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/export.h>
21 #include <linux/bug.h>
22 #include <linux/kthread.h>
23 #include <linux/stop_machine.h>
24 #include <linux/mutex.h>
25 #include <linux/gfp.h>
26 #include <linux/suspend.h>
27 #include <linux/lockdep.h>
28 #include <linux/tick.h>
29 #include <linux/irq.h>
30 #include <linux/nmi.h>
31 #include <linux/smpboot.h>
32 #include <linux/relay.h>
33 #include <linux/slab.h>
34 #include <linux/percpu-rwsem.h>
36 #include <trace/events/power.h>
37 #define CREATE_TRACE_POINTS
38 #include <trace/events/cpuhp.h>
43 * cpuhp_cpu_state - Per cpu hotplug state storage
44 * @state: The current cpu state
45 * @target: The target state
46 * @thread: Pointer to the hotplug thread
47 * @should_run: Thread should execute
48 * @rollback: Perform a rollback
49 * @single: Single callback invocation
50 * @bringup: Single callback bringup or teardown selector
51 * @cb_state: The state for a single callback (install/uninstall)
52 * @result: Result of the operation
53 * @done_up: Signal completion to the issuer of the task for cpu-up
54 * @done_down: Signal completion to the issuer of the task for cpu-down
56 struct cpuhp_cpu_state {
57 enum cpuhp_state state;
58 enum cpuhp_state target;
59 enum cpuhp_state fail;
61 struct task_struct *thread;
66 struct hlist_node *node;
67 struct hlist_node *last;
68 enum cpuhp_state cb_state;
70 struct completion done_up;
71 struct completion done_down;
75 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
76 .fail = CPUHP_INVALID,
80 cpumask_t cpus_booted_once_mask;
83 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
84 static struct lockdep_map cpuhp_state_up_map =
85 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
86 static struct lockdep_map cpuhp_state_down_map =
87 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
90 static inline void cpuhp_lock_acquire(bool bringup)
92 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
95 static inline void cpuhp_lock_release(bool bringup)
97 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
101 static inline void cpuhp_lock_acquire(bool bringup) { }
102 static inline void cpuhp_lock_release(bool bringup) { }
107 * cpuhp_step - Hotplug state machine step
108 * @name: Name of the step
109 * @startup: Startup function of the step
110 * @teardown: Teardown function of the step
111 * @cant_stop: Bringup/teardown can't be stopped at this step
116 int (*single)(unsigned int cpu);
117 int (*multi)(unsigned int cpu,
118 struct hlist_node *node);
121 int (*single)(unsigned int cpu);
122 int (*multi)(unsigned int cpu,
123 struct hlist_node *node);
125 struct hlist_head list;
130 static DEFINE_MUTEX(cpuhp_state_mutex);
131 static struct cpuhp_step cpuhp_hp_states[];
133 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
135 return cpuhp_hp_states + state;
139 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
140 * @cpu: The cpu for which the callback should be invoked
141 * @state: The state to do callbacks for
142 * @bringup: True if the bringup callback should be invoked
143 * @node: For multi-instance, do a single entry callback for install/remove
144 * @lastp: For multi-instance rollback, remember how far we got
146 * Called from cpu hotplug and from the state register machinery.
148 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
149 bool bringup, struct hlist_node *node,
150 struct hlist_node **lastp)
152 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
153 struct cpuhp_step *step = cpuhp_get_step(state);
154 int (*cbm)(unsigned int cpu, struct hlist_node *node);
155 int (*cb)(unsigned int cpu);
158 if (st->fail == state) {
159 st->fail = CPUHP_INVALID;
161 if (!(bringup ? step->startup.single : step->teardown.single))
167 if (!step->multi_instance) {
168 WARN_ON_ONCE(lastp && *lastp);
169 cb = bringup ? step->startup.single : step->teardown.single;
172 trace_cpuhp_enter(cpu, st->target, state, cb);
174 trace_cpuhp_exit(cpu, st->state, state, ret);
177 cbm = bringup ? step->startup.multi : step->teardown.multi;
181 /* Single invocation for instance add/remove */
183 WARN_ON_ONCE(lastp && *lastp);
184 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
185 ret = cbm(cpu, node);
186 trace_cpuhp_exit(cpu, st->state, state, ret);
190 /* State transition. Invoke on all instances */
192 hlist_for_each(node, &step->list) {
193 if (lastp && node == *lastp)
196 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
197 ret = cbm(cpu, node);
198 trace_cpuhp_exit(cpu, st->state, state, ret);
212 /* Rollback the instances if one failed */
213 cbm = !bringup ? step->startup.multi : step->teardown.multi;
217 hlist_for_each(node, &step->list) {
221 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
222 ret = cbm(cpu, node);
223 trace_cpuhp_exit(cpu, st->state, state, ret);
225 * Rollback must not fail,
233 static bool cpuhp_is_ap_state(enum cpuhp_state state)
236 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
237 * purposes as that state is handled explicitly in cpu_down.
239 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
242 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
244 struct completion *done = bringup ? &st->done_up : &st->done_down;
245 wait_for_completion(done);
248 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
250 struct completion *done = bringup ? &st->done_up : &st->done_down;
255 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
257 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
259 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
262 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
263 static DEFINE_MUTEX(cpu_add_remove_lock);
264 bool cpuhp_tasks_frozen;
265 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
268 * The following two APIs (cpu_maps_update_begin/done) must be used when
269 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
271 void cpu_maps_update_begin(void)
273 mutex_lock(&cpu_add_remove_lock);
276 void cpu_maps_update_done(void)
278 mutex_unlock(&cpu_add_remove_lock);
282 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
283 * Should always be manipulated under cpu_add_remove_lock
285 static int cpu_hotplug_disabled;
287 #ifdef CONFIG_HOTPLUG_CPU
289 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
291 void cpus_read_lock(void)
293 percpu_down_read(&cpu_hotplug_lock);
295 EXPORT_SYMBOL_GPL(cpus_read_lock);
297 int cpus_read_trylock(void)
299 return percpu_down_read_trylock(&cpu_hotplug_lock);
301 EXPORT_SYMBOL_GPL(cpus_read_trylock);
303 void cpus_read_unlock(void)
305 percpu_up_read(&cpu_hotplug_lock);
307 EXPORT_SYMBOL_GPL(cpus_read_unlock);
309 void cpus_write_lock(void)
311 percpu_down_write(&cpu_hotplug_lock);
314 void cpus_write_unlock(void)
316 percpu_up_write(&cpu_hotplug_lock);
319 void lockdep_assert_cpus_held(void)
322 * We can't have hotplug operations before userspace starts running,
323 * and some init codepaths will knowingly not take the hotplug lock.
324 * This is all valid, so mute lockdep until it makes sense to report
327 if (system_state < SYSTEM_RUNNING)
330 percpu_rwsem_assert_held(&cpu_hotplug_lock);
333 static void lockdep_acquire_cpus_lock(void)
335 rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
338 static void lockdep_release_cpus_lock(void)
340 rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
344 * Wait for currently running CPU hotplug operations to complete (if any) and
345 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
346 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
347 * hotplug path before performing hotplug operations. So acquiring that lock
348 * guarantees mutual exclusion from any currently running hotplug operations.
350 void cpu_hotplug_disable(void)
352 cpu_maps_update_begin();
353 cpu_hotplug_disabled++;
354 cpu_maps_update_done();
356 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
358 static void __cpu_hotplug_enable(void)
360 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
362 cpu_hotplug_disabled--;
365 void cpu_hotplug_enable(void)
367 cpu_maps_update_begin();
368 __cpu_hotplug_enable();
369 cpu_maps_update_done();
371 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
375 static void lockdep_acquire_cpus_lock(void)
379 static void lockdep_release_cpus_lock(void)
383 #endif /* CONFIG_HOTPLUG_CPU */
386 * Architectures that need SMT-specific errata handling during SMT hotplug
387 * should override this.
389 void __weak arch_smt_update(void) { }
391 #ifdef CONFIG_HOTPLUG_SMT
392 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
394 void __init cpu_smt_disable(bool force)
396 if (!cpu_smt_possible())
400 pr_info("SMT: Force disabled\n");
401 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
403 pr_info("SMT: disabled\n");
404 cpu_smt_control = CPU_SMT_DISABLED;
409 * The decision whether SMT is supported can only be done after the full
410 * CPU identification. Called from architecture code.
412 void __init cpu_smt_check_topology(void)
414 if (!topology_smt_supported())
415 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
418 static int __init smt_cmdline_disable(char *str)
420 cpu_smt_disable(str && !strcmp(str, "force"));
423 early_param("nosmt", smt_cmdline_disable);
425 static inline bool cpu_smt_allowed(unsigned int cpu)
427 if (cpu_smt_control == CPU_SMT_ENABLED)
430 if (topology_is_primary_thread(cpu))
434 * On x86 it's required to boot all logical CPUs at least once so
435 * that the init code can get a chance to set CR4.MCE on each
436 * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
437 * core will shutdown the machine.
439 return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
442 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
443 bool cpu_smt_possible(void)
445 return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
446 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
448 EXPORT_SYMBOL_GPL(cpu_smt_possible);
450 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
453 static inline enum cpuhp_state
454 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
456 enum cpuhp_state prev_state = st->state;
458 st->rollback = false;
463 st->bringup = st->state < target;
469 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
474 * If we have st->last we need to undo partial multi_instance of this
475 * state first. Otherwise start undo at the previous state.
484 st->target = prev_state;
485 st->bringup = !st->bringup;
488 /* Regular hotplug invocation of the AP hotplug thread */
489 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
491 if (!st->single && st->state == st->target)
496 * Make sure the above stores are visible before should_run becomes
497 * true. Paired with the mb() above in cpuhp_thread_fun()
500 st->should_run = true;
501 wake_up_process(st->thread);
502 wait_for_ap_thread(st, st->bringup);
505 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
507 enum cpuhp_state prev_state;
510 prev_state = cpuhp_set_state(st, target);
512 if ((ret = st->result)) {
513 cpuhp_reset_state(st, prev_state);
520 static int bringup_wait_for_ap(unsigned int cpu)
522 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
524 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
525 wait_for_ap_thread(st, true);
526 if (WARN_ON_ONCE((!cpu_online(cpu))))
529 /* Unpark the hotplug thread of the target cpu */
530 kthread_unpark(st->thread);
533 * SMT soft disabling on X86 requires to bring the CPU out of the
534 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
535 * CPU marked itself as booted_once in notify_cpu_starting() so the
536 * cpu_smt_allowed() check will now return false if this is not the
539 if (!cpu_smt_allowed(cpu))
542 if (st->target <= CPUHP_AP_ONLINE_IDLE)
545 return cpuhp_kick_ap(st, st->target);
548 static int bringup_cpu(unsigned int cpu)
550 struct task_struct *idle = idle_thread_get(cpu);
554 * Some architectures have to walk the irq descriptors to
555 * setup the vector space for the cpu which comes online.
556 * Prevent irq alloc/free across the bringup.
560 /* Arch-specific enabling code. */
561 ret = __cpu_up(cpu, idle);
565 return bringup_wait_for_ap(cpu);
568 static int finish_cpu(unsigned int cpu)
570 struct task_struct *idle = idle_thread_get(cpu);
571 struct mm_struct *mm = idle->active_mm;
574 * idle_task_exit() will have switched to &init_mm, now
575 * clean up any remaining active_mm state.
578 idle->active_mm = &init_mm;
584 * Hotplug state machine related functions
587 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
589 for (st->state--; st->state > st->target; st->state--)
590 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
593 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
595 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
598 * When CPU hotplug is disabled, then taking the CPU down is not
599 * possible because takedown_cpu() and the architecture and
600 * subsystem specific mechanisms are not available. So the CPU
601 * which would be completely unplugged again needs to stay around
602 * in the current state.
604 return st->state <= CPUHP_BRINGUP_CPU;
607 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
608 enum cpuhp_state target)
610 enum cpuhp_state prev_state = st->state;
613 while (st->state < target) {
615 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
617 if (can_rollback_cpu(st)) {
618 st->target = prev_state;
619 undo_cpu_up(cpu, st);
628 * The cpu hotplug threads manage the bringup and teardown of the cpus
630 static void cpuhp_create(unsigned int cpu)
632 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
634 init_completion(&st->done_up);
635 init_completion(&st->done_down);
638 static int cpuhp_should_run(unsigned int cpu)
640 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
642 return st->should_run;
646 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
647 * callbacks when a state gets [un]installed at runtime.
649 * Each invocation of this function by the smpboot thread does a single AP
652 * It has 3 modes of operation:
653 * - single: runs st->cb_state
654 * - up: runs ++st->state, while st->state < st->target
655 * - down: runs st->state--, while st->state > st->target
657 * When complete or on error, should_run is cleared and the completion is fired.
659 static void cpuhp_thread_fun(unsigned int cpu)
661 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
662 bool bringup = st->bringup;
663 enum cpuhp_state state;
665 if (WARN_ON_ONCE(!st->should_run))
669 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
670 * that if we see ->should_run we also see the rest of the state.
675 * The BP holds the hotplug lock, but we're now running on the AP,
676 * ensure that anybody asserting the lock is held, will actually find
679 lockdep_acquire_cpus_lock();
680 cpuhp_lock_acquire(bringup);
683 state = st->cb_state;
684 st->should_run = false;
689 st->should_run = (st->state < st->target);
690 WARN_ON_ONCE(st->state > st->target);
694 st->should_run = (st->state > st->target);
695 WARN_ON_ONCE(st->state < st->target);
699 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
701 if (cpuhp_is_atomic_state(state)) {
703 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
707 * STARTING/DYING must not fail!
709 WARN_ON_ONCE(st->result);
711 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
716 * If we fail on a rollback, we're up a creek without no
717 * paddle, no way forward, no way back. We loose, thanks for
720 WARN_ON_ONCE(st->rollback);
721 st->should_run = false;
724 cpuhp_lock_release(bringup);
725 lockdep_release_cpus_lock();
728 complete_ap_thread(st, bringup);
731 /* Invoke a single callback on a remote cpu */
733 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
734 struct hlist_node *node)
736 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
739 if (!cpu_online(cpu))
742 cpuhp_lock_acquire(false);
743 cpuhp_lock_release(false);
745 cpuhp_lock_acquire(true);
746 cpuhp_lock_release(true);
749 * If we are up and running, use the hotplug thread. For early calls
750 * we invoke the thread function directly.
753 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
755 st->rollback = false;
759 st->bringup = bringup;
760 st->cb_state = state;
766 * If we failed and did a partial, do a rollback.
768 if ((ret = st->result) && st->last) {
770 st->bringup = !bringup;
776 * Clean up the leftovers so the next hotplug operation wont use stale
779 st->node = st->last = NULL;
783 static int cpuhp_kick_ap_work(unsigned int cpu)
785 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
786 enum cpuhp_state prev_state = st->state;
789 cpuhp_lock_acquire(false);
790 cpuhp_lock_release(false);
792 cpuhp_lock_acquire(true);
793 cpuhp_lock_release(true);
795 trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
796 ret = cpuhp_kick_ap(st, st->target);
797 trace_cpuhp_exit(cpu, st->state, prev_state, ret);
802 static struct smp_hotplug_thread cpuhp_threads = {
803 .store = &cpuhp_state.thread,
804 .create = &cpuhp_create,
805 .thread_should_run = cpuhp_should_run,
806 .thread_fn = cpuhp_thread_fun,
807 .thread_comm = "cpuhp/%u",
811 void __init cpuhp_threads_init(void)
813 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
814 kthread_unpark(this_cpu_read(cpuhp_state.thread));
817 #ifdef CONFIG_HOTPLUG_CPU
818 #ifndef arch_clear_mm_cpumask_cpu
819 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
823 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
826 * This function walks all processes, finds a valid mm struct for each one and
827 * then clears a corresponding bit in mm's cpumask. While this all sounds
828 * trivial, there are various non-obvious corner cases, which this function
829 * tries to solve in a safe manner.
831 * Also note that the function uses a somewhat relaxed locking scheme, so it may
832 * be called only for an already offlined CPU.
834 void clear_tasks_mm_cpumask(int cpu)
836 struct task_struct *p;
839 * This function is called after the cpu is taken down and marked
840 * offline, so its not like new tasks will ever get this cpu set in
841 * their mm mask. -- Peter Zijlstra
842 * Thus, we may use rcu_read_lock() here, instead of grabbing
843 * full-fledged tasklist_lock.
845 WARN_ON(cpu_online(cpu));
847 for_each_process(p) {
848 struct task_struct *t;
851 * Main thread might exit, but other threads may still have
852 * a valid mm. Find one.
854 t = find_lock_task_mm(p);
857 arch_clear_mm_cpumask_cpu(cpu, t->mm);
863 /* Take this CPU down. */
864 static int take_cpu_down(void *_param)
866 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
867 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
868 int err, cpu = smp_processor_id();
871 /* Ensure this CPU doesn't handle any more interrupts. */
872 err = __cpu_disable();
877 * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
878 * do this step again.
880 WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
882 /* Invoke the former CPU_DYING callbacks */
883 for (; st->state > target; st->state--) {
884 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
886 * DYING must not fail!
891 /* Give up timekeeping duties */
892 tick_handover_do_timer();
893 /* Remove CPU from timer broadcasting */
894 tick_offline_cpu(cpu);
895 /* Park the stopper thread */
896 stop_machine_park(cpu);
900 static int takedown_cpu(unsigned int cpu)
902 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
905 /* Park the smpboot threads */
906 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
909 * Prevent irq alloc/free while the dying cpu reorganizes the
910 * interrupt affinities.
915 * So now all preempt/rcu users must observe !cpu_active().
917 err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
919 /* CPU refused to die */
921 /* Unpark the hotplug thread so we can rollback there */
922 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
925 BUG_ON(cpu_online(cpu));
928 * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
929 * all runnable tasks from the CPU, there's only the idle task left now
930 * that the migration thread is done doing the stop_machine thing.
932 * Wait for the stop thread to go away.
934 wait_for_ap_thread(st, false);
935 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
937 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
940 hotplug_cpu__broadcast_tick_pull(cpu);
941 /* This actually kills the CPU. */
944 tick_cleanup_dead_cpu(cpu);
945 rcutree_migrate_callbacks(cpu);
949 static void cpuhp_complete_idle_dead(void *arg)
951 struct cpuhp_cpu_state *st = arg;
953 complete_ap_thread(st, false);
956 void cpuhp_report_idle_dead(void)
958 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
960 BUG_ON(st->state != CPUHP_AP_OFFLINE);
961 rcu_report_dead(smp_processor_id());
962 st->state = CPUHP_AP_IDLE_DEAD;
964 * We cannot call complete after rcu_report_dead() so we delegate it
967 smp_call_function_single(cpumask_first(cpu_online_mask),
968 cpuhp_complete_idle_dead, st, 0);
971 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
973 for (st->state++; st->state < st->target; st->state++)
974 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
977 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
978 enum cpuhp_state target)
980 enum cpuhp_state prev_state = st->state;
983 for (; st->state > target; st->state--) {
984 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
986 st->target = prev_state;
987 if (st->state < prev_state)
988 undo_cpu_down(cpu, st);
995 /* Requires cpu_add_remove_lock to be held */
996 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
997 enum cpuhp_state target)
999 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1000 int prev_state, ret = 0;
1002 if (num_online_cpus() == 1)
1005 if (!cpu_present(cpu))
1010 cpuhp_tasks_frozen = tasks_frozen;
1012 prev_state = cpuhp_set_state(st, target);
1014 * If the current CPU state is in the range of the AP hotplug thread,
1015 * then we need to kick the thread.
1017 if (st->state > CPUHP_TEARDOWN_CPU) {
1018 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1019 ret = cpuhp_kick_ap_work(cpu);
1021 * The AP side has done the error rollback already. Just
1022 * return the error code..
1028 * We might have stopped still in the range of the AP hotplug
1029 * thread. Nothing to do anymore.
1031 if (st->state > CPUHP_TEARDOWN_CPU)
1034 st->target = target;
1037 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1038 * to do the further cleanups.
1040 ret = cpuhp_down_callbacks(cpu, st, target);
1041 if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1042 cpuhp_reset_state(st, prev_state);
1043 __cpuhp_kick_ap(st);
1047 cpus_write_unlock();
1049 * Do post unplug cleanup. This is still protected against
1050 * concurrent CPU hotplug via cpu_add_remove_lock.
1052 lockup_detector_cleanup();
1057 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1059 if (cpu_hotplug_disabled)
1061 return _cpu_down(cpu, 0, target);
1064 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1068 cpu_maps_update_begin();
1069 err = cpu_down_maps_locked(cpu, target);
1070 cpu_maps_update_done();
1075 * cpu_device_down - Bring down a cpu device
1076 * @dev: Pointer to the cpu device to offline
1078 * This function is meant to be used by device core cpu subsystem only.
1080 * Other subsystems should use remove_cpu() instead.
1082 int cpu_device_down(struct device *dev)
1084 return cpu_down(dev->id, CPUHP_OFFLINE);
1087 int remove_cpu(unsigned int cpu)
1091 lock_device_hotplug();
1092 ret = device_offline(get_cpu_device(cpu));
1093 unlock_device_hotplug();
1097 EXPORT_SYMBOL_GPL(remove_cpu);
1099 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1104 cpu_maps_update_begin();
1107 * Make certain the cpu I'm about to reboot on is online.
1109 * This is inline to what migrate_to_reboot_cpu() already do.
1111 if (!cpu_online(primary_cpu))
1112 primary_cpu = cpumask_first(cpu_online_mask);
1114 for_each_online_cpu(cpu) {
1115 if (cpu == primary_cpu)
1118 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1120 pr_err("Failed to offline CPU%d - error=%d",
1127 * Ensure all but the reboot CPU are offline.
1129 BUG_ON(num_online_cpus() > 1);
1132 * Make sure the CPUs won't be enabled by someone else after this
1133 * point. Kexec will reboot to a new kernel shortly resetting
1134 * everything along the way.
1136 cpu_hotplug_disabled++;
1138 cpu_maps_update_done();
1142 #define takedown_cpu NULL
1143 #endif /*CONFIG_HOTPLUG_CPU*/
1146 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1147 * @cpu: cpu that just started
1149 * It must be called by the arch code on the new cpu, before the new cpu
1150 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1152 void notify_cpu_starting(unsigned int cpu)
1154 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1155 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1158 rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1159 cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1160 while (st->state < target) {
1162 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1164 * STARTING must not fail!
1171 * Called from the idle task. Wake up the controlling task which brings the
1172 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1173 * online bringup to the hotplug thread.
1175 void cpuhp_online_idle(enum cpuhp_state state)
1177 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1179 /* Happens for the boot cpu */
1180 if (state != CPUHP_AP_ONLINE_IDLE)
1184 * Unpart the stopper thread before we start the idle loop (and start
1185 * scheduling); this ensures the stopper task is always available.
1187 stop_machine_unpark(smp_processor_id());
1189 st->state = CPUHP_AP_ONLINE_IDLE;
1190 complete_ap_thread(st, true);
1193 /* Requires cpu_add_remove_lock to be held */
1194 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1196 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1197 struct task_struct *idle;
1202 if (!cpu_present(cpu)) {
1208 * The caller of cpu_up() might have raced with another
1209 * caller. Nothing to do.
1211 if (st->state >= target)
1214 if (st->state == CPUHP_OFFLINE) {
1215 /* Let it fail before we try to bring the cpu up */
1216 idle = idle_thread_get(cpu);
1218 ret = PTR_ERR(idle);
1223 cpuhp_tasks_frozen = tasks_frozen;
1225 cpuhp_set_state(st, target);
1227 * If the current CPU state is in the range of the AP hotplug thread,
1228 * then we need to kick the thread once more.
1230 if (st->state > CPUHP_BRINGUP_CPU) {
1231 ret = cpuhp_kick_ap_work(cpu);
1233 * The AP side has done the error rollback already. Just
1234 * return the error code..
1241 * Try to reach the target state. We max out on the BP at
1242 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1243 * responsible for bringing it up to the target state.
1245 target = min((int)target, CPUHP_BRINGUP_CPU);
1246 ret = cpuhp_up_callbacks(cpu, st, target);
1248 cpus_write_unlock();
1253 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1257 if (!cpu_possible(cpu)) {
1258 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1260 #if defined(CONFIG_IA64)
1261 pr_err("please check additional_cpus= boot parameter\n");
1266 err = try_online_node(cpu_to_node(cpu));
1270 cpu_maps_update_begin();
1272 if (cpu_hotplug_disabled) {
1276 if (!cpu_smt_allowed(cpu)) {
1281 err = _cpu_up(cpu, 0, target);
1283 cpu_maps_update_done();
1288 * cpu_device_up - Bring up a cpu device
1289 * @dev: Pointer to the cpu device to online
1291 * This function is meant to be used by device core cpu subsystem only.
1293 * Other subsystems should use add_cpu() instead.
1295 int cpu_device_up(struct device *dev)
1297 return cpu_up(dev->id, CPUHP_ONLINE);
1300 int add_cpu(unsigned int cpu)
1304 lock_device_hotplug();
1305 ret = device_online(get_cpu_device(cpu));
1306 unlock_device_hotplug();
1310 EXPORT_SYMBOL_GPL(add_cpu);
1313 * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1314 * @sleep_cpu: The cpu we hibernated on and should be brought up.
1316 * On some architectures like arm64, we can hibernate on any CPU, but on
1317 * wake up the CPU we hibernated on might be offline as a side effect of
1318 * using maxcpus= for example.
1320 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1324 if (!cpu_online(sleep_cpu)) {
1325 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1326 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1328 pr_err("Failed to bring hibernate-CPU up!\n");
1335 void bringup_nonboot_cpus(unsigned int setup_max_cpus)
1339 for_each_present_cpu(cpu) {
1340 if (num_online_cpus() >= setup_max_cpus)
1342 if (!cpu_online(cpu))
1343 cpu_up(cpu, CPUHP_ONLINE);
1347 #ifdef CONFIG_PM_SLEEP_SMP
1348 static cpumask_var_t frozen_cpus;
1350 int freeze_secondary_cpus(int primary)
1354 cpu_maps_update_begin();
1355 if (primary == -1) {
1356 primary = cpumask_first(cpu_online_mask);
1357 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1358 primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1360 if (!cpu_online(primary))
1361 primary = cpumask_first(cpu_online_mask);
1365 * We take down all of the non-boot CPUs in one shot to avoid races
1366 * with the userspace trying to use the CPU hotplug at the same time
1368 cpumask_clear(frozen_cpus);
1370 pr_info("Disabling non-boot CPUs ...\n");
1371 for_each_online_cpu(cpu) {
1375 if (pm_wakeup_pending()) {
1376 pr_info("Wakeup pending. Abort CPU freeze\n");
1381 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1382 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1383 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1385 cpumask_set_cpu(cpu, frozen_cpus);
1387 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1393 BUG_ON(num_online_cpus() > 1);
1395 pr_err("Non-boot CPUs are not disabled\n");
1398 * Make sure the CPUs won't be enabled by someone else. We need to do
1399 * this even in case of failure as all freeze_secondary_cpus() users are
1400 * supposed to do thaw_secondary_cpus() on the failure path.
1402 cpu_hotplug_disabled++;
1404 cpu_maps_update_done();
1408 void __weak arch_thaw_secondary_cpus_begin(void)
1412 void __weak arch_thaw_secondary_cpus_end(void)
1416 void thaw_secondary_cpus(void)
1420 /* Allow everyone to use the CPU hotplug again */
1421 cpu_maps_update_begin();
1422 __cpu_hotplug_enable();
1423 if (cpumask_empty(frozen_cpus))
1426 pr_info("Enabling non-boot CPUs ...\n");
1428 arch_thaw_secondary_cpus_begin();
1430 for_each_cpu(cpu, frozen_cpus) {
1431 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1432 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1433 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1435 pr_info("CPU%d is up\n", cpu);
1438 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1441 arch_thaw_secondary_cpus_end();
1443 cpumask_clear(frozen_cpus);
1445 cpu_maps_update_done();
1448 static int __init alloc_frozen_cpus(void)
1450 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1454 core_initcall(alloc_frozen_cpus);
1457 * When callbacks for CPU hotplug notifications are being executed, we must
1458 * ensure that the state of the system with respect to the tasks being frozen
1459 * or not, as reported by the notification, remains unchanged *throughout the
1460 * duration* of the execution of the callbacks.
1461 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1463 * This synchronization is implemented by mutually excluding regular CPU
1464 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1465 * Hibernate notifications.
1468 cpu_hotplug_pm_callback(struct notifier_block *nb,
1469 unsigned long action, void *ptr)
1473 case PM_SUSPEND_PREPARE:
1474 case PM_HIBERNATION_PREPARE:
1475 cpu_hotplug_disable();
1478 case PM_POST_SUSPEND:
1479 case PM_POST_HIBERNATION:
1480 cpu_hotplug_enable();
1491 static int __init cpu_hotplug_pm_sync_init(void)
1494 * cpu_hotplug_pm_callback has higher priority than x86
1495 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1496 * to disable cpu hotplug to avoid cpu hotplug race.
1498 pm_notifier(cpu_hotplug_pm_callback, 0);
1501 core_initcall(cpu_hotplug_pm_sync_init);
1503 #endif /* CONFIG_PM_SLEEP_SMP */
1507 #endif /* CONFIG_SMP */
1509 /* Boot processor state steps */
1510 static struct cpuhp_step cpuhp_hp_states[] = {
1513 .startup.single = NULL,
1514 .teardown.single = NULL,
1517 [CPUHP_CREATE_THREADS]= {
1518 .name = "threads:prepare",
1519 .startup.single = smpboot_create_threads,
1520 .teardown.single = NULL,
1523 [CPUHP_PERF_PREPARE] = {
1524 .name = "perf:prepare",
1525 .startup.single = perf_event_init_cpu,
1526 .teardown.single = perf_event_exit_cpu,
1528 [CPUHP_WORKQUEUE_PREP] = {
1529 .name = "workqueue:prepare",
1530 .startup.single = workqueue_prepare_cpu,
1531 .teardown.single = NULL,
1533 [CPUHP_HRTIMERS_PREPARE] = {
1534 .name = "hrtimers:prepare",
1535 .startup.single = hrtimers_prepare_cpu,
1536 .teardown.single = hrtimers_dead_cpu,
1538 [CPUHP_SMPCFD_PREPARE] = {
1539 .name = "smpcfd:prepare",
1540 .startup.single = smpcfd_prepare_cpu,
1541 .teardown.single = smpcfd_dead_cpu,
1543 [CPUHP_RELAY_PREPARE] = {
1544 .name = "relay:prepare",
1545 .startup.single = relay_prepare_cpu,
1546 .teardown.single = NULL,
1548 [CPUHP_SLAB_PREPARE] = {
1549 .name = "slab:prepare",
1550 .startup.single = slab_prepare_cpu,
1551 .teardown.single = slab_dead_cpu,
1553 [CPUHP_RCUTREE_PREP] = {
1554 .name = "RCU/tree:prepare",
1555 .startup.single = rcutree_prepare_cpu,
1556 .teardown.single = rcutree_dead_cpu,
1559 * On the tear-down path, timers_dead_cpu() must be invoked
1560 * before blk_mq_queue_reinit_notify() from notify_dead(),
1561 * otherwise a RCU stall occurs.
1563 [CPUHP_TIMERS_PREPARE] = {
1564 .name = "timers:prepare",
1565 .startup.single = timers_prepare_cpu,
1566 .teardown.single = timers_dead_cpu,
1568 /* Kicks the plugged cpu into life */
1569 [CPUHP_BRINGUP_CPU] = {
1570 .name = "cpu:bringup",
1571 .startup.single = bringup_cpu,
1572 .teardown.single = finish_cpu,
1575 /* Final state before CPU kills itself */
1576 [CPUHP_AP_IDLE_DEAD] = {
1577 .name = "idle:dead",
1580 * Last state before CPU enters the idle loop to die. Transient state
1581 * for synchronization.
1583 [CPUHP_AP_OFFLINE] = {
1584 .name = "ap:offline",
1587 /* First state is scheduler control. Interrupts are disabled */
1588 [CPUHP_AP_SCHED_STARTING] = {
1589 .name = "sched:starting",
1590 .startup.single = sched_cpu_starting,
1591 .teardown.single = sched_cpu_dying,
1593 [CPUHP_AP_RCUTREE_DYING] = {
1594 .name = "RCU/tree:dying",
1595 .startup.single = NULL,
1596 .teardown.single = rcutree_dying_cpu,
1598 [CPUHP_AP_SMPCFD_DYING] = {
1599 .name = "smpcfd:dying",
1600 .startup.single = NULL,
1601 .teardown.single = smpcfd_dying_cpu,
1603 /* Entry state on starting. Interrupts enabled from here on. Transient
1604 * state for synchronsization */
1605 [CPUHP_AP_ONLINE] = {
1606 .name = "ap:online",
1609 * Handled on controll processor until the plugged processor manages
1612 [CPUHP_TEARDOWN_CPU] = {
1613 .name = "cpu:teardown",
1614 .startup.single = NULL,
1615 .teardown.single = takedown_cpu,
1618 /* Handle smpboot threads park/unpark */
1619 [CPUHP_AP_SMPBOOT_THREADS] = {
1620 .name = "smpboot/threads:online",
1621 .startup.single = smpboot_unpark_threads,
1622 .teardown.single = smpboot_park_threads,
1624 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1625 .name = "irq/affinity:online",
1626 .startup.single = irq_affinity_online_cpu,
1627 .teardown.single = NULL,
1629 [CPUHP_AP_PERF_ONLINE] = {
1630 .name = "perf:online",
1631 .startup.single = perf_event_init_cpu,
1632 .teardown.single = perf_event_exit_cpu,
1634 [CPUHP_AP_WATCHDOG_ONLINE] = {
1635 .name = "lockup_detector:online",
1636 .startup.single = lockup_detector_online_cpu,
1637 .teardown.single = lockup_detector_offline_cpu,
1639 [CPUHP_AP_WORKQUEUE_ONLINE] = {
1640 .name = "workqueue:online",
1641 .startup.single = workqueue_online_cpu,
1642 .teardown.single = workqueue_offline_cpu,
1644 [CPUHP_AP_RCUTREE_ONLINE] = {
1645 .name = "RCU/tree:online",
1646 .startup.single = rcutree_online_cpu,
1647 .teardown.single = rcutree_offline_cpu,
1651 * The dynamically registered state space is here
1655 /* Last state is scheduler control setting the cpu active */
1656 [CPUHP_AP_ACTIVE] = {
1657 .name = "sched:active",
1658 .startup.single = sched_cpu_activate,
1659 .teardown.single = sched_cpu_deactivate,
1663 /* CPU is fully up and running. */
1666 .startup.single = NULL,
1667 .teardown.single = NULL,
1671 /* Sanity check for callbacks */
1672 static int cpuhp_cb_check(enum cpuhp_state state)
1674 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1680 * Returns a free for dynamic slot assignment of the Online state. The states
1681 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1682 * by having no name assigned.
1684 static int cpuhp_reserve_state(enum cpuhp_state state)
1686 enum cpuhp_state i, end;
1687 struct cpuhp_step *step;
1690 case CPUHP_AP_ONLINE_DYN:
1691 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1692 end = CPUHP_AP_ONLINE_DYN_END;
1694 case CPUHP_BP_PREPARE_DYN:
1695 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1696 end = CPUHP_BP_PREPARE_DYN_END;
1702 for (i = state; i <= end; i++, step++) {
1706 WARN(1, "No more dynamic states available for CPU hotplug\n");
1710 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1711 int (*startup)(unsigned int cpu),
1712 int (*teardown)(unsigned int cpu),
1713 bool multi_instance)
1715 /* (Un)Install the callbacks for further cpu hotplug operations */
1716 struct cpuhp_step *sp;
1720 * If name is NULL, then the state gets removed.
1722 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1723 * the first allocation from these dynamic ranges, so the removal
1724 * would trigger a new allocation and clear the wrong (already
1725 * empty) state, leaving the callbacks of the to be cleared state
1726 * dangling, which causes wreckage on the next hotplug operation.
1728 if (name && (state == CPUHP_AP_ONLINE_DYN ||
1729 state == CPUHP_BP_PREPARE_DYN)) {
1730 ret = cpuhp_reserve_state(state);
1735 sp = cpuhp_get_step(state);
1736 if (name && sp->name)
1739 sp->startup.single = startup;
1740 sp->teardown.single = teardown;
1742 sp->multi_instance = multi_instance;
1743 INIT_HLIST_HEAD(&sp->list);
1747 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1749 return cpuhp_get_step(state)->teardown.single;
1753 * Call the startup/teardown function for a step either on the AP or
1754 * on the current CPU.
1756 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1757 struct hlist_node *node)
1759 struct cpuhp_step *sp = cpuhp_get_step(state);
1763 * If there's nothing to do, we done.
1764 * Relies on the union for multi_instance.
1766 if ((bringup && !sp->startup.single) ||
1767 (!bringup && !sp->teardown.single))
1770 * The non AP bound callbacks can fail on bringup. On teardown
1771 * e.g. module removal we crash for now.
1774 if (cpuhp_is_ap_state(state))
1775 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1777 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1779 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1781 BUG_ON(ret && !bringup);
1786 * Called from __cpuhp_setup_state on a recoverable failure.
1788 * Note: The teardown callbacks for rollback are not allowed to fail!
1790 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1791 struct hlist_node *node)
1795 /* Roll back the already executed steps on the other cpus */
1796 for_each_present_cpu(cpu) {
1797 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1798 int cpustate = st->state;
1800 if (cpu >= failedcpu)
1803 /* Did we invoke the startup call on that cpu ? */
1804 if (cpustate >= state)
1805 cpuhp_issue_call(cpu, state, false, node);
1809 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1810 struct hlist_node *node,
1813 struct cpuhp_step *sp;
1817 lockdep_assert_cpus_held();
1819 sp = cpuhp_get_step(state);
1820 if (sp->multi_instance == false)
1823 mutex_lock(&cpuhp_state_mutex);
1825 if (!invoke || !sp->startup.multi)
1829 * Try to call the startup callback for each present cpu
1830 * depending on the hotplug state of the cpu.
1832 for_each_present_cpu(cpu) {
1833 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1834 int cpustate = st->state;
1836 if (cpustate < state)
1839 ret = cpuhp_issue_call(cpu, state, true, node);
1841 if (sp->teardown.multi)
1842 cpuhp_rollback_install(cpu, state, node);
1848 hlist_add_head(node, &sp->list);
1850 mutex_unlock(&cpuhp_state_mutex);
1854 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1860 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1864 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1867 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1868 * @state: The state to setup
1869 * @invoke: If true, the startup function is invoked for cpus where
1870 * cpu state >= @state
1871 * @startup: startup callback function
1872 * @teardown: teardown callback function
1873 * @multi_instance: State is set up for multiple instances which get
1876 * The caller needs to hold cpus read locked while calling this function.
1879 * Positive state number if @state is CPUHP_AP_ONLINE_DYN
1880 * 0 for all other states
1881 * On failure: proper (negative) error code
1883 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1884 const char *name, bool invoke,
1885 int (*startup)(unsigned int cpu),
1886 int (*teardown)(unsigned int cpu),
1887 bool multi_instance)
1892 lockdep_assert_cpus_held();
1894 if (cpuhp_cb_check(state) || !name)
1897 mutex_lock(&cpuhp_state_mutex);
1899 ret = cpuhp_store_callbacks(state, name, startup, teardown,
1902 dynstate = state == CPUHP_AP_ONLINE_DYN;
1903 if (ret > 0 && dynstate) {
1908 if (ret || !invoke || !startup)
1912 * Try to call the startup callback for each present cpu
1913 * depending on the hotplug state of the cpu.
1915 for_each_present_cpu(cpu) {
1916 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1917 int cpustate = st->state;
1919 if (cpustate < state)
1922 ret = cpuhp_issue_call(cpu, state, true, NULL);
1925 cpuhp_rollback_install(cpu, state, NULL);
1926 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1931 mutex_unlock(&cpuhp_state_mutex);
1933 * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1934 * dynamically allocated state in case of success.
1936 if (!ret && dynstate)
1940 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1942 int __cpuhp_setup_state(enum cpuhp_state state,
1943 const char *name, bool invoke,
1944 int (*startup)(unsigned int cpu),
1945 int (*teardown)(unsigned int cpu),
1946 bool multi_instance)
1951 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1952 teardown, multi_instance);
1956 EXPORT_SYMBOL(__cpuhp_setup_state);
1958 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1959 struct hlist_node *node, bool invoke)
1961 struct cpuhp_step *sp = cpuhp_get_step(state);
1964 BUG_ON(cpuhp_cb_check(state));
1966 if (!sp->multi_instance)
1970 mutex_lock(&cpuhp_state_mutex);
1972 if (!invoke || !cpuhp_get_teardown_cb(state))
1975 * Call the teardown callback for each present cpu depending
1976 * on the hotplug state of the cpu. This function is not
1977 * allowed to fail currently!
1979 for_each_present_cpu(cpu) {
1980 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1981 int cpustate = st->state;
1983 if (cpustate >= state)
1984 cpuhp_issue_call(cpu, state, false, node);
1989 mutex_unlock(&cpuhp_state_mutex);
1994 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1997 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1998 * @state: The state to remove
1999 * @invoke: If true, the teardown function is invoked for cpus where
2000 * cpu state >= @state
2002 * The caller needs to hold cpus read locked while calling this function.
2003 * The teardown callback is currently not allowed to fail. Think
2004 * about module removal!
2006 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2008 struct cpuhp_step *sp = cpuhp_get_step(state);
2011 BUG_ON(cpuhp_cb_check(state));
2013 lockdep_assert_cpus_held();
2015 mutex_lock(&cpuhp_state_mutex);
2016 if (sp->multi_instance) {
2017 WARN(!hlist_empty(&sp->list),
2018 "Error: Removing state %d which has instances left.\n",
2023 if (!invoke || !cpuhp_get_teardown_cb(state))
2027 * Call the teardown callback for each present cpu depending
2028 * on the hotplug state of the cpu. This function is not
2029 * allowed to fail currently!
2031 for_each_present_cpu(cpu) {
2032 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2033 int cpustate = st->state;
2035 if (cpustate >= state)
2036 cpuhp_issue_call(cpu, state, false, NULL);
2039 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2040 mutex_unlock(&cpuhp_state_mutex);
2042 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2044 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2047 __cpuhp_remove_state_cpuslocked(state, invoke);
2050 EXPORT_SYMBOL(__cpuhp_remove_state);
2052 #ifdef CONFIG_HOTPLUG_SMT
2053 static void cpuhp_offline_cpu_device(unsigned int cpu)
2055 struct device *dev = get_cpu_device(cpu);
2057 dev->offline = true;
2058 /* Tell user space about the state change */
2059 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2062 static void cpuhp_online_cpu_device(unsigned int cpu)
2064 struct device *dev = get_cpu_device(cpu);
2066 dev->offline = false;
2067 /* Tell user space about the state change */
2068 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2071 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2075 cpu_maps_update_begin();
2076 for_each_online_cpu(cpu) {
2077 if (topology_is_primary_thread(cpu))
2079 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2083 * As this needs to hold the cpu maps lock it's impossible
2084 * to call device_offline() because that ends up calling
2085 * cpu_down() which takes cpu maps lock. cpu maps lock
2086 * needs to be held as this might race against in kernel
2087 * abusers of the hotplug machinery (thermal management).
2089 * So nothing would update device:offline state. That would
2090 * leave the sysfs entry stale and prevent onlining after
2091 * smt control has been changed to 'off' again. This is
2092 * called under the sysfs hotplug lock, so it is properly
2093 * serialized against the regular offline usage.
2095 cpuhp_offline_cpu_device(cpu);
2098 cpu_smt_control = ctrlval;
2099 cpu_maps_update_done();
2103 int cpuhp_smt_enable(void)
2107 cpu_maps_update_begin();
2108 cpu_smt_control = CPU_SMT_ENABLED;
2109 for_each_present_cpu(cpu) {
2110 /* Skip online CPUs and CPUs on offline nodes */
2111 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2113 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2116 /* See comment in cpuhp_smt_disable() */
2117 cpuhp_online_cpu_device(cpu);
2119 cpu_maps_update_done();
2124 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2125 static ssize_t show_cpuhp_state(struct device *dev,
2126 struct device_attribute *attr, char *buf)
2128 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2130 return sprintf(buf, "%d\n", st->state);
2132 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
2134 static ssize_t write_cpuhp_target(struct device *dev,
2135 struct device_attribute *attr,
2136 const char *buf, size_t count)
2138 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2139 struct cpuhp_step *sp;
2142 ret = kstrtoint(buf, 10, &target);
2146 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2147 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2150 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2154 ret = lock_device_hotplug_sysfs();
2158 mutex_lock(&cpuhp_state_mutex);
2159 sp = cpuhp_get_step(target);
2160 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2161 mutex_unlock(&cpuhp_state_mutex);
2165 if (st->state < target)
2166 ret = cpu_up(dev->id, target);
2168 ret = cpu_down(dev->id, target);
2170 unlock_device_hotplug();
2171 return ret ? ret : count;
2174 static ssize_t show_cpuhp_target(struct device *dev,
2175 struct device_attribute *attr, char *buf)
2177 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2179 return sprintf(buf, "%d\n", st->target);
2181 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
2184 static ssize_t write_cpuhp_fail(struct device *dev,
2185 struct device_attribute *attr,
2186 const char *buf, size_t count)
2188 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2189 struct cpuhp_step *sp;
2192 ret = kstrtoint(buf, 10, &fail);
2196 if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2200 * Cannot fail STARTING/DYING callbacks.
2202 if (cpuhp_is_atomic_state(fail))
2206 * Cannot fail anything that doesn't have callbacks.
2208 mutex_lock(&cpuhp_state_mutex);
2209 sp = cpuhp_get_step(fail);
2210 if (!sp->startup.single && !sp->teardown.single)
2212 mutex_unlock(&cpuhp_state_mutex);
2221 static ssize_t show_cpuhp_fail(struct device *dev,
2222 struct device_attribute *attr, char *buf)
2224 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2226 return sprintf(buf, "%d\n", st->fail);
2229 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2231 static struct attribute *cpuhp_cpu_attrs[] = {
2232 &dev_attr_state.attr,
2233 &dev_attr_target.attr,
2234 &dev_attr_fail.attr,
2238 static const struct attribute_group cpuhp_cpu_attr_group = {
2239 .attrs = cpuhp_cpu_attrs,
2244 static ssize_t show_cpuhp_states(struct device *dev,
2245 struct device_attribute *attr, char *buf)
2247 ssize_t cur, res = 0;
2250 mutex_lock(&cpuhp_state_mutex);
2251 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2252 struct cpuhp_step *sp = cpuhp_get_step(i);
2255 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2260 mutex_unlock(&cpuhp_state_mutex);
2263 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2265 static struct attribute *cpuhp_cpu_root_attrs[] = {
2266 &dev_attr_states.attr,
2270 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2271 .attrs = cpuhp_cpu_root_attrs,
2276 #ifdef CONFIG_HOTPLUG_SMT
2279 __store_smt_control(struct device *dev, struct device_attribute *attr,
2280 const char *buf, size_t count)
2284 if (sysfs_streq(buf, "on"))
2285 ctrlval = CPU_SMT_ENABLED;
2286 else if (sysfs_streq(buf, "off"))
2287 ctrlval = CPU_SMT_DISABLED;
2288 else if (sysfs_streq(buf, "forceoff"))
2289 ctrlval = CPU_SMT_FORCE_DISABLED;
2293 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2296 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2299 ret = lock_device_hotplug_sysfs();
2303 if (ctrlval != cpu_smt_control) {
2305 case CPU_SMT_ENABLED:
2306 ret = cpuhp_smt_enable();
2308 case CPU_SMT_DISABLED:
2309 case CPU_SMT_FORCE_DISABLED:
2310 ret = cpuhp_smt_disable(ctrlval);
2315 unlock_device_hotplug();
2316 return ret ? ret : count;
2319 #else /* !CONFIG_HOTPLUG_SMT */
2321 __store_smt_control(struct device *dev, struct device_attribute *attr,
2322 const char *buf, size_t count)
2326 #endif /* CONFIG_HOTPLUG_SMT */
2328 static const char *smt_states[] = {
2329 [CPU_SMT_ENABLED] = "on",
2330 [CPU_SMT_DISABLED] = "off",
2331 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2332 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2333 [CPU_SMT_NOT_IMPLEMENTED] = "notimplemented",
2337 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2339 const char *state = smt_states[cpu_smt_control];
2341 return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2345 store_smt_control(struct device *dev, struct device_attribute *attr,
2346 const char *buf, size_t count)
2348 return __store_smt_control(dev, attr, buf, count);
2350 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2353 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2355 return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2357 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2359 static struct attribute *cpuhp_smt_attrs[] = {
2360 &dev_attr_control.attr,
2361 &dev_attr_active.attr,
2365 static const struct attribute_group cpuhp_smt_attr_group = {
2366 .attrs = cpuhp_smt_attrs,
2371 static int __init cpu_smt_sysfs_init(void)
2373 return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2374 &cpuhp_smt_attr_group);
2377 static int __init cpuhp_sysfs_init(void)
2381 ret = cpu_smt_sysfs_init();
2385 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2386 &cpuhp_cpu_root_attr_group);
2390 for_each_possible_cpu(cpu) {
2391 struct device *dev = get_cpu_device(cpu);
2395 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2401 device_initcall(cpuhp_sysfs_init);
2402 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2405 * cpu_bit_bitmap[] is a special, "compressed" data structure that
2406 * represents all NR_CPUS bits binary values of 1<<nr.
2408 * It is used by cpumask_of() to get a constant address to a CPU
2409 * mask value that has a single bit set only.
2412 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2413 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
2414 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2415 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2416 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2418 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2420 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
2421 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
2422 #if BITS_PER_LONG > 32
2423 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
2424 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
2427 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2429 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2430 EXPORT_SYMBOL(cpu_all_bits);
2432 #ifdef CONFIG_INIT_ALL_POSSIBLE
2433 struct cpumask __cpu_possible_mask __read_mostly
2436 struct cpumask __cpu_possible_mask __read_mostly;
2438 EXPORT_SYMBOL(__cpu_possible_mask);
2440 struct cpumask __cpu_online_mask __read_mostly;
2441 EXPORT_SYMBOL(__cpu_online_mask);
2443 struct cpumask __cpu_present_mask __read_mostly;
2444 EXPORT_SYMBOL(__cpu_present_mask);
2446 struct cpumask __cpu_active_mask __read_mostly;
2447 EXPORT_SYMBOL(__cpu_active_mask);
2449 atomic_t __num_online_cpus __read_mostly;
2450 EXPORT_SYMBOL(__num_online_cpus);
2452 void init_cpu_present(const struct cpumask *src)
2454 cpumask_copy(&__cpu_present_mask, src);
2457 void init_cpu_possible(const struct cpumask *src)
2459 cpumask_copy(&__cpu_possible_mask, src);
2462 void init_cpu_online(const struct cpumask *src)
2464 cpumask_copy(&__cpu_online_mask, src);
2467 void set_cpu_online(unsigned int cpu, bool online)
2470 * atomic_inc/dec() is required to handle the horrid abuse of this
2471 * function by the reboot and kexec code which invoke it from
2472 * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2473 * regular CPU hotplug is properly serialized.
2475 * Note, that the fact that __num_online_cpus is of type atomic_t
2476 * does not protect readers which are not serialized against
2477 * concurrent hotplug operations.
2480 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2481 atomic_inc(&__num_online_cpus);
2483 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2484 atomic_dec(&__num_online_cpus);
2489 * Activate the first processor.
2491 void __init boot_cpu_init(void)
2493 int cpu = smp_processor_id();
2495 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2496 set_cpu_online(cpu, true);
2497 set_cpu_active(cpu, true);
2498 set_cpu_present(cpu, true);
2499 set_cpu_possible(cpu, true);
2502 __boot_cpu_id = cpu;
2507 * Must be called _AFTER_ setting up the per_cpu areas
2509 void __init boot_cpu_hotplug_init(void)
2512 cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2514 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2518 * These are used for a global "mitigations=" cmdline option for toggling
2519 * optional CPU mitigations.
2521 enum cpu_mitigations {
2522 CPU_MITIGATIONS_OFF,
2523 CPU_MITIGATIONS_AUTO,
2524 CPU_MITIGATIONS_AUTO_NOSMT,
2527 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2528 CPU_MITIGATIONS_AUTO;
2530 static int __init mitigations_parse_cmdline(char *arg)
2532 if (!strcmp(arg, "off"))
2533 cpu_mitigations = CPU_MITIGATIONS_OFF;
2534 else if (!strcmp(arg, "auto"))
2535 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2536 else if (!strcmp(arg, "auto,nosmt"))
2537 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2539 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2544 early_param("mitigations", mitigations_parse_cmdline);
2546 /* mitigations=off */
2547 bool cpu_mitigations_off(void)
2549 return cpu_mitigations == CPU_MITIGATIONS_OFF;
2551 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2553 /* mitigations=auto,nosmt */
2554 bool cpu_mitigations_auto_nosmt(void)
2556 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2558 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);