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/sched/smt.h>
12 #include <linux/unistd.h>
13 #include <linux/cpu.h>
14 #include <linux/oom.h>
15 #include <linux/rcupdate.h>
16 #include <linux/export.h>
17 #include <linux/bug.h>
18 #include <linux/kthread.h>
19 #include <linux/stop_machine.h>
20 #include <linux/mutex.h>
21 #include <linux/gfp.h>
22 #include <linux/suspend.h>
23 #include <linux/lockdep.h>
24 #include <linux/tick.h>
25 #include <linux/irq.h>
26 #include <linux/smpboot.h>
27 #include <linux/relay.h>
28 #include <linux/slab.h>
30 #include <trace/events/power.h>
31 #define CREATE_TRACE_POINTS
32 #include <trace/events/cpuhp.h>
37 * cpuhp_cpu_state - Per cpu hotplug state storage
38 * @state: The current cpu state
39 * @target: The target state
40 * @thread: Pointer to the hotplug thread
41 * @should_run: Thread should execute
42 * @rollback: Perform a rollback
43 * @single: Single callback invocation
44 * @bringup: Single callback bringup or teardown selector
45 * @cb_state: The state for a single callback (install/uninstall)
46 * @result: Result of the operation
47 * @done: Signal completion to the issuer of the task
49 struct cpuhp_cpu_state {
50 enum cpuhp_state state;
51 enum cpuhp_state target;
53 struct task_struct *thread;
59 struct hlist_node *node;
60 enum cpuhp_state cb_state;
62 struct completion done;
66 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
68 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
69 static struct lock_class_key cpuhp_state_key;
70 static struct lockdep_map cpuhp_state_lock_map =
71 STATIC_LOCKDEP_MAP_INIT("cpuhp_state", &cpuhp_state_key);
75 * cpuhp_step - Hotplug state machine step
76 * @name: Name of the step
77 * @startup: Startup function of the step
78 * @teardown: Teardown function of the step
79 * @skip_onerr: Do not invoke the functions on error rollback
80 * Will go away once the notifiers are gone
81 * @cant_stop: Bringup/teardown can't be stopped at this step
86 int (*single)(unsigned int cpu);
87 int (*multi)(unsigned int cpu,
88 struct hlist_node *node);
91 int (*single)(unsigned int cpu);
92 int (*multi)(unsigned int cpu,
93 struct hlist_node *node);
95 struct hlist_head list;
101 static DEFINE_MUTEX(cpuhp_state_mutex);
102 static struct cpuhp_step cpuhp_bp_states[];
103 static struct cpuhp_step cpuhp_ap_states[];
105 static bool cpuhp_is_ap_state(enum cpuhp_state state)
108 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
109 * purposes as that state is handled explicitly in cpu_down.
111 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
114 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
116 struct cpuhp_step *sp;
118 sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
123 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
124 * @cpu: The cpu for which the callback should be invoked
125 * @step: The step in the state machine
126 * @bringup: True if the bringup callback should be invoked
128 * Called from cpu hotplug and from the state register machinery.
130 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
131 bool bringup, struct hlist_node *node)
133 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
134 struct cpuhp_step *step = cpuhp_get_step(state);
135 int (*cbm)(unsigned int cpu, struct hlist_node *node);
136 int (*cb)(unsigned int cpu);
139 if (!step->multi_instance) {
140 cb = bringup ? step->startup.single : step->teardown.single;
143 trace_cpuhp_enter(cpu, st->target, state, cb);
145 trace_cpuhp_exit(cpu, st->state, state, ret);
148 cbm = bringup ? step->startup.multi : step->teardown.multi;
152 /* Single invocation for instance add/remove */
154 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
155 ret = cbm(cpu, node);
156 trace_cpuhp_exit(cpu, st->state, state, ret);
160 /* State transition. Invoke on all instances */
162 hlist_for_each(node, &step->list) {
163 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
164 ret = cbm(cpu, node);
165 trace_cpuhp_exit(cpu, st->state, state, ret);
172 /* Rollback the instances if one failed */
173 cbm = !bringup ? step->startup.multi : step->teardown.multi;
177 hlist_for_each(node, &step->list) {
186 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
187 static DEFINE_MUTEX(cpu_add_remove_lock);
188 bool cpuhp_tasks_frozen;
189 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
192 * The following two APIs (cpu_maps_update_begin/done) must be used when
193 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
194 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
195 * hotplug callback (un)registration performed using __register_cpu_notifier()
196 * or __unregister_cpu_notifier().
198 void cpu_maps_update_begin(void)
200 mutex_lock(&cpu_add_remove_lock);
202 EXPORT_SYMBOL(cpu_notifier_register_begin);
204 void cpu_maps_update_done(void)
206 mutex_unlock(&cpu_add_remove_lock);
208 EXPORT_SYMBOL(cpu_notifier_register_done);
210 static RAW_NOTIFIER_HEAD(cpu_chain);
212 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
213 * Should always be manipulated under cpu_add_remove_lock
215 static int cpu_hotplug_disabled;
217 #ifdef CONFIG_HOTPLUG_CPU
220 struct task_struct *active_writer;
221 /* wait queue to wake up the active_writer */
222 wait_queue_head_t wq;
223 /* verifies that no writer will get active while readers are active */
226 * Also blocks the new readers during
227 * an ongoing cpu hotplug operation.
231 #ifdef CONFIG_DEBUG_LOCK_ALLOC
232 struct lockdep_map dep_map;
235 .active_writer = NULL,
236 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
237 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
238 #ifdef CONFIG_DEBUG_LOCK_ALLOC
239 .dep_map = STATIC_LOCKDEP_MAP_INIT("cpu_hotplug.dep_map", &cpu_hotplug.dep_map),
243 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
244 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
245 #define cpuhp_lock_acquire_tryread() \
246 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
247 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
248 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
251 void get_online_cpus(void)
254 if (cpu_hotplug.active_writer == current)
256 cpuhp_lock_acquire_read();
257 mutex_lock(&cpu_hotplug.lock);
258 atomic_inc(&cpu_hotplug.refcount);
259 mutex_unlock(&cpu_hotplug.lock);
261 EXPORT_SYMBOL_GPL(get_online_cpus);
263 void put_online_cpus(void)
267 if (cpu_hotplug.active_writer == current)
270 refcount = atomic_dec_return(&cpu_hotplug.refcount);
271 if (WARN_ON(refcount < 0)) /* try to fix things up */
272 atomic_inc(&cpu_hotplug.refcount);
274 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
275 wake_up(&cpu_hotplug.wq);
277 cpuhp_lock_release();
280 EXPORT_SYMBOL_GPL(put_online_cpus);
283 * This ensures that the hotplug operation can begin only when the
284 * refcount goes to zero.
286 * Note that during a cpu-hotplug operation, the new readers, if any,
287 * will be blocked by the cpu_hotplug.lock
289 * Since cpu_hotplug_begin() is always called after invoking
290 * cpu_maps_update_begin(), we can be sure that only one writer is active.
292 * Note that theoretically, there is a possibility of a livelock:
293 * - Refcount goes to zero, last reader wakes up the sleeping
295 * - Last reader unlocks the cpu_hotplug.lock.
296 * - A new reader arrives at this moment, bumps up the refcount.
297 * - The writer acquires the cpu_hotplug.lock finds the refcount
298 * non zero and goes to sleep again.
300 * However, this is very difficult to achieve in practice since
301 * get_online_cpus() not an api which is called all that often.
304 void cpu_hotplug_begin(void)
308 cpu_hotplug.active_writer = current;
309 cpuhp_lock_acquire();
312 mutex_lock(&cpu_hotplug.lock);
313 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
314 if (likely(!atomic_read(&cpu_hotplug.refcount)))
316 mutex_unlock(&cpu_hotplug.lock);
319 finish_wait(&cpu_hotplug.wq, &wait);
322 void cpu_hotplug_done(void)
324 cpu_hotplug.active_writer = NULL;
325 mutex_unlock(&cpu_hotplug.lock);
326 cpuhp_lock_release();
330 * Wait for currently running CPU hotplug operations to complete (if any) and
331 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
332 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
333 * hotplug path before performing hotplug operations. So acquiring that lock
334 * guarantees mutual exclusion from any currently running hotplug operations.
336 void cpu_hotplug_disable(void)
338 cpu_maps_update_begin();
339 cpu_hotplug_disabled++;
340 cpu_maps_update_done();
342 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
344 static void __cpu_hotplug_enable(void)
346 if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
348 cpu_hotplug_disabled--;
351 void cpu_hotplug_enable(void)
353 cpu_maps_update_begin();
354 __cpu_hotplug_enable();
355 cpu_maps_update_done();
357 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
358 #endif /* CONFIG_HOTPLUG_CPU */
361 * Architectures that need SMT-specific errata handling during SMT hotplug
362 * should override this.
364 void __weak arch_smt_update(void) { }
366 #ifdef CONFIG_HOTPLUG_SMT
367 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
368 EXPORT_SYMBOL_GPL(cpu_smt_control);
370 static bool cpu_smt_available __read_mostly;
372 void __init cpu_smt_disable(bool force)
374 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
375 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
379 pr_info("SMT: Force disabled\n");
380 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
382 pr_info("SMT: disabled\n");
383 cpu_smt_control = CPU_SMT_DISABLED;
388 * The decision whether SMT is supported can only be done after the full
389 * CPU identification. Called from architecture code before non boot CPUs
392 void __init cpu_smt_check_topology_early(void)
394 if (!topology_smt_supported())
395 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
399 * If SMT was disabled by BIOS, detect it here, after the CPUs have been
400 * brought online. This ensures the smt/l1tf sysfs entries are consistent
401 * with reality. cpu_smt_available is set to true during the bringup of non
402 * boot CPUs when a SMT sibling is detected. Note, this may overwrite
403 * cpu_smt_control's previous setting.
405 void __init cpu_smt_check_topology(void)
407 if (!cpu_smt_available)
408 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
411 static int __init smt_cmdline_disable(char *str)
413 cpu_smt_disable(str && !strcmp(str, "force"));
416 early_param("nosmt", smt_cmdline_disable);
418 static inline bool cpu_smt_allowed(unsigned int cpu)
420 if (topology_is_primary_thread(cpu))
424 * If the CPU is not a 'primary' thread and the booted_once bit is
425 * set then the processor has SMT support. Store this information
426 * for the late check of SMT support in cpu_smt_check_topology().
428 if (per_cpu(cpuhp_state, cpu).booted_once)
429 cpu_smt_available = true;
431 if (cpu_smt_control == CPU_SMT_ENABLED)
435 * On x86 it's required to boot all logical CPUs at least once so
436 * that the init code can get a chance to set CR4.MCE on each
437 * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
438 * core will shutdown the machine.
440 return !per_cpu(cpuhp_state, cpu).booted_once;
443 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
446 /* Need to know about CPUs going up/down? */
447 int register_cpu_notifier(struct notifier_block *nb)
450 cpu_maps_update_begin();
451 ret = raw_notifier_chain_register(&cpu_chain, nb);
452 cpu_maps_update_done();
456 int __register_cpu_notifier(struct notifier_block *nb)
458 return raw_notifier_chain_register(&cpu_chain, nb);
461 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
464 unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
465 void *hcpu = (void *)(long)cpu;
469 ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
472 return notifier_to_errno(ret);
475 static int cpu_notify(unsigned long val, unsigned int cpu)
477 return __cpu_notify(val, cpu, -1, NULL);
480 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
482 BUG_ON(cpu_notify(val, cpu));
485 /* Notifier wrappers for transitioning to state machine */
486 static int notify_prepare(unsigned int cpu)
491 ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
494 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
496 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
501 static int notify_online(unsigned int cpu)
503 cpu_notify(CPU_ONLINE, cpu);
507 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st);
509 static int bringup_wait_for_ap(unsigned int cpu)
511 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
513 /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
514 wait_for_completion(&st->done);
515 if (WARN_ON_ONCE((!cpu_online(cpu))))
518 /* Unpark the hotplug thread of the target cpu */
519 kthread_unpark(st->thread);
522 * SMT soft disabling on X86 requires to bring the CPU out of the
523 * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit. The
524 * CPU marked itself as booted_once in cpu_notify_starting() so the
525 * cpu_smt_allowed() check will now return false if this is not the
528 if (!cpu_smt_allowed(cpu))
531 /* Should we go further up ? */
532 if (st->target > CPUHP_AP_ONLINE_IDLE) {
533 __cpuhp_kick_ap_work(st);
534 wait_for_completion(&st->done);
539 static int bringup_cpu(unsigned int cpu)
541 struct task_struct *idle = idle_thread_get(cpu);
545 * Some architectures have to walk the irq descriptors to
546 * setup the vector space for the cpu which comes online.
547 * Prevent irq alloc/free across the bringup.
551 /* Arch-specific enabling code. */
552 ret = __cpu_up(cpu, idle);
555 cpu_notify(CPU_UP_CANCELED, cpu);
558 return bringup_wait_for_ap(cpu);
562 * Hotplug state machine related functions
564 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
566 for (st->state++; st->state < st->target; st->state++) {
567 struct cpuhp_step *step = cpuhp_get_step(st->state);
569 if (!step->skip_onerr)
570 cpuhp_invoke_callback(cpu, st->state, true, NULL);
574 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
575 enum cpuhp_state target)
577 enum cpuhp_state prev_state = st->state;
580 for (; st->state > target; st->state--) {
581 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
583 st->target = prev_state;
584 undo_cpu_down(cpu, st);
591 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
593 for (st->state--; st->state > st->target; st->state--) {
594 struct cpuhp_step *step = cpuhp_get_step(st->state);
596 if (!step->skip_onerr)
597 cpuhp_invoke_callback(cpu, st->state, false, NULL);
601 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
603 if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
606 * When CPU hotplug is disabled, then taking the CPU down is not
607 * possible because takedown_cpu() and the architecture and
608 * subsystem specific mechanisms are not available. So the CPU
609 * which would be completely unplugged again needs to stay around
610 * in the current state.
612 return st->state <= CPUHP_BRINGUP_CPU;
615 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
616 enum cpuhp_state target)
618 enum cpuhp_state prev_state = st->state;
621 while (st->state < target) {
623 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
625 if (can_rollback_cpu(st)) {
626 st->target = prev_state;
627 undo_cpu_up(cpu, st);
636 * The cpu hotplug threads manage the bringup and teardown of the cpus
638 static void cpuhp_create(unsigned int cpu)
640 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
642 init_completion(&st->done);
645 static int cpuhp_should_run(unsigned int cpu)
647 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
649 return st->should_run;
652 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
653 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
655 enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
657 return cpuhp_down_callbacks(cpu, st, target);
660 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
661 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
663 return cpuhp_up_callbacks(cpu, st, st->target);
667 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
668 * callbacks when a state gets [un]installed at runtime.
670 static void cpuhp_thread_fun(unsigned int cpu)
672 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
676 * Paired with the mb() in cpuhp_kick_ap_work and
677 * cpuhp_invoke_ap_callback, so the work set is consistent visible.
683 st->should_run = false;
685 lock_map_acquire(&cpuhp_state_lock_map);
686 /* Single callback invocation for [un]install ? */
688 if (st->cb_state < CPUHP_AP_ONLINE) {
690 ret = cpuhp_invoke_callback(cpu, st->cb_state,
691 st->bringup, st->node);
694 ret = cpuhp_invoke_callback(cpu, st->cb_state,
695 st->bringup, st->node);
697 } else if (st->rollback) {
698 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
700 undo_cpu_down(cpu, st);
702 * This is a momentary workaround to keep the notifier users
703 * happy. Will go away once we got rid of the notifiers.
705 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
706 st->rollback = false;
708 /* Cannot happen .... */
709 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
711 /* Regular hotplug work */
712 if (st->state < st->target)
713 ret = cpuhp_ap_online(cpu, st);
714 else if (st->state > st->target)
715 ret = cpuhp_ap_offline(cpu, st);
717 lock_map_release(&cpuhp_state_lock_map);
722 /* Invoke a single callback on a remote cpu */
724 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
725 struct hlist_node *node)
727 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
729 if (!cpu_online(cpu))
732 lock_map_acquire(&cpuhp_state_lock_map);
733 lock_map_release(&cpuhp_state_lock_map);
736 * If we are up and running, use the hotplug thread. For early calls
737 * we invoke the thread function directly.
740 return cpuhp_invoke_callback(cpu, state, bringup, node);
742 st->cb_state = state;
744 st->bringup = bringup;
748 * Make sure the above stores are visible before should_run becomes
749 * true. Paired with the mb() above in cpuhp_thread_fun()
752 st->should_run = true;
753 wake_up_process(st->thread);
754 wait_for_completion(&st->done);
758 /* Regular hotplug invocation of the AP hotplug thread */
759 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
764 * Make sure the above stores are visible before should_run becomes
765 * true. Paired with the mb() above in cpuhp_thread_fun()
768 st->should_run = true;
769 wake_up_process(st->thread);
772 static int cpuhp_kick_ap_work(unsigned int cpu)
774 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
775 enum cpuhp_state state = st->state;
777 trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
778 lock_map_acquire(&cpuhp_state_lock_map);
779 lock_map_release(&cpuhp_state_lock_map);
780 __cpuhp_kick_ap_work(st);
781 wait_for_completion(&st->done);
782 trace_cpuhp_exit(cpu, st->state, state, st->result);
786 static struct smp_hotplug_thread cpuhp_threads = {
787 .store = &cpuhp_state.thread,
788 .create = &cpuhp_create,
789 .thread_should_run = cpuhp_should_run,
790 .thread_fn = cpuhp_thread_fun,
791 .thread_comm = "cpuhp/%u",
795 void __init cpuhp_threads_init(void)
797 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
798 kthread_unpark(this_cpu_read(cpuhp_state.thread));
801 EXPORT_SYMBOL(register_cpu_notifier);
802 EXPORT_SYMBOL(__register_cpu_notifier);
803 void unregister_cpu_notifier(struct notifier_block *nb)
805 cpu_maps_update_begin();
806 raw_notifier_chain_unregister(&cpu_chain, nb);
807 cpu_maps_update_done();
809 EXPORT_SYMBOL(unregister_cpu_notifier);
811 void __unregister_cpu_notifier(struct notifier_block *nb)
813 raw_notifier_chain_unregister(&cpu_chain, nb);
815 EXPORT_SYMBOL(__unregister_cpu_notifier);
817 #ifdef CONFIG_HOTPLUG_CPU
818 #ifndef arch_clear_mm_cpumask_cpu
819 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
823 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
826 * This function walks all processes, finds a valid mm struct for each one and
827 * then clears a corresponding bit in mm's cpumask. While this all sounds
828 * trivial, there are various non-obvious corner cases, which this function
829 * tries to solve in a safe manner.
831 * Also note that the function uses a somewhat relaxed locking scheme, so it may
832 * be called only for an already offlined CPU.
834 void clear_tasks_mm_cpumask(int cpu)
836 struct task_struct *p;
839 * This function is called after the cpu is taken down and marked
840 * offline, so its not like new tasks will ever get this cpu set in
841 * their mm mask. -- Peter Zijlstra
842 * Thus, we may use rcu_read_lock() here, instead of grabbing
843 * full-fledged tasklist_lock.
845 WARN_ON(cpu_online(cpu));
847 for_each_process(p) {
848 struct task_struct *t;
851 * Main thread might exit, but other threads may still have
852 * a valid mm. Find one.
854 t = find_lock_task_mm(p);
857 arch_clear_mm_cpumask_cpu(cpu, t->mm);
863 static inline void check_for_tasks(int dead_cpu)
865 struct task_struct *g, *p;
867 read_lock(&tasklist_lock);
868 for_each_process_thread(g, p) {
872 * We do the check with unlocked task_rq(p)->lock.
873 * Order the reading to do not warn about a task,
874 * which was running on this cpu in the past, and
875 * it's just been woken on another cpu.
878 if (task_cpu(p) != dead_cpu)
881 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
882 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
884 read_unlock(&tasklist_lock);
887 static int notify_down_prepare(unsigned int cpu)
889 int err, nr_calls = 0;
891 err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
894 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
895 pr_warn("%s: attempt to take down CPU %u failed\n",
901 /* Take this CPU down. */
902 static int take_cpu_down(void *_param)
904 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
905 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
906 int err, cpu = smp_processor_id();
908 /* Ensure this CPU doesn't handle any more interrupts. */
909 err = __cpu_disable();
914 * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
915 * do this step again.
917 WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
919 /* Invoke the former CPU_DYING callbacks */
920 for (; st->state > target; st->state--)
921 cpuhp_invoke_callback(cpu, st->state, false, NULL);
923 /* Give up timekeeping duties */
924 tick_handover_do_timer();
925 /* Park the stopper thread */
926 stop_machine_park(cpu);
930 static int takedown_cpu(unsigned int cpu)
932 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
935 /* Park the smpboot threads */
936 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
939 * Prevent irq alloc/free while the dying cpu reorganizes the
940 * interrupt affinities.
945 * So now all preempt/rcu users must observe !cpu_active().
947 err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
949 /* CPU refused to die */
951 /* Unpark the hotplug thread so we can rollback there */
952 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
955 BUG_ON(cpu_online(cpu));
958 * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
959 * runnable tasks from the cpu, there's only the idle task left now
960 * that the migration thread is done doing the stop_machine thing.
962 * Wait for the stop thread to go away.
964 wait_for_completion(&st->done);
965 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
967 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
970 hotplug_cpu__broadcast_tick_pull(cpu);
971 /* This actually kills the CPU. */
974 tick_cleanup_dead_cpu(cpu);
978 static int notify_dead(unsigned int cpu)
980 cpu_notify_nofail(CPU_DEAD, cpu);
981 check_for_tasks(cpu);
985 static void cpuhp_complete_idle_dead(void *arg)
987 struct cpuhp_cpu_state *st = arg;
992 void cpuhp_report_idle_dead(void)
994 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
996 BUG_ON(st->state != CPUHP_AP_OFFLINE);
997 rcu_report_dead(smp_processor_id());
998 st->state = CPUHP_AP_IDLE_DEAD;
1000 * We cannot call complete after rcu_report_dead() so we delegate it
1003 smp_call_function_single(cpumask_first(cpu_online_mask),
1004 cpuhp_complete_idle_dead, st, 0);
1008 #define notify_down_prepare NULL
1009 #define takedown_cpu NULL
1010 #define notify_dead NULL
1013 #ifdef CONFIG_HOTPLUG_CPU
1015 /* Requires cpu_add_remove_lock to be held */
1016 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1017 enum cpuhp_state target)
1019 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1020 int prev_state, ret = 0;
1021 bool hasdied = false;
1023 if (num_online_cpus() == 1)
1026 if (!cpu_present(cpu))
1029 cpu_hotplug_begin();
1031 cpuhp_tasks_frozen = tasks_frozen;
1033 prev_state = st->state;
1034 st->target = target;
1036 * If the current CPU state is in the range of the AP hotplug thread,
1037 * then we need to kick the thread.
1039 if (st->state > CPUHP_TEARDOWN_CPU) {
1040 ret = cpuhp_kick_ap_work(cpu);
1042 * The AP side has done the error rollback already. Just
1043 * return the error code..
1049 * We might have stopped still in the range of the AP hotplug
1050 * thread. Nothing to do anymore.
1052 if (st->state > CPUHP_TEARDOWN_CPU)
1056 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1057 * to do the further cleanups.
1059 ret = cpuhp_down_callbacks(cpu, st, target);
1060 if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1061 st->target = prev_state;
1062 st->rollback = true;
1063 cpuhp_kick_ap_work(cpu);
1066 hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
1069 /* This post dead nonsense must die */
1070 if (!ret && hasdied)
1071 cpu_notify_nofail(CPU_POST_DEAD, cpu);
1076 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1078 if (cpu_hotplug_disabled)
1080 return _cpu_down(cpu, 0, target);
1083 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1087 cpu_maps_update_begin();
1088 err = cpu_down_maps_locked(cpu, target);
1089 cpu_maps_update_done();
1092 int cpu_down(unsigned int cpu)
1094 return do_cpu_down(cpu, CPUHP_OFFLINE);
1096 EXPORT_SYMBOL(cpu_down);
1097 #endif /*CONFIG_HOTPLUG_CPU*/
1100 * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1101 * @cpu: cpu that just started
1103 * It must be called by the arch code on the new cpu, before the new cpu
1104 * enables interrupts and before the "boot" cpu returns from __cpu_up().
1106 void notify_cpu_starting(unsigned int cpu)
1108 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1109 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1111 rcu_cpu_starting(cpu); /* Enables RCU usage on this CPU. */
1112 st->booted_once = true;
1113 while (st->state < target) {
1115 cpuhp_invoke_callback(cpu, st->state, true, NULL);
1120 * Called from the idle task. Wake up the controlling task which brings the
1121 * hotplug thread of the upcoming CPU up and then delegates the rest of the
1122 * online bringup to the hotplug thread.
1124 void cpuhp_online_idle(enum cpuhp_state state)
1126 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1128 /* Happens for the boot cpu */
1129 if (state != CPUHP_AP_ONLINE_IDLE)
1133 * Unpart the stopper thread before we start the idle loop (and start
1134 * scheduling); this ensures the stopper task is always available.
1136 stop_machine_unpark(smp_processor_id());
1138 st->state = CPUHP_AP_ONLINE_IDLE;
1139 complete(&st->done);
1142 /* Requires cpu_add_remove_lock to be held */
1143 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1145 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1146 struct task_struct *idle;
1149 cpu_hotplug_begin();
1151 if (!cpu_present(cpu)) {
1157 * The caller of do_cpu_up might have raced with another
1158 * caller. Ignore it for now.
1160 if (st->state >= target)
1163 if (st->state == CPUHP_OFFLINE) {
1164 /* Let it fail before we try to bring the cpu up */
1165 idle = idle_thread_get(cpu);
1167 ret = PTR_ERR(idle);
1172 cpuhp_tasks_frozen = tasks_frozen;
1174 st->target = target;
1176 * If the current CPU state is in the range of the AP hotplug thread,
1177 * then we need to kick the thread once more.
1179 if (st->state > CPUHP_BRINGUP_CPU) {
1180 ret = cpuhp_kick_ap_work(cpu);
1182 * The AP side has done the error rollback already. Just
1183 * return the error code..
1190 * Try to reach the target state. We max out on the BP at
1191 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1192 * responsible for bringing it up to the target state.
1194 target = min((int)target, CPUHP_BRINGUP_CPU);
1195 ret = cpuhp_up_callbacks(cpu, st, target);
1202 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1206 if (!cpu_possible(cpu)) {
1207 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1209 #if defined(CONFIG_IA64)
1210 pr_err("please check additional_cpus= boot parameter\n");
1215 err = try_online_node(cpu_to_node(cpu));
1219 cpu_maps_update_begin();
1221 if (cpu_hotplug_disabled) {
1225 if (!cpu_smt_allowed(cpu)) {
1230 err = _cpu_up(cpu, 0, target);
1232 cpu_maps_update_done();
1236 int cpu_up(unsigned int cpu)
1238 return do_cpu_up(cpu, CPUHP_ONLINE);
1240 EXPORT_SYMBOL_GPL(cpu_up);
1242 #ifdef CONFIG_PM_SLEEP_SMP
1243 static cpumask_var_t frozen_cpus;
1245 int freeze_secondary_cpus(int primary)
1249 cpu_maps_update_begin();
1250 if (!cpu_online(primary))
1251 primary = cpumask_first(cpu_online_mask);
1253 * We take down all of the non-boot CPUs in one shot to avoid races
1254 * with the userspace trying to use the CPU hotplug at the same time
1256 cpumask_clear(frozen_cpus);
1258 pr_info("Disabling non-boot CPUs ...\n");
1259 for_each_online_cpu(cpu) {
1262 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1263 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1264 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1266 cpumask_set_cpu(cpu, frozen_cpus);
1268 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1274 BUG_ON(num_online_cpus() > 1);
1276 pr_err("Non-boot CPUs are not disabled\n");
1279 * Make sure the CPUs won't be enabled by someone else. We need to do
1280 * this even in case of failure as all disable_nonboot_cpus() users are
1281 * supposed to do enable_nonboot_cpus() on the failure path.
1283 cpu_hotplug_disabled++;
1285 cpu_maps_update_done();
1289 void __weak arch_enable_nonboot_cpus_begin(void)
1293 void __weak arch_enable_nonboot_cpus_end(void)
1297 void enable_nonboot_cpus(void)
1301 /* Allow everyone to use the CPU hotplug again */
1302 cpu_maps_update_begin();
1303 __cpu_hotplug_enable();
1304 if (cpumask_empty(frozen_cpus))
1307 pr_info("Enabling non-boot CPUs ...\n");
1309 arch_enable_nonboot_cpus_begin();
1311 for_each_cpu(cpu, frozen_cpus) {
1312 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1313 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1314 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1316 pr_info("CPU%d is up\n", cpu);
1319 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1322 arch_enable_nonboot_cpus_end();
1324 cpumask_clear(frozen_cpus);
1326 cpu_maps_update_done();
1329 static int __init alloc_frozen_cpus(void)
1331 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1335 core_initcall(alloc_frozen_cpus);
1338 * When callbacks for CPU hotplug notifications are being executed, we must
1339 * ensure that the state of the system with respect to the tasks being frozen
1340 * or not, as reported by the notification, remains unchanged *throughout the
1341 * duration* of the execution of the callbacks.
1342 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1344 * This synchronization is implemented by mutually excluding regular CPU
1345 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1346 * Hibernate notifications.
1349 cpu_hotplug_pm_callback(struct notifier_block *nb,
1350 unsigned long action, void *ptr)
1354 case PM_SUSPEND_PREPARE:
1355 case PM_HIBERNATION_PREPARE:
1356 cpu_hotplug_disable();
1359 case PM_POST_SUSPEND:
1360 case PM_POST_HIBERNATION:
1361 cpu_hotplug_enable();
1372 static int __init cpu_hotplug_pm_sync_init(void)
1375 * cpu_hotplug_pm_callback has higher priority than x86
1376 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1377 * to disable cpu hotplug to avoid cpu hotplug race.
1379 pm_notifier(cpu_hotplug_pm_callback, 0);
1382 core_initcall(cpu_hotplug_pm_sync_init);
1384 #endif /* CONFIG_PM_SLEEP_SMP */
1386 #endif /* CONFIG_SMP */
1388 /* Boot processor state steps */
1389 static struct cpuhp_step cpuhp_bp_states[] = {
1392 .startup.single = NULL,
1393 .teardown.single = NULL,
1396 [CPUHP_CREATE_THREADS]= {
1397 .name = "threads:prepare",
1398 .startup.single = smpboot_create_threads,
1399 .teardown.single = NULL,
1402 [CPUHP_PERF_PREPARE] = {
1403 .name = "perf:prepare",
1404 .startup.single = perf_event_init_cpu,
1405 .teardown.single = perf_event_exit_cpu,
1407 [CPUHP_WORKQUEUE_PREP] = {
1408 .name = "workqueue:prepare",
1409 .startup.single = workqueue_prepare_cpu,
1410 .teardown.single = NULL,
1412 [CPUHP_HRTIMERS_PREPARE] = {
1413 .name = "hrtimers:prepare",
1414 .startup.single = hrtimers_prepare_cpu,
1415 .teardown.single = hrtimers_dead_cpu,
1417 [CPUHP_SMPCFD_PREPARE] = {
1418 .name = "smpcfd:prepare",
1419 .startup.single = smpcfd_prepare_cpu,
1420 .teardown.single = smpcfd_dead_cpu,
1422 [CPUHP_RELAY_PREPARE] = {
1423 .name = "relay:prepare",
1424 .startup.single = relay_prepare_cpu,
1425 .teardown.single = NULL,
1427 [CPUHP_SLAB_PREPARE] = {
1428 .name = "slab:prepare",
1429 .startup.single = slab_prepare_cpu,
1430 .teardown.single = slab_dead_cpu,
1432 [CPUHP_RCUTREE_PREP] = {
1433 .name = "RCU/tree:prepare",
1434 .startup.single = rcutree_prepare_cpu,
1435 .teardown.single = rcutree_dead_cpu,
1438 * Preparatory and dead notifiers. Will be replaced once the notifiers
1439 * are converted to states.
1441 [CPUHP_NOTIFY_PREPARE] = {
1442 .name = "notify:prepare",
1443 .startup.single = notify_prepare,
1444 .teardown.single = notify_dead,
1449 * On the tear-down path, timers_dead_cpu() must be invoked
1450 * before blk_mq_queue_reinit_notify() from notify_dead(),
1451 * otherwise a RCU stall occurs.
1453 [CPUHP_TIMERS_PREPARE] = {
1454 .name = "timers:dead",
1455 .startup.single = timers_prepare_cpu,
1456 .teardown.single = timers_dead_cpu,
1458 /* Kicks the plugged cpu into life */
1459 [CPUHP_BRINGUP_CPU] = {
1460 .name = "cpu:bringup",
1461 .startup.single = bringup_cpu,
1462 .teardown.single = NULL,
1466 * Handled on controll processor until the plugged processor manages
1469 [CPUHP_TEARDOWN_CPU] = {
1470 .name = "cpu:teardown",
1471 .startup.single = NULL,
1472 .teardown.single = takedown_cpu,
1476 [CPUHP_BRINGUP_CPU] = { },
1480 /* Application processor state steps */
1481 static struct cpuhp_step cpuhp_ap_states[] = {
1483 /* Final state before CPU kills itself */
1484 [CPUHP_AP_IDLE_DEAD] = {
1485 .name = "idle:dead",
1488 * Last state before CPU enters the idle loop to die. Transient state
1489 * for synchronization.
1491 [CPUHP_AP_OFFLINE] = {
1492 .name = "ap:offline",
1495 /* First state is scheduler control. Interrupts are disabled */
1496 [CPUHP_AP_SCHED_STARTING] = {
1497 .name = "sched:starting",
1498 .startup.single = sched_cpu_starting,
1499 .teardown.single = sched_cpu_dying,
1501 [CPUHP_AP_RCUTREE_DYING] = {
1502 .name = "RCU/tree:dying",
1503 .startup.single = NULL,
1504 .teardown.single = rcutree_dying_cpu,
1506 [CPUHP_AP_SMPCFD_DYING] = {
1507 .name = "smpcfd:dying",
1508 .startup.single = NULL,
1509 .teardown.single = smpcfd_dying_cpu,
1511 /* Entry state on starting. Interrupts enabled from here on. Transient
1512 * state for synchronsization */
1513 [CPUHP_AP_ONLINE] = {
1514 .name = "ap:online",
1516 /* Handle smpboot threads park/unpark */
1517 [CPUHP_AP_SMPBOOT_THREADS] = {
1518 .name = "smpboot/threads:online",
1519 .startup.single = smpboot_unpark_threads,
1520 .teardown.single = smpboot_park_threads,
1522 [CPUHP_AP_PERF_ONLINE] = {
1523 .name = "perf:online",
1524 .startup.single = perf_event_init_cpu,
1525 .teardown.single = perf_event_exit_cpu,
1527 [CPUHP_AP_WORKQUEUE_ONLINE] = {
1528 .name = "workqueue:online",
1529 .startup.single = workqueue_online_cpu,
1530 .teardown.single = workqueue_offline_cpu,
1532 [CPUHP_AP_RCUTREE_ONLINE] = {
1533 .name = "RCU/tree:online",
1534 .startup.single = rcutree_online_cpu,
1535 .teardown.single = rcutree_offline_cpu,
1539 * Online/down_prepare notifiers. Will be removed once the notifiers
1540 * are converted to states.
1542 [CPUHP_AP_NOTIFY_ONLINE] = {
1543 .name = "notify:online",
1544 .startup.single = notify_online,
1545 .teardown.single = notify_down_prepare,
1550 * The dynamically registered state space is here
1554 /* Last state is scheduler control setting the cpu active */
1555 [CPUHP_AP_ACTIVE] = {
1556 .name = "sched:active",
1557 .startup.single = sched_cpu_activate,
1558 .teardown.single = sched_cpu_deactivate,
1562 /* CPU is fully up and running. */
1565 .startup.single = NULL,
1566 .teardown.single = NULL,
1570 /* Sanity check for callbacks */
1571 static int cpuhp_cb_check(enum cpuhp_state state)
1573 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1578 static void cpuhp_store_callbacks(enum cpuhp_state state,
1580 int (*startup)(unsigned int cpu),
1581 int (*teardown)(unsigned int cpu),
1582 bool multi_instance)
1584 /* (Un)Install the callbacks for further cpu hotplug operations */
1585 struct cpuhp_step *sp;
1587 sp = cpuhp_get_step(state);
1588 sp->startup.single = startup;
1589 sp->teardown.single = teardown;
1591 sp->multi_instance = multi_instance;
1592 INIT_HLIST_HEAD(&sp->list);
1595 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1597 return cpuhp_get_step(state)->teardown.single;
1601 * Call the startup/teardown function for a step either on the AP or
1602 * on the current CPU.
1604 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1605 struct hlist_node *node)
1607 struct cpuhp_step *sp = cpuhp_get_step(state);
1610 if ((bringup && !sp->startup.single) ||
1611 (!bringup && !sp->teardown.single))
1614 * The non AP bound callbacks can fail on bringup. On teardown
1615 * e.g. module removal we crash for now.
1618 if (cpuhp_is_ap_state(state))
1619 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1621 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1623 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1625 BUG_ON(ret && !bringup);
1630 * Called from __cpuhp_setup_state on a recoverable failure.
1632 * Note: The teardown callbacks for rollback are not allowed to fail!
1634 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1635 struct hlist_node *node)
1639 /* Roll back the already executed steps on the other cpus */
1640 for_each_present_cpu(cpu) {
1641 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1642 int cpustate = st->state;
1644 if (cpu >= failedcpu)
1647 /* Did we invoke the startup call on that cpu ? */
1648 if (cpustate >= state)
1649 cpuhp_issue_call(cpu, state, false, node);
1654 * Returns a free for dynamic slot assignment of the Online state. The states
1655 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1656 * by having no name assigned.
1658 static int cpuhp_reserve_state(enum cpuhp_state state)
1662 for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1663 if (cpuhp_ap_states[i].name)
1666 cpuhp_ap_states[i].name = "Reserved";
1669 WARN(1, "No more dynamic states available for CPU hotplug\n");
1673 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1676 struct cpuhp_step *sp;
1680 sp = cpuhp_get_step(state);
1681 if (sp->multi_instance == false)
1685 mutex_lock(&cpuhp_state_mutex);
1687 if (!invoke || !sp->startup.multi)
1691 * Try to call the startup callback for each present cpu
1692 * depending on the hotplug state of the cpu.
1694 for_each_present_cpu(cpu) {
1695 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1696 int cpustate = st->state;
1698 if (cpustate < state)
1701 ret = cpuhp_issue_call(cpu, state, true, node);
1703 if (sp->teardown.multi)
1704 cpuhp_rollback_install(cpu, state, node);
1710 hlist_add_head(node, &sp->list);
1713 mutex_unlock(&cpuhp_state_mutex);
1717 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1720 * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1721 * @state: The state to setup
1722 * @invoke: If true, the startup function is invoked for cpus where
1723 * cpu state >= @state
1724 * @startup: startup callback function
1725 * @teardown: teardown callback function
1727 * Returns 0 if successful, otherwise a proper error code
1729 int __cpuhp_setup_state(enum cpuhp_state state,
1730 const char *name, bool invoke,
1731 int (*startup)(unsigned int cpu),
1732 int (*teardown)(unsigned int cpu),
1733 bool multi_instance)
1738 if (cpuhp_cb_check(state) || !name)
1742 mutex_lock(&cpuhp_state_mutex);
1744 /* currently assignments for the ONLINE state are possible */
1745 if (state == CPUHP_AP_ONLINE_DYN) {
1747 ret = cpuhp_reserve_state(state);
1753 cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1755 if (!invoke || !startup)
1759 * Try to call the startup callback for each present cpu
1760 * depending on the hotplug state of the cpu.
1762 for_each_present_cpu(cpu) {
1763 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1764 int cpustate = st->state;
1766 if (cpustate < state)
1769 ret = cpuhp_issue_call(cpu, state, true, NULL);
1772 cpuhp_rollback_install(cpu, state, NULL);
1773 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1778 mutex_unlock(&cpuhp_state_mutex);
1781 if (!ret && dyn_state)
1785 EXPORT_SYMBOL(__cpuhp_setup_state);
1787 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1788 struct hlist_node *node, bool invoke)
1790 struct cpuhp_step *sp = cpuhp_get_step(state);
1793 BUG_ON(cpuhp_cb_check(state));
1795 if (!sp->multi_instance)
1799 mutex_lock(&cpuhp_state_mutex);
1801 if (!invoke || !cpuhp_get_teardown_cb(state))
1804 * Call the teardown callback for each present cpu depending
1805 * on the hotplug state of the cpu. This function is not
1806 * allowed to fail currently!
1808 for_each_present_cpu(cpu) {
1809 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1810 int cpustate = st->state;
1812 if (cpustate >= state)
1813 cpuhp_issue_call(cpu, state, false, node);
1818 mutex_unlock(&cpuhp_state_mutex);
1823 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1825 * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1826 * @state: The state to remove
1827 * @invoke: If true, the teardown function is invoked for cpus where
1828 * cpu state >= @state
1830 * The teardown callback is currently not allowed to fail. Think
1831 * about module removal!
1833 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1835 struct cpuhp_step *sp = cpuhp_get_step(state);
1838 BUG_ON(cpuhp_cb_check(state));
1841 mutex_lock(&cpuhp_state_mutex);
1843 if (sp->multi_instance) {
1844 WARN(!hlist_empty(&sp->list),
1845 "Error: Removing state %d which has instances left.\n",
1850 if (!invoke || !cpuhp_get_teardown_cb(state))
1854 * Call the teardown callback for each present cpu depending
1855 * on the hotplug state of the cpu. This function is not
1856 * allowed to fail currently!
1858 for_each_present_cpu(cpu) {
1859 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1860 int cpustate = st->state;
1862 if (cpustate >= state)
1863 cpuhp_issue_call(cpu, state, false, NULL);
1866 cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1867 mutex_unlock(&cpuhp_state_mutex);
1870 EXPORT_SYMBOL(__cpuhp_remove_state);
1872 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1873 static ssize_t show_cpuhp_state(struct device *dev,
1874 struct device_attribute *attr, char *buf)
1876 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1878 return sprintf(buf, "%d\n", st->state);
1880 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1882 static ssize_t write_cpuhp_target(struct device *dev,
1883 struct device_attribute *attr,
1884 const char *buf, size_t count)
1886 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1887 struct cpuhp_step *sp;
1890 ret = kstrtoint(buf, 10, &target);
1894 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1895 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1898 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1902 ret = lock_device_hotplug_sysfs();
1906 mutex_lock(&cpuhp_state_mutex);
1907 sp = cpuhp_get_step(target);
1908 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1909 mutex_unlock(&cpuhp_state_mutex);
1913 if (st->state < target)
1914 ret = do_cpu_up(dev->id, target);
1916 ret = do_cpu_down(dev->id, target);
1918 unlock_device_hotplug();
1919 return ret ? ret : count;
1922 static ssize_t show_cpuhp_target(struct device *dev,
1923 struct device_attribute *attr, char *buf)
1925 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1927 return sprintf(buf, "%d\n", st->target);
1929 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1931 static struct attribute *cpuhp_cpu_attrs[] = {
1932 &dev_attr_state.attr,
1933 &dev_attr_target.attr,
1937 static struct attribute_group cpuhp_cpu_attr_group = {
1938 .attrs = cpuhp_cpu_attrs,
1943 static ssize_t show_cpuhp_states(struct device *dev,
1944 struct device_attribute *attr, char *buf)
1946 ssize_t cur, res = 0;
1949 mutex_lock(&cpuhp_state_mutex);
1950 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1951 struct cpuhp_step *sp = cpuhp_get_step(i);
1954 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1959 mutex_unlock(&cpuhp_state_mutex);
1962 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1964 static struct attribute *cpuhp_cpu_root_attrs[] = {
1965 &dev_attr_states.attr,
1969 static struct attribute_group cpuhp_cpu_root_attr_group = {
1970 .attrs = cpuhp_cpu_root_attrs,
1975 #ifdef CONFIG_HOTPLUG_SMT
1977 static const char *smt_states[] = {
1978 [CPU_SMT_ENABLED] = "on",
1979 [CPU_SMT_DISABLED] = "off",
1980 [CPU_SMT_FORCE_DISABLED] = "forceoff",
1981 [CPU_SMT_NOT_SUPPORTED] = "notsupported",
1985 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
1987 return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
1990 static void cpuhp_offline_cpu_device(unsigned int cpu)
1992 struct device *dev = get_cpu_device(cpu);
1994 dev->offline = true;
1995 /* Tell user space about the state change */
1996 kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
1999 static void cpuhp_online_cpu_device(unsigned int cpu)
2001 struct device *dev = get_cpu_device(cpu);
2003 dev->offline = false;
2004 /* Tell user space about the state change */
2005 kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2008 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2012 cpu_maps_update_begin();
2013 for_each_online_cpu(cpu) {
2014 if (topology_is_primary_thread(cpu))
2016 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2020 * As this needs to hold the cpu maps lock it's impossible
2021 * to call device_offline() because that ends up calling
2022 * cpu_down() which takes cpu maps lock. cpu maps lock
2023 * needs to be held as this might race against in kernel
2024 * abusers of the hotplug machinery (thermal management).
2026 * So nothing would update device:offline state. That would
2027 * leave the sysfs entry stale and prevent onlining after
2028 * smt control has been changed to 'off' again. This is
2029 * called under the sysfs hotplug lock, so it is properly
2030 * serialized against the regular offline usage.
2032 cpuhp_offline_cpu_device(cpu);
2035 cpu_smt_control = ctrlval;
2036 cpu_maps_update_done();
2040 int cpuhp_smt_enable(void)
2044 cpu_maps_update_begin();
2045 cpu_smt_control = CPU_SMT_ENABLED;
2046 for_each_present_cpu(cpu) {
2047 /* Skip online CPUs and CPUs on offline nodes */
2048 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2050 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2053 /* See comment in cpuhp_smt_disable() */
2054 cpuhp_online_cpu_device(cpu);
2056 cpu_maps_update_done();
2061 store_smt_control(struct device *dev, struct device_attribute *attr,
2062 const char *buf, size_t count)
2066 if (sysfs_streq(buf, "on"))
2067 ctrlval = CPU_SMT_ENABLED;
2068 else if (sysfs_streq(buf, "off"))
2069 ctrlval = CPU_SMT_DISABLED;
2070 else if (sysfs_streq(buf, "forceoff"))
2071 ctrlval = CPU_SMT_FORCE_DISABLED;
2075 if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2078 if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2081 ret = lock_device_hotplug_sysfs();
2085 if (ctrlval != cpu_smt_control) {
2087 case CPU_SMT_ENABLED:
2088 ret = cpuhp_smt_enable();
2090 case CPU_SMT_DISABLED:
2091 case CPU_SMT_FORCE_DISABLED:
2092 ret = cpuhp_smt_disable(ctrlval);
2097 unlock_device_hotplug();
2098 return ret ? ret : count;
2100 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2103 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2105 bool active = topology_max_smt_threads() > 1;
2107 return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2109 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2111 static struct attribute *cpuhp_smt_attrs[] = {
2112 &dev_attr_control.attr,
2113 &dev_attr_active.attr,
2117 static const struct attribute_group cpuhp_smt_attr_group = {
2118 .attrs = cpuhp_smt_attrs,
2123 static int __init cpu_smt_state_init(void)
2125 return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2126 &cpuhp_smt_attr_group);
2130 static inline int cpu_smt_state_init(void) { return 0; }
2133 static int __init cpuhp_sysfs_init(void)
2137 ret = cpu_smt_state_init();
2141 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2142 &cpuhp_cpu_root_attr_group);
2146 for_each_possible_cpu(cpu) {
2147 struct device *dev = get_cpu_device(cpu);
2151 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2157 device_initcall(cpuhp_sysfs_init);
2161 * cpu_bit_bitmap[] is a special, "compressed" data structure that
2162 * represents all NR_CPUS bits binary values of 1<<nr.
2164 * It is used by cpumask_of() to get a constant address to a CPU
2165 * mask value that has a single bit set only.
2168 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2169 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
2170 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2171 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2172 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2174 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2176 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
2177 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
2178 #if BITS_PER_LONG > 32
2179 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
2180 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
2183 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2185 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2186 EXPORT_SYMBOL(cpu_all_bits);
2188 #ifdef CONFIG_INIT_ALL_POSSIBLE
2189 struct cpumask __cpu_possible_mask __read_mostly
2192 struct cpumask __cpu_possible_mask __read_mostly;
2194 EXPORT_SYMBOL(__cpu_possible_mask);
2196 struct cpumask __cpu_online_mask __read_mostly;
2197 EXPORT_SYMBOL(__cpu_online_mask);
2199 struct cpumask __cpu_present_mask __read_mostly;
2200 EXPORT_SYMBOL(__cpu_present_mask);
2202 struct cpumask __cpu_active_mask __read_mostly;
2203 EXPORT_SYMBOL(__cpu_active_mask);
2205 void init_cpu_present(const struct cpumask *src)
2207 cpumask_copy(&__cpu_present_mask, src);
2210 void init_cpu_possible(const struct cpumask *src)
2212 cpumask_copy(&__cpu_possible_mask, src);
2215 void init_cpu_online(const struct cpumask *src)
2217 cpumask_copy(&__cpu_online_mask, src);
2221 * Activate the first processor.
2223 void __init boot_cpu_init(void)
2225 int cpu = smp_processor_id();
2227 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2228 set_cpu_online(cpu, true);
2229 set_cpu_active(cpu, true);
2230 set_cpu_present(cpu, true);
2231 set_cpu_possible(cpu, true);
2235 * Must be called _AFTER_ setting up the per_cpu areas
2237 void __init boot_cpu_hotplug_init(void)
2240 this_cpu_write(cpuhp_state.booted_once, true);
2242 this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2246 * These are used for a global "mitigations=" cmdline option for toggling
2247 * optional CPU mitigations.
2249 enum cpu_mitigations {
2250 CPU_MITIGATIONS_OFF,
2251 CPU_MITIGATIONS_AUTO,
2252 CPU_MITIGATIONS_AUTO_NOSMT,
2255 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2256 CPU_MITIGATIONS_AUTO;
2258 static int __init mitigations_parse_cmdline(char *arg)
2260 if (!strcmp(arg, "off"))
2261 cpu_mitigations = CPU_MITIGATIONS_OFF;
2262 else if (!strcmp(arg, "auto"))
2263 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2264 else if (!strcmp(arg, "auto,nosmt"))
2265 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2267 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2272 early_param("mitigations", mitigations_parse_cmdline);
2274 /* mitigations=off */
2275 bool cpu_mitigations_off(void)
2277 return cpu_mitigations == CPU_MITIGATIONS_OFF;
2279 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2281 /* mitigations=auto,nosmt */
2282 bool cpu_mitigations_auto_nosmt(void)
2284 return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2286 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);