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 int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
595 enum cpuhp_state target)
597 enum cpuhp_state prev_state = st->state;
600 while (st->state < target) {
602 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
604 st->target = prev_state;
605 undo_cpu_up(cpu, st);
613 * The cpu hotplug threads manage the bringup and teardown of the cpus
615 static void cpuhp_create(unsigned int cpu)
617 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
619 init_completion(&st->done);
622 static int cpuhp_should_run(unsigned int cpu)
624 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
626 return st->should_run;
629 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
630 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
632 enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
634 return cpuhp_down_callbacks(cpu, st, target);
637 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
638 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
640 return cpuhp_up_callbacks(cpu, st, st->target);
644 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
645 * callbacks when a state gets [un]installed at runtime.
647 static void cpuhp_thread_fun(unsigned int cpu)
649 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
653 * Paired with the mb() in cpuhp_kick_ap_work and
654 * cpuhp_invoke_ap_callback, so the work set is consistent visible.
660 st->should_run = false;
662 lock_map_acquire(&cpuhp_state_lock_map);
663 /* Single callback invocation for [un]install ? */
665 if (st->cb_state < CPUHP_AP_ONLINE) {
667 ret = cpuhp_invoke_callback(cpu, st->cb_state,
668 st->bringup, st->node);
671 ret = cpuhp_invoke_callback(cpu, st->cb_state,
672 st->bringup, st->node);
674 } else if (st->rollback) {
675 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
677 undo_cpu_down(cpu, st);
679 * This is a momentary workaround to keep the notifier users
680 * happy. Will go away once we got rid of the notifiers.
682 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
683 st->rollback = false;
685 /* Cannot happen .... */
686 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
688 /* Regular hotplug work */
689 if (st->state < st->target)
690 ret = cpuhp_ap_online(cpu, st);
691 else if (st->state > st->target)
692 ret = cpuhp_ap_offline(cpu, st);
694 lock_map_release(&cpuhp_state_lock_map);
699 /* Invoke a single callback on a remote cpu */
701 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
702 struct hlist_node *node)
704 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
706 if (!cpu_online(cpu))
709 lock_map_acquire(&cpuhp_state_lock_map);
710 lock_map_release(&cpuhp_state_lock_map);
713 * If we are up and running, use the hotplug thread. For early calls
714 * we invoke the thread function directly.
717 return cpuhp_invoke_callback(cpu, state, bringup, node);
719 st->cb_state = state;
721 st->bringup = bringup;
725 * Make sure the above stores are visible before should_run becomes
726 * true. Paired with the mb() above in cpuhp_thread_fun()
729 st->should_run = true;
730 wake_up_process(st->thread);
731 wait_for_completion(&st->done);
735 /* Regular hotplug invocation of the AP hotplug thread */
736 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
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);
749 static int cpuhp_kick_ap_work(unsigned int cpu)
751 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
752 enum cpuhp_state state = st->state;
754 trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
755 lock_map_acquire(&cpuhp_state_lock_map);
756 lock_map_release(&cpuhp_state_lock_map);
757 __cpuhp_kick_ap_work(st);
758 wait_for_completion(&st->done);
759 trace_cpuhp_exit(cpu, st->state, state, st->result);
763 static struct smp_hotplug_thread cpuhp_threads = {
764 .store = &cpuhp_state.thread,
765 .create = &cpuhp_create,
766 .thread_should_run = cpuhp_should_run,
767 .thread_fn = cpuhp_thread_fun,
768 .thread_comm = "cpuhp/%u",
772 void __init cpuhp_threads_init(void)
774 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
775 kthread_unpark(this_cpu_read(cpuhp_state.thread));
778 EXPORT_SYMBOL(register_cpu_notifier);
779 EXPORT_SYMBOL(__register_cpu_notifier);
780 void unregister_cpu_notifier(struct notifier_block *nb)
782 cpu_maps_update_begin();
783 raw_notifier_chain_unregister(&cpu_chain, nb);
784 cpu_maps_update_done();
786 EXPORT_SYMBOL(unregister_cpu_notifier);
788 void __unregister_cpu_notifier(struct notifier_block *nb)
790 raw_notifier_chain_unregister(&cpu_chain, nb);
792 EXPORT_SYMBOL(__unregister_cpu_notifier);
794 #ifdef CONFIG_HOTPLUG_CPU
796 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
799 * This function walks all processes, finds a valid mm struct for each one and
800 * then clears a corresponding bit in mm's cpumask. While this all sounds
801 * trivial, there are various non-obvious corner cases, which this function
802 * tries to solve in a safe manner.
804 * Also note that the function uses a somewhat relaxed locking scheme, so it may
805 * be called only for an already offlined CPU.
807 void clear_tasks_mm_cpumask(int cpu)
809 struct task_struct *p;
812 * This function is called after the cpu is taken down and marked
813 * offline, so its not like new tasks will ever get this cpu set in
814 * their mm mask. -- Peter Zijlstra
815 * Thus, we may use rcu_read_lock() here, instead of grabbing
816 * full-fledged tasklist_lock.
818 WARN_ON(cpu_online(cpu));
820 for_each_process(p) {
821 struct task_struct *t;
824 * Main thread might exit, but other threads may still have
825 * a valid mm. Find one.
827 t = find_lock_task_mm(p);
830 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
836 static inline void check_for_tasks(int dead_cpu)
838 struct task_struct *g, *p;
840 read_lock(&tasklist_lock);
841 for_each_process_thread(g, p) {
845 * We do the check with unlocked task_rq(p)->lock.
846 * Order the reading to do not warn about a task,
847 * which was running on this cpu in the past, and
848 * it's just been woken on another cpu.
851 if (task_cpu(p) != dead_cpu)
854 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
855 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
857 read_unlock(&tasklist_lock);
860 static int notify_down_prepare(unsigned int cpu)
862 int err, nr_calls = 0;
864 err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
867 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
868 pr_warn("%s: attempt to take down CPU %u failed\n",
874 /* Take this CPU down. */
875 static int take_cpu_down(void *_param)
877 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
878 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
879 int err, cpu = smp_processor_id();
881 /* Ensure this CPU doesn't handle any more interrupts. */
882 err = __cpu_disable();
887 * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
888 * do this step again.
890 WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
892 /* Invoke the former CPU_DYING callbacks */
893 for (; st->state > target; st->state--)
894 cpuhp_invoke_callback(cpu, st->state, false, NULL);
896 /* Give up timekeeping duties */
897 tick_handover_do_timer();
898 /* Park the stopper thread */
899 stop_machine_park(cpu);
903 static int takedown_cpu(unsigned int cpu)
905 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
908 /* Park the smpboot threads */
909 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
912 * Prevent irq alloc/free while the dying cpu reorganizes the
913 * interrupt affinities.
918 * So now all preempt/rcu users must observe !cpu_active().
920 err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
922 /* CPU refused to die */
924 /* Unpark the hotplug thread so we can rollback there */
925 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
928 BUG_ON(cpu_online(cpu));
931 * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
932 * runnable tasks from the cpu, there's only the idle task left now
933 * that the migration thread is done doing the stop_machine thing.
935 * Wait for the stop thread to go away.
937 wait_for_completion(&st->done);
938 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
940 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
943 hotplug_cpu__broadcast_tick_pull(cpu);
944 /* This actually kills the CPU. */
947 tick_cleanup_dead_cpu(cpu);
951 static int notify_dead(unsigned int cpu)
953 cpu_notify_nofail(CPU_DEAD, cpu);
954 check_for_tasks(cpu);
958 static void cpuhp_complete_idle_dead(void *arg)
960 struct cpuhp_cpu_state *st = arg;
965 void cpuhp_report_idle_dead(void)
967 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
969 BUG_ON(st->state != CPUHP_AP_OFFLINE);
970 rcu_report_dead(smp_processor_id());
971 st->state = CPUHP_AP_IDLE_DEAD;
973 * We cannot call complete after rcu_report_dead() so we delegate it
976 smp_call_function_single(cpumask_first(cpu_online_mask),
977 cpuhp_complete_idle_dead, st, 0);
981 #define notify_down_prepare NULL
982 #define takedown_cpu NULL
983 #define notify_dead NULL
986 #ifdef CONFIG_HOTPLUG_CPU
988 /* Requires cpu_add_remove_lock to be held */
989 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
990 enum cpuhp_state target)
992 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
993 int prev_state, ret = 0;
994 bool hasdied = false;
996 if (num_online_cpus() == 1)
999 if (!cpu_present(cpu))
1002 cpu_hotplug_begin();
1004 cpuhp_tasks_frozen = tasks_frozen;
1006 prev_state = st->state;
1007 st->target = target;
1009 * If the current CPU state is in the range of the AP hotplug thread,
1010 * then we need to kick the thread.
1012 if (st->state > CPUHP_TEARDOWN_CPU) {
1013 ret = cpuhp_kick_ap_work(cpu);
1015 * The AP side has done the error rollback already. Just
1016 * return the error code..
1022 * We might have stopped still in the range of the AP hotplug
1023 * thread. Nothing to do anymore.
1025 if (st->state > CPUHP_TEARDOWN_CPU)
1029 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1030 * to do the further cleanups.
1032 ret = cpuhp_down_callbacks(cpu, st, target);
1033 if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1034 st->target = prev_state;
1035 st->rollback = true;
1036 cpuhp_kick_ap_work(cpu);
1039 hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
1042 /* This post dead nonsense must die */
1043 if (!ret && hasdied)
1044 cpu_notify_nofail(CPU_POST_DEAD, cpu);
1048 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1050 if (cpu_hotplug_disabled)
1052 return _cpu_down(cpu, 0, target);
1055 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1059 cpu_maps_update_begin();
1060 err = cpu_down_maps_locked(cpu, target);
1061 cpu_maps_update_done();
1064 int cpu_down(unsigned int cpu)
1066 return do_cpu_down(cpu, CPUHP_OFFLINE);
1068 EXPORT_SYMBOL(cpu_down);
1069 #endif /*CONFIG_HOTPLUG_CPU*/
1072 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1073 * @cpu: cpu that just started
1075 * It must be called by the arch code on the new cpu, before the new cpu
1076 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1078 void notify_cpu_starting(unsigned int cpu)
1080 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1081 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1083 rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1084 st->booted_once = true;
1085 while (st->state < target) {
1087 cpuhp_invoke_callback(cpu, st->state, true, NULL);
1092 * Called from the idle task. Wake up the controlling task which brings the
1093 * stopper and the hotplug thread of the upcoming CPU up and then delegates
1094 * the rest of the online bringup to the hotplug thread.
1096 void cpuhp_online_idle(enum cpuhp_state state)
1098 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1100 /* Happens for the boot cpu */
1101 if (state != CPUHP_AP_ONLINE_IDLE)
1104 st->state = CPUHP_AP_ONLINE_IDLE;
1105 complete(&st->done);
1108 /* Requires cpu_add_remove_lock to be held */
1109 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1111 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1112 struct task_struct *idle;
1115 cpu_hotplug_begin();
1117 if (!cpu_present(cpu)) {
1123 * The caller of do_cpu_up might have raced with another
1124 * caller. Ignore it for now.
1126 if (st->state >= target)
1129 if (st->state == CPUHP_OFFLINE) {
1130 /* Let it fail before we try to bring the cpu up */
1131 idle = idle_thread_get(cpu);
1133 ret = PTR_ERR(idle);
1138 cpuhp_tasks_frozen = tasks_frozen;
1140 st->target = target;
1142 * If the current CPU state is in the range of the AP hotplug thread,
1143 * then we need to kick the thread once more.
1145 if (st->state > CPUHP_BRINGUP_CPU) {
1146 ret = cpuhp_kick_ap_work(cpu);
1148 * The AP side has done the error rollback already. Just
1149 * return the error code..
1156 * Try to reach the target state. We max out on the BP at
1157 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1158 * responsible for bringing it up to the target state.
1160 target = min((int)target, CPUHP_BRINGUP_CPU);
1161 ret = cpuhp_up_callbacks(cpu, st, target);
1167 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1171 if (!cpu_possible(cpu)) {
1172 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1174 #if defined(CONFIG_IA64)
1175 pr_err("please check additional_cpus= boot parameter\n");
1180 err = try_online_node(cpu_to_node(cpu));
1184 cpu_maps_update_begin();
1186 if (cpu_hotplug_disabled) {
1190 if (!cpu_smt_allowed(cpu)) {
1195 err = _cpu_up(cpu, 0, target);
1197 cpu_maps_update_done();
1201 int cpu_up(unsigned int cpu)
1203 return do_cpu_up(cpu, CPUHP_ONLINE);
1205 EXPORT_SYMBOL_GPL(cpu_up);
1207 #ifdef CONFIG_PM_SLEEP_SMP
1208 static cpumask_var_t frozen_cpus;
1210 int freeze_secondary_cpus(int primary)
1214 cpu_maps_update_begin();
1215 if (!cpu_online(primary))
1216 primary = cpumask_first(cpu_online_mask);
1218 * We take down all of the non-boot CPUs in one shot to avoid races
1219 * with the userspace trying to use the CPU hotplug at the same time
1221 cpumask_clear(frozen_cpus);
1223 pr_info("Disabling non-boot CPUs ...\n");
1224 for_each_online_cpu(cpu) {
1227 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1228 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1229 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1231 cpumask_set_cpu(cpu, frozen_cpus);
1233 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1239 BUG_ON(num_online_cpus() > 1);
1241 pr_err("Non-boot CPUs are not disabled\n");
1244 * Make sure the CPUs won't be enabled by someone else. We need to do
1245 * this even in case of failure as all disable_nonboot_cpus() users are
1246 * supposed to do enable_nonboot_cpus() on the failure path.
1248 cpu_hotplug_disabled++;
1250 cpu_maps_update_done();
1254 void __weak arch_enable_nonboot_cpus_begin(void)
1258 void __weak arch_enable_nonboot_cpus_end(void)
1262 void enable_nonboot_cpus(void)
1266 /* Allow everyone to use the CPU hotplug again */
1267 cpu_maps_update_begin();
1268 __cpu_hotplug_enable();
1269 if (cpumask_empty(frozen_cpus))
1272 pr_info("Enabling non-boot CPUs ...\n");
1274 arch_enable_nonboot_cpus_begin();
1276 for_each_cpu(cpu, frozen_cpus) {
1277 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1278 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1279 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1281 pr_info("CPU%d is up\n", cpu);
1284 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1287 arch_enable_nonboot_cpus_end();
1289 cpumask_clear(frozen_cpus);
1291 cpu_maps_update_done();
1294 static int __init alloc_frozen_cpus(void)
1296 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1300 core_initcall(alloc_frozen_cpus);
1303 * When callbacks for CPU hotplug notifications are being executed, we must
1304 * ensure that the state of the system with respect to the tasks being frozen
1305 * or not, as reported by the notification, remains unchanged *throughout the
1306 * duration* of the execution of the callbacks.
1307 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1309 * This synchronization is implemented by mutually excluding regular CPU
1310 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1311 * Hibernate notifications.
1314 cpu_hotplug_pm_callback(struct notifier_block *nb,
1315 unsigned long action, void *ptr)
1319 case PM_SUSPEND_PREPARE:
1320 case PM_HIBERNATION_PREPARE:
1321 cpu_hotplug_disable();
1324 case PM_POST_SUSPEND:
1325 case PM_POST_HIBERNATION:
1326 cpu_hotplug_enable();
1337 static int __init cpu_hotplug_pm_sync_init(void)
1340 * cpu_hotplug_pm_callback has higher priority than x86
1341 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1342 * to disable cpu hotplug to avoid cpu hotplug race.
1344 pm_notifier(cpu_hotplug_pm_callback, 0);
1347 core_initcall(cpu_hotplug_pm_sync_init);
1349 #endif /* CONFIG_PM_SLEEP_SMP */
1351 #endif /* CONFIG_SMP */
1353 /* Boot processor state steps */
1354 static struct cpuhp_step cpuhp_bp_states[] = {
1357 .startup.single = NULL,
1358 .teardown.single = NULL,
1361 [CPUHP_CREATE_THREADS]= {
1362 .name = "threads:prepare",
1363 .startup.single = smpboot_create_threads,
1364 .teardown.single = NULL,
1367 [CPUHP_PERF_PREPARE] = {
1368 .name = "perf:prepare",
1369 .startup.single = perf_event_init_cpu,
1370 .teardown.single = perf_event_exit_cpu,
1372 [CPUHP_WORKQUEUE_PREP] = {
1373 .name = "workqueue:prepare",
1374 .startup.single = workqueue_prepare_cpu,
1375 .teardown.single = NULL,
1377 [CPUHP_HRTIMERS_PREPARE] = {
1378 .name = "hrtimers:prepare",
1379 .startup.single = hrtimers_prepare_cpu,
1380 .teardown.single = hrtimers_dead_cpu,
1382 [CPUHP_SMPCFD_PREPARE] = {
1383 .name = "smpcfd:prepare",
1384 .startup.single = smpcfd_prepare_cpu,
1385 .teardown.single = smpcfd_dead_cpu,
1387 [CPUHP_RELAY_PREPARE] = {
1388 .name = "relay:prepare",
1389 .startup.single = relay_prepare_cpu,
1390 .teardown.single = NULL,
1392 [CPUHP_SLAB_PREPARE] = {
1393 .name = "slab:prepare",
1394 .startup.single = slab_prepare_cpu,
1395 .teardown.single = slab_dead_cpu,
1397 [CPUHP_RCUTREE_PREP] = {
1398 .name = "RCU/tree:prepare",
1399 .startup.single = rcutree_prepare_cpu,
1400 .teardown.single = rcutree_dead_cpu,
1403 * Preparatory and dead notifiers. Will be replaced once the notifiers
1404 * are converted to states.
1406 [CPUHP_NOTIFY_PREPARE] = {
1407 .name = "notify:prepare",
1408 .startup.single = notify_prepare,
1409 .teardown.single = notify_dead,
1414 * On the tear-down path, timers_dead_cpu() must be invoked
1415 * before blk_mq_queue_reinit_notify() from notify_dead(),
1416 * otherwise a RCU stall occurs.
1418 [CPUHP_TIMERS_PREPARE] = {
1419 .name = "timers:dead",
1420 .startup.single = timers_prepare_cpu,
1421 .teardown.single = timers_dead_cpu,
1423 /* Kicks the plugged cpu into life */
1424 [CPUHP_BRINGUP_CPU] = {
1425 .name = "cpu:bringup",
1426 .startup.single = bringup_cpu,
1427 .teardown.single = NULL,
1431 * Handled on controll processor until the plugged processor manages
1434 [CPUHP_TEARDOWN_CPU] = {
1435 .name = "cpu:teardown",
1436 .startup.single = NULL,
1437 .teardown.single = takedown_cpu,
1441 [CPUHP_BRINGUP_CPU] = { },
1445 /* Application processor state steps */
1446 static struct cpuhp_step cpuhp_ap_states[] = {
1448 /* Final state before CPU kills itself */
1449 [CPUHP_AP_IDLE_DEAD] = {
1450 .name = "idle:dead",
1453 * Last state before CPU enters the idle loop to die. Transient state
1454 * for synchronization.
1456 [CPUHP_AP_OFFLINE] = {
1457 .name = "ap:offline",
1460 /* First state is scheduler control. Interrupts are disabled */
1461 [CPUHP_AP_SCHED_STARTING] = {
1462 .name = "sched:starting",
1463 .startup.single = sched_cpu_starting,
1464 .teardown.single = sched_cpu_dying,
1466 [CPUHP_AP_RCUTREE_DYING] = {
1467 .name = "RCU/tree:dying",
1468 .startup.single = NULL,
1469 .teardown.single = rcutree_dying_cpu,
1471 [CPUHP_AP_SMPCFD_DYING] = {
1472 .name = "smpcfd:dying",
1473 .startup.single = NULL,
1474 .teardown.single = smpcfd_dying_cpu,
1476 /* Entry state on starting. Interrupts enabled from here on. Transient
1477 * state for synchronsization */
1478 [CPUHP_AP_ONLINE] = {
1479 .name = "ap:online",
1481 /* Handle smpboot threads park/unpark */
1482 [CPUHP_AP_SMPBOOT_THREADS] = {
1483 .name = "smpboot/threads:online",
1484 .startup.single = smpboot_unpark_threads,
1485 .teardown.single = smpboot_park_threads,
1487 [CPUHP_AP_PERF_ONLINE] = {
1488 .name = "perf:online",
1489 .startup.single = perf_event_init_cpu,
1490 .teardown.single = perf_event_exit_cpu,
1492 [CPUHP_AP_WORKQUEUE_ONLINE] = {
1493 .name = "workqueue:online",
1494 .startup.single = workqueue_online_cpu,
1495 .teardown.single = workqueue_offline_cpu,
1497 [CPUHP_AP_RCUTREE_ONLINE] = {
1498 .name = "RCU/tree:online",
1499 .startup.single = rcutree_online_cpu,
1500 .teardown.single = rcutree_offline_cpu,
1504 * Online/down_prepare notifiers. Will be removed once the notifiers
1505 * are converted to states.
1507 [CPUHP_AP_NOTIFY_ONLINE] = {
1508 .name = "notify:online",
1509 .startup.single = notify_online,
1510 .teardown.single = notify_down_prepare,
1515 * The dynamically registered state space is here
1519 /* Last state is scheduler control setting the cpu active */
1520 [CPUHP_AP_ACTIVE] = {
1521 .name = "sched:active",
1522 .startup.single = sched_cpu_activate,
1523 .teardown.single = sched_cpu_deactivate,
1527 /* CPU is fully up and running. */
1530 .startup.single = NULL,
1531 .teardown.single = NULL,
1535 /* Sanity check for callbacks */
1536 static int cpuhp_cb_check(enum cpuhp_state state)
1538 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1543 static void cpuhp_store_callbacks(enum cpuhp_state state,
1545 int (*startup)(unsigned int cpu),
1546 int (*teardown)(unsigned int cpu),
1547 bool multi_instance)
1549 /* (Un)Install the callbacks for further cpu hotplug operations */
1550 struct cpuhp_step *sp;
1552 sp = cpuhp_get_step(state);
1553 sp->startup.single = startup;
1554 sp->teardown.single = teardown;
1556 sp->multi_instance = multi_instance;
1557 INIT_HLIST_HEAD(&sp->list);
1560 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1562 return cpuhp_get_step(state)->teardown.single;
1566 * Call the startup/teardown function for a step either on the AP or
1567 * on the current CPU.
1569 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1570 struct hlist_node *node)
1572 struct cpuhp_step *sp = cpuhp_get_step(state);
1575 if ((bringup && !sp->startup.single) ||
1576 (!bringup && !sp->teardown.single))
1579 * The non AP bound callbacks can fail on bringup. On teardown
1580 * e.g. module removal we crash for now.
1583 if (cpuhp_is_ap_state(state))
1584 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1586 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1588 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1590 BUG_ON(ret && !bringup);
1595 * Called from __cpuhp_setup_state on a recoverable failure.
1597 * Note: The teardown callbacks for rollback are not allowed to fail!
1599 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1600 struct hlist_node *node)
1604 /* Roll back the already executed steps on the other cpus */
1605 for_each_present_cpu(cpu) {
1606 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1607 int cpustate = st->state;
1609 if (cpu >= failedcpu)
1612 /* Did we invoke the startup call on that cpu ? */
1613 if (cpustate >= state)
1614 cpuhp_issue_call(cpu, state, false, node);
1619 * Returns a free for dynamic slot assignment of the Online state. The states
1620 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1621 * by having no name assigned.
1623 static int cpuhp_reserve_state(enum cpuhp_state state)
1627 for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1628 if (cpuhp_ap_states[i].name)
1631 cpuhp_ap_states[i].name = "Reserved";
1634 WARN(1, "No more dynamic states available for CPU hotplug\n");
1638 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1641 struct cpuhp_step *sp;
1645 sp = cpuhp_get_step(state);
1646 if (sp->multi_instance == false)
1650 mutex_lock(&cpuhp_state_mutex);
1652 if (!invoke || !sp->startup.multi)
1656 * Try to call the startup callback for each present cpu
1657 * depending on the hotplug state of the cpu.
1659 for_each_present_cpu(cpu) {
1660 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1661 int cpustate = st->state;
1663 if (cpustate < state)
1666 ret = cpuhp_issue_call(cpu, state, true, node);
1668 if (sp->teardown.multi)
1669 cpuhp_rollback_install(cpu, state, node);
1675 hlist_add_head(node, &sp->list);
1678 mutex_unlock(&cpuhp_state_mutex);
1682 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1685 * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1686 * @state: The state to setup
1687 * @invoke: If true, the startup function is invoked for cpus where
1688 * cpu state >= @state
1689 * @startup: startup callback function
1690 * @teardown: teardown callback function
1692 * Returns 0 if successful, otherwise a proper error code
1694 int __cpuhp_setup_state(enum cpuhp_state state,
1695 const char *name, bool invoke,
1696 int (*startup)(unsigned int cpu),
1697 int (*teardown)(unsigned int cpu),
1698 bool multi_instance)
1703 if (cpuhp_cb_check(state) || !name)
1707 mutex_lock(&cpuhp_state_mutex);
1709 /* currently assignments for the ONLINE state are possible */
1710 if (state == CPUHP_AP_ONLINE_DYN) {
1712 ret = cpuhp_reserve_state(state);
1718 cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1720 if (!invoke || !startup)
1724 * Try to call the startup callback for each present cpu
1725 * depending on the hotplug state of the cpu.
1727 for_each_present_cpu(cpu) {
1728 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1729 int cpustate = st->state;
1731 if (cpustate < state)
1734 ret = cpuhp_issue_call(cpu, state, true, NULL);
1737 cpuhp_rollback_install(cpu, state, NULL);
1738 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1743 mutex_unlock(&cpuhp_state_mutex);
1746 if (!ret && dyn_state)
1750 EXPORT_SYMBOL(__cpuhp_setup_state);
1752 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1753 struct hlist_node *node, bool invoke)
1755 struct cpuhp_step *sp = cpuhp_get_step(state);
1758 BUG_ON(cpuhp_cb_check(state));
1760 if (!sp->multi_instance)
1764 mutex_lock(&cpuhp_state_mutex);
1766 if (!invoke || !cpuhp_get_teardown_cb(state))
1769 * Call the teardown callback for each present cpu depending
1770 * on the hotplug state of the cpu. This function is not
1771 * allowed to fail currently!
1773 for_each_present_cpu(cpu) {
1774 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1775 int cpustate = st->state;
1777 if (cpustate >= state)
1778 cpuhp_issue_call(cpu, state, false, node);
1783 mutex_unlock(&cpuhp_state_mutex);
1788 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1790 * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1791 * @state: The state to remove
1792 * @invoke: If true, the teardown function is invoked for cpus where
1793 * cpu state >= @state
1795 * The teardown callback is currently not allowed to fail. Think
1796 * about module removal!
1798 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1800 struct cpuhp_step *sp = cpuhp_get_step(state);
1803 BUG_ON(cpuhp_cb_check(state));
1806 mutex_lock(&cpuhp_state_mutex);
1808 if (sp->multi_instance) {
1809 WARN(!hlist_empty(&sp->list),
1810 "Error: Removing state %d which has instances left.\n",
1815 if (!invoke || !cpuhp_get_teardown_cb(state))
1819 * Call the teardown callback for each present cpu depending
1820 * on the hotplug state of the cpu. This function is not
1821 * allowed to fail currently!
1823 for_each_present_cpu(cpu) {
1824 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1825 int cpustate = st->state;
1827 if (cpustate >= state)
1828 cpuhp_issue_call(cpu, state, false, NULL);
1831 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1832 mutex_unlock(&cpuhp_state_mutex);
1835 EXPORT_SYMBOL(__cpuhp_remove_state);
1837 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1838 static ssize_t show_cpuhp_state(struct device *dev,
1839 struct device_attribute *attr, char *buf)
1841 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1843 return sprintf(buf, "%d\n", st->state);
1845 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1847 static ssize_t write_cpuhp_target(struct device *dev,
1848 struct device_attribute *attr,
1849 const char *buf, size_t count)
1851 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1852 struct cpuhp_step *sp;
1855 ret = kstrtoint(buf, 10, &target);
1859 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1860 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1863 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1867 ret = lock_device_hotplug_sysfs();
1871 mutex_lock(&cpuhp_state_mutex);
1872 sp = cpuhp_get_step(target);
1873 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1874 mutex_unlock(&cpuhp_state_mutex);
1878 if (st->state < target)
1879 ret = do_cpu_up(dev->id, target);
1881 ret = do_cpu_down(dev->id, target);
1883 unlock_device_hotplug();
1884 return ret ? ret : count;
1887 static ssize_t show_cpuhp_target(struct device *dev,
1888 struct device_attribute *attr, char *buf)
1890 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1892 return sprintf(buf, "%d\n", st->target);
1894 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1896 static struct attribute *cpuhp_cpu_attrs[] = {
1897 &dev_attr_state.attr,
1898 &dev_attr_target.attr,
1902 static struct attribute_group cpuhp_cpu_attr_group = {
1903 .attrs = cpuhp_cpu_attrs,
1908 static ssize_t show_cpuhp_states(struct device *dev,
1909 struct device_attribute *attr, char *buf)
1911 ssize_t cur, res = 0;
1914 mutex_lock(&cpuhp_state_mutex);
1915 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1916 struct cpuhp_step *sp = cpuhp_get_step(i);
1919 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1924 mutex_unlock(&cpuhp_state_mutex);
1927 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1929 static struct attribute *cpuhp_cpu_root_attrs[] = {
1930 &dev_attr_states.attr,
1934 static struct attribute_group cpuhp_cpu_root_attr_group = {
1935 .attrs = cpuhp_cpu_root_attrs,
1940 #ifdef CONFIG_HOTPLUG_SMT
1942 static const char *smt_states[] = {
1943 [CPU_SMT_ENABLED] = "on",
1944 [CPU_SMT_DISABLED] = "off",
1945 [CPU_SMT_FORCE_DISABLED] = "forceoff",
1946 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
1950 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
1952 return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
1955 static void cpuhp_offline_cpu_device(unsigned int cpu)
1957 struct device *dev = get_cpu_device(cpu);
1959 dev->offline = true;
1960 /* Tell user space about the state change */
1961 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
1964 static void cpuhp_online_cpu_device(unsigned int cpu)
1966 struct device *dev = get_cpu_device(cpu);
1968 dev->offline = false;
1969 /* Tell user space about the state change */
1970 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
1973 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
1977 cpu_maps_update_begin();
1978 for_each_online_cpu(cpu) {
1979 if (topology_is_primary_thread(cpu))
1981 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1985 * As this needs to hold the cpu maps lock it's impossible
1986 * to call device_offline() because that ends up calling
1987 * cpu_down() which takes cpu maps lock. cpu maps lock
1988 * needs to be held as this might race against in kernel
1989 * abusers of the hotplug machinery (thermal management).
1991 * So nothing would update device:offline state. That would
1992 * leave the sysfs entry stale and prevent onlining after
1993 * smt control has been changed to 'off' again. This is
1994 * called under the sysfs hotplug lock, so it is properly
1995 * serialized against the regular offline usage.
1997 cpuhp_offline_cpu_device(cpu);
2000 cpu_smt_control = ctrlval;
2001 cpu_maps_update_done();
2005 static int cpuhp_smt_enable(void)
2009 cpu_maps_update_begin();
2010 cpu_smt_control = CPU_SMT_ENABLED;
2011 for_each_present_cpu(cpu) {
2012 /* Skip online CPUs and CPUs on offline nodes */
2013 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2015 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2018 /* See comment in cpuhp_smt_disable() */
2019 cpuhp_online_cpu_device(cpu);
2021 cpu_maps_update_done();
2026 store_smt_control(struct device *dev, struct device_attribute *attr,
2027 const char *buf, size_t count)
2031 if (sysfs_streq(buf, "on"))
2032 ctrlval = CPU_SMT_ENABLED;
2033 else if (sysfs_streq(buf, "off"))
2034 ctrlval = CPU_SMT_DISABLED;
2035 else if (sysfs_streq(buf, "forceoff"))
2036 ctrlval = CPU_SMT_FORCE_DISABLED;
2040 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2043 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2046 ret = lock_device_hotplug_sysfs();
2050 if (ctrlval != cpu_smt_control) {
2052 case CPU_SMT_ENABLED:
2053 ret = cpuhp_smt_enable();
2055 case CPU_SMT_DISABLED:
2056 case CPU_SMT_FORCE_DISABLED:
2057 ret = cpuhp_smt_disable(ctrlval);
2062 unlock_device_hotplug();
2063 return ret ? ret : count;
2065 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2068 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2070 bool active = topology_max_smt_threads() > 1;
2072 return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2074 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2076 static struct attribute *cpuhp_smt_attrs[] = {
2077 &dev_attr_control.attr,
2078 &dev_attr_active.attr,
2082 static const struct attribute_group cpuhp_smt_attr_group = {
2083 .attrs = cpuhp_smt_attrs,
2088 static int __init cpu_smt_state_init(void)
2090 return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2091 &cpuhp_smt_attr_group);
2095 static inline int cpu_smt_state_init(void) { return 0; }
2098 static int __init cpuhp_sysfs_init(void)
2102 ret = cpu_smt_state_init();
2106 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2107 &cpuhp_cpu_root_attr_group);
2111 for_each_possible_cpu(cpu) {
2112 struct device *dev = get_cpu_device(cpu);
2116 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2122 device_initcall(cpuhp_sysfs_init);
2126 * cpu_bit_bitmap[] is a special, "compressed" data structure that
2127 * represents all NR_CPUS bits binary values of 1<<nr.
2129 * It is used by cpumask_of() to get a constant address to a CPU
2130 * mask value that has a single bit set only.
2133 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2134 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
2135 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2136 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2137 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2139 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2141 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
2142 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
2143 #if BITS_PER_LONG > 32
2144 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
2145 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
2148 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2150 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2151 EXPORT_SYMBOL(cpu_all_bits);
2153 #ifdef CONFIG_INIT_ALL_POSSIBLE
2154 struct cpumask __cpu_possible_mask __read_mostly
2157 struct cpumask __cpu_possible_mask __read_mostly;
2159 EXPORT_SYMBOL(__cpu_possible_mask);
2161 struct cpumask __cpu_online_mask __read_mostly;
2162 EXPORT_SYMBOL(__cpu_online_mask);
2164 struct cpumask __cpu_present_mask __read_mostly;
2165 EXPORT_SYMBOL(__cpu_present_mask);
2167 struct cpumask __cpu_active_mask __read_mostly;
2168 EXPORT_SYMBOL(__cpu_active_mask);
2170 void init_cpu_present(const struct cpumask *src)
2172 cpumask_copy(&__cpu_present_mask, src);
2175 void init_cpu_possible(const struct cpumask *src)
2177 cpumask_copy(&__cpu_possible_mask, src);
2180 void init_cpu_online(const struct cpumask *src)
2182 cpumask_copy(&__cpu_online_mask, src);
2186 * Activate the first processor.
2188 void __init boot_cpu_init(void)
2190 int cpu = smp_processor_id();
2192 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2193 set_cpu_online(cpu, true);
2194 set_cpu_active(cpu, true);
2195 set_cpu_present(cpu, true);
2196 set_cpu_possible(cpu, true);
2200 * Must be called _AFTER_ setting up the per_cpu areas
2202 void __init boot_cpu_hotplug_init(void)
2205 this_cpu_write(cpuhp_state.booted_once, true);
2207 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);