2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
4 * This code is licenced under the GPL.
6 #include <linux/proc_fs.h>
8 #include <linux/init.h>
9 #include <linux/notifier.h>
10 #include <linux/sched/signal.h>
11 #include <linux/sched/hotplug.h>
12 #include <linux/sched/task.h>
13 #include <linux/unistd.h>
14 #include <linux/cpu.h>
15 #include <linux/oom.h>
16 #include <linux/rcupdate.h>
17 #include <linux/export.h>
18 #include <linux/bug.h>
19 #include <linux/kthread.h>
20 #include <linux/stop_machine.h>
21 #include <linux/mutex.h>
22 #include <linux/gfp.h>
23 #include <linux/suspend.h>
24 #include <linux/lockdep.h>
25 #include <linux/tick.h>
26 #include <linux/irq.h>
27 #include <linux/nmi.h>
28 #include <linux/smpboot.h>
29 #include <linux/relay.h>
30 #include <linux/slab.h>
31 #include <linux/percpu-rwsem.h>
33 #include <trace/events/power.h>
34 #define CREATE_TRACE_POINTS
35 #include <trace/events/cpuhp.h>
40 * cpuhp_cpu_state - Per cpu hotplug state storage
41 * @state: The current cpu state
42 * @target: The target state
43 * @thread: Pointer to the hotplug thread
44 * @should_run: Thread should execute
45 * @rollback: Perform a rollback
46 * @single: Single callback invocation
47 * @bringup: Single callback bringup or teardown selector
48 * @cb_state: The state for a single callback (install/uninstall)
49 * @result: Result of the operation
50 * @done_up: Signal completion to the issuer of the task for cpu-up
51 * @done_down: Signal completion to the issuer of the task for cpu-down
53 struct cpuhp_cpu_state {
54 enum cpuhp_state state;
55 enum cpuhp_state target;
56 enum cpuhp_state fail;
58 struct task_struct *thread;
64 struct hlist_node *node;
65 struct hlist_node *last;
66 enum cpuhp_state cb_state;
68 struct completion done_up;
69 struct completion done_down;
73 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
74 .fail = CPUHP_INVALID,
77 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
78 static struct lockdep_map cpuhp_state_up_map =
79 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
80 static struct lockdep_map cpuhp_state_down_map =
81 STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
84 static void inline cpuhp_lock_acquire(bool bringup)
86 lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
89 static void inline cpuhp_lock_release(bool bringup)
91 lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
95 static void inline cpuhp_lock_acquire(bool bringup) { }
96 static void inline cpuhp_lock_release(bool bringup) { }
101 * cpuhp_step - Hotplug state machine step
102 * @name: Name of the step
103 * @startup: Startup function of the step
104 * @teardown: Teardown function of the step
105 * @skip_onerr: Do not invoke the functions on error rollback
106 * Will go away once the notifiers are gone
107 * @cant_stop: Bringup/teardown can't be stopped at this step
112 int (*single)(unsigned int cpu);
113 int (*multi)(unsigned int cpu,
114 struct hlist_node *node);
117 int (*single)(unsigned int cpu);
118 int (*multi)(unsigned int cpu,
119 struct hlist_node *node);
121 struct hlist_head list;
127 static DEFINE_MUTEX(cpuhp_state_mutex);
128 static struct cpuhp_step cpuhp_bp_states[];
129 static struct cpuhp_step cpuhp_ap_states[];
131 static bool cpuhp_is_ap_state(enum cpuhp_state state)
134 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
135 * purposes as that state is handled explicitly in cpu_down.
137 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
140 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
142 struct cpuhp_step *sp;
144 sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
149 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
150 * @cpu: The cpu for which the callback should be invoked
151 * @state: The state to do callbacks for
152 * @bringup: True if the bringup callback should be invoked
153 * @node: For multi-instance, do a single entry callback for install/remove
154 * @lastp: For multi-instance rollback, remember how far we got
156 * Called from cpu hotplug and from the state register machinery.
158 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
159 bool bringup, struct hlist_node *node,
160 struct hlist_node **lastp)
162 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
163 struct cpuhp_step *step = cpuhp_get_step(state);
164 int (*cbm)(unsigned int cpu, struct hlist_node *node);
165 int (*cb)(unsigned int cpu);
168 if (st->fail == state) {
169 st->fail = CPUHP_INVALID;
171 if (!(bringup ? step->startup.single : step->teardown.single))
177 if (!step->multi_instance) {
178 WARN_ON_ONCE(lastp && *lastp);
179 cb = bringup ? step->startup.single : step->teardown.single;
182 trace_cpuhp_enter(cpu, st->target, state, cb);
184 trace_cpuhp_exit(cpu, st->state, state, ret);
187 cbm = bringup ? step->startup.multi : step->teardown.multi;
191 /* Single invocation for instance add/remove */
193 WARN_ON_ONCE(lastp && *lastp);
194 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
195 ret = cbm(cpu, node);
196 trace_cpuhp_exit(cpu, st->state, state, ret);
200 /* State transition. Invoke on all instances */
202 hlist_for_each(node, &step->list) {
203 if (lastp && node == *lastp)
206 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
207 ret = cbm(cpu, node);
208 trace_cpuhp_exit(cpu, st->state, state, ret);
222 /* Rollback the instances if one failed */
223 cbm = !bringup ? step->startup.multi : step->teardown.multi;
227 hlist_for_each(node, &step->list) {
231 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
232 ret = cbm(cpu, node);
233 trace_cpuhp_exit(cpu, st->state, state, ret);
235 * Rollback must not fail,
243 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
245 struct completion *done = bringup ? &st->done_up : &st->done_down;
246 wait_for_completion(done);
249 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
251 struct completion *done = bringup ? &st->done_up : &st->done_down;
256 * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
258 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
260 return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
263 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
264 static DEFINE_MUTEX(cpu_add_remove_lock);
265 bool cpuhp_tasks_frozen;
266 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
269 * The following two APIs (cpu_maps_update_begin/done) must be used when
270 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
272 void cpu_maps_update_begin(void)
274 mutex_lock(&cpu_add_remove_lock);
277 void cpu_maps_update_done(void)
279 mutex_unlock(&cpu_add_remove_lock);
283 * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
284 * Should always be manipulated under cpu_add_remove_lock
286 static int cpu_hotplug_disabled;
288 #ifdef CONFIG_HOTPLUG_CPU
290 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
292 void cpus_read_lock(void)
294 percpu_down_read(&cpu_hotplug_lock);
296 EXPORT_SYMBOL_GPL(cpus_read_lock);
298 void cpus_read_unlock(void)
300 percpu_up_read(&cpu_hotplug_lock);
302 EXPORT_SYMBOL_GPL(cpus_read_unlock);
304 void cpus_write_lock(void)
306 percpu_down_write(&cpu_hotplug_lock);
309 void cpus_write_unlock(void)
311 percpu_up_write(&cpu_hotplug_lock);
314 void lockdep_assert_cpus_held(void)
316 percpu_rwsem_assert_held(&cpu_hotplug_lock);
320 * Wait for currently running CPU hotplug operations to complete (if any) and
321 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
322 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
323 * hotplug path before performing hotplug operations. So acquiring that lock
324 * guarantees mutual exclusion from any currently running hotplug operations.
326 void cpu_hotplug_disable(void)
328 cpu_maps_update_begin();
329 cpu_hotplug_disabled++;
330 cpu_maps_update_done();
332 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
334 static void __cpu_hotplug_enable(void)
336 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
338 cpu_hotplug_disabled--;
341 void cpu_hotplug_enable(void)
343 cpu_maps_update_begin();
344 __cpu_hotplug_enable();
345 cpu_maps_update_done();
347 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
348 #endif /* CONFIG_HOTPLUG_CPU */
350 #ifdef CONFIG_HOTPLUG_SMT
351 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
352 EXPORT_SYMBOL_GPL(cpu_smt_control);
354 static bool cpu_smt_available __read_mostly;
356 void __init cpu_smt_disable(bool force)
358 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
359 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
363 pr_info("SMT: Force disabled\n");
364 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
366 cpu_smt_control = CPU_SMT_DISABLED;
371 * The decision whether SMT is supported can only be done after the full
372 * CPU identification. Called from architecture code before non boot CPUs
375 void __init cpu_smt_check_topology_early(void)
377 if (!topology_smt_supported())
378 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
382 * If SMT was disabled by BIOS, detect it here, after the CPUs have been
383 * brought online. This ensures the smt/l1tf sysfs entries are consistent
384 * with reality. cpu_smt_available is set to true during the bringup of non
385 * boot CPUs when a SMT sibling is detected. Note, this may overwrite
386 * cpu_smt_control's previous setting.
388 void __init cpu_smt_check_topology(void)
390 if (!cpu_smt_available)
391 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
394 static int __init smt_cmdline_disable(char *str)
396 cpu_smt_disable(str && !strcmp(str, "force"));
399 early_param("nosmt", smt_cmdline_disable);
401 static inline bool cpu_smt_allowed(unsigned int cpu)
403 if (topology_is_primary_thread(cpu))
407 * If the CPU is not a 'primary' thread and the booted_once bit is
408 * set then the processor has SMT support. Store this information
409 * for the late check of SMT support in cpu_smt_check_topology().
411 if (per_cpu(cpuhp_state, cpu).booted_once)
412 cpu_smt_available = true;
414 if (cpu_smt_control == CPU_SMT_ENABLED)
418 * On x86 it's required to boot all logical CPUs at least once so
419 * that the init code can get a chance to set CR4.MCE on each
420 * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
421 * core will shutdown the machine.
423 return !per_cpu(cpuhp_state, cpu).booted_once;
426 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
429 static inline enum cpuhp_state
430 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
432 enum cpuhp_state prev_state = st->state;
434 st->rollback = false;
439 st->bringup = st->state < target;
445 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
450 * If we have st->last we need to undo partial multi_instance of this
451 * state first. Otherwise start undo at the previous state.
460 st->target = prev_state;
461 st->bringup = !st->bringup;
464 /* Regular hotplug invocation of the AP hotplug thread */
465 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
467 if (!st->single && st->state == st->target)
472 * Make sure the above stores are visible before should_run becomes
473 * true. Paired with the mb() above in cpuhp_thread_fun()
476 st->should_run = true;
477 wake_up_process(st->thread);
478 wait_for_ap_thread(st, st->bringup);
481 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
483 enum cpuhp_state prev_state;
486 prev_state = cpuhp_set_state(st, target);
488 if ((ret = st->result)) {
489 cpuhp_reset_state(st, prev_state);
496 static int bringup_wait_for_ap(unsigned int cpu)
498 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
500 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
501 wait_for_ap_thread(st, true);
502 if (WARN_ON_ONCE((!cpu_online(cpu))))
505 /* Unpark the stopper thread and the hotplug thread of the target cpu */
506 stop_machine_unpark(cpu);
507 kthread_unpark(st->thread);
510 * SMT soft disabling on X86 requires to bring the CPU out of the
511 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
512 * CPU marked itself as booted_once in cpu_notify_starting() so the
513 * cpu_smt_allowed() check will now return false if this is not the
516 if (!cpu_smt_allowed(cpu))
519 if (st->target <= CPUHP_AP_ONLINE_IDLE)
522 return cpuhp_kick_ap(st, st->target);
525 static int bringup_cpu(unsigned int cpu)
527 struct task_struct *idle = idle_thread_get(cpu);
531 * Some architectures have to walk the irq descriptors to
532 * setup the vector space for the cpu which comes online.
533 * Prevent irq alloc/free across the bringup.
537 /* Arch-specific enabling code. */
538 ret = __cpu_up(cpu, idle);
542 return bringup_wait_for_ap(cpu);
546 * Hotplug state machine related functions
549 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
551 for (st->state--; st->state > st->target; st->state--) {
552 struct cpuhp_step *step = cpuhp_get_step(st->state);
554 if (!step->skip_onerr)
555 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
559 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
560 enum cpuhp_state target)
562 enum cpuhp_state prev_state = st->state;
565 while (st->state < target) {
567 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
569 st->target = prev_state;
570 undo_cpu_up(cpu, st);
578 * The cpu hotplug threads manage the bringup and teardown of the cpus
580 static void cpuhp_create(unsigned int cpu)
582 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
584 init_completion(&st->done_up);
585 init_completion(&st->done_down);
588 static int cpuhp_should_run(unsigned int cpu)
590 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
592 return st->should_run;
596 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
597 * callbacks when a state gets [un]installed at runtime.
599 * Each invocation of this function by the smpboot thread does a single AP
602 * It has 3 modes of operation:
603 * - single: runs st->cb_state
604 * - up: runs ++st->state, while st->state < st->target
605 * - down: runs st->state--, while st->state > st->target
607 * When complete or on error, should_run is cleared and the completion is fired.
609 static void cpuhp_thread_fun(unsigned int cpu)
611 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
612 bool bringup = st->bringup;
613 enum cpuhp_state state;
615 if (WARN_ON_ONCE(!st->should_run))
619 * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
620 * that if we see ->should_run we also see the rest of the state.
624 cpuhp_lock_acquire(bringup);
627 state = st->cb_state;
628 st->should_run = false;
633 st->should_run = (st->state < st->target);
634 WARN_ON_ONCE(st->state > st->target);
638 st->should_run = (st->state > st->target);
639 WARN_ON_ONCE(st->state < st->target);
643 WARN_ON_ONCE(!cpuhp_is_ap_state(state));
646 struct cpuhp_step *step = cpuhp_get_step(state);
647 if (step->skip_onerr)
651 if (cpuhp_is_atomic_state(state)) {
653 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
657 * STARTING/DYING must not fail!
659 WARN_ON_ONCE(st->result);
661 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
666 * If we fail on a rollback, we're up a creek without no
667 * paddle, no way forward, no way back. We loose, thanks for
670 WARN_ON_ONCE(st->rollback);
671 st->should_run = false;
675 cpuhp_lock_release(bringup);
678 complete_ap_thread(st, bringup);
681 /* Invoke a single callback on a remote cpu */
683 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
684 struct hlist_node *node)
686 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
689 if (!cpu_online(cpu))
692 cpuhp_lock_acquire(false);
693 cpuhp_lock_release(false);
695 cpuhp_lock_acquire(true);
696 cpuhp_lock_release(true);
699 * If we are up and running, use the hotplug thread. For early calls
700 * we invoke the thread function directly.
703 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
705 st->rollback = false;
709 st->bringup = bringup;
710 st->cb_state = state;
716 * If we failed and did a partial, do a rollback.
718 if ((ret = st->result) && st->last) {
720 st->bringup = !bringup;
726 * Clean up the leftovers so the next hotplug operation wont use stale
729 st->node = st->last = NULL;
733 static int cpuhp_kick_ap_work(unsigned int cpu)
735 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
736 enum cpuhp_state prev_state = st->state;
739 cpuhp_lock_acquire(false);
740 cpuhp_lock_release(false);
742 cpuhp_lock_acquire(true);
743 cpuhp_lock_release(true);
745 trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
746 ret = cpuhp_kick_ap(st, st->target);
747 trace_cpuhp_exit(cpu, st->state, prev_state, ret);
752 static struct smp_hotplug_thread cpuhp_threads = {
753 .store = &cpuhp_state.thread,
754 .create = &cpuhp_create,
755 .thread_should_run = cpuhp_should_run,
756 .thread_fn = cpuhp_thread_fun,
757 .thread_comm = "cpuhp/%u",
761 void __init cpuhp_threads_init(void)
763 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
764 kthread_unpark(this_cpu_read(cpuhp_state.thread));
767 #ifdef CONFIG_HOTPLUG_CPU
769 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
772 * This function walks all processes, finds a valid mm struct for each one and
773 * then clears a corresponding bit in mm's cpumask. While this all sounds
774 * trivial, there are various non-obvious corner cases, which this function
775 * tries to solve in a safe manner.
777 * Also note that the function uses a somewhat relaxed locking scheme, so it may
778 * be called only for an already offlined CPU.
780 void clear_tasks_mm_cpumask(int cpu)
782 struct task_struct *p;
785 * This function is called after the cpu is taken down and marked
786 * offline, so its not like new tasks will ever get this cpu set in
787 * their mm mask. -- Peter Zijlstra
788 * Thus, we may use rcu_read_lock() here, instead of grabbing
789 * full-fledged tasklist_lock.
791 WARN_ON(cpu_online(cpu));
793 for_each_process(p) {
794 struct task_struct *t;
797 * Main thread might exit, but other threads may still have
798 * a valid mm. Find one.
800 t = find_lock_task_mm(p);
803 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
809 /* Take this CPU down. */
810 static int take_cpu_down(void *_param)
812 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
813 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
814 int err, cpu = smp_processor_id();
817 /* Ensure this CPU doesn't handle any more interrupts. */
818 err = __cpu_disable();
823 * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
824 * do this step again.
826 WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
828 /* Invoke the former CPU_DYING callbacks */
829 for (; st->state > target; st->state--) {
830 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
832 * DYING must not fail!
837 /* Give up timekeeping duties */
838 tick_handover_do_timer();
839 /* Park the stopper thread */
840 stop_machine_park(cpu);
844 static int takedown_cpu(unsigned int cpu)
846 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
849 /* Park the smpboot threads */
850 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
853 * Prevent irq alloc/free while the dying cpu reorganizes the
854 * interrupt affinities.
859 * So now all preempt/rcu users must observe !cpu_active().
861 err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
863 /* CPU refused to die */
865 /* Unpark the hotplug thread so we can rollback there */
866 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
869 BUG_ON(cpu_online(cpu));
872 * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
873 * runnable tasks from the cpu, there's only the idle task left now
874 * that the migration thread is done doing the stop_machine thing.
876 * Wait for the stop thread to go away.
878 wait_for_ap_thread(st, false);
879 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
881 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
884 hotplug_cpu__broadcast_tick_pull(cpu);
885 /* This actually kills the CPU. */
888 tick_cleanup_dead_cpu(cpu);
889 rcutree_migrate_callbacks(cpu);
893 static void cpuhp_complete_idle_dead(void *arg)
895 struct cpuhp_cpu_state *st = arg;
897 complete_ap_thread(st, false);
900 void cpuhp_report_idle_dead(void)
902 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
904 BUG_ON(st->state != CPUHP_AP_OFFLINE);
905 rcu_report_dead(smp_processor_id());
906 st->state = CPUHP_AP_IDLE_DEAD;
908 * We cannot call complete after rcu_report_dead() so we delegate it
911 smp_call_function_single(cpumask_first(cpu_online_mask),
912 cpuhp_complete_idle_dead, st, 0);
915 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
917 for (st->state++; st->state < st->target; st->state++) {
918 struct cpuhp_step *step = cpuhp_get_step(st->state);
920 if (!step->skip_onerr)
921 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
925 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
926 enum cpuhp_state target)
928 enum cpuhp_state prev_state = st->state;
931 for (; st->state > target; st->state--) {
932 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
934 st->target = prev_state;
935 if (st->state < prev_state)
936 undo_cpu_down(cpu, st);
943 /* Requires cpu_add_remove_lock to be held */
944 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
945 enum cpuhp_state target)
947 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
948 int prev_state, ret = 0;
950 if (num_online_cpus() == 1)
953 if (!cpu_present(cpu))
958 cpuhp_tasks_frozen = tasks_frozen;
960 prev_state = cpuhp_set_state(st, target);
962 * If the current CPU state is in the range of the AP hotplug thread,
963 * then we need to kick the thread.
965 if (st->state > CPUHP_TEARDOWN_CPU) {
966 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
967 ret = cpuhp_kick_ap_work(cpu);
969 * The AP side has done the error rollback already. Just
970 * return the error code..
976 * We might have stopped still in the range of the AP hotplug
977 * thread. Nothing to do anymore.
979 if (st->state > CPUHP_TEARDOWN_CPU)
985 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
986 * to do the further cleanups.
988 ret = cpuhp_down_callbacks(cpu, st, target);
989 if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
990 cpuhp_reset_state(st, prev_state);
997 * Do post unplug cleanup. This is still protected against
998 * concurrent CPU hotplug via cpu_add_remove_lock.
1000 lockup_detector_cleanup();
1004 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1006 if (cpu_hotplug_disabled)
1008 return _cpu_down(cpu, 0, target);
1011 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1015 cpu_maps_update_begin();
1016 err = cpu_down_maps_locked(cpu, target);
1017 cpu_maps_update_done();
1021 int cpu_down(unsigned int cpu)
1023 return do_cpu_down(cpu, CPUHP_OFFLINE);
1025 EXPORT_SYMBOL(cpu_down);
1028 #define takedown_cpu NULL
1029 #endif /*CONFIG_HOTPLUG_CPU*/
1032 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1033 * @cpu: cpu that just started
1035 * It must be called by the arch code on the new cpu, before the new cpu
1036 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1038 void notify_cpu_starting(unsigned int cpu)
1040 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1041 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1044 rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1045 st->booted_once = true;
1046 while (st->state < target) {
1048 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1050 * STARTING must not fail!
1057 * Called from the idle task. Wake up the controlling task which brings the
1058 * stopper and the hotplug thread of the upcoming CPU up and then delegates
1059 * the rest of the online bringup to the hotplug thread.
1061 void cpuhp_online_idle(enum cpuhp_state state)
1063 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1065 /* Happens for the boot cpu */
1066 if (state != CPUHP_AP_ONLINE_IDLE)
1069 st->state = CPUHP_AP_ONLINE_IDLE;
1070 complete_ap_thread(st, true);
1073 /* Requires cpu_add_remove_lock to be held */
1074 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1076 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1077 struct task_struct *idle;
1082 if (!cpu_present(cpu)) {
1088 * The caller of do_cpu_up might have raced with another
1089 * caller. Ignore it for now.
1091 if (st->state >= target)
1094 if (st->state == CPUHP_OFFLINE) {
1095 /* Let it fail before we try to bring the cpu up */
1096 idle = idle_thread_get(cpu);
1098 ret = PTR_ERR(idle);
1103 cpuhp_tasks_frozen = tasks_frozen;
1105 cpuhp_set_state(st, target);
1107 * If the current CPU state is in the range of the AP hotplug thread,
1108 * then we need to kick the thread once more.
1110 if (st->state > CPUHP_BRINGUP_CPU) {
1111 ret = cpuhp_kick_ap_work(cpu);
1113 * The AP side has done the error rollback already. Just
1114 * return the error code..
1121 * Try to reach the target state. We max out on the BP at
1122 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1123 * responsible for bringing it up to the target state.
1125 target = min((int)target, CPUHP_BRINGUP_CPU);
1126 ret = cpuhp_up_callbacks(cpu, st, target);
1128 cpus_write_unlock();
1132 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1136 if (!cpu_possible(cpu)) {
1137 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1139 #if defined(CONFIG_IA64)
1140 pr_err("please check additional_cpus= boot parameter\n");
1145 err = try_online_node(cpu_to_node(cpu));
1149 cpu_maps_update_begin();
1151 if (cpu_hotplug_disabled) {
1155 if (!cpu_smt_allowed(cpu)) {
1160 err = _cpu_up(cpu, 0, target);
1162 cpu_maps_update_done();
1166 int cpu_up(unsigned int cpu)
1168 return do_cpu_up(cpu, CPUHP_ONLINE);
1170 EXPORT_SYMBOL_GPL(cpu_up);
1172 #ifdef CONFIG_PM_SLEEP_SMP
1173 static cpumask_var_t frozen_cpus;
1175 int freeze_secondary_cpus(int primary)
1179 cpu_maps_update_begin();
1180 if (!cpu_online(primary))
1181 primary = cpumask_first(cpu_online_mask);
1183 * We take down all of the non-boot CPUs in one shot to avoid races
1184 * with the userspace trying to use the CPU hotplug at the same time
1186 cpumask_clear(frozen_cpus);
1188 pr_info("Disabling non-boot CPUs ...\n");
1189 for_each_online_cpu(cpu) {
1192 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1193 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1194 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1196 cpumask_set_cpu(cpu, frozen_cpus);
1198 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1204 BUG_ON(num_online_cpus() > 1);
1206 pr_err("Non-boot CPUs are not disabled\n");
1209 * Make sure the CPUs won't be enabled by someone else. We need to do
1210 * this even in case of failure as all disable_nonboot_cpus() users are
1211 * supposed to do enable_nonboot_cpus() on the failure path.
1213 cpu_hotplug_disabled++;
1215 cpu_maps_update_done();
1219 void __weak arch_enable_nonboot_cpus_begin(void)
1223 void __weak arch_enable_nonboot_cpus_end(void)
1227 void enable_nonboot_cpus(void)
1231 /* Allow everyone to use the CPU hotplug again */
1232 cpu_maps_update_begin();
1233 __cpu_hotplug_enable();
1234 if (cpumask_empty(frozen_cpus))
1237 pr_info("Enabling non-boot CPUs ...\n");
1239 arch_enable_nonboot_cpus_begin();
1241 for_each_cpu(cpu, frozen_cpus) {
1242 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1243 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1244 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1246 pr_info("CPU%d is up\n", cpu);
1249 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1252 arch_enable_nonboot_cpus_end();
1254 cpumask_clear(frozen_cpus);
1256 cpu_maps_update_done();
1259 static int __init alloc_frozen_cpus(void)
1261 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1265 core_initcall(alloc_frozen_cpus);
1268 * When callbacks for CPU hotplug notifications are being executed, we must
1269 * ensure that the state of the system with respect to the tasks being frozen
1270 * or not, as reported by the notification, remains unchanged *throughout the
1271 * duration* of the execution of the callbacks.
1272 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1274 * This synchronization is implemented by mutually excluding regular CPU
1275 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1276 * Hibernate notifications.
1279 cpu_hotplug_pm_callback(struct notifier_block *nb,
1280 unsigned long action, void *ptr)
1284 case PM_SUSPEND_PREPARE:
1285 case PM_HIBERNATION_PREPARE:
1286 cpu_hotplug_disable();
1289 case PM_POST_SUSPEND:
1290 case PM_POST_HIBERNATION:
1291 cpu_hotplug_enable();
1302 static int __init cpu_hotplug_pm_sync_init(void)
1305 * cpu_hotplug_pm_callback has higher priority than x86
1306 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1307 * to disable cpu hotplug to avoid cpu hotplug race.
1309 pm_notifier(cpu_hotplug_pm_callback, 0);
1312 core_initcall(cpu_hotplug_pm_sync_init);
1314 #endif /* CONFIG_PM_SLEEP_SMP */
1318 #endif /* CONFIG_SMP */
1320 /* Boot processor state steps */
1321 static struct cpuhp_step cpuhp_bp_states[] = {
1324 .startup.single = NULL,
1325 .teardown.single = NULL,
1328 [CPUHP_CREATE_THREADS]= {
1329 .name = "threads:prepare",
1330 .startup.single = smpboot_create_threads,
1331 .teardown.single = NULL,
1334 [CPUHP_PERF_PREPARE] = {
1335 .name = "perf:prepare",
1336 .startup.single = perf_event_init_cpu,
1337 .teardown.single = perf_event_exit_cpu,
1339 [CPUHP_WORKQUEUE_PREP] = {
1340 .name = "workqueue:prepare",
1341 .startup.single = workqueue_prepare_cpu,
1342 .teardown.single = NULL,
1344 [CPUHP_HRTIMERS_PREPARE] = {
1345 .name = "hrtimers:prepare",
1346 .startup.single = hrtimers_prepare_cpu,
1347 .teardown.single = hrtimers_dead_cpu,
1349 [CPUHP_SMPCFD_PREPARE] = {
1350 .name = "smpcfd:prepare",
1351 .startup.single = smpcfd_prepare_cpu,
1352 .teardown.single = smpcfd_dead_cpu,
1354 [CPUHP_RELAY_PREPARE] = {
1355 .name = "relay:prepare",
1356 .startup.single = relay_prepare_cpu,
1357 .teardown.single = NULL,
1359 [CPUHP_SLAB_PREPARE] = {
1360 .name = "slab:prepare",
1361 .startup.single = slab_prepare_cpu,
1362 .teardown.single = slab_dead_cpu,
1364 [CPUHP_RCUTREE_PREP] = {
1365 .name = "RCU/tree:prepare",
1366 .startup.single = rcutree_prepare_cpu,
1367 .teardown.single = rcutree_dead_cpu,
1370 * On the tear-down path, timers_dead_cpu() must be invoked
1371 * before blk_mq_queue_reinit_notify() from notify_dead(),
1372 * otherwise a RCU stall occurs.
1374 [CPUHP_TIMERS_PREPARE] = {
1375 .name = "timers:dead",
1376 .startup.single = timers_prepare_cpu,
1377 .teardown.single = timers_dead_cpu,
1379 /* Kicks the plugged cpu into life */
1380 [CPUHP_BRINGUP_CPU] = {
1381 .name = "cpu:bringup",
1382 .startup.single = bringup_cpu,
1383 .teardown.single = NULL,
1387 * Handled on controll processor until the plugged processor manages
1390 [CPUHP_TEARDOWN_CPU] = {
1391 .name = "cpu:teardown",
1392 .startup.single = NULL,
1393 .teardown.single = takedown_cpu,
1397 [CPUHP_BRINGUP_CPU] = { },
1401 /* Application processor state steps */
1402 static struct cpuhp_step cpuhp_ap_states[] = {
1404 /* Final state before CPU kills itself */
1405 [CPUHP_AP_IDLE_DEAD] = {
1406 .name = "idle:dead",
1409 * Last state before CPU enters the idle loop to die. Transient state
1410 * for synchronization.
1412 [CPUHP_AP_OFFLINE] = {
1413 .name = "ap:offline",
1416 /* First state is scheduler control. Interrupts are disabled */
1417 [CPUHP_AP_SCHED_STARTING] = {
1418 .name = "sched:starting",
1419 .startup.single = sched_cpu_starting,
1420 .teardown.single = sched_cpu_dying,
1422 [CPUHP_AP_RCUTREE_DYING] = {
1423 .name = "RCU/tree:dying",
1424 .startup.single = NULL,
1425 .teardown.single = rcutree_dying_cpu,
1427 [CPUHP_AP_SMPCFD_DYING] = {
1428 .name = "smpcfd:dying",
1429 .startup.single = NULL,
1430 .teardown.single = smpcfd_dying_cpu,
1432 /* Entry state on starting. Interrupts enabled from here on. Transient
1433 * state for synchronsization */
1434 [CPUHP_AP_ONLINE] = {
1435 .name = "ap:online",
1437 /* Handle smpboot threads park/unpark */
1438 [CPUHP_AP_SMPBOOT_THREADS] = {
1439 .name = "smpboot/threads:online",
1440 .startup.single = smpboot_unpark_threads,
1441 .teardown.single = smpboot_park_threads,
1443 [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1444 .name = "irq/affinity:online",
1445 .startup.single = irq_affinity_online_cpu,
1446 .teardown.single = NULL,
1448 [CPUHP_AP_PERF_ONLINE] = {
1449 .name = "perf:online",
1450 .startup.single = perf_event_init_cpu,
1451 .teardown.single = perf_event_exit_cpu,
1453 [CPUHP_AP_WORKQUEUE_ONLINE] = {
1454 .name = "workqueue:online",
1455 .startup.single = workqueue_online_cpu,
1456 .teardown.single = workqueue_offline_cpu,
1458 [CPUHP_AP_RCUTREE_ONLINE] = {
1459 .name = "RCU/tree:online",
1460 .startup.single = rcutree_online_cpu,
1461 .teardown.single = rcutree_offline_cpu,
1465 * The dynamically registered state space is here
1469 /* Last state is scheduler control setting the cpu active */
1470 [CPUHP_AP_ACTIVE] = {
1471 .name = "sched:active",
1472 .startup.single = sched_cpu_activate,
1473 .teardown.single = sched_cpu_deactivate,
1477 /* CPU is fully up and running. */
1480 .startup.single = NULL,
1481 .teardown.single = NULL,
1485 /* Sanity check for callbacks */
1486 static int cpuhp_cb_check(enum cpuhp_state state)
1488 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1494 * Returns a free for dynamic slot assignment of the Online state. The states
1495 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1496 * by having no name assigned.
1498 static int cpuhp_reserve_state(enum cpuhp_state state)
1500 enum cpuhp_state i, end;
1501 struct cpuhp_step *step;
1504 case CPUHP_AP_ONLINE_DYN:
1505 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1506 end = CPUHP_AP_ONLINE_DYN_END;
1508 case CPUHP_BP_PREPARE_DYN:
1509 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1510 end = CPUHP_BP_PREPARE_DYN_END;
1516 for (i = state; i <= end; i++, step++) {
1520 WARN(1, "No more dynamic states available for CPU hotplug\n");
1524 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1525 int (*startup)(unsigned int cpu),
1526 int (*teardown)(unsigned int cpu),
1527 bool multi_instance)
1529 /* (Un)Install the callbacks for further cpu hotplug operations */
1530 struct cpuhp_step *sp;
1534 * If name is NULL, then the state gets removed.
1536 * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1537 * the first allocation from these dynamic ranges, so the removal
1538 * would trigger a new allocation and clear the wrong (already
1539 * empty) state, leaving the callbacks of the to be cleared state
1540 * dangling, which causes wreckage on the next hotplug operation.
1542 if (name && (state == CPUHP_AP_ONLINE_DYN ||
1543 state == CPUHP_BP_PREPARE_DYN)) {
1544 ret = cpuhp_reserve_state(state);
1549 sp = cpuhp_get_step(state);
1550 if (name && sp->name)
1553 sp->startup.single = startup;
1554 sp->teardown.single = teardown;
1556 sp->multi_instance = multi_instance;
1557 INIT_HLIST_HEAD(&sp->list);
1561 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1563 return cpuhp_get_step(state)->teardown.single;
1567 * Call the startup/teardown function for a step either on the AP or
1568 * on the current CPU.
1570 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1571 struct hlist_node *node)
1573 struct cpuhp_step *sp = cpuhp_get_step(state);
1577 * If there's nothing to do, we done.
1578 * Relies on the union for multi_instance.
1580 if ((bringup && !sp->startup.single) ||
1581 (!bringup && !sp->teardown.single))
1584 * The non AP bound callbacks can fail on bringup. On teardown
1585 * e.g. module removal we crash for now.
1588 if (cpuhp_is_ap_state(state))
1589 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1591 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1593 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1595 BUG_ON(ret && !bringup);
1600 * Called from __cpuhp_setup_state on a recoverable failure.
1602 * Note: The teardown callbacks for rollback are not allowed to fail!
1604 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1605 struct hlist_node *node)
1609 /* Roll back the already executed steps on the other cpus */
1610 for_each_present_cpu(cpu) {
1611 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1612 int cpustate = st->state;
1614 if (cpu >= failedcpu)
1617 /* Did we invoke the startup call on that cpu ? */
1618 if (cpustate >= state)
1619 cpuhp_issue_call(cpu, state, false, node);
1623 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1624 struct hlist_node *node,
1627 struct cpuhp_step *sp;
1631 lockdep_assert_cpus_held();
1633 sp = cpuhp_get_step(state);
1634 if (sp->multi_instance == false)
1637 mutex_lock(&cpuhp_state_mutex);
1639 if (!invoke || !sp->startup.multi)
1643 * Try to call the startup callback for each present cpu
1644 * depending on the hotplug state of the cpu.
1646 for_each_present_cpu(cpu) {
1647 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1648 int cpustate = st->state;
1650 if (cpustate < state)
1653 ret = cpuhp_issue_call(cpu, state, true, node);
1655 if (sp->teardown.multi)
1656 cpuhp_rollback_install(cpu, state, node);
1662 hlist_add_head(node, &sp->list);
1664 mutex_unlock(&cpuhp_state_mutex);
1668 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1674 ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1678 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1681 * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1682 * @state: The state to setup
1683 * @invoke: If true, the startup function is invoked for cpus where
1684 * cpu state >= @state
1685 * @startup: startup callback function
1686 * @teardown: teardown callback function
1687 * @multi_instance: State is set up for multiple instances which get
1690 * The caller needs to hold cpus read locked while calling this function.
1693 * Positive state number if @state is CPUHP_AP_ONLINE_DYN
1694 * 0 for all other states
1695 * On failure: proper (negative) error code
1697 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1698 const char *name, bool invoke,
1699 int (*startup)(unsigned int cpu),
1700 int (*teardown)(unsigned int cpu),
1701 bool multi_instance)
1706 lockdep_assert_cpus_held();
1708 if (cpuhp_cb_check(state) || !name)
1711 mutex_lock(&cpuhp_state_mutex);
1713 ret = cpuhp_store_callbacks(state, name, startup, teardown,
1716 dynstate = state == CPUHP_AP_ONLINE_DYN;
1717 if (ret > 0 && dynstate) {
1722 if (ret || !invoke || !startup)
1726 * Try to call the startup callback for each present cpu
1727 * depending on the hotplug state of the cpu.
1729 for_each_present_cpu(cpu) {
1730 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1731 int cpustate = st->state;
1733 if (cpustate < state)
1736 ret = cpuhp_issue_call(cpu, state, true, NULL);
1739 cpuhp_rollback_install(cpu, state, NULL);
1740 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1745 mutex_unlock(&cpuhp_state_mutex);
1747 * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1748 * dynamically allocated state in case of success.
1750 if (!ret && dynstate)
1754 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1756 int __cpuhp_setup_state(enum cpuhp_state state,
1757 const char *name, bool invoke,
1758 int (*startup)(unsigned int cpu),
1759 int (*teardown)(unsigned int cpu),
1760 bool multi_instance)
1765 ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1766 teardown, multi_instance);
1770 EXPORT_SYMBOL(__cpuhp_setup_state);
1772 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1773 struct hlist_node *node, bool invoke)
1775 struct cpuhp_step *sp = cpuhp_get_step(state);
1778 BUG_ON(cpuhp_cb_check(state));
1780 if (!sp->multi_instance)
1784 mutex_lock(&cpuhp_state_mutex);
1786 if (!invoke || !cpuhp_get_teardown_cb(state))
1789 * Call the teardown callback for each present cpu depending
1790 * on the hotplug state of the cpu. This function is not
1791 * allowed to fail currently!
1793 for_each_present_cpu(cpu) {
1794 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1795 int cpustate = st->state;
1797 if (cpustate >= state)
1798 cpuhp_issue_call(cpu, state, false, node);
1803 mutex_unlock(&cpuhp_state_mutex);
1808 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1811 * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1812 * @state: The state to remove
1813 * @invoke: If true, the teardown function is invoked for cpus where
1814 * cpu state >= @state
1816 * The caller needs to hold cpus read locked while calling this function.
1817 * The teardown callback is currently not allowed to fail. Think
1818 * about module removal!
1820 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1822 struct cpuhp_step *sp = cpuhp_get_step(state);
1825 BUG_ON(cpuhp_cb_check(state));
1827 lockdep_assert_cpus_held();
1829 mutex_lock(&cpuhp_state_mutex);
1830 if (sp->multi_instance) {
1831 WARN(!hlist_empty(&sp->list),
1832 "Error: Removing state %d which has instances left.\n",
1837 if (!invoke || !cpuhp_get_teardown_cb(state))
1841 * Call the teardown callback for each present cpu depending
1842 * on the hotplug state of the cpu. This function is not
1843 * allowed to fail currently!
1845 for_each_present_cpu(cpu) {
1846 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1847 int cpustate = st->state;
1849 if (cpustate >= state)
1850 cpuhp_issue_call(cpu, state, false, NULL);
1853 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1854 mutex_unlock(&cpuhp_state_mutex);
1856 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1858 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1861 __cpuhp_remove_state_cpuslocked(state, invoke);
1864 EXPORT_SYMBOL(__cpuhp_remove_state);
1866 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1867 static ssize_t show_cpuhp_state(struct device *dev,
1868 struct device_attribute *attr, char *buf)
1870 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1872 return sprintf(buf, "%d\n", st->state);
1874 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1876 static ssize_t write_cpuhp_target(struct device *dev,
1877 struct device_attribute *attr,
1878 const char *buf, size_t count)
1880 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1881 struct cpuhp_step *sp;
1884 ret = kstrtoint(buf, 10, &target);
1888 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1889 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1892 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1896 ret = lock_device_hotplug_sysfs();
1900 mutex_lock(&cpuhp_state_mutex);
1901 sp = cpuhp_get_step(target);
1902 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1903 mutex_unlock(&cpuhp_state_mutex);
1907 if (st->state < target)
1908 ret = do_cpu_up(dev->id, target);
1910 ret = do_cpu_down(dev->id, target);
1912 unlock_device_hotplug();
1913 return ret ? ret : count;
1916 static ssize_t show_cpuhp_target(struct device *dev,
1917 struct device_attribute *attr, char *buf)
1919 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1921 return sprintf(buf, "%d\n", st->target);
1923 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1926 static ssize_t write_cpuhp_fail(struct device *dev,
1927 struct device_attribute *attr,
1928 const char *buf, size_t count)
1930 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1931 struct cpuhp_step *sp;
1934 ret = kstrtoint(buf, 10, &fail);
1939 * Cannot fail STARTING/DYING callbacks.
1941 if (cpuhp_is_atomic_state(fail))
1945 * Cannot fail anything that doesn't have callbacks.
1947 mutex_lock(&cpuhp_state_mutex);
1948 sp = cpuhp_get_step(fail);
1949 if (!sp->startup.single && !sp->teardown.single)
1951 mutex_unlock(&cpuhp_state_mutex);
1960 static ssize_t show_cpuhp_fail(struct device *dev,
1961 struct device_attribute *attr, char *buf)
1963 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1965 return sprintf(buf, "%d\n", st->fail);
1968 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1970 static struct attribute *cpuhp_cpu_attrs[] = {
1971 &dev_attr_state.attr,
1972 &dev_attr_target.attr,
1973 &dev_attr_fail.attr,
1977 static const struct attribute_group cpuhp_cpu_attr_group = {
1978 .attrs = cpuhp_cpu_attrs,
1983 static ssize_t show_cpuhp_states(struct device *dev,
1984 struct device_attribute *attr, char *buf)
1986 ssize_t cur, res = 0;
1989 mutex_lock(&cpuhp_state_mutex);
1990 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1991 struct cpuhp_step *sp = cpuhp_get_step(i);
1994 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1999 mutex_unlock(&cpuhp_state_mutex);
2002 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2004 static struct attribute *cpuhp_cpu_root_attrs[] = {
2005 &dev_attr_states.attr,
2009 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2010 .attrs = cpuhp_cpu_root_attrs,
2015 #ifdef CONFIG_HOTPLUG_SMT
2017 static const char *smt_states[] = {
2018 [CPU_SMT_ENABLED] = "on",
2019 [CPU_SMT_DISABLED] = "off",
2020 [CPU_SMT_FORCE_DISABLED] = "forceoff",
2021 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
2025 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2027 return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2030 static void cpuhp_offline_cpu_device(unsigned int cpu)
2032 struct device *dev = get_cpu_device(cpu);
2034 dev->offline = true;
2035 /* Tell user space about the state change */
2036 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2039 static void cpuhp_online_cpu_device(unsigned int cpu)
2041 struct device *dev = get_cpu_device(cpu);
2043 dev->offline = false;
2044 /* Tell user space about the state change */
2045 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2048 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2052 cpu_maps_update_begin();
2053 for_each_online_cpu(cpu) {
2054 if (topology_is_primary_thread(cpu))
2056 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2060 * As this needs to hold the cpu maps lock it's impossible
2061 * to call device_offline() because that ends up calling
2062 * cpu_down() which takes cpu maps lock. cpu maps lock
2063 * needs to be held as this might race against in kernel
2064 * abusers of the hotplug machinery (thermal management).
2066 * So nothing would update device:offline state. That would
2067 * leave the sysfs entry stale and prevent onlining after
2068 * smt control has been changed to 'off' again. This is
2069 * called under the sysfs hotplug lock, so it is properly
2070 * serialized against the regular offline usage.
2072 cpuhp_offline_cpu_device(cpu);
2075 cpu_smt_control = ctrlval;
2076 cpu_maps_update_done();
2080 static int cpuhp_smt_enable(void)
2084 cpu_maps_update_begin();
2085 cpu_smt_control = CPU_SMT_ENABLED;
2086 for_each_present_cpu(cpu) {
2087 /* Skip online CPUs and CPUs on offline nodes */
2088 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2090 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2093 /* See comment in cpuhp_smt_disable() */
2094 cpuhp_online_cpu_device(cpu);
2096 cpu_maps_update_done();
2101 store_smt_control(struct device *dev, struct device_attribute *attr,
2102 const char *buf, size_t count)
2106 if (sysfs_streq(buf, "on"))
2107 ctrlval = CPU_SMT_ENABLED;
2108 else if (sysfs_streq(buf, "off"))
2109 ctrlval = CPU_SMT_DISABLED;
2110 else if (sysfs_streq(buf, "forceoff"))
2111 ctrlval = CPU_SMT_FORCE_DISABLED;
2115 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2118 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2121 ret = lock_device_hotplug_sysfs();
2125 if (ctrlval != cpu_smt_control) {
2127 case CPU_SMT_ENABLED:
2128 ret = cpuhp_smt_enable();
2130 case CPU_SMT_DISABLED:
2131 case CPU_SMT_FORCE_DISABLED:
2132 ret = cpuhp_smt_disable(ctrlval);
2137 unlock_device_hotplug();
2138 return ret ? ret : count;
2140 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2143 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2145 bool active = topology_max_smt_threads() > 1;
2147 return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2149 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2151 static struct attribute *cpuhp_smt_attrs[] = {
2152 &dev_attr_control.attr,
2153 &dev_attr_active.attr,
2157 static const struct attribute_group cpuhp_smt_attr_group = {
2158 .attrs = cpuhp_smt_attrs,
2163 static int __init cpu_smt_state_init(void)
2165 return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2166 &cpuhp_smt_attr_group);
2170 static inline int cpu_smt_state_init(void) { return 0; }
2173 static int __init cpuhp_sysfs_init(void)
2177 ret = cpu_smt_state_init();
2181 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2182 &cpuhp_cpu_root_attr_group);
2186 for_each_possible_cpu(cpu) {
2187 struct device *dev = get_cpu_device(cpu);
2191 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2197 device_initcall(cpuhp_sysfs_init);
2201 * cpu_bit_bitmap[] is a special, "compressed" data structure that
2202 * represents all NR_CPUS bits binary values of 1<<nr.
2204 * It is used by cpumask_of() to get a constant address to a CPU
2205 * mask value that has a single bit set only.
2208 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2209 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
2210 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2211 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2212 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2214 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2216 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
2217 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
2218 #if BITS_PER_LONG > 32
2219 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
2220 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
2223 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2225 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2226 EXPORT_SYMBOL(cpu_all_bits);
2228 #ifdef CONFIG_INIT_ALL_POSSIBLE
2229 struct cpumask __cpu_possible_mask __read_mostly
2232 struct cpumask __cpu_possible_mask __read_mostly;
2234 EXPORT_SYMBOL(__cpu_possible_mask);
2236 struct cpumask __cpu_online_mask __read_mostly;
2237 EXPORT_SYMBOL(__cpu_online_mask);
2239 struct cpumask __cpu_present_mask __read_mostly;
2240 EXPORT_SYMBOL(__cpu_present_mask);
2242 struct cpumask __cpu_active_mask __read_mostly;
2243 EXPORT_SYMBOL(__cpu_active_mask);
2245 void init_cpu_present(const struct cpumask *src)
2247 cpumask_copy(&__cpu_present_mask, src);
2250 void init_cpu_possible(const struct cpumask *src)
2252 cpumask_copy(&__cpu_possible_mask, src);
2255 void init_cpu_online(const struct cpumask *src)
2257 cpumask_copy(&__cpu_online_mask, src);
2261 * Activate the first processor.
2263 void __init boot_cpu_init(void)
2265 int cpu = smp_processor_id();
2267 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2268 set_cpu_online(cpu, true);
2269 set_cpu_active(cpu, true);
2270 set_cpu_present(cpu, true);
2271 set_cpu_possible(cpu, true);
2274 __boot_cpu_id = cpu;
2279 * Must be called _AFTER_ setting up the per_cpu areas
2281 void __init boot_cpu_hotplug_init(void)
2284 this_cpu_write(cpuhp_state.booted_once, true);
2286 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);