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.h>
11 #include <linux/unistd.h>
12 #include <linux/cpu.h>
13 #include <linux/oom.h>
14 #include <linux/rcupdate.h>
15 #include <linux/export.h>
16 #include <linux/bug.h>
17 #include <linux/kthread.h>
18 #include <linux/stop_machine.h>
19 #include <linux/mutex.h>
20 #include <linux/gfp.h>
21 #include <linux/suspend.h>
22 #include <linux/lockdep.h>
23 #include <linux/tick.h>
24 #include <linux/irq.h>
25 #include <linux/smpboot.h>
26 #include <linux/relay.h>
27 #include <linux/slab.h>
29 #include <trace/events/power.h>
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/cpuhp.h>
36 * cpuhp_cpu_state - Per cpu hotplug state storage
37 * @state: The current cpu state
38 * @target: The target state
39 * @thread: Pointer to the hotplug thread
40 * @should_run: Thread should execute
41 * @rollback: Perform a rollback
42 * @single: Single callback invocation
43 * @bringup: Single callback bringup or teardown selector
44 * @cb_state: The state for a single callback (install/uninstall)
45 * @result: Result of the operation
46 * @done: Signal completion to the issuer of the task
48 struct cpuhp_cpu_state {
49 enum cpuhp_state state;
50 enum cpuhp_state target;
52 struct task_struct *thread;
58 struct hlist_node *node;
59 enum cpuhp_state cb_state;
61 struct completion done;
65 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
67 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
68 static struct lock_class_key cpuhp_state_key;
69 static struct lockdep_map cpuhp_state_lock_map =
70 STATIC_LOCKDEP_MAP_INIT("cpuhp_state", &cpuhp_state_key);
74 * cpuhp_step - Hotplug state machine step
75 * @name: Name of the step
76 * @startup: Startup function of the step
77 * @teardown: Teardown function of the step
78 * @skip_onerr: Do not invoke the functions on error rollback
79 * Will go away once the notifiers are gone
80 * @cant_stop: Bringup/teardown can't be stopped at this step
85 int (*single)(unsigned int cpu);
86 int (*multi)(unsigned int cpu,
87 struct hlist_node *node);
90 int (*single)(unsigned int cpu);
91 int (*multi)(unsigned int cpu,
92 struct hlist_node *node);
94 struct hlist_head list;
100 static DEFINE_MUTEX(cpuhp_state_mutex);
101 static struct cpuhp_step cpuhp_bp_states[];
102 static struct cpuhp_step cpuhp_ap_states[];
104 static bool cpuhp_is_ap_state(enum cpuhp_state state)
107 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
108 * purposes as that state is handled explicitly in cpu_down.
110 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
113 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
115 struct cpuhp_step *sp;
117 sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
122 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
123 * @cpu: The cpu for which the callback should be invoked
124 * @step: The step in the state machine
125 * @bringup: True if the bringup callback should be invoked
127 * Called from cpu hotplug and from the state register machinery.
129 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
130 bool bringup, struct hlist_node *node)
132 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
133 struct cpuhp_step *step = cpuhp_get_step(state);
134 int (*cbm)(unsigned int cpu, struct hlist_node *node);
135 int (*cb)(unsigned int cpu);
138 if (!step->multi_instance) {
139 cb = bringup ? step->startup.single : step->teardown.single;
142 trace_cpuhp_enter(cpu, st->target, state, cb);
144 trace_cpuhp_exit(cpu, st->state, state, ret);
147 cbm = bringup ? step->startup.multi : step->teardown.multi;
151 /* Single invocation for instance add/remove */
153 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
154 ret = cbm(cpu, node);
155 trace_cpuhp_exit(cpu, st->state, state, ret);
159 /* State transition. Invoke on all instances */
161 hlist_for_each(node, &step->list) {
162 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
163 ret = cbm(cpu, node);
164 trace_cpuhp_exit(cpu, st->state, state, ret);
171 /* Rollback the instances if one failed */
172 cbm = !bringup ? step->startup.multi : step->teardown.multi;
176 hlist_for_each(node, &step->list) {
185 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
186 static DEFINE_MUTEX(cpu_add_remove_lock);
187 bool cpuhp_tasks_frozen;
188 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
191 * The following two APIs (cpu_maps_update_begin/done) must be used when
192 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
193 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
194 * hotplug callback (un)registration performed using __register_cpu_notifier()
195 * or __unregister_cpu_notifier().
197 void cpu_maps_update_begin(void)
199 mutex_lock(&cpu_add_remove_lock);
201 EXPORT_SYMBOL(cpu_notifier_register_begin);
203 void cpu_maps_update_done(void)
205 mutex_unlock(&cpu_add_remove_lock);
207 EXPORT_SYMBOL(cpu_notifier_register_done);
209 static RAW_NOTIFIER_HEAD(cpu_chain);
211 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
212 * Should always be manipulated under cpu_add_remove_lock
214 static int cpu_hotplug_disabled;
216 #ifdef CONFIG_HOTPLUG_CPU
219 struct task_struct *active_writer;
220 /* wait queue to wake up the active_writer */
221 wait_queue_head_t wq;
222 /* verifies that no writer will get active while readers are active */
225 * Also blocks the new readers during
226 * an ongoing cpu hotplug operation.
230 #ifdef CONFIG_DEBUG_LOCK_ALLOC
231 struct lockdep_map dep_map;
234 .active_writer = NULL,
235 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
236 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
237 #ifdef CONFIG_DEBUG_LOCK_ALLOC
238 .dep_map = STATIC_LOCKDEP_MAP_INIT("cpu_hotplug.dep_map", &cpu_hotplug.dep_map),
242 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
243 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
244 #define cpuhp_lock_acquire_tryread() \
245 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
246 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
247 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
250 void get_online_cpus(void)
253 if (cpu_hotplug.active_writer == current)
255 cpuhp_lock_acquire_read();
256 mutex_lock(&cpu_hotplug.lock);
257 atomic_inc(&cpu_hotplug.refcount);
258 mutex_unlock(&cpu_hotplug.lock);
260 EXPORT_SYMBOL_GPL(get_online_cpus);
262 void put_online_cpus(void)
266 if (cpu_hotplug.active_writer == current)
269 refcount = atomic_dec_return(&cpu_hotplug.refcount);
270 if (WARN_ON(refcount < 0)) /* try to fix things up */
271 atomic_inc(&cpu_hotplug.refcount);
273 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
274 wake_up(&cpu_hotplug.wq);
276 cpuhp_lock_release();
279 EXPORT_SYMBOL_GPL(put_online_cpus);
282 * This ensures that the hotplug operation can begin only when the
283 * refcount goes to zero.
285 * Note that during a cpu-hotplug operation, the new readers, if any,
286 * will be blocked by the cpu_hotplug.lock
288 * Since cpu_hotplug_begin() is always called after invoking
289 * cpu_maps_update_begin(), we can be sure that only one writer is active.
291 * Note that theoretically, there is a possibility of a livelock:
292 * - Refcount goes to zero, last reader wakes up the sleeping
294 * - Last reader unlocks the cpu_hotplug.lock.
295 * - A new reader arrives at this moment, bumps up the refcount.
296 * - The writer acquires the cpu_hotplug.lock finds the refcount
297 * non zero and goes to sleep again.
299 * However, this is very difficult to achieve in practice since
300 * get_online_cpus() not an api which is called all that often.
303 void cpu_hotplug_begin(void)
307 cpu_hotplug.active_writer = current;
308 cpuhp_lock_acquire();
311 mutex_lock(&cpu_hotplug.lock);
312 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
313 if (likely(!atomic_read(&cpu_hotplug.refcount)))
315 mutex_unlock(&cpu_hotplug.lock);
318 finish_wait(&cpu_hotplug.wq, &wait);
321 void cpu_hotplug_done(void)
323 cpu_hotplug.active_writer = NULL;
324 mutex_unlock(&cpu_hotplug.lock);
325 cpuhp_lock_release();
329 * Wait for currently running CPU hotplug operations to complete (if any) and
330 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
331 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
332 * hotplug path before performing hotplug operations. So acquiring that lock
333 * guarantees mutual exclusion from any currently running hotplug operations.
335 void cpu_hotplug_disable(void)
337 cpu_maps_update_begin();
338 cpu_hotplug_disabled++;
339 cpu_maps_update_done();
341 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
343 static void __cpu_hotplug_enable(void)
345 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
347 cpu_hotplug_disabled--;
350 void cpu_hotplug_enable(void)
352 cpu_maps_update_begin();
353 __cpu_hotplug_enable();
354 cpu_maps_update_done();
356 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
357 #endif /* CONFIG_HOTPLUG_CPU */
359 #ifdef CONFIG_HOTPLUG_SMT
360 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
361 EXPORT_SYMBOL_GPL(cpu_smt_control);
363 static bool cpu_smt_available __read_mostly;
365 void __init cpu_smt_disable(bool force)
367 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
368 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
372 pr_info("SMT: Force disabled\n");
373 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
375 cpu_smt_control = CPU_SMT_DISABLED;
380 * The decision whether SMT is supported can only be done after the full
381 * CPU identification. Called from architecture code before non boot CPUs
384 void __init cpu_smt_check_topology_early(void)
386 if (!topology_smt_supported())
387 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
391 * If SMT was disabled by BIOS, detect it here, after the CPUs have been
392 * brought online. This ensures the smt/l1tf sysfs entries are consistent
393 * with reality. cpu_smt_available is set to true during the bringup of non
394 * boot CPUs when a SMT sibling is detected. Note, this may overwrite
395 * cpu_smt_control's previous setting.
397 void __init cpu_smt_check_topology(void)
399 if (!cpu_smt_available)
400 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
403 static int __init smt_cmdline_disable(char *str)
405 cpu_smt_disable(str && !strcmp(str, "force"));
408 early_param("nosmt", smt_cmdline_disable);
410 static inline bool cpu_smt_allowed(unsigned int cpu)
412 if (topology_is_primary_thread(cpu))
416 * If the CPU is not a 'primary' thread and the booted_once bit is
417 * set then the processor has SMT support. Store this information
418 * for the late check of SMT support in cpu_smt_check_topology().
420 if (per_cpu(cpuhp_state, cpu).booted_once)
421 cpu_smt_available = true;
423 if (cpu_smt_control == CPU_SMT_ENABLED)
427 * On x86 it's required to boot all logical CPUs at least once so
428 * that the init code can get a chance to set CR4.MCE on each
429 * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
430 * core will shutdown the machine.
432 return !per_cpu(cpuhp_state, cpu).booted_once;
435 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
438 /* Need to know about CPUs going up/down? */
439 int register_cpu_notifier(struct notifier_block *nb)
442 cpu_maps_update_begin();
443 ret = raw_notifier_chain_register(&cpu_chain, nb);
444 cpu_maps_update_done();
448 int __register_cpu_notifier(struct notifier_block *nb)
450 return raw_notifier_chain_register(&cpu_chain, nb);
453 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
456 unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
457 void *hcpu = (void *)(long)cpu;
461 ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
464 return notifier_to_errno(ret);
467 static int cpu_notify(unsigned long val, unsigned int cpu)
469 return __cpu_notify(val, cpu, -1, NULL);
472 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
474 BUG_ON(cpu_notify(val, cpu));
477 /* Notifier wrappers for transitioning to state machine */
478 static int notify_prepare(unsigned int cpu)
483 ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
486 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
488 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
493 static int notify_online(unsigned int cpu)
495 cpu_notify(CPU_ONLINE, cpu);
499 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st);
501 static int bringup_wait_for_ap(unsigned int cpu)
503 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
505 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
506 wait_for_completion(&st->done);
507 if (WARN_ON_ONCE((!cpu_online(cpu))))
510 /* Unpark the stopper thread and the hotplug thread of the target cpu */
511 stop_machine_unpark(cpu);
512 kthread_unpark(st->thread);
515 * SMT soft disabling on X86 requires to bring the CPU out of the
516 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
517 * CPU marked itself as booted_once in cpu_notify_starting() so the
518 * cpu_smt_allowed() check will now return false if this is not the
521 if (!cpu_smt_allowed(cpu))
524 /* Should we go further up ? */
525 if (st->target > CPUHP_AP_ONLINE_IDLE) {
526 __cpuhp_kick_ap_work(st);
527 wait_for_completion(&st->done);
532 static int bringup_cpu(unsigned int cpu)
534 struct task_struct *idle = idle_thread_get(cpu);
538 * Some architectures have to walk the irq descriptors to
539 * setup the vector space for the cpu which comes online.
540 * Prevent irq alloc/free across the bringup.
544 /* Arch-specific enabling code. */
545 ret = __cpu_up(cpu, idle);
548 cpu_notify(CPU_UP_CANCELED, cpu);
551 return bringup_wait_for_ap(cpu);
555 * Hotplug state machine related functions
557 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
559 for (st->state++; st->state < st->target; st->state++) {
560 struct cpuhp_step *step = cpuhp_get_step(st->state);
562 if (!step->skip_onerr)
563 cpuhp_invoke_callback(cpu, st->state, true, NULL);
567 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
568 enum cpuhp_state target)
570 enum cpuhp_state prev_state = st->state;
573 for (; st->state > target; st->state--) {
574 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
576 st->target = prev_state;
577 undo_cpu_down(cpu, st);
584 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
586 for (st->state--; st->state > st->target; st->state--) {
587 struct cpuhp_step *step = cpuhp_get_step(st->state);
589 if (!step->skip_onerr)
590 cpuhp_invoke_callback(cpu, st->state, false, NULL);
594 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
596 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
599 * When CPU hotplug is disabled, then taking the CPU down is not
600 * possible because takedown_cpu() and the architecture and
601 * subsystem specific mechanisms are not available. So the CPU
602 * which would be completely unplugged again needs to stay around
603 * in the current state.
605 return st->state <= CPUHP_BRINGUP_CPU;
608 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
609 enum cpuhp_state target)
611 enum cpuhp_state prev_state = st->state;
614 while (st->state < target) {
616 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
618 if (can_rollback_cpu(st)) {
619 st->target = prev_state;
620 undo_cpu_up(cpu, st);
629 * The cpu hotplug threads manage the bringup and teardown of the cpus
631 static void cpuhp_create(unsigned int cpu)
633 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
635 init_completion(&st->done);
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;
645 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
646 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
648 enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
650 return cpuhp_down_callbacks(cpu, st, target);
653 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
654 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
656 return cpuhp_up_callbacks(cpu, st, st->target);
660 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
661 * callbacks when a state gets [un]installed at runtime.
663 static void cpuhp_thread_fun(unsigned int cpu)
665 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
669 * Paired with the mb() in cpuhp_kick_ap_work and
670 * cpuhp_invoke_ap_callback, so the work set is consistent visible.
676 st->should_run = false;
678 lock_map_acquire(&cpuhp_state_lock_map);
679 /* Single callback invocation for [un]install ? */
681 if (st->cb_state < CPUHP_AP_ONLINE) {
683 ret = cpuhp_invoke_callback(cpu, st->cb_state,
684 st->bringup, st->node);
687 ret = cpuhp_invoke_callback(cpu, st->cb_state,
688 st->bringup, st->node);
690 } else if (st->rollback) {
691 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
693 undo_cpu_down(cpu, st);
695 * This is a momentary workaround to keep the notifier users
696 * happy. Will go away once we got rid of the notifiers.
698 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
699 st->rollback = false;
701 /* Cannot happen .... */
702 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
704 /* Regular hotplug work */
705 if (st->state < st->target)
706 ret = cpuhp_ap_online(cpu, st);
707 else if (st->state > st->target)
708 ret = cpuhp_ap_offline(cpu, st);
710 lock_map_release(&cpuhp_state_lock_map);
715 /* Invoke a single callback on a remote cpu */
717 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
718 struct hlist_node *node)
720 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
722 if (!cpu_online(cpu))
725 lock_map_acquire(&cpuhp_state_lock_map);
726 lock_map_release(&cpuhp_state_lock_map);
729 * If we are up and running, use the hotplug thread. For early calls
730 * we invoke the thread function directly.
733 return cpuhp_invoke_callback(cpu, state, bringup, node);
735 st->cb_state = state;
737 st->bringup = bringup;
741 * Make sure the above stores are visible before should_run becomes
742 * true. Paired with the mb() above in cpuhp_thread_fun()
745 st->should_run = true;
746 wake_up_process(st->thread);
747 wait_for_completion(&st->done);
751 /* Regular hotplug invocation of the AP hotplug thread */
752 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
757 * Make sure the above stores are visible before should_run becomes
758 * true. Paired with the mb() above in cpuhp_thread_fun()
761 st->should_run = true;
762 wake_up_process(st->thread);
765 static int cpuhp_kick_ap_work(unsigned int cpu)
767 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
768 enum cpuhp_state state = st->state;
770 trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
771 lock_map_acquire(&cpuhp_state_lock_map);
772 lock_map_release(&cpuhp_state_lock_map);
773 __cpuhp_kick_ap_work(st);
774 wait_for_completion(&st->done);
775 trace_cpuhp_exit(cpu, st->state, state, st->result);
779 static struct smp_hotplug_thread cpuhp_threads = {
780 .store = &cpuhp_state.thread,
781 .create = &cpuhp_create,
782 .thread_should_run = cpuhp_should_run,
783 .thread_fn = cpuhp_thread_fun,
784 .thread_comm = "cpuhp/%u",
788 void __init cpuhp_threads_init(void)
790 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
791 kthread_unpark(this_cpu_read(cpuhp_state.thread));
794 EXPORT_SYMBOL(register_cpu_notifier);
795 EXPORT_SYMBOL(__register_cpu_notifier);
796 void unregister_cpu_notifier(struct notifier_block *nb)
798 cpu_maps_update_begin();
799 raw_notifier_chain_unregister(&cpu_chain, nb);
800 cpu_maps_update_done();
802 EXPORT_SYMBOL(unregister_cpu_notifier);
804 void __unregister_cpu_notifier(struct notifier_block *nb)
806 raw_notifier_chain_unregister(&cpu_chain, nb);
808 EXPORT_SYMBOL(__unregister_cpu_notifier);
810 #ifdef CONFIG_HOTPLUG_CPU
812 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
815 * This function walks all processes, finds a valid mm struct for each one and
816 * then clears a corresponding bit in mm's cpumask. While this all sounds
817 * trivial, there are various non-obvious corner cases, which this function
818 * tries to solve in a safe manner.
820 * Also note that the function uses a somewhat relaxed locking scheme, so it may
821 * be called only for an already offlined CPU.
823 void clear_tasks_mm_cpumask(int cpu)
825 struct task_struct *p;
828 * This function is called after the cpu is taken down and marked
829 * offline, so its not like new tasks will ever get this cpu set in
830 * their mm mask. -- Peter Zijlstra
831 * Thus, we may use rcu_read_lock() here, instead of grabbing
832 * full-fledged tasklist_lock.
834 WARN_ON(cpu_online(cpu));
836 for_each_process(p) {
837 struct task_struct *t;
840 * Main thread might exit, but other threads may still have
841 * a valid mm. Find one.
843 t = find_lock_task_mm(p);
846 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
852 static inline void check_for_tasks(int dead_cpu)
854 struct task_struct *g, *p;
856 read_lock(&tasklist_lock);
857 for_each_process_thread(g, p) {
861 * We do the check with unlocked task_rq(p)->lock.
862 * Order the reading to do not warn about a task,
863 * which was running on this cpu in the past, and
864 * it's just been woken on another cpu.
867 if (task_cpu(p) != dead_cpu)
870 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
871 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
873 read_unlock(&tasklist_lock);
876 static int notify_down_prepare(unsigned int cpu)
878 int err, nr_calls = 0;
880 err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
883 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
884 pr_warn("%s: attempt to take down CPU %u failed\n",
890 /* Take this CPU down. */
891 static int take_cpu_down(void *_param)
893 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
894 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
895 int err, cpu = smp_processor_id();
897 /* Ensure this CPU doesn't handle any more interrupts. */
898 err = __cpu_disable();
903 * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
904 * do this step again.
906 WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
908 /* Invoke the former CPU_DYING callbacks */
909 for (; st->state > target; st->state--)
910 cpuhp_invoke_callback(cpu, st->state, false, NULL);
912 /* Give up timekeeping duties */
913 tick_handover_do_timer();
914 /* Park the stopper thread */
915 stop_machine_park(cpu);
919 static int takedown_cpu(unsigned int cpu)
921 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
924 /* Park the smpboot threads */
925 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
928 * Prevent irq alloc/free while the dying cpu reorganizes the
929 * interrupt affinities.
934 * So now all preempt/rcu users must observe !cpu_active().
936 err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
938 /* CPU refused to die */
940 /* Unpark the hotplug thread so we can rollback there */
941 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
944 BUG_ON(cpu_online(cpu));
947 * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
948 * runnable tasks from the cpu, there's only the idle task left now
949 * that the migration thread is done doing the stop_machine thing.
951 * Wait for the stop thread to go away.
953 wait_for_completion(&st->done);
954 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
956 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
959 hotplug_cpu__broadcast_tick_pull(cpu);
960 /* This actually kills the CPU. */
963 tick_cleanup_dead_cpu(cpu);
967 static int notify_dead(unsigned int cpu)
969 cpu_notify_nofail(CPU_DEAD, cpu);
970 check_for_tasks(cpu);
974 static void cpuhp_complete_idle_dead(void *arg)
976 struct cpuhp_cpu_state *st = arg;
981 void cpuhp_report_idle_dead(void)
983 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
985 BUG_ON(st->state != CPUHP_AP_OFFLINE);
986 rcu_report_dead(smp_processor_id());
987 st->state = CPUHP_AP_IDLE_DEAD;
989 * We cannot call complete after rcu_report_dead() so we delegate it
992 smp_call_function_single(cpumask_first(cpu_online_mask),
993 cpuhp_complete_idle_dead, st, 0);
997 #define notify_down_prepare NULL
998 #define takedown_cpu NULL
999 #define notify_dead NULL
1002 #ifdef CONFIG_HOTPLUG_CPU
1004 /* Requires cpu_add_remove_lock to be held */
1005 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1006 enum cpuhp_state target)
1008 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1009 int prev_state, ret = 0;
1010 bool hasdied = false;
1012 if (num_online_cpus() == 1)
1015 if (!cpu_present(cpu))
1018 cpu_hotplug_begin();
1020 cpuhp_tasks_frozen = tasks_frozen;
1022 prev_state = st->state;
1023 st->target = target;
1025 * If the current CPU state is in the range of the AP hotplug thread,
1026 * then we need to kick the thread.
1028 if (st->state > CPUHP_TEARDOWN_CPU) {
1029 ret = cpuhp_kick_ap_work(cpu);
1031 * The AP side has done the error rollback already. Just
1032 * return the error code..
1038 * We might have stopped still in the range of the AP hotplug
1039 * thread. Nothing to do anymore.
1041 if (st->state > CPUHP_TEARDOWN_CPU)
1045 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1046 * to do the further cleanups.
1048 ret = cpuhp_down_callbacks(cpu, st, target);
1049 if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1050 st->target = prev_state;
1051 st->rollback = true;
1052 cpuhp_kick_ap_work(cpu);
1055 hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
1058 /* This post dead nonsense must die */
1059 if (!ret && hasdied)
1060 cpu_notify_nofail(CPU_POST_DEAD, cpu);
1064 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1066 if (cpu_hotplug_disabled)
1068 return _cpu_down(cpu, 0, target);
1071 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1075 cpu_maps_update_begin();
1076 err = cpu_down_maps_locked(cpu, target);
1077 cpu_maps_update_done();
1080 int cpu_down(unsigned int cpu)
1082 return do_cpu_down(cpu, CPUHP_OFFLINE);
1084 EXPORT_SYMBOL(cpu_down);
1085 #endif /*CONFIG_HOTPLUG_CPU*/
1088 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1089 * @cpu: cpu that just started
1091 * It must be called by the arch code on the new cpu, before the new cpu
1092 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1094 void notify_cpu_starting(unsigned int cpu)
1096 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1097 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1099 rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1100 st->booted_once = true;
1101 while (st->state < target) {
1103 cpuhp_invoke_callback(cpu, st->state, true, NULL);
1108 * Called from the idle task. Wake up the controlling task which brings the
1109 * stopper and the hotplug thread of the upcoming CPU up and then delegates
1110 * the rest of the online bringup to the hotplug thread.
1112 void cpuhp_online_idle(enum cpuhp_state state)
1114 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1116 /* Happens for the boot cpu */
1117 if (state != CPUHP_AP_ONLINE_IDLE)
1120 st->state = CPUHP_AP_ONLINE_IDLE;
1121 complete(&st->done);
1124 /* Requires cpu_add_remove_lock to be held */
1125 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1127 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1128 struct task_struct *idle;
1131 cpu_hotplug_begin();
1133 if (!cpu_present(cpu)) {
1139 * The caller of do_cpu_up might have raced with another
1140 * caller. Ignore it for now.
1142 if (st->state >= target)
1145 if (st->state == CPUHP_OFFLINE) {
1146 /* Let it fail before we try to bring the cpu up */
1147 idle = idle_thread_get(cpu);
1149 ret = PTR_ERR(idle);
1154 cpuhp_tasks_frozen = tasks_frozen;
1156 st->target = target;
1158 * If the current CPU state is in the range of the AP hotplug thread,
1159 * then we need to kick the thread once more.
1161 if (st->state > CPUHP_BRINGUP_CPU) {
1162 ret = cpuhp_kick_ap_work(cpu);
1164 * The AP side has done the error rollback already. Just
1165 * return the error code..
1172 * Try to reach the target state. We max out on the BP at
1173 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1174 * responsible for bringing it up to the target state.
1176 target = min((int)target, CPUHP_BRINGUP_CPU);
1177 ret = cpuhp_up_callbacks(cpu, st, target);
1183 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1187 if (!cpu_possible(cpu)) {
1188 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1190 #if defined(CONFIG_IA64)
1191 pr_err("please check additional_cpus= boot parameter\n");
1196 err = try_online_node(cpu_to_node(cpu));
1200 cpu_maps_update_begin();
1202 if (cpu_hotplug_disabled) {
1206 if (!cpu_smt_allowed(cpu)) {
1211 err = _cpu_up(cpu, 0, target);
1213 cpu_maps_update_done();
1217 int cpu_up(unsigned int cpu)
1219 return do_cpu_up(cpu, CPUHP_ONLINE);
1221 EXPORT_SYMBOL_GPL(cpu_up);
1223 #ifdef CONFIG_PM_SLEEP_SMP
1224 static cpumask_var_t frozen_cpus;
1226 int freeze_secondary_cpus(int primary)
1230 cpu_maps_update_begin();
1231 if (!cpu_online(primary))
1232 primary = cpumask_first(cpu_online_mask);
1234 * We take down all of the non-boot CPUs in one shot to avoid races
1235 * with the userspace trying to use the CPU hotplug at the same time
1237 cpumask_clear(frozen_cpus);
1239 pr_info("Disabling non-boot CPUs ...\n");
1240 for_each_online_cpu(cpu) {
1243 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1244 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1245 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1247 cpumask_set_cpu(cpu, frozen_cpus);
1249 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1255 BUG_ON(num_online_cpus() > 1);
1257 pr_err("Non-boot CPUs are not disabled\n");
1260 * Make sure the CPUs won't be enabled by someone else. We need to do
1261 * this even in case of failure as all disable_nonboot_cpus() users are
1262 * supposed to do enable_nonboot_cpus() on the failure path.
1264 cpu_hotplug_disabled++;
1266 cpu_maps_update_done();
1270 void __weak arch_enable_nonboot_cpus_begin(void)
1274 void __weak arch_enable_nonboot_cpus_end(void)
1278 void enable_nonboot_cpus(void)
1282 /* Allow everyone to use the CPU hotplug again */
1283 cpu_maps_update_begin();
1284 __cpu_hotplug_enable();
1285 if (cpumask_empty(frozen_cpus))
1288 pr_info("Enabling non-boot CPUs ...\n");
1290 arch_enable_nonboot_cpus_begin();
1292 for_each_cpu(cpu, frozen_cpus) {
1293 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1294 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1295 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1297 pr_info("CPU%d is up\n", cpu);
1300 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1303 arch_enable_nonboot_cpus_end();
1305 cpumask_clear(frozen_cpus);
1307 cpu_maps_update_done();
1310 static int __init alloc_frozen_cpus(void)
1312 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1316 core_initcall(alloc_frozen_cpus);
1319 * When callbacks for CPU hotplug notifications are being executed, we must
1320 * ensure that the state of the system with respect to the tasks being frozen
1321 * or not, as reported by the notification, remains unchanged *throughout the
1322 * duration* of the execution of the callbacks.
1323 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1325 * This synchronization is implemented by mutually excluding regular CPU
1326 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1327 * Hibernate notifications.
1330 cpu_hotplug_pm_callback(struct notifier_block *nb,
1331 unsigned long action, void *ptr)
1335 case PM_SUSPEND_PREPARE:
1336 case PM_HIBERNATION_PREPARE:
1337 cpu_hotplug_disable();
1340 case PM_POST_SUSPEND:
1341 case PM_POST_HIBERNATION:
1342 cpu_hotplug_enable();
1353 static int __init cpu_hotplug_pm_sync_init(void)
1356 * cpu_hotplug_pm_callback has higher priority than x86
1357 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1358 * to disable cpu hotplug to avoid cpu hotplug race.
1360 pm_notifier(cpu_hotplug_pm_callback, 0);
1363 core_initcall(cpu_hotplug_pm_sync_init);
1365 #endif /* CONFIG_PM_SLEEP_SMP */
1367 #endif /* CONFIG_SMP */
1369 /* Boot processor state steps */
1370 static struct cpuhp_step cpuhp_bp_states[] = {
1373 .startup.single = NULL,
1374 .teardown.single = NULL,
1377 [CPUHP_CREATE_THREADS]= {
1378 .name = "threads:prepare",
1379 .startup.single = smpboot_create_threads,
1380 .teardown.single = NULL,
1383 [CPUHP_PERF_PREPARE] = {
1384 .name = "perf:prepare",
1385 .startup.single = perf_event_init_cpu,
1386 .teardown.single = perf_event_exit_cpu,
1388 [CPUHP_WORKQUEUE_PREP] = {
1389 .name = "workqueue:prepare",
1390 .startup.single = workqueue_prepare_cpu,
1391 .teardown.single = NULL,
1393 [CPUHP_HRTIMERS_PREPARE] = {
1394 .name = "hrtimers:prepare",
1395 .startup.single = hrtimers_prepare_cpu,
1396 .teardown.single = hrtimers_dead_cpu,
1398 [CPUHP_SMPCFD_PREPARE] = {
1399 .name = "smpcfd:prepare",
1400 .startup.single = smpcfd_prepare_cpu,
1401 .teardown.single = smpcfd_dead_cpu,
1403 [CPUHP_RELAY_PREPARE] = {
1404 .name = "relay:prepare",
1405 .startup.single = relay_prepare_cpu,
1406 .teardown.single = NULL,
1408 [CPUHP_SLAB_PREPARE] = {
1409 .name = "slab:prepare",
1410 .startup.single = slab_prepare_cpu,
1411 .teardown.single = slab_dead_cpu,
1413 [CPUHP_RCUTREE_PREP] = {
1414 .name = "RCU/tree:prepare",
1415 .startup.single = rcutree_prepare_cpu,
1416 .teardown.single = rcutree_dead_cpu,
1419 * Preparatory and dead notifiers. Will be replaced once the notifiers
1420 * are converted to states.
1422 [CPUHP_NOTIFY_PREPARE] = {
1423 .name = "notify:prepare",
1424 .startup.single = notify_prepare,
1425 .teardown.single = notify_dead,
1430 * On the tear-down path, timers_dead_cpu() must be invoked
1431 * before blk_mq_queue_reinit_notify() from notify_dead(),
1432 * otherwise a RCU stall occurs.
1434 [CPUHP_TIMERS_PREPARE] = {
1435 .name = "timers:dead",
1436 .startup.single = timers_prepare_cpu,
1437 .teardown.single = timers_dead_cpu,
1439 /* Kicks the plugged cpu into life */
1440 [CPUHP_BRINGUP_CPU] = {
1441 .name = "cpu:bringup",
1442 .startup.single = bringup_cpu,
1443 .teardown.single = NULL,
1447 * Handled on controll processor until the plugged processor manages
1450 [CPUHP_TEARDOWN_CPU] = {
1451 .name = "cpu:teardown",
1452 .startup.single = NULL,
1453 .teardown.single = takedown_cpu,
1457 [CPUHP_BRINGUP_CPU] = { },
1461 /* Application processor state steps */
1462 static struct cpuhp_step cpuhp_ap_states[] = {
1464 /* Final state before CPU kills itself */
1465 [CPUHP_AP_IDLE_DEAD] = {
1466 .name = "idle:dead",
1469 * Last state before CPU enters the idle loop to die. Transient state
1470 * for synchronization.
1472 [CPUHP_AP_OFFLINE] = {
1473 .name = "ap:offline",
1476 /* First state is scheduler control. Interrupts are disabled */
1477 [CPUHP_AP_SCHED_STARTING] = {
1478 .name = "sched:starting",
1479 .startup.single = sched_cpu_starting,
1480 .teardown.single = sched_cpu_dying,
1482 [CPUHP_AP_RCUTREE_DYING] = {
1483 .name = "RCU/tree:dying",
1484 .startup.single = NULL,
1485 .teardown.single = rcutree_dying_cpu,
1487 [CPUHP_AP_SMPCFD_DYING] = {
1488 .name = "smpcfd:dying",
1489 .startup.single = NULL,
1490 .teardown.single = smpcfd_dying_cpu,
1492 /* Entry state on starting. Interrupts enabled from here on. Transient
1493 * state for synchronsization */
1494 [CPUHP_AP_ONLINE] = {
1495 .name = "ap:online",
1497 /* Handle smpboot threads park/unpark */
1498 [CPUHP_AP_SMPBOOT_THREADS] = {
1499 .name = "smpboot/threads:online",
1500 .startup.single = smpboot_unpark_threads,
1501 .teardown.single = smpboot_park_threads,
1503 [CPUHP_AP_PERF_ONLINE] = {
1504 .name = "perf:online",
1505 .startup.single = perf_event_init_cpu,
1506 .teardown.single = perf_event_exit_cpu,
1508 [CPUHP_AP_WORKQUEUE_ONLINE] = {
1509 .name = "workqueue:online",
1510 .startup.single = workqueue_online_cpu,
1511 .teardown.single = workqueue_offline_cpu,
1513 [CPUHP_AP_RCUTREE_ONLINE] = {
1514 .name = "RCU/tree:online",
1515 .startup.single = rcutree_online_cpu,
1516 .teardown.single = rcutree_offline_cpu,
1520 * Online/down_prepare notifiers. Will be removed once the notifiers
1521 * are converted to states.
1523 [CPUHP_AP_NOTIFY_ONLINE] = {
1524 .name = "notify:online",
1525 .startup.single = notify_online,
1526 .teardown.single = notify_down_prepare,
1531 * The dynamically registered state space is here
1535 /* Last state is scheduler control setting the cpu active */
1536 [CPUHP_AP_ACTIVE] = {
1537 .name = "sched:active",
1538 .startup.single = sched_cpu_activate,
1539 .teardown.single = sched_cpu_deactivate,
1543 /* CPU is fully up and running. */
1546 .startup.single = NULL,
1547 .teardown.single = NULL,
1551 /* Sanity check for callbacks */
1552 static int cpuhp_cb_check(enum cpuhp_state state)
1554 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1559 static void cpuhp_store_callbacks(enum cpuhp_state state,
1561 int (*startup)(unsigned int cpu),
1562 int (*teardown)(unsigned int cpu),
1563 bool multi_instance)
1565 /* (Un)Install the callbacks for further cpu hotplug operations */
1566 struct cpuhp_step *sp;
1568 sp = cpuhp_get_step(state);
1569 sp->startup.single = startup;
1570 sp->teardown.single = teardown;
1572 sp->multi_instance = multi_instance;
1573 INIT_HLIST_HEAD(&sp->list);
1576 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1578 return cpuhp_get_step(state)->teardown.single;
1582 * Call the startup/teardown function for a step either on the AP or
1583 * on the current CPU.
1585 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1586 struct hlist_node *node)
1588 struct cpuhp_step *sp = cpuhp_get_step(state);
1591 if ((bringup && !sp->startup.single) ||
1592 (!bringup && !sp->teardown.single))
1595 * The non AP bound callbacks can fail on bringup. On teardown
1596 * e.g. module removal we crash for now.
1599 if (cpuhp_is_ap_state(state))
1600 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1602 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1604 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1606 BUG_ON(ret && !bringup);
1611 * Called from __cpuhp_setup_state on a recoverable failure.
1613 * Note: The teardown callbacks for rollback are not allowed to fail!
1615 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1616 struct hlist_node *node)
1620 /* Roll back the already executed steps on the other cpus */
1621 for_each_present_cpu(cpu) {
1622 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1623 int cpustate = st->state;
1625 if (cpu >= failedcpu)
1628 /* Did we invoke the startup call on that cpu ? */
1629 if (cpustate >= state)
1630 cpuhp_issue_call(cpu, state, false, node);
1635 * Returns a free for dynamic slot assignment of the Online state. The states
1636 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1637 * by having no name assigned.
1639 static int cpuhp_reserve_state(enum cpuhp_state state)
1643 for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1644 if (cpuhp_ap_states[i].name)
1647 cpuhp_ap_states[i].name = "Reserved";
1650 WARN(1, "No more dynamic states available for CPU hotplug\n");
1654 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1657 struct cpuhp_step *sp;
1661 sp = cpuhp_get_step(state);
1662 if (sp->multi_instance == false)
1666 mutex_lock(&cpuhp_state_mutex);
1668 if (!invoke || !sp->startup.multi)
1672 * Try to call the startup callback for each present cpu
1673 * depending on the hotplug state of the cpu.
1675 for_each_present_cpu(cpu) {
1676 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1677 int cpustate = st->state;
1679 if (cpustate < state)
1682 ret = cpuhp_issue_call(cpu, state, true, node);
1684 if (sp->teardown.multi)
1685 cpuhp_rollback_install(cpu, state, node);
1691 hlist_add_head(node, &sp->list);
1694 mutex_unlock(&cpuhp_state_mutex);
1698 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1701 * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1702 * @state: The state to setup
1703 * @invoke: If true, the startup function is invoked for cpus where
1704 * cpu state >= @state
1705 * @startup: startup callback function
1706 * @teardown: teardown callback function
1708 * Returns 0 if successful, otherwise a proper error code
1710 int __cpuhp_setup_state(enum cpuhp_state state,
1711 const char *name, bool invoke,
1712 int (*startup)(unsigned int cpu),
1713 int (*teardown)(unsigned int cpu),
1714 bool multi_instance)
1719 if (cpuhp_cb_check(state) || !name)
1723 mutex_lock(&cpuhp_state_mutex);
1725 /* currently assignments for the ONLINE state are possible */
1726 if (state == CPUHP_AP_ONLINE_DYN) {
1728 ret = cpuhp_reserve_state(state);
1734 cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1736 if (!invoke || !startup)
1740 * Try to call the startup callback for each present cpu
1741 * depending on the hotplug state of the cpu.
1743 for_each_present_cpu(cpu) {
1744 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1745 int cpustate = st->state;
1747 if (cpustate < state)
1750 ret = cpuhp_issue_call(cpu, state, true, NULL);
1753 cpuhp_rollback_install(cpu, state, NULL);
1754 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1759 mutex_unlock(&cpuhp_state_mutex);
1762 if (!ret && dyn_state)
1766 EXPORT_SYMBOL(__cpuhp_setup_state);
1768 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1769 struct hlist_node *node, bool invoke)
1771 struct cpuhp_step *sp = cpuhp_get_step(state);
1774 BUG_ON(cpuhp_cb_check(state));
1776 if (!sp->multi_instance)
1780 mutex_lock(&cpuhp_state_mutex);
1782 if (!invoke || !cpuhp_get_teardown_cb(state))
1785 * Call the teardown callback for each present cpu depending
1786 * on the hotplug state of the cpu. This function is not
1787 * allowed to fail currently!
1789 for_each_present_cpu(cpu) {
1790 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1791 int cpustate = st->state;
1793 if (cpustate >= state)
1794 cpuhp_issue_call(cpu, state, false, node);
1799 mutex_unlock(&cpuhp_state_mutex);
1804 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1806 * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1807 * @state: The state to remove
1808 * @invoke: If true, the teardown function is invoked for cpus where
1809 * cpu state >= @state
1811 * The teardown callback is currently not allowed to fail. Think
1812 * about module removal!
1814 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1816 struct cpuhp_step *sp = cpuhp_get_step(state);
1819 BUG_ON(cpuhp_cb_check(state));
1822 mutex_lock(&cpuhp_state_mutex);
1824 if (sp->multi_instance) {
1825 WARN(!hlist_empty(&sp->list),
1826 "Error: Removing state %d which has instances left.\n",
1831 if (!invoke || !cpuhp_get_teardown_cb(state))
1835 * Call the teardown callback for each present cpu depending
1836 * on the hotplug state of the cpu. This function is not
1837 * allowed to fail currently!
1839 for_each_present_cpu(cpu) {
1840 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1841 int cpustate = st->state;
1843 if (cpustate >= state)
1844 cpuhp_issue_call(cpu, state, false, NULL);
1847 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1848 mutex_unlock(&cpuhp_state_mutex);
1851 EXPORT_SYMBOL(__cpuhp_remove_state);
1853 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1854 static ssize_t show_cpuhp_state(struct device *dev,
1855 struct device_attribute *attr, char *buf)
1857 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1859 return sprintf(buf, "%d\n", st->state);
1861 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1863 static ssize_t write_cpuhp_target(struct device *dev,
1864 struct device_attribute *attr,
1865 const char *buf, size_t count)
1867 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1868 struct cpuhp_step *sp;
1871 ret = kstrtoint(buf, 10, &target);
1875 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1876 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1879 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1883 ret = lock_device_hotplug_sysfs();
1887 mutex_lock(&cpuhp_state_mutex);
1888 sp = cpuhp_get_step(target);
1889 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1890 mutex_unlock(&cpuhp_state_mutex);
1894 if (st->state < target)
1895 ret = do_cpu_up(dev->id, target);
1897 ret = do_cpu_down(dev->id, target);
1899 unlock_device_hotplug();
1900 return ret ? ret : count;
1903 static ssize_t show_cpuhp_target(struct device *dev,
1904 struct device_attribute *attr, char *buf)
1906 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1908 return sprintf(buf, "%d\n", st->target);
1910 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1912 static struct attribute *cpuhp_cpu_attrs[] = {
1913 &dev_attr_state.attr,
1914 &dev_attr_target.attr,
1918 static struct attribute_group cpuhp_cpu_attr_group = {
1919 .attrs = cpuhp_cpu_attrs,
1924 static ssize_t show_cpuhp_states(struct device *dev,
1925 struct device_attribute *attr, char *buf)
1927 ssize_t cur, res = 0;
1930 mutex_lock(&cpuhp_state_mutex);
1931 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1932 struct cpuhp_step *sp = cpuhp_get_step(i);
1935 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1940 mutex_unlock(&cpuhp_state_mutex);
1943 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1945 static struct attribute *cpuhp_cpu_root_attrs[] = {
1946 &dev_attr_states.attr,
1950 static struct attribute_group cpuhp_cpu_root_attr_group = {
1951 .attrs = cpuhp_cpu_root_attrs,
1956 #ifdef CONFIG_HOTPLUG_SMT
1958 static const char *smt_states[] = {
1959 [CPU_SMT_ENABLED] = "on",
1960 [CPU_SMT_DISABLED] = "off",
1961 [CPU_SMT_FORCE_DISABLED] = "forceoff",
1962 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
1966 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
1968 return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
1971 static void cpuhp_offline_cpu_device(unsigned int cpu)
1973 struct device *dev = get_cpu_device(cpu);
1975 dev->offline = true;
1976 /* Tell user space about the state change */
1977 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
1980 static void cpuhp_online_cpu_device(unsigned int cpu)
1982 struct device *dev = get_cpu_device(cpu);
1984 dev->offline = false;
1985 /* Tell user space about the state change */
1986 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
1989 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
1993 cpu_maps_update_begin();
1994 for_each_online_cpu(cpu) {
1995 if (topology_is_primary_thread(cpu))
1997 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2001 * As this needs to hold the cpu maps lock it's impossible
2002 * to call device_offline() because that ends up calling
2003 * cpu_down() which takes cpu maps lock. cpu maps lock
2004 * needs to be held as this might race against in kernel
2005 * abusers of the hotplug machinery (thermal management).
2007 * So nothing would update device:offline state. That would
2008 * leave the sysfs entry stale and prevent onlining after
2009 * smt control has been changed to 'off' again. This is
2010 * called under the sysfs hotplug lock, so it is properly
2011 * serialized against the regular offline usage.
2013 cpuhp_offline_cpu_device(cpu);
2016 cpu_smt_control = ctrlval;
2017 cpu_maps_update_done();
2021 static int cpuhp_smt_enable(void)
2025 cpu_maps_update_begin();
2026 cpu_smt_control = CPU_SMT_ENABLED;
2027 for_each_present_cpu(cpu) {
2028 /* Skip online CPUs and CPUs on offline nodes */
2029 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2031 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2034 /* See comment in cpuhp_smt_disable() */
2035 cpuhp_online_cpu_device(cpu);
2037 cpu_maps_update_done();
2042 store_smt_control(struct device *dev, struct device_attribute *attr,
2043 const char *buf, size_t count)
2047 if (sysfs_streq(buf, "on"))
2048 ctrlval = CPU_SMT_ENABLED;
2049 else if (sysfs_streq(buf, "off"))
2050 ctrlval = CPU_SMT_DISABLED;
2051 else if (sysfs_streq(buf, "forceoff"))
2052 ctrlval = CPU_SMT_FORCE_DISABLED;
2056 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2059 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2062 ret = lock_device_hotplug_sysfs();
2066 if (ctrlval != cpu_smt_control) {
2068 case CPU_SMT_ENABLED:
2069 ret = cpuhp_smt_enable();
2071 case CPU_SMT_DISABLED:
2072 case CPU_SMT_FORCE_DISABLED:
2073 ret = cpuhp_smt_disable(ctrlval);
2078 unlock_device_hotplug();
2079 return ret ? ret : count;
2081 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2084 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2086 bool active = topology_max_smt_threads() > 1;
2088 return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2090 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2092 static struct attribute *cpuhp_smt_attrs[] = {
2093 &dev_attr_control.attr,
2094 &dev_attr_active.attr,
2098 static const struct attribute_group cpuhp_smt_attr_group = {
2099 .attrs = cpuhp_smt_attrs,
2104 static int __init cpu_smt_state_init(void)
2106 return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2107 &cpuhp_smt_attr_group);
2111 static inline int cpu_smt_state_init(void) { return 0; }
2114 static int __init cpuhp_sysfs_init(void)
2118 ret = cpu_smt_state_init();
2122 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2123 &cpuhp_cpu_root_attr_group);
2127 for_each_possible_cpu(cpu) {
2128 struct device *dev = get_cpu_device(cpu);
2132 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2138 device_initcall(cpuhp_sysfs_init);
2142 * cpu_bit_bitmap[] is a special, "compressed" data structure that
2143 * represents all NR_CPUS bits binary values of 1<<nr.
2145 * It is used by cpumask_of() to get a constant address to a CPU
2146 * mask value that has a single bit set only.
2149 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2150 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
2151 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2152 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2153 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2155 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2157 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
2158 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
2159 #if BITS_PER_LONG > 32
2160 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
2161 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
2164 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2166 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2167 EXPORT_SYMBOL(cpu_all_bits);
2169 #ifdef CONFIG_INIT_ALL_POSSIBLE
2170 struct cpumask __cpu_possible_mask __read_mostly
2173 struct cpumask __cpu_possible_mask __read_mostly;
2175 EXPORT_SYMBOL(__cpu_possible_mask);
2177 struct cpumask __cpu_online_mask __read_mostly;
2178 EXPORT_SYMBOL(__cpu_online_mask);
2180 struct cpumask __cpu_present_mask __read_mostly;
2181 EXPORT_SYMBOL(__cpu_present_mask);
2183 struct cpumask __cpu_active_mask __read_mostly;
2184 EXPORT_SYMBOL(__cpu_active_mask);
2186 void init_cpu_present(const struct cpumask *src)
2188 cpumask_copy(&__cpu_present_mask, src);
2191 void init_cpu_possible(const struct cpumask *src)
2193 cpumask_copy(&__cpu_possible_mask, src);
2196 void init_cpu_online(const struct cpumask *src)
2198 cpumask_copy(&__cpu_online_mask, src);
2202 * Activate the first processor.
2204 void __init boot_cpu_init(void)
2206 int cpu = smp_processor_id();
2208 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2209 set_cpu_online(cpu, true);
2210 set_cpu_active(cpu, true);
2211 set_cpu_present(cpu, true);
2212 set_cpu_possible(cpu, true);
2216 * Must be called _AFTER_ setting up the per_cpu areas
2218 void __init boot_cpu_hotplug_init(void)
2221 this_cpu_write(cpuhp_state.booted_once, true);
2223 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);