cpu_add_remove_lock
481 static int cpu_hotplug_disabled;
483 #ifdef CONFIG_HOTPLUG_CPU
485 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
487 void cpus_read_lock(void)
489 percpu_down_read(&cpu_hotplug_lock);
491 EXPORT_SYMBOL_GPL(cpus_read_lock);
493 int cpus_read_trylock(void)
495 return percpu_down_read_trylock(&cpu_hotplug_lock);
497 EXPORT_SYMBOL_GPL(cpus_read_trylock);
499 void cpus_read_unlock(void)
501 percpu_up_read(&cpu_hotplug_lock);
503 EXPORT_SYMBOL_GPL(cpus_read_unlock);
505 void cpus_write_lock(void)
507 percpu_down_write(&cpu_hotplug_lock);
510 void cpus_write_unlock(void)
512 percpu_up_write(&cpu_hotplug_lock);
515 void lockdep_assert_cpus_held(void)
518 * We can't have hotplug operations before userspace starts running,
519 * and some init codepaths will knowingly not take the hotplug lock.
520 * This is all valid, so mute lockdep until it makes sense to report
523 if (system_state < SYSTEM_RUNNING)
526 percpu_rwsem_assert_held(&cpu_hotplug_lock);
529 #ifdef CONFIG_LOCKDEP
530 int lockdep_is_cpus_held(void)
532 return percpu_rwsem_is_held(&cpu_hotplug_lock);
536 static void lockdep_acquire_cpus_lock(void)
538 rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
541 static void lockdep_release_cpus_lock(void)
543 rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
547 * Wait for currently running CPU hotplug operations to complete (if any) and
548 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
549 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
550 * hotplug path before performing hotplug operations. So acquiring that lock
551 * guarantees mutual exclusion from any currently running hotplug operations.
553 void cpu_hotplug_disable(void)
555 cpu_maps_update_begin();
556 cpu_hotplug_disabled++;
557 cpu_maps_update_done();
559 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
561 static void __cpu_hotplug_enable(void)
563 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
565 cpu_hotplug_disabled--;
568 void cpu_hotplug_enable(void)
570 cpu_maps_update_begin();
571 __cpu_hotplug_enable();
572 cpu_maps_update_done();
574 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
578 static void lockdep_acquire_cpus_lock(void)
582 static void lockdep_release_cpus_lock(void)
586 #endif /* CONFIG_HOTPLUG_CPU */
589 * Architectures that need SMT-specific errata handling during SMT hotplug
590 * should override this.
592 void __weak arch_smt_update(void) { }
594 #ifdef CONFIG_HOTPLUG_SMT
596 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
597 static unsigned int cpu_smt_max_threads __ro_after_init;
598 unsigned int cpu_smt_num_threads __read_mostly = UINT_MAX;
600 void __init cpu_smt_disable(bool force)
602 if (!cpu_smt_possible())
606 pr_info("SMT: Force disabled\n");
607 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
609 pr_info("SMT: disabled\n");
610 cpu_smt_control = CPU_SMT_DISABLED;
612 cpu_smt_num_threads = 1;
616 * The decision whether SMT is supported can only be done after the full
617 * CPU identification. Called from architecture code.
619 void __init cpu_smt_set_num_threads(unsigned int num_threads,
620 unsigned int max_threads)
622 WARN_ON(!num_threads || (num_threads > max_threads));
624 if (max_threads == 1)
625 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
627 cpu_smt_max_threads = max_threads;
630 * If SMT has been disabled via the kernel command line or SMT is
631 * not supported, set cpu_smt_num_threads to 1 for consistency.
632 * If enabled, take the architecture requested number of threads
633 * to bring up into account.
635 if (cpu_smt_control != CPU_SMT_ENABLED)
636 cpu_smt_num_threads = 1;
637 else if (num_threads < cpu_smt_num_threads)
638 cpu_smt_num_threads = num_threads;
641 static int __init smt_cmdline_disable(char *str)
643 cpu_smt_disable(str && !strcmp(str, "force"));
646 early_param("nosmt", smt_cmdline_disable);
649 * For Archicture supporting partial SMT states check if the thread is allowed.
650 * Otherwise this has already been checked through cpu_smt_max_threads when
651 * setting the SMT level.
653 static inline bool cpu_smt_thread_allowed(unsigned int cpu)
655 #ifdef CONFIG_SMT_NUM_THREADS_DYNAMIC
656 return topology_smt_thread_allowed(cpu);
662 static inline bool cpu_bootable(unsigned int cpu)
664 if (cpu_smt_control == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
667 /* All CPUs are bootable if controls are not configured */
668 if (cpu_smt_control == CPU_SMT_NOT_IMPLEMENTED)
671 /* All CPUs are bootable if CPU is not SMT capable */
672 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
675 if (topology_is_primary_thread(cpu))
679 * On x86 it's required to boot all logical CPUs at least once so
680 * that the init code can get a chance to set CR4.MCE on each
681 * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
682 * core will shutdown the machine.
684 return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
687 /* Returns true if SMT is supported and not forcefully (irreversibly) disabled */
688 bool cpu_smt_possible(void)
690 return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
691 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
693 EXPORT_SYMBOL_GPL(cpu_smt_possible);
696 static inline bool cpu_bootable(unsigned int cpu) { return true; }
699 static inline enum cpuhp_state
700 cpuhp_set_state(int cpu, struct cpuhp_cpu_state *st, enum cpuhp_state target)
702 enum cpuhp_state prev_state = st->state;
703 bool bringup = st->state < target;
705 st->rollback = false;
710 st->bringup = bringup;
711 if (cpu_dying(cpu) != !bringup)
712 set_cpu_dying(cpu, !bringup);
718 cpuhp_reset_state(int cpu, struct cpuhp_cpu_state *st,
719 enum cpuhp_state prev_state)
721 bool bringup = !st->bringup;
723 st->target = prev_state;
726 * Already rolling back. No need invert the bringup value or to change
735 * If we have st->last we need to undo partial multi_instance of this
736 * state first. Otherwise start undo at the previous state.
745 st->bringup = bringup;
746 if (cpu_dying(cpu) != !bringup)
747 set_cpu_dying(cpu, !bringup);
750 /* Regular hotplug invocation of the AP hotplug thread */
751 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
753 if (!st->single && st->state == st->target)
758 * Make sure the above stores are visible before should_run becomes
759 * true. Paired with the mb() above in cpuhp_thread_fun()
762 st->should_run = true;
763 wake_up_process(st->thread);
764 wait_for_ap_thread(st, st->bringup);
767 static int cpuhp_kick_ap(int cpu, struct cpuhp_cpu_state *st,
768 enum cpuhp_state target)
770 enum cpuhp_state prev_state;
773 prev_state = cpuhp_set_state(cpu, st, target);
775 if ((ret = st->result)) {
776 cpuhp_reset_state(cpu, st, prev_state);
783 static int bringup_wait_for_ap_online(unsigned int cpu)
785 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
787 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
788 wait_for_ap_thread(st, true);
789 if (WARN_ON_ONCE((!cpu_online(cpu))))
792 /* Unpark the hotplug thread of the target cpu */
793 kthread_unpark(st->thread);
796 * SMT soft disabling on X86 requires to bring the CPU out of the
797 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
798 * CPU marked itself as booted_once in notify_cpu_starting() so the
799 * cpu_bootable() check will now return false if this is not the
802 if (!cpu_bootable(cpu))
807 #ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
808 static int cpuhp_kick_ap_alive(unsigned int cpu)
810 if (!cpuhp_can_boot_ap(cpu))
813 return arch_cpuhp_kick_ap_alive(cpu, idle_thread_get(cpu));
816 static int cpuhp_bringup_ap(unsigned int cpu)
818 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
822 * Some architectures have to walk the irq descriptors to
823 * setup the vector space for the cpu which comes online.
824 * Prevent irq alloc/free across the bringup.
828 ret = cpuhp_bp_sync_alive(cpu);
832 ret = bringup_wait_for_ap_online(cpu);
838 if (st->target <= CPUHP_AP_ONLINE_IDLE)
841 return cpuhp_kick_ap(cpu, st, st->target);
848 static int bringup_cpu(unsigned int cpu)
850 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
851 struct task_struct *idle = idle_thread_get(cpu);
854 if (!cpuhp_can_boot_ap(cpu))
858 * Some architectures have to walk the irq descriptors to
859 * setup the vector space for the cpu which comes online.
861 * Prevent irq alloc/free across the bringup by acquiring the
862 * sparse irq lock. Hold it until the upcoming CPU completes the
863 * startup in cpuhp_online_idle() which allows to avoid
864 * intermediate synchronization points in the architecture code.
868 ret = __cpu_up(cpu, idle);
872 ret = cpuhp_bp_sync_alive(cpu);
876 ret = bringup_wait_for_ap_online(cpu);
882 if (st->target <= CPUHP_AP_ONLINE_IDLE)
885 return cpuhp_kick_ap(cpu, st, st->target);
893 static int finish_cpu(unsigned int cpu)
895 struct task_struct *idle = idle_thread_get(cpu);
896 struct mm_struct *mm = idle->active_mm;
899 * idle_task_exit() will have switched to &init_mm, now
900 * clean up any remaining active_mm state.
903 idle->active_mm = &init_mm;
909 * Hotplug state machine related functions
913 * Get the next state to run. Empty ones will be skipped. Returns true if a
916 * st->state will be modified ahead of time, to match state_to_run, as if it
919 static bool cpuhp_next_state(bool bringup,
920 enum cpuhp_state *state_to_run,
921 struct cpuhp_cpu_state *st,
922 enum cpuhp_state target)
926 if (st->state >= target)
929 *state_to_run = ++st->state;
931 if (st->state <= target)
934 *state_to_run = st->state--;
937 if (!cpuhp_step_empty(bringup, cpuhp_get_step(*state_to_run)))
944 static int __cpuhp_invoke_callback_range(bool bringup,
946 struct cpuhp_cpu_state *st,
947 enum cpuhp_state target,
950 enum cpuhp_state state;
953 while (cpuhp_next_state(bringup, &state, st, target)) {
956 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
961 pr_warn("CPU %u %s state %s (%d) failed (%d)\n",
962 cpu, bringup ? "UP" : "DOWN",
963 cpuhp_get_step(st->state)->name,
975 static inline int cpuhp_invoke_callback_range(bool bringup,
977 struct cpuhp_cpu_state *st,
978 enum cpuhp_state target)
980 return __cpuhp_invoke_callback_range(bringup, cpu, st, target, false);
983 static inline void cpuhp_invoke_callback_range_nofail(bool bringup,
985 struct cpuhp_cpu_state *st,
986 enum cpuhp_state target)
988 __cpuhp_invoke_callback_range(bringup, cpu, st, target, true);
991 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
993 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
996 * When CPU hotplug is disabled, then taking the CPU down is not
997 * possible because takedown_cpu() and the architecture and
998 * subsystem specific mechanisms are not available. So the CPU
999 * which would be completely unplugged again needs to stay around
1000 * in the current state.
1002 return st->state <= CPUHP_BRINGUP_CPU;
1005 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1006 enum cpuhp_state target)
1008 enum cpuhp_state prev_state = st->state;
1011 ret = cpuhp_invoke_callback_range(true, cpu, st, target);
1013 pr_debug("CPU UP failed (%d) CPU %u state %s (%d)\n",
1014 ret, cpu, cpuhp_get_step(st->state)->name,
1017 cpuhp_reset_state(cpu, st, prev_state);
1018 if (can_rollback_cpu(st))
1019 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
1026 * The cpu hotplug threads manage the bringup and teardown of the cpus
1028 static int cpuhp_should_run(unsigned int cpu)
1030 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1032 return st->should_run;
1036 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
1037 * callbacks when a state gets [un]installed at runtime.
1039 * Each invocation of this function by the smpboot thread does a single AP
1042 * It has 3 modes of operation:
1043 * - single: runs st->cb_state
1044 * - up: runs ++st->state, while st->state < st->target
1045 * - down: runs st->state--, while st->state > st->target
1047 * When complete or on error, should_run is cleared and the completion is fired.
1049 static void cpuhp_thread_fun(unsigned int cpu)
1051 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1052 bool bringup = st->bringup;
1053 enum cpuhp_state state;
1055 if (WARN_ON_ONCE(!st->should_run))
1059 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
1060 * that if we see ->should_run we also see the rest of the state.
1065 * The BP holds the hotplug lock, but we're now running on the AP,
1066 * ensure that anybody asserting the lock is held, will actually find
1069 lockdep_acquire_cpus_lock();
1070 cpuhp_lock_acquire(bringup);
1073 state = st->cb_state;
1074 st->should_run = false;
1076 st->should_run = cpuhp_next_state(bringup, &state, st, st->target);
1077 if (!st->should_run)
1081 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
1083 if (cpuhp_is_atomic_state(state)) {
1084 local_irq_disable();
1085 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1089 * STARTING/DYING must not fail!
1091 WARN_ON_ONCE(st->result);
1093 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
1098 * If we fail on a rollback, we're up a creek without no
1099 * paddle, no way forward, no way back. We loose, thanks for
1102 WARN_ON_ONCE(st->rollback);
1103 st->should_run = false;
1107 cpuhp_lock_release(bringup);
1108 lockdep_release_cpus_lock();
1110 if (!st->should_run)
1111 complete_ap_thread(st, bringup);
1114 /* Invoke a single callback on a remote cpu */
1116 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
1117 struct hlist_node *node)
1119 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1122 if (!cpu_online(cpu))
1125 cpuhp_lock_acquire(false);
1126 cpuhp_lock_release(false);
1128 cpuhp_lock_acquire(true);
1129 cpuhp_lock_release(true);
1132 * If we are up and running, use the hotplug thread. For early calls
1133 * we invoke the thread function directly.
1136 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1138 st->rollback = false;
1142 st->bringup = bringup;
1143 st->cb_state = state;
1146 __cpuhp_kick_ap(st);
1149 * If we failed and did a partial, do a rollback.
1151 if ((ret = st->result) && st->last) {
1152 st->rollback = true;
1153 st->bringup = !bringup;
1155 __cpuhp_kick_ap(st);
1159 * Clean up the leftovers so the next hotplug operation wont use stale
1162 st->node = st->last = NULL;
1166 static int cpuhp_kick_ap_work(unsigned int cpu)
1168 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1169 enum cpuhp_state prev_state = st->state;
1172 cpuhp_lock_acquire(false);
1173 cpuhp_lock_release(false);
1175 cpuhp_lock_acquire(true);
1176 cpuhp_lock_release(true);
1178 trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
1179 ret = cpuhp_kick_ap(cpu, st, st->target);
1180 trace_cpuhp_exit(cpu, st->state, prev_state, ret);
1185 static struct smp_hotplug_thread cpuhp_threads = {
1186 .store = &cpuhp_state.thread,
1187 .thread_should_run = cpuhp_should_run,
1188 .thread_fn = cpuhp_thread_fun,
1189 .thread_comm = "cpuhp/%u",
1190 .selfparking = true,
1193 static __init void cpuhp_init_state(void)
1195 struct cpuhp_cpu_state *st;
1198 for_each_possible_cpu(cpu) {
1199 st = per_cpu_ptr(&cpuhp_state, cpu);
1200 init_completion(&st->done_up);
1201 init_completion(&st->done_down);
1205 void __init cpuhp_threads_init(void)
1208 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
1209 kthread_unpark(this_cpu_read(cpuhp_state.thread));
1214 * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
1217 * The operation is still serialized against concurrent CPU hotplug via
1218 * cpu_add_remove_lock, i.e. CPU map protection. But it is _not_
1219 * serialized against other hotplug related activity like adding or
1220 * removing of state callbacks and state instances, which invoke either the
1221 * startup or the teardown callback of the affected state.
1223 * This is required for subsystems which are unfixable vs. CPU hotplug and
1224 * evade lock inversion problems by scheduling work which has to be
1225 * completed _before_ cpu_up()/_cpu_down() returns.
1227 * Don't even think about adding anything to this for any new code or even
1228 * drivers. It's only purpose is to keep existing lock order trainwrecks
1231 * For cpu_down() there might be valid reasons to finish cleanups which are
1232 * not required to be done under cpu_hotplug_lock, but that's a different
1233 * story and would be not invoked via this.
1235 static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
1238 * cpusets delegate hotplug operations to a worker to "solve" the
1239 * lock order problems. Wait for the worker, but only if tasks are
1240 * _not_ frozen (suspend, hibernate) as that would wait forever.
1242 * The wait is required because otherwise the hotplug operation
1243 * returns with inconsistent state, which could even be observed in
1244 * user space when a new CPU is brought up. The CPU plug uevent
1245 * would be delivered and user space reacting on it would fail to
1246 * move tasks to the newly plugged CPU up to the point where the
1247 * work has finished because up to that point the newly plugged CPU
1248 * is not assignable in cpusets/cgroups. On unplug that's not
1249 * necessarily a visible issue, but it is still inconsistent state,
1250 * which is the real problem which needs to be "fixed". This can't
1251 * prevent the transient state between scheduling the work and
1252 * returning from waiting for it.
1255 cpuset_wait_for_hotplug();
1258 #ifdef CONFIG_HOTPLUG_CPU
1259 #ifndef arch_clear_mm_cpumask_cpu
1260 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
1264 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
1267 * This function walks all processes, finds a valid mm struct for each one and
1268 * then clears a corresponding bit in mm's cpumask. While this all sounds
1269 * trivial, there are various non-obvious corner cases, which this function
1270 * tries to solve in a safe manner.
1272 * Also note that the function uses a somewhat relaxed locking scheme, so it may
1273 * be called only for an already offlined CPU.
1275 void clear_tasks_mm_cpumask(int cpu)
1277 struct task_struct *p;
1280 * This function is called after the cpu is taken down and marked
1281 * offline, so its not like new tasks will ever get this cpu set in
1282 * their mm mask. -- Peter Zijlstra
1283 * Thus, we may use rcu_read_lock() here, instead of grabbing
1284 * full-fledged tasklist_lock.
1286 WARN_ON(cpu_online(cpu));
1288 for_each_process(p) {
1289 struct task_struct *t;
1292 * Main thread might exit, but other threads may still have
1293 * a valid mm. Find one.
1295 t = find_lock_task_mm(p);
1298 arch_clear_mm_cpumask_cpu(cpu, t->mm);
1304 /* Take this CPU down. */
1305 static int take_cpu_down(void *_param)
1307 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1308 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
1309 int err, cpu = smp_processor_id();
1311 /* Ensure this CPU doesn't handle any more interrupts. */
1312 err = __cpu_disable();
1317 * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
1318 * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
1320 WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
1323 * Invoke the former CPU_DYING callbacks. DYING must not fail!
1325 cpuhp_invoke_callback_range_nofail(false, cpu, st, target);
1327 /* Give up timekeeping duties */
1328 tick_handover_do_timer();
1329 /* Remove CPU from timer broadcasting */
1330 tick_offline_cpu(cpu);
1331 /* Park the stopper thread */
1332 stop_machine_park(cpu);
1336 static int takedown_cpu(unsigned int cpu)
1338 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1341 /* Park the smpboot threads */
1342 kthread_park(st->thread);
1345 * Prevent irq alloc/free while the dying cpu reorganizes the
1346 * interrupt affinities.
1351 * So now all preempt/rcu users must observe !cpu_active().
1353 err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
1355 /* CPU refused to die */
1356 irq_unlock_sparse();
1357 /* Unpark the hotplug thread so we can rollback there */
1358 kthread_unpark(st->thread);
1361 BUG_ON(cpu_online(cpu));
1364 * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1365 * all runnable tasks from the CPU, there's only the idle task left now
1366 * that the migration thread is done doing the stop_machine thing.
1368 * Wait for the stop thread to go away.
1370 wait_for_ap_thread(st, false);
1371 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1373 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1374 irq_unlock_sparse();
1376 hotplug_cpu__broadcast_tick_pull(cpu);
1377 /* This actually kills the CPU. */
1380 cpuhp_bp_sync_dead(cpu);
1382 tick_cleanup_dead_cpu(cpu);
1385 * Callbacks must be re-integrated right away to the RCU state machine.
1386 * Otherwise an RCU callback could block a further teardown function
1387 * waiting for its completion.
1389 rcutree_migrate_callbacks(cpu);
1394 static void cpuhp_complete_idle_dead(void *arg)
1396 struct cpuhp_cpu_state *st = arg;
1398 complete_ap_thread(st, false);
1401 void cpuhp_report_idle_dead(void)
1403 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1405 BUG_ON(st->state != CPUHP_AP_OFFLINE);
1406 rcutree_report_cpu_dead();
1407 st->state = CPUHP_AP_IDLE_DEAD;
1409 * We cannot call complete after rcutree_report_cpu_dead() so we delegate it
1412 smp_call_function_single(cpumask_first(cpu_online_mask),
1413 cpuhp_complete_idle_dead, st, 0);
1416 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1417 enum cpuhp_state target)
1419 enum cpuhp_state prev_state = st->state;
1422 ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1424 pr_debug("CPU DOWN failed (%d) CPU %u state %s (%d)\n",
1425 ret, cpu, cpuhp_get_step(st->state)->name,
1428 cpuhp_reset_state(cpu, st, prev_state);
1430 if (st->state < prev_state)
1431 WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1438 /* Requires cpu_add_remove_lock to be held */
1439 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1440 enum cpuhp_state target)
1442 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1443 int prev_state, ret = 0;
1445 if (num_online_cpus() == 1)
1448 if (!cpu_present(cpu))
1453 cpuhp_tasks_frozen = tasks_frozen;
1455 prev_state = cpuhp_set_state(cpu, st, target);
1457 * If the current CPU state is in the range of the AP hotplug thread,
1458 * then we need to kick the thread.
1460 if (st->state > CPUHP_TEARDOWN_CPU) {
1461 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1462 ret = cpuhp_kick_ap_work(cpu);
1464 * The AP side has done the error rollback already. Just
1465 * return the error code..
1471 * We might have stopped still in the range of the AP hotplug
1472 * thread. Nothing to do anymore.
1474 if (st->state > CPUHP_TEARDOWN_CPU)
1477 st->target = target;
1480 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1481 * to do the further cleanups.
1483 ret = cpuhp_down_callbacks(cpu, st, target);
1484 if (ret && st->state < prev_state) {
1485 if (st->state == CPUHP_TEARDOWN_CPU) {
1486 cpuhp_reset_state(cpu, st, prev_state);
1487 __cpuhp_kick_ap(st);
1489 WARN(1, "DEAD callback error for CPU%d", cpu);
1494 cpus_write_unlock();
1496 * Do post unplug cleanup. This is still protected against
1497 * concurrent CPU hotplug via cpu_add_remove_lock.
1499 lockup_detector_cleanup();
1501 cpu_up_down_serialize_trainwrecks(tasks_frozen);
1505 struct cpu_down_work {
1507 enum cpuhp_state target;
1510 static long __cpu_down_maps_locked(void *arg)
1512 struct cpu_down_work *work = arg;
1514 return _cpu_down(work->cpu, 0, work->target);
1517 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1519 struct cpu_down_work work = { .cpu = cpu, .target = target, };
1522 * If the platform does not support hotplug, report it explicitly to
1523 * differentiate it from a transient offlining failure.
1525 if (cc_platform_has(CC_ATTR_HOTPLUG_DISABLED))
1527 if (cpu_hotplug_disabled)
1531 * Ensure that the control task does not run on the to be offlined
1532 * CPU to prevent a deadlock against cfs_b->period_timer.
1533 * Also keep at least one housekeeping cpu onlined to avoid generating
1534 * an empty sched_domain span.
1536 for_each_cpu_and(cpu, cpu_online_mask, housekeeping_cpumask(HK_TYPE_DOMAIN)) {
1537 if (cpu != work.cpu)
1538 return work_on_cpu(cpu, __cpu_down_maps_locked, &work);
1543 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1547 cpu_maps_update_begin();
1548 err = cpu_down_maps_locked(cpu, target);
1549 cpu_maps_update_done();
1554 * cpu_device_down - Bring down a cpu device
1555 * @dev: Pointer to the cpu device to offline
1557 * This function is meant to be used by device core cpu subsystem only.
1559 * Other subsystems should use remove_cpu() instead.
1561 * Return: %0 on success or a negative errno code
1563 int cpu_device_down(struct device *dev)
1565 return cpu_down(dev->id, CPUHP_OFFLINE);
1568 int remove_cpu(unsigned int cpu)
1572 lock_device_hotplug();
1573 ret = device_offline(get_cpu_device(cpu));
1574 unlock_device_hotplug();
1578 EXPORT_SYMBOL_GPL(remove_cpu);
1580 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1585 cpu_maps_update_begin();
1588 * Make certain the cpu I'm about to reboot on is online.
1590 * This is inline to what migrate_to_reboot_cpu() already do.
1592 if (!cpu_online(primary_cpu))
1593 primary_cpu = cpumask_first(cpu_online_mask);
1595 for_each_online_cpu(cpu) {
1596 if (cpu == primary_cpu)
1599 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1601 pr_err("Failed to offline CPU%d - error=%d",
1608 * Ensure all but the reboot CPU are offline.
1610 BUG_ON(num_online_cpus() > 1);
1613 * Make sure the CPUs won't be enabled by someone else after this
1614 * point. Kexec will reboot to a new kernel shortly resetting
1615 * everything along the way.
1617 cpu_hotplug_disabled++;
1619 cpu_maps_update_done();
1623 #define takedown_cpu NULL
1624 #endif /*CONFIG_HOTPLUG_CPU*/
1627 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1628 * @cpu: cpu that just started
1630 * It must be called by the arch code on the new cpu, before the new cpu
1631 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1633 void notify_cpu_starting(unsigned int cpu)
1635 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1636 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1638 rcutree_report_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1639 cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1642 * STARTING must not fail!
1644 cpuhp_invoke_callback_range_nofail(true, cpu, st, target);
1648 * Called from the idle task. Wake up the controlling task which brings the
1649 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1650 * online bringup to the hotplug thread.
1652 void cpuhp_online_idle(enum cpuhp_state state)
1654 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1656 /* Happens for the boot cpu */
1657 if (state != CPUHP_AP_ONLINE_IDLE)
1660 cpuhp_ap_update_sync_state(SYNC_STATE_ONLINE);
1663 * Unpark the stopper thread before we start the idle loop (and start
1664 * scheduling); this ensures the stopper task is always available.
1666 stop_machine_unpark(smp_processor_id());
1668 st->state = CPUHP_AP_ONLINE_IDLE;
1669 complete_ap_thread(st, true);
1672 /* Requires cpu_add_remove_lock to be held */
1673 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1675 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1676 struct task_struct *idle;
1681 if (!cpu_present(cpu)) {
1687 * The caller of cpu_up() might have raced with another
1688 * caller. Nothing to do.
1690 if (st->state >= target)
1693 if (st->state == CPUHP_OFFLINE) {
1694 /* Let it fail before we try to bring the cpu up */
1695 idle = idle_thread_get(cpu);
1697 ret = PTR_ERR(idle);
1702 * Reset stale stack state from the last time this CPU was online.
1704 scs_task_reset(idle);
1705 kasan_unpoison_task_stack(idle);
1708 cpuhp_tasks_frozen = tasks_frozen;
1710 cpuhp_set_state(cpu, st, target);
1712 * If the current CPU state is in the range of the AP hotplug thread,
1713 * then we need to kick the thread once more.
1715 if (st->state > CPUHP_BRINGUP_CPU) {
1716 ret = cpuhp_kick_ap_work(cpu);
1718 * The AP side has done the error rollback already. Just
1719 * return the error code..
1726 * Try to reach the target state. We max out on the BP at
1727 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1728 * responsible for bringing it up to the target state.
1730 target = min((int)target, CPUHP_BRINGUP_CPU);
1731 ret = cpuhp_up_callbacks(cpu, st, target);
1733 cpus_write_unlock();
1735 cpu_up_down_serialize_trainwrecks(tasks_frozen);
1739 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1743 if (!cpu_possible(cpu)) {
1744 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1749 err = try_online_node(cpu_to_node(cpu));
1753 cpu_maps_update_begin();
1755 if (cpu_hotplug_disabled) {
1759 if (!cpu_bootable(cpu)) {
1764 err = _cpu_up(cpu, 0, target);
1766 cpu_maps_update_done();
1771 * cpu_device_up - Bring up a cpu device
1772 * @dev: Pointer to the cpu device to online
1774 * This function is meant to be used by device core cpu subsystem only.
1776 * Other subsystems should use add_cpu() instead.
1778 * Return: %0 on success or a negative errno code
1780 int cpu_device_up(struct device *dev)
1782 return cpu_up(dev->id, CPUHP_ONLINE);
1785 int add_cpu(unsigned int cpu)
1789 lock_device_hotplug();
1790 ret = device_online(get_cpu_device(cpu));
1791 unlock_device_hotplug();
1795 EXPORT_SYMBOL_GPL(add_cpu);
1798 * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1799 * @sleep_cpu: The cpu we hibernated on and should be brought up.
1801 * On some architectures like arm64, we can hibernate on any CPU, but on
1802 * wake up the CPU we hibernated on might be offline as a side effect of
1803 * using maxcpus= for example.
1805 * Return: %0 on success or a negative errno code
1807 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1811 if (!cpu_online(sleep_cpu)) {
1812 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1813 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1815 pr_err("Failed to bring hibernate-CPU up!\n");
1822 static void __init cpuhp_bringup_mask(const struct cpumask *mask, unsigned int ncpus,
1823 enum cpuhp_state target)
1827 for_each_cpu(cpu, mask) {
1828 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1830 if (cpu_up(cpu, target) && can_rollback_cpu(st)) {
1832 * If this failed then cpu_up() might have only
1833 * rolled back to CPUHP_BP_KICK_AP for the final
1834 * online. Clean it up. NOOP if already rolled back.
1836 WARN_ON(cpuhp_invoke_callback_range(false, cpu, st, CPUHP_OFFLINE));
1844 #ifdef CONFIG_HOTPLUG_PARALLEL
1845 static bool __cpuhp_parallel_bringup __ro_after_init = true;
1847 static int __init parallel_bringup_parse_param(char *arg)
1849 return kstrtobool(arg, &__cpuhp_parallel_bringup);
1851 early_param("cpuhp.parallel", parallel_bringup_parse_param);
1853 static inline bool cpuhp_smt_aware(void)
1855 return cpu_smt_max_threads > 1;
1858 static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
1860 return cpu_primary_thread_mask;
1864 * On architectures which have enabled parallel bringup this invokes all BP
1865 * prepare states for each of the to be onlined APs first. The last state
1866 * sends the startup IPI to the APs. The APs proceed through the low level
1867 * bringup code in parallel and then wait for the control CPU to release
1868 * them one by one for the final onlining procedure.
1870 * This avoids waiting for each AP to respond to the startup IPI in
1871 * CPUHP_BRINGUP_CPU.
1873 static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
1875 const struct cpumask *mask = cpu_present_mask;
1877 if (__cpuhp_parallel_bringup)
1878 __cpuhp_parallel_bringup = arch_cpuhp_init_parallel_bringup();
1879 if (!__cpuhp_parallel_bringup)
1882 if (cpuhp_smt_aware()) {
1883 const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
1884 static struct cpumask tmp_mask __initdata;
1887 * X86 requires to prevent that SMT siblings stopped while
1888 * the primary thread does a microcode update for various
1889 * reasons. Bring the primary threads up first.
1891 cpumask_and(&tmp_mask, mask, pmask);
1892 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_BP_KICK_AP);
1893 cpuhp_bringup_mask(&tmp_mask, ncpus, CPUHP_ONLINE);
1894 /* Account for the online CPUs */
1895 ncpus -= num_online_cpus();
1898 /* Create the mask for secondary CPUs */
1899 cpumask_andnot(&tmp_mask, mask, pmask);
1903 /* Bring the not-yet started CPUs up */
1904 cpuhp_bringup_mask(mask, ncpus, CPUHP_BP_KICK_AP);
1905 cpuhp_bringup_mask(mask, ncpus, CPUHP_ONLINE);
1909 static inline bool cpuhp_bringup_cpus_parallel(unsigned int ncpus) { return false; }
1910 #endif /* CONFIG_HOTPLUG_PARALLEL */
1912 void __init bringup_nonboot_cpus(unsigned int setup_max_cpus)
1914 /* Try parallel bringup optimization if enabled */
1915 if (cpuhp_bringup_cpus_parallel(setup_max_cpus))
1918 /* Full per CPU serialized bringup */
1919 cpuhp_bringup_mask(cpu_present_mask, setup_max_cpus, CPUHP_ONLINE);
1922 #ifdef CONFIG_PM_SLEEP_SMP
1923 static cpumask_var_t frozen_cpus;
1925 int freeze_secondary_cpus(int primary)
1929 cpu_maps_update_begin();
1930 if (primary == -1) {
1931 primary = cpumask_first(cpu_online_mask);
1932 if (!housekeeping_cpu(primary, HK_TYPE_TIMER))
1933 primary = housekeeping_any_cpu(HK_TYPE_TIMER);
1935 if (!cpu_online(primary))
1936 primary = cpumask_first(cpu_online_mask);
1940 * We take down all of the non-boot CPUs in one shot to avoid races
1941 * with the userspace trying to use the CPU hotplug at the same time
1943 cpumask_clear(frozen_cpus);
1945 pr_info("Disabling non-boot CPUs ...\n");
1946 for_each_online_cpu(cpu) {
1950 if (pm_wakeup_pending()) {
1951 pr_info("Wakeup pending. Abort CPU freeze\n");
1956 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1957 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1958 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1960 cpumask_set_cpu(cpu, frozen_cpus);
1962 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1968 BUG_ON(num_online_cpus() > 1);
1970 pr_err("Non-boot CPUs are not disabled\n");
1973 * Make sure the CPUs won't be enabled by someone else. We need to do
1974 * this even in case of failure as all freeze_secondary_cpus() users are
1975 * supposed to do thaw_secondary_cpus() on the failure path.
1977 cpu_hotplug_disabled++;
1979 cpu_maps_update_done();
1983 void __weak arch_thaw_secondary_cpus_begin(void)
1987 void __weak arch_thaw_secondary_cpus_end(void)
1991 void thaw_secondary_cpus(void)
1995 /* Allow everyone to use the CPU hotplug again */
1996 cpu_maps_update_begin();
1997 __cpu_hotplug_enable();
1998 if (cpumask_empty(frozen_cpus))
2001 pr_info("Enabling non-boot CPUs ...\n");
2003 arch_thaw_secondary_cpus_begin();
2005 for_each_cpu(cpu, frozen_cpus) {
2006 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
2007 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
2008 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
2010 pr_info("CPU%d is up\n", cpu);
2013 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
2016 arch_thaw_secondary_cpus_end();
2018 cpumask_clear(frozen_cpus);
2020 cpu_maps_update_done();
2023 static int __init alloc_frozen_cpus(void)
2025 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
2029 core_initcall(alloc_frozen_cpus);
2032 * When callbacks for CPU hotplug notifications are being executed, we must
2033 * ensure that the state of the system with respect to the tasks being frozen
2034 * or not, as reported by the notification, remains unchanged *throughout the
2035 * duration* of the execution of the callbacks.
2036 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
2038 * This synchronization is implemented by mutually excluding regular CPU
2039 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
2040 * Hibernate notifications.
2043 cpu_hotplug_pm_callback(struct notifier_block *nb,
2044 unsigned long action, void *ptr)
2048 case PM_SUSPEND_PREPARE:
2049 case PM_HIBERNATION_PREPARE:
2050 cpu_hotplug_disable();
2053 case PM_POST_SUSPEND:
2054 case PM_POST_HIBERNATION:
2055 cpu_hotplug_enable();
2066 static int __init cpu_hotplug_pm_sync_init(void)
2069 * cpu_hotplug_pm_callback has higher priority than x86
2070 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
2071 * to disable cpu hotplug to avoid cpu hotplug race.
2073 pm_notifier(cpu_hotplug_pm_callback, 0);
2076 core_initcall(cpu_hotplug_pm_sync_init);
2078 #endif /* CONFIG_PM_SLEEP_SMP */
2082 #endif /* CONFIG_SMP */
2084 /* Boot processor state steps */
2085 static struct cpuhp_step cpuhp_hp_states[] = {
2088 .startup.single = NULL,
2089 .teardown.single = NULL,
2092 [CPUHP_CREATE_THREADS]= {
2093 .name = "threads:prepare",
2094 .startup.single = smpboot_create_threads,
2095 .teardown.single = NULL,
2098 [CPUHP_PERF_PREPARE] = {
2099 .name = "perf:prepare",
2100 .startup.single = perf_event_init_cpu,
2101 .teardown.single = perf_event_exit_cpu,
2103 [CPUHP_RANDOM_PREPARE] = {
2104 .name = "random:prepare",
2105 .startup.single = random_prepare_cpu,
2106 .teardown.single = NULL,
2108 [CPUHP_WORKQUEUE_PREP] = {
2109 .name = "workqueue:prepare",
2110 .startup.single = workqueue_prepare_cpu,
2111 .teardown.single = NULL,
2113 [CPUHP_HRTIMERS_PREPARE] = {
2114 .name = "hrtimers:prepare",
2115 .startup.single = hrtimers_prepare_cpu,
2116 .teardown.single = hrtimers_dead_cpu,
2118 [CPUHP_SMPCFD_PREPARE] = {
2119 .name = "smpcfd:prepare",
2120 .startup.single = smpcfd_prepare_cpu,
2121 .teardown.single = smpcfd_dead_cpu,
2123 [CPUHP_RELAY_PREPARE] = {
2124 .name = "relay:prepare",
2125 .startup.single = relay_prepare_cpu,
2126 .teardown.single = NULL,
2128 [CPUHP_SLAB_PREPARE] = {
2129 .name = "slab:prepare",
2130 .startup.single = slab_prepare_cpu,
2131 .teardown.single = slab_dead_cpu,
2133 [CPUHP_RCUTREE_PREP] = {
2134 .name = "RCU/tree:prepare",
2135 .startup.single = rcutree_prepare_cpu,
2136 .teardown.single = rcutree_dead_cpu,
2139 * On the tear-down path, timers_dead_cpu() must be invoked
2140 * before blk_mq_queue_reinit_notify() from notify_dead(),
2141 * otherwise a RCU stall occurs.
2143 [CPUHP_TIMERS_PREPARE] = {
2144 .name = "timers:prepare",
2145 .startup.single = timers_prepare_cpu,
2146 .teardown.single = timers_dead_cpu,
2149 #ifdef CONFIG_HOTPLUG_SPLIT_STARTUP
2151 * Kicks the AP alive. AP will wait in cpuhp_ap_sync_alive() until
2152 * the next step will release it.
2154 [CPUHP_BP_KICK_AP] = {
2155 .name = "cpu:kick_ap",
2156 .startup.single = cpuhp_kick_ap_alive,
2160 * Waits for the AP to reach cpuhp_ap_sync_alive() and then
2161 * releases it for the complete bringup.
2163 [CPUHP_BRINGUP_CPU] = {
2164 .name = "cpu:bringup",
2165 .startup.single = cpuhp_bringup_ap,
2166 .teardown.single = finish_cpu,
2171 * All-in-one CPU bringup state which includes the kick alive.
2173 [CPUHP_BRINGUP_CPU] = {
2174 .name = "cpu:bringup",
2175 .startup.single = bringup_cpu,
2176 .teardown.single = finish_cpu,
2180 /* Final state before CPU kills itself */
2181 [CPUHP_AP_IDLE_DEAD] = {
2182 .name = "idle:dead",
2185 * Last state before CPU enters the idle loop to die. Transient state
2186 * for synchronization.
2188 [CPUHP_AP_OFFLINE] = {
2189 .name = "ap:offline",
2192 /* First state is scheduler control. Interrupts are disabled */
2193 [CPUHP_AP_SCHED_STARTING] = {
2194 .name = "sched:starting",
2195 .startup.single = sched_cpu_starting,
2196 .teardown.single = sched_cpu_dying,
2198 [CPUHP_AP_RCUTREE_DYING] = {
2199 .name = "RCU/tree:dying",
2200 .startup.single = NULL,
2201 .teardown.single = rcutree_dying_cpu,
2203 [CPUHP_AP_SMPCFD_DYING] = {
2204 .name = "smpcfd:dying",
2205 .startup.single = NULL,
2206 .teardown.single = smpcfd_dying_cpu,
2208 /* Entry state on starting. Interrupts enabled from here on. Transient
2209 * state for synchronsization */
2210 [CPUHP_AP_ONLINE] = {
2211 .name = "ap:online",
2214 * Handled on control processor until the plugged processor manages
2217 [CPUHP_TEARDOWN_CPU] = {
2218 .name = "cpu:teardown",
2219 .startup.single = NULL,
2220 .teardown.single = takedown_cpu,
2224 [CPUHP_AP_SCHED_WAIT_EMPTY] = {
2225 .name = "sched:waitempty",
2226 .startup.single = NULL,
2227 .teardown.single = sched_cpu_wait_empty,
2230 /* Handle smpboot threads park/unpark */
2231 [CPUHP_AP_SMPBOOT_THREADS] = {
2232 .name = "smpboot/threads:online",
2233 .startup.single = smpboot_unpark_threads,
2234 .teardown.single = smpboot_park_threads,
2236 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
2237 .name = "irq/affinity:online",
2238 .startup.single = irq_affinity_online_cpu,
2239 .teardown.single = NULL,
2241 [CPUHP_AP_PERF_ONLINE] = {
2242 .name = "perf:online",
2243 .startup.single = perf_event_init_cpu,
2244 .teardown.single = perf_event_exit_cpu,
2246 [CPUHP_AP_WATCHDOG_ONLINE] = {
2247 .name = "lockup_detector:online",
2248 .startup.single = lockup_detector_online_cpu,
2249 .teardown.single = lockup_detector_offline_cpu,
2251 [CPUHP_AP_WORKQUEUE_ONLINE] = {
2252 .name = "workqueue:online",
2253 .startup.single = workqueue_online_cpu,
2254 .teardown.single = workqueue_offline_cpu,
2256 [CPUHP_AP_RANDOM_ONLINE] = {
2257 .name = "random:online",
2258 .startup.single = random_online_cpu,
2259 .teardown.single = NULL,
2261 [CPUHP_AP_RCUTREE_ONLINE] = {
2262 .name = "RCU/tree:online",
2263 .startup.single = rcutree_online_cpu,
2264 .teardown.single = rcutree_offline_cpu,
2268 * The dynamically registered state space is here
2272 /* Last state is scheduler control setting the cpu active */
2273 [CPUHP_AP_ACTIVE] = {
2274 .name = "sched:active",
2275 .startup.single = sched_cpu_activate,
2276 .teardown.single = sched_cpu_deactivate,
2280 /* CPU is fully up and running. */
2283 .startup.single = NULL,
2284 .teardown.single = NULL,
2288 /* Sanity check for callbacks */
2289 static int cpuhp_cb_check(enum cpuhp_state state)
2291 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
2297 * Returns a free for dynamic slot assignment of the Online state. The states
2298 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
2299 * by having no name assigned.
2301 static int cpuhp_reserve_state(enum cpuhp_state state)
2303 enum cpuhp_state i, end;
2304 struct cpuhp_step *step;
2307 case CPUHP_AP_ONLINE_DYN:
2308 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
2309 end = CPUHP_AP_ONLINE_DYN_END;
2311 case CPUHP_BP_PREPARE_DYN:
2312 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
2313 end = CPUHP_BP_PREPARE_DYN_END;
2319 for (i = state; i <= end; i++, step++) {
2323 WARN(1, "No more dynamic states available for CPU hotplug\n");
2327 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
2328 int (*startup)(unsigned int cpu),
2329 int (*teardown)(unsigned int cpu),
2330 bool multi_instance)
2332 /* (Un)Install the callbacks for further cpu hotplug operations */
2333 struct cpuhp_step *sp;
2337 * If name is NULL, then the state gets removed.
2339 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
2340 * the first allocation from these dynamic ranges, so the removal
2341 * would trigger a new allocation and clear the wrong (already
2342 * empty) state, leaving the callbacks of the to be cleared state
2343 * dangling, which causes wreckage on the next hotplug operation.
2345 if (name && (state == CPUHP_AP_ONLINE_DYN ||
2346 state == CPUHP_BP_PREPARE_DYN)) {
2347 ret = cpuhp_reserve_state(state);
2352 sp = cpuhp_get_step(state);
2353 if (name && sp->name)
2356 sp->startup.single = startup;
2357 sp->teardown.single = teardown;
2359 sp->multi_instance = multi_instance;
2360 INIT_HLIST_HEAD(&sp->list);
2364 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
2366 return cpuhp_get_step(state)->teardown.single;
2370 * Call the startup/teardown function for a step either on the AP or
2371 * on the current CPU.
2373 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
2374 struct hlist_node *node)
2376 struct cpuhp_step *sp = cpuhp_get_step(state);
2380 * If there's nothing to do, we done.
2381 * Relies on the union for multi_instance.
2383 if (cpuhp_step_empty(bringup, sp))
2386 * The non AP bound callbacks can fail on bringup. On teardown
2387 * e.g. module removal we crash for now.
2390 if (cpuhp_is_ap_state(state))
2391 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
2393 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2395 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
2397 BUG_ON(ret && !bringup);
2402 * Called from __cpuhp_setup_state on a recoverable failure.
2404 * Note: The teardown callbacks for rollback are not allowed to fail!
2406 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
2407 struct hlist_node *node)
2411 /* Roll back the already executed steps on the other cpus */
2412 for_each_present_cpu(cpu) {
2413 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2414 int cpustate = st->state;
2416 if (cpu >= failedcpu)
2419 /* Did we invoke the startup call on that cpu ? */
2420 if (cpustate >= state)
2421 cpuhp_issue_call(cpu, state, false, node);
2425 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
2426 struct hlist_node *node,
2429 struct cpuhp_step *sp;
2433 lockdep_assert_cpus_held();
2435 sp = cpuhp_get_step(state);
2436 if (sp->multi_instance == false)
2439 mutex_lock(&cpuhp_state_mutex);
2441 if (!invoke || !sp->startup.multi)
2445 * Try to call the startup callback for each present cpu
2446 * depending on the hotplug state of the cpu.
2448 for_each_present_cpu(cpu) {
2449 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2450 int cpustate = st->state;
2452 if (cpustate < state)
2455 ret = cpuhp_issue_call(cpu, state, true, node);
2457 if (sp->teardown.multi)
2458 cpuhp_rollback_install(cpu, state, node);
2464 hlist_add_head(node, &sp->list);
2466 mutex_unlock(&cpuhp_state_mutex);
2470 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
2476 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
2480 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
2483 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
2484 * @state: The state to setup
2485 * @name: Name of the step
2486 * @invoke: If true, the startup function is invoked for cpus where
2487 * cpu state >= @state
2488 * @startup: startup callback function
2489 * @teardown: teardown callback function
2490 * @multi_instance: State is set up for multiple instances which get
2493 * The caller needs to hold cpus read locked while calling this function.
2496 * Positive state number if @state is CPUHP_AP_ONLINE_DYN;
2497 * 0 for all other states
2498 * On failure: proper (negative) error code
2500 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
2501 const char *name, bool invoke,
2502 int (*startup)(unsigned int cpu),
2503 int (*teardown)(unsigned int cpu),
2504 bool multi_instance)
2509 lockdep_assert_cpus_held();
2511 if (cpuhp_cb_check(state) || !name)
2514 mutex_lock(&cpuhp_state_mutex);
2516 ret = cpuhp_store_callbacks(state, name, startup, teardown,
2519 dynstate = state == CPUHP_AP_ONLINE_DYN;
2520 if (ret > 0 && dynstate) {
2525 if (ret || !invoke || !startup)
2529 * Try to call the startup callback for each present cpu
2530 * depending on the hotplug state of the cpu.
2532 for_each_present_cpu(cpu) {
2533 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2534 int cpustate = st->state;
2536 if (cpustate < state)
2539 ret = cpuhp_issue_call(cpu, state, true, NULL);
2542 cpuhp_rollback_install(cpu, state, NULL);
2543 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2548 mutex_unlock(&cpuhp_state_mutex);
2550 * If the requested state is CPUHP_AP_ONLINE_DYN, return the
2551 * dynamically allocated state in case of success.
2553 if (!ret && dynstate)
2557 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2559 int __cpuhp_setup_state(enum cpuhp_state state,
2560 const char *name, bool invoke,
2561 int (*startup)(unsigned int cpu),
2562 int (*teardown)(unsigned int cpu),
2563 bool multi_instance)
2568 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2569 teardown, multi_instance);
2573 EXPORT_SYMBOL(__cpuhp_setup_state);
2575 int __cpuhp_state_remove_instance(enum cpuhp_state state,
2576 struct hlist_node *node, bool invoke)
2578 struct cpuhp_step *sp = cpuhp_get_step(state);
2581 BUG_ON(cpuhp_cb_check(state));
2583 if (!sp->multi_instance)
2587 mutex_lock(&cpuhp_state_mutex);
2589 if (!invoke || !cpuhp_get_teardown_cb(state))
2592 * Call the teardown callback for each present cpu depending
2593 * on the hotplug state of the cpu. This function is not
2594 * allowed to fail currently!
2596 for_each_present_cpu(cpu) {
2597 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2598 int cpustate = st->state;
2600 if (cpustate >= state)
2601 cpuhp_issue_call(cpu, state, false, node);
2606 mutex_unlock(&cpuhp_state_mutex);
2611 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2614 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2615 * @state: The state to remove
2616 * @invoke: If true, the teardown function is invoked for cpus where
2617 * cpu state >= @state
2619 * The caller needs to hold cpus read locked while calling this function.
2620 * The teardown callback is currently not allowed to fail. Think
2621 * about module removal!
2623 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2625 struct cpuhp_step *sp = cpuhp_get_step(state);
2628 BUG_ON(cpuhp_cb_check(state));
2630 lockdep_assert_cpus_held();
2632 mutex_lock(&cpuhp_state_mutex);
2633 if (sp->multi_instance) {
2634 WARN(!hlist_empty(&sp->list),
2635 "Error: Removing state %d which has instances left.\n",
2640 if (!invoke || !cpuhp_get_teardown_cb(state))
2644 * Call the teardown callback for each present cpu depending
2645 * on the hotplug state of the cpu. This function is not
2646 * allowed to fail currently!
2648 for_each_present_cpu(cpu) {
2649 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2650 int cpustate = st->state;
2652 if (cpustate >= state)
2653 cpuhp_issue_call(cpu, state, false, NULL);
2656 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2657 mutex_unlock(&cpuhp_state_mutex);
2659 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2661 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2664 __cpuhp_remove_state_cpuslocked(state, invoke);
2667 EXPORT_SYMBOL(__cpuhp_remove_state);
2669 #ifdef CONFIG_HOTPLUG_SMT
2670 static void cpuhp_offline_cpu_device(unsigned int cpu)
2672 struct device *dev = get_cpu_device(cpu);
2674 dev->offline = true;
2675 /* Tell user space about the state change */
2676 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2679 static void cpuhp_online_cpu_device(unsigned int cpu)
2681 struct device *dev = get_cpu_device(cpu);
2683 dev->offline = false;
2684 /* Tell user space about the state change */
2685 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2688 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2692 cpu_maps_update_begin();
2693 for_each_online_cpu(cpu) {
2694 if (topology_is_primary_thread(cpu))
2697 * Disable can be called with CPU_SMT_ENABLED when changing
2698 * from a higher to lower number of SMT threads per core.
2700 if (ctrlval == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
2702 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2706 * As this needs to hold the cpu maps lock it's impossible
2707 * to call device_offline() because that ends up calling
2708 * cpu_down() which takes cpu maps lock. cpu maps lock
2709 * needs to be held as this might race against in kernel
2710 * abusers of the hotplug machinery (thermal management).
2712 * So nothing would update device:offline state. That would
2713 * leave the sysfs entry stale and prevent onlining after
2714 * smt control has been changed to 'off' again. This is
2715 * called under the sysfs hotplug lock, so it is properly
2716 * serialized against the regular offline usage.
2718 cpuhp_offline_cpu_device(cpu);
2721 cpu_smt_control = ctrlval;
2722 cpu_maps_update_done();
2726 int cpuhp_smt_enable(void)
2730 cpu_maps_update_begin();
2731 cpu_smt_control = CPU_SMT_ENABLED;
2732 for_each_present_cpu(cpu) {
2733 /* Skip online CPUs and CPUs on offline nodes */
2734 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2736 if (!cpu_smt_thread_allowed(cpu))
2738 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2741 /* See comment in cpuhp_smt_disable() */
2742 cpuhp_online_cpu_device(cpu);
2744 cpu_maps_update_done();
2749 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2750 static ssize_t state_show(struct device *dev,
2751 struct device_attribute *attr, char *buf)
2753 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2755 return sprintf(buf, "%d\n", st->state);
2757 static DEVICE_ATTR_RO(state);
2759 static ssize_t target_store(struct device *dev, struct device_attribute *attr,
2760 const char *buf, size_t count)
2762 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2763 struct cpuhp_step *sp;
2766 ret = kstrtoint(buf, 10, &target);
2770 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2771 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2774 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2778 ret = lock_device_hotplug_sysfs();
2782 mutex_lock(&cpuhp_state_mutex);
2783 sp = cpuhp_get_step(target);
2784 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2785 mutex_unlock(&cpuhp_state_mutex);
2789 if (st->state < target)
2790 ret = cpu_up(dev->id, target);
2791 else if (st->state > target)
2792 ret = cpu_down(dev->id, target);
2793 else if (WARN_ON(st->target != target))
2794 st->target = target;
2796 unlock_device_hotplug();
2797 return ret ? ret : count;
2800 static ssize_t target_show(struct device *dev,
2801 struct device_attribute *attr, char *buf)
2803 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2805 return sprintf(buf, "%d\n", st->target);
2807 static DEVICE_ATTR_RW(target);
2809 static ssize_t fail_store(struct device *dev, struct device_attribute *attr,
2810 const char *buf, size_t count)
2812 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2813 struct cpuhp_step *sp;
2816 ret = kstrtoint(buf, 10, &fail);
2820 if (fail == CPUHP_INVALID) {
2825 if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2829 * Cannot fail STARTING/DYING callbacks.
2831 if (cpuhp_is_atomic_state(fail))
2835 * DEAD callbacks cannot fail...
2836 * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2837 * triggering STARTING callbacks, a failure in this state would
2840 if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2844 * Cannot fail anything that doesn't have callbacks.
2846 mutex_lock(&cpuhp_state_mutex);
2847 sp = cpuhp_get_step(fail);
2848 if (!sp->startup.single && !sp->teardown.single)
2850 mutex_unlock(&cpuhp_state_mutex);
2859 static ssize_t fail_show(struct device *dev,
2860 struct device_attribute *attr, char *buf)
2862 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2864 return sprintf(buf, "%d\n", st->fail);
2867 static DEVICE_ATTR_RW(fail);
2869 static struct attribute *cpuhp_cpu_attrs[] = {
2870 &dev_attr_state.attr,
2871 &dev_attr_target.attr,
2872 &dev_attr_fail.attr,
2876 static const struct attribute_group cpuhp_cpu_attr_group = {
2877 .attrs = cpuhp_cpu_attrs,
2882 static ssize_t states_show(struct device *dev,
2883 struct device_attribute *attr, char *buf)
2885 ssize_t cur, res = 0;
2888 mutex_lock(&cpuhp_state_mutex);
2889 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2890 struct cpuhp_step *sp = cpuhp_get_step(i);
2893 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2898 mutex_unlock(&cpuhp_state_mutex);
2901 static DEVICE_ATTR_RO(states);
2903 static struct attribute *cpuhp_cpu_root_attrs[] = {
2904 &dev_attr_states.attr,
2908 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2909 .attrs = cpuhp_cpu_root_attrs,
2914 #ifdef CONFIG_HOTPLUG_SMT
2916 static bool cpu_smt_num_threads_valid(unsigned int threads)
2918 if (IS_ENABLED(CONFIG_SMT_NUM_THREADS_DYNAMIC))
2919 return threads >= 1 && threads <= cpu_smt_max_threads;
2920 return threads == 1 || threads == cpu_smt_max_threads;
2924 __store_smt_control(struct device *dev, struct device_attribute *attr,
2925 const char *buf, size_t count)
2927 int ctrlval, ret, num_threads, orig_threads;
2930 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2933 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2936 if (sysfs_streq(buf, "on")) {
2937 ctrlval = CPU_SMT_ENABLED;
2938 num_threads = cpu_smt_max_threads;
2939 } else if (sysfs_streq(buf, "off")) {
2940 ctrlval = CPU_SMT_DISABLED;
2942 } else if (sysfs_streq(buf, "forceoff")) {
2943 ctrlval = CPU_SMT_FORCE_DISABLED;
2945 } else if (kstrtoint(buf, 10, &num_threads) == 0) {
2946 if (num_threads == 1)
2947 ctrlval = CPU_SMT_DISABLED;
2948 else if (cpu_smt_num_threads_valid(num_threads))
2949 ctrlval = CPU_SMT_ENABLED;
2956 ret = lock_device_hotplug_sysfs();
2960 orig_threads = cpu_smt_num_threads;
2961 cpu_smt_num_threads = num_threads;
2963 force_off = ctrlval != cpu_smt_control && ctrlval == CPU_SMT_FORCE_DISABLED;
2965 if (num_threads > orig_threads)
2966 ret = cpuhp_smt_enable();
2967 else if (num_threads < orig_threads || force_off)
2968 ret = cpuhp_smt_disable(ctrlval);
2970 unlock_device_hotplug();
2971 return ret ? ret : count;
2974 #else /* !CONFIG_HOTPLUG_SMT */
2976 __store_smt_control(struct device *dev, struct device_attribute *attr,
2977 const char *buf, size_t count)
2981 #endif /* CONFIG_HOTPLUG_SMT */
2983 static const char *smt_states[] = {
2984 [CPU_SMT_ENABLED] = "on",
2985 [CPU_SMT_DISABLED] = "off",
2986 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2987 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2988 [CPU_SMT_NOT_IMPLEMENTED] = "notimplemented",
2991 static ssize_t control_show(struct device *dev,
2992 struct device_attribute *attr, char *buf)
2994 const char *state = smt_states[cpu_smt_control];
2996 #ifdef CONFIG_HOTPLUG_SMT
2998 * If SMT is enabled but not all threads are enabled then show the
2999 * number of threads. If all threads are enabled show "on". Otherwise
3000 * show the state name.
3002 if (cpu_smt_control == CPU_SMT_ENABLED &&
3003 cpu_smt_num_threads != cpu_smt_max_threads)
3004 return sysfs_emit(buf, "%d\n", cpu_smt_num_threads);
3007 return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
3010 static ssize_t control_store(struct device *dev, struct device_attribute *attr,
3011 const char *buf, size_t count)
3013 return __store_smt_control(dev, attr, buf, count);
3015 static DEVICE_ATTR_RW(control);
3017 static ssize_t active_show(struct device *dev,
3018 struct device_attribute *attr, char *buf)
3020 return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
3022 static DEVICE_ATTR_RO(active);
3024 static struct attribute *cpuhp_smt_attrs[] = {
3025 &dev_attr_control.attr,
3026 &dev_attr_active.attr,
3030 static const struct attribute_group cpuhp_smt_attr_group = {
3031 .attrs = cpuhp_smt_attrs,
3036 static int __init cpu_smt_sysfs_init(void)
3038 struct device *dev_root;
3041 dev_root = bus_get_dev_root(&cpu_subsys);
3043 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_smt_attr_group);
3044 put_device(dev_root);
3049 static int __init cpuhp_sysfs_init(void)
3051 struct device *dev_root;
3054 ret = cpu_smt_sysfs_init();
3058 dev_root = bus_get_dev_root(&cpu_subsys);
3060 ret = sysfs_create_group(&dev_root->kobj, &cpuhp_cpu_root_attr_group);
3061 put_device(dev_root);
3066 for_each_possible_cpu(cpu) {
3067 struct device *dev = get_cpu_device(cpu);
3071 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
3077 device_initcall(cpuhp_sysfs_init);
3078 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
3081 * cpu_bit_bitmap[] is a special, "compressed" data structure that
3082 * represents all NR_CPUS bits binary values of 1<<nr.
3084 * It is used by cpumask_of() to get a constant address to a CPU
3085 * mask value that has a single bit set only.
3088 /* cpu_bit_bitmap[0] is empty - so we can back into it */
3089 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
3090 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
3091 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
3092 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
3094 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
3096 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
3097 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
3098 #if BITS_PER_LONG > 32
3099 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
3100 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
3103 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
3105 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
3106 EXPORT_SYMBOL(cpu_all_bits);
3108 #ifdef CONFIG_INIT_ALL_POSSIBLE
3109 struct cpumask __cpu_possible_mask __read_mostly
3112 struct cpumask __cpu_possible_mask __read_mostly;
3114 EXPORT_SYMBOL(__cpu_possible_mask);
3116 struct cpumask __cpu_online_mask __read_mostly;
3117 EXPORT_SYMBOL(__cpu_online_mask);
3119 struct cpumask __cpu_present_mask __read_mostly;
3120 EXPORT_SYMBOL(__cpu_present_mask);
3122 struct cpumask __cpu_active_mask __read_mostly;
3123 EXPORT_SYMBOL(__cpu_active_mask);
3125 struct cpumask __cpu_dying_mask __read_mostly;
3126 EXPORT_SYMBOL(__cpu_dying_mask);
3128 atomic_t __num_online_cpus __read_mostly;
3129 EXPORT_SYMBOL(__num_online_cpus);
3131 void init_cpu_present(const struct cpumask *src)
3133 cpumask_copy(&__cpu_present_mask, src);
3136 void init_cpu_possible(const struct cpumask *src)
3138 cpumask_copy(&__cpu_possible_mask, src);
3141 void init_cpu_online(const struct cpumask *src)
3143 cpumask_copy(&__cpu_online_mask, src);
3146 void set_cpu_online(unsigned int cpu, bool online)
3149 * atomic_inc/dec() is required to handle the horrid abuse of this
3150 * function by the reboot and kexec code which invoke it from
3151 * IPI/NMI broadcasts when shutting down CPUs. Invocation from
3152 * regular CPU hotplug is properly serialized.
3154 * Note, that the fact that __num_online_cpus is of type atomic_t
3155 * does not protect readers which are not serialized against
3156 * concurrent hotplug operations.
3159 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
3160 atomic_inc(&__num_online_cpus);
3162 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
3163 atomic_dec(&__num_online_cpus);
3168 * Activate the first processor.
3170 void __init boot_cpu_init(void)
3172 int cpu = smp_processor_id();
3174 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
3175 set_cpu_online(cpu, true);
3176 set_cpu_active(cpu, true);
3177 set_cpu_present(cpu, true);
3178 set_cpu_possible(cpu, true);
3181 __boot_cpu_id = cpu;
3186 * Must be called _AFTER_ setting up the per_cpu areas
3188 void __init boot_cpu_hotplug_init(void)
3191 cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
3192 atomic_set(this_cpu_ptr(&cpuhp_state.ap_sync_state), SYNC_STATE_ONLINE);
3194 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
3195 this_cpu_write(cpuhp_state.target, CPUHP_ONLINE);
3199 * These are used for a global "mitigations=" cmdline option for toggling
3200 * optional CPU mitigations.
3202 enum cpu_mitigations {
3203 CPU_MITIGATIONS_OFF,
3204 CPU_MITIGATIONS_AUTO,
3205 CPU_MITIGATIONS_AUTO_NOSMT,
3208 static enum cpu_mitigations cpu_mitigations __ro_after_init =
3209 CPU_MITIGATIONS_AUTO;
3211 static int __init mitigations_parse_cmdline(char *arg)
3213 if (!strcmp(arg, "off"))
3214 cpu_mitigations = CPU_MITIGATIONS_OFF;
3215 else if (!strcmp(arg, "auto"))
3216 cpu_mitigations = CPU_MITIGATIONS_AUTO;
3217 else if (!strcmp(arg, "auto,nosmt"))
3218 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
3220 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
3225 early_param("mitigations", mitigations_parse_cmdline);
3227 /* mitigations=off */
3228 bool cpu_mitigations_off(void)
3230 return cpu_mitigations == CPU_MITIGATIONS_OFF;
3232 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
3234 /* mitigations=auto,nosmt */
3235 bool cpu_mitigations_auto_nosmt(void)
3237 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
3239 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);