fc28e17940e02e84f23404818913958af97b3ffd
[sfrench/cifs-2.6.git] / kernel / cpu.c
1 /* CPU control.
2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
3  *
4  * This code is licenced under the GPL.
5  */
6 #include <linux/proc_fs.h>
7 #include <linux/smp.h>
8 #include <linux/init.h>
9 #include <linux/notifier.h>
10 #include <linux/sched/signal.h>
11 #include <linux/sched/hotplug.h>
12 #include <linux/sched/isolation.h>
13 #include <linux/sched/task.h>
14 #include <linux/sched/smt.h>
15 #include <linux/unistd.h>
16 #include <linux/cpu.h>
17 #include <linux/oom.h>
18 #include <linux/rcupdate.h>
19 #include <linux/export.h>
20 #include <linux/bug.h>
21 #include <linux/kthread.h>
22 #include <linux/stop_machine.h>
23 #include <linux/mutex.h>
24 #include <linux/gfp.h>
25 #include <linux/suspend.h>
26 #include <linux/lockdep.h>
27 #include <linux/tick.h>
28 #include <linux/irq.h>
29 #include <linux/nmi.h>
30 #include <linux/smpboot.h>
31 #include <linux/relay.h>
32 #include <linux/slab.h>
33 #include <linux/percpu-rwsem.h>
34
35 #include <trace/events/power.h>
36 #define CREATE_TRACE_POINTS
37 #include <trace/events/cpuhp.h>
38
39 #include "smpboot.h"
40
41 /**
42  * cpuhp_cpu_state - Per cpu hotplug state storage
43  * @state:      The current cpu state
44  * @target:     The target state
45  * @thread:     Pointer to the hotplug thread
46  * @should_run: Thread should execute
47  * @rollback:   Perform a rollback
48  * @single:     Single callback invocation
49  * @bringup:    Single callback bringup or teardown selector
50  * @cb_state:   The state for a single callback (install/uninstall)
51  * @result:     Result of the operation
52  * @done_up:    Signal completion to the issuer of the task for cpu-up
53  * @done_down:  Signal completion to the issuer of the task for cpu-down
54  */
55 struct cpuhp_cpu_state {
56         enum cpuhp_state        state;
57         enum cpuhp_state        target;
58         enum cpuhp_state        fail;
59 #ifdef CONFIG_SMP
60         struct task_struct      *thread;
61         bool                    should_run;
62         bool                    rollback;
63         bool                    single;
64         bool                    bringup;
65         struct hlist_node       *node;
66         struct hlist_node       *last;
67         enum cpuhp_state        cb_state;
68         int                     result;
69         struct completion       done_up;
70         struct completion       done_down;
71 #endif
72 };
73
74 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
75         .fail = CPUHP_INVALID,
76 };
77
78 #ifdef CONFIG_SMP
79 cpumask_t cpus_booted_once_mask;
80 #endif
81
82 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
83 static struct lockdep_map cpuhp_state_up_map =
84         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
85 static struct lockdep_map cpuhp_state_down_map =
86         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
87
88
89 static inline void cpuhp_lock_acquire(bool bringup)
90 {
91         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
92 }
93
94 static inline void cpuhp_lock_release(bool bringup)
95 {
96         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
97 }
98 #else
99
100 static inline void cpuhp_lock_acquire(bool bringup) { }
101 static inline void cpuhp_lock_release(bool bringup) { }
102
103 #endif
104
105 /**
106  * cpuhp_step - Hotplug state machine step
107  * @name:       Name of the step
108  * @startup:    Startup function of the step
109  * @teardown:   Teardown function of the step
110  * @cant_stop:  Bringup/teardown can't be stopped at this step
111  */
112 struct cpuhp_step {
113         const char              *name;
114         union {
115                 int             (*single)(unsigned int cpu);
116                 int             (*multi)(unsigned int cpu,
117                                          struct hlist_node *node);
118         } startup;
119         union {
120                 int             (*single)(unsigned int cpu);
121                 int             (*multi)(unsigned int cpu,
122                                          struct hlist_node *node);
123         } teardown;
124         struct hlist_head       list;
125         bool                    cant_stop;
126         bool                    multi_instance;
127 };
128
129 static DEFINE_MUTEX(cpuhp_state_mutex);
130 static struct cpuhp_step cpuhp_hp_states[];
131
132 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
133 {
134         return cpuhp_hp_states + state;
135 }
136
137 /**
138  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
139  * @cpu:        The cpu for which the callback should be invoked
140  * @state:      The state to do callbacks for
141  * @bringup:    True if the bringup callback should be invoked
142  * @node:       For multi-instance, do a single entry callback for install/remove
143  * @lastp:      For multi-instance rollback, remember how far we got
144  *
145  * Called from cpu hotplug and from the state register machinery.
146  */
147 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
148                                  bool bringup, struct hlist_node *node,
149                                  struct hlist_node **lastp)
150 {
151         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
152         struct cpuhp_step *step = cpuhp_get_step(state);
153         int (*cbm)(unsigned int cpu, struct hlist_node *node);
154         int (*cb)(unsigned int cpu);
155         int ret, cnt;
156
157         if (st->fail == state) {
158                 st->fail = CPUHP_INVALID;
159
160                 if (!(bringup ? step->startup.single : step->teardown.single))
161                         return 0;
162
163                 return -EAGAIN;
164         }
165
166         if (!step->multi_instance) {
167                 WARN_ON_ONCE(lastp && *lastp);
168                 cb = bringup ? step->startup.single : step->teardown.single;
169                 if (!cb)
170                         return 0;
171                 trace_cpuhp_enter(cpu, st->target, state, cb);
172                 ret = cb(cpu);
173                 trace_cpuhp_exit(cpu, st->state, state, ret);
174                 return ret;
175         }
176         cbm = bringup ? step->startup.multi : step->teardown.multi;
177         if (!cbm)
178                 return 0;
179
180         /* Single invocation for instance add/remove */
181         if (node) {
182                 WARN_ON_ONCE(lastp && *lastp);
183                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
184                 ret = cbm(cpu, node);
185                 trace_cpuhp_exit(cpu, st->state, state, ret);
186                 return ret;
187         }
188
189         /* State transition. Invoke on all instances */
190         cnt = 0;
191         hlist_for_each(node, &step->list) {
192                 if (lastp && node == *lastp)
193                         break;
194
195                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
196                 ret = cbm(cpu, node);
197                 trace_cpuhp_exit(cpu, st->state, state, ret);
198                 if (ret) {
199                         if (!lastp)
200                                 goto err;
201
202                         *lastp = node;
203                         return ret;
204                 }
205                 cnt++;
206         }
207         if (lastp)
208                 *lastp = NULL;
209         return 0;
210 err:
211         /* Rollback the instances if one failed */
212         cbm = !bringup ? step->startup.multi : step->teardown.multi;
213         if (!cbm)
214                 return ret;
215
216         hlist_for_each(node, &step->list) {
217                 if (!cnt--)
218                         break;
219
220                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
221                 ret = cbm(cpu, node);
222                 trace_cpuhp_exit(cpu, st->state, state, ret);
223                 /*
224                  * Rollback must not fail,
225                  */
226                 WARN_ON_ONCE(ret);
227         }
228         return ret;
229 }
230
231 #ifdef CONFIG_SMP
232 static bool cpuhp_is_ap_state(enum cpuhp_state state)
233 {
234         /*
235          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
236          * purposes as that state is handled explicitly in cpu_down.
237          */
238         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
239 }
240
241 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
242 {
243         struct completion *done = bringup ? &st->done_up : &st->done_down;
244         wait_for_completion(done);
245 }
246
247 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
248 {
249         struct completion *done = bringup ? &st->done_up : &st->done_down;
250         complete(done);
251 }
252
253 /*
254  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
255  */
256 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
257 {
258         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
259 }
260
261 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
262 static DEFINE_MUTEX(cpu_add_remove_lock);
263 bool cpuhp_tasks_frozen;
264 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
265
266 /*
267  * The following two APIs (cpu_maps_update_begin/done) must be used when
268  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
269  */
270 void cpu_maps_update_begin(void)
271 {
272         mutex_lock(&cpu_add_remove_lock);
273 }
274
275 void cpu_maps_update_done(void)
276 {
277         mutex_unlock(&cpu_add_remove_lock);
278 }
279
280 /*
281  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
282  * Should always be manipulated under cpu_add_remove_lock
283  */
284 static int cpu_hotplug_disabled;
285
286 #ifdef CONFIG_HOTPLUG_CPU
287
288 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
289
290 void cpus_read_lock(void)
291 {
292         percpu_down_read(&cpu_hotplug_lock);
293 }
294 EXPORT_SYMBOL_GPL(cpus_read_lock);
295
296 int cpus_read_trylock(void)
297 {
298         return percpu_down_read_trylock(&cpu_hotplug_lock);
299 }
300 EXPORT_SYMBOL_GPL(cpus_read_trylock);
301
302 void cpus_read_unlock(void)
303 {
304         percpu_up_read(&cpu_hotplug_lock);
305 }
306 EXPORT_SYMBOL_GPL(cpus_read_unlock);
307
308 void cpus_write_lock(void)
309 {
310         percpu_down_write(&cpu_hotplug_lock);
311 }
312
313 void cpus_write_unlock(void)
314 {
315         percpu_up_write(&cpu_hotplug_lock);
316 }
317
318 void lockdep_assert_cpus_held(void)
319 {
320         /*
321          * We can't have hotplug operations before userspace starts running,
322          * and some init codepaths will knowingly not take the hotplug lock.
323          * This is all valid, so mute lockdep until it makes sense to report
324          * unheld locks.
325          */
326         if (system_state < SYSTEM_RUNNING)
327                 return;
328
329         percpu_rwsem_assert_held(&cpu_hotplug_lock);
330 }
331
332 static void lockdep_acquire_cpus_lock(void)
333 {
334         rwsem_acquire(&cpu_hotplug_lock.rw_sem.dep_map, 0, 0, _THIS_IP_);
335 }
336
337 static void lockdep_release_cpus_lock(void)
338 {
339         rwsem_release(&cpu_hotplug_lock.rw_sem.dep_map, 1, _THIS_IP_);
340 }
341
342 /*
343  * Wait for currently running CPU hotplug operations to complete (if any) and
344  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
345  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
346  * hotplug path before performing hotplug operations. So acquiring that lock
347  * guarantees mutual exclusion from any currently running hotplug operations.
348  */
349 void cpu_hotplug_disable(void)
350 {
351         cpu_maps_update_begin();
352         cpu_hotplug_disabled++;
353         cpu_maps_update_done();
354 }
355 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
356
357 static void __cpu_hotplug_enable(void)
358 {
359         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
360                 return;
361         cpu_hotplug_disabled--;
362 }
363
364 void cpu_hotplug_enable(void)
365 {
366         cpu_maps_update_begin();
367         __cpu_hotplug_enable();
368         cpu_maps_update_done();
369 }
370 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
371
372 #else
373
374 static void lockdep_acquire_cpus_lock(void)
375 {
376 }
377
378 static void lockdep_release_cpus_lock(void)
379 {
380 }
381
382 #endif  /* CONFIG_HOTPLUG_CPU */
383
384 /*
385  * Architectures that need SMT-specific errata handling during SMT hotplug
386  * should override this.
387  */
388 void __weak arch_smt_update(void) { }
389
390 #ifdef CONFIG_HOTPLUG_SMT
391 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
392
393 void __init cpu_smt_disable(bool force)
394 {
395         if (!cpu_smt_possible())
396                 return;
397
398         if (force) {
399                 pr_info("SMT: Force disabled\n");
400                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
401         } else {
402                 pr_info("SMT: disabled\n");
403                 cpu_smt_control = CPU_SMT_DISABLED;
404         }
405 }
406
407 /*
408  * The decision whether SMT is supported can only be done after the full
409  * CPU identification. Called from architecture code.
410  */
411 void __init cpu_smt_check_topology(void)
412 {
413         if (!topology_smt_supported())
414                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
415 }
416
417 static int __init smt_cmdline_disable(char *str)
418 {
419         cpu_smt_disable(str && !strcmp(str, "force"));
420         return 0;
421 }
422 early_param("nosmt", smt_cmdline_disable);
423
424 static inline bool cpu_smt_allowed(unsigned int cpu)
425 {
426         if (cpu_smt_control == CPU_SMT_ENABLED)
427                 return true;
428
429         if (topology_is_primary_thread(cpu))
430                 return true;
431
432         /*
433          * On x86 it's required to boot all logical CPUs at least once so
434          * that the init code can get a chance to set CR4.MCE on each
435          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
436          * core will shutdown the machine.
437          */
438         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
439 }
440
441 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
442 bool cpu_smt_possible(void)
443 {
444         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
445                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
446 }
447 EXPORT_SYMBOL_GPL(cpu_smt_possible);
448 #else
449 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
450 #endif
451
452 static inline enum cpuhp_state
453 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
454 {
455         enum cpuhp_state prev_state = st->state;
456
457         st->rollback = false;
458         st->last = NULL;
459
460         st->target = target;
461         st->single = false;
462         st->bringup = st->state < target;
463
464         return prev_state;
465 }
466
467 static inline void
468 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
469 {
470         st->rollback = true;
471
472         /*
473          * If we have st->last we need to undo partial multi_instance of this
474          * state first. Otherwise start undo at the previous state.
475          */
476         if (!st->last) {
477                 if (st->bringup)
478                         st->state--;
479                 else
480                         st->state++;
481         }
482
483         st->target = prev_state;
484         st->bringup = !st->bringup;
485 }
486
487 /* Regular hotplug invocation of the AP hotplug thread */
488 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
489 {
490         if (!st->single && st->state == st->target)
491                 return;
492
493         st->result = 0;
494         /*
495          * Make sure the above stores are visible before should_run becomes
496          * true. Paired with the mb() above in cpuhp_thread_fun()
497          */
498         smp_mb();
499         st->should_run = true;
500         wake_up_process(st->thread);
501         wait_for_ap_thread(st, st->bringup);
502 }
503
504 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
505 {
506         enum cpuhp_state prev_state;
507         int ret;
508
509         prev_state = cpuhp_set_state(st, target);
510         __cpuhp_kick_ap(st);
511         if ((ret = st->result)) {
512                 cpuhp_reset_state(st, prev_state);
513                 __cpuhp_kick_ap(st);
514         }
515
516         return ret;
517 }
518
519 static int bringup_wait_for_ap(unsigned int cpu)
520 {
521         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
522
523         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
524         wait_for_ap_thread(st, true);
525         if (WARN_ON_ONCE((!cpu_online(cpu))))
526                 return -ECANCELED;
527
528         /* Unpark the stopper thread and the hotplug thread of the target cpu */
529         stop_machine_unpark(cpu);
530         kthread_unpark(st->thread);
531
532         /*
533          * SMT soft disabling on X86 requires to bring the CPU out of the
534          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
535          * CPU marked itself as booted_once in notify_cpu_starting() so the
536          * cpu_smt_allowed() check will now return false if this is not the
537          * primary sibling.
538          */
539         if (!cpu_smt_allowed(cpu))
540                 return -ECANCELED;
541
542         if (st->target <= CPUHP_AP_ONLINE_IDLE)
543                 return 0;
544
545         return cpuhp_kick_ap(st, st->target);
546 }
547
548 static int bringup_cpu(unsigned int cpu)
549 {
550         struct task_struct *idle = idle_thread_get(cpu);
551         int ret;
552
553         /*
554          * Some architectures have to walk the irq descriptors to
555          * setup the vector space for the cpu which comes online.
556          * Prevent irq alloc/free across the bringup.
557          */
558         irq_lock_sparse();
559
560         /* Arch-specific enabling code. */
561         ret = __cpu_up(cpu, idle);
562         irq_unlock_sparse();
563         if (ret)
564                 return ret;
565         return bringup_wait_for_ap(cpu);
566 }
567
568 /*
569  * Hotplug state machine related functions
570  */
571
572 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
573 {
574         for (st->state--; st->state > st->target; st->state--)
575                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
576 }
577
578 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
579 {
580         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
581                 return true;
582         /*
583          * When CPU hotplug is disabled, then taking the CPU down is not
584          * possible because takedown_cpu() and the architecture and
585          * subsystem specific mechanisms are not available. So the CPU
586          * which would be completely unplugged again needs to stay around
587          * in the current state.
588          */
589         return st->state <= CPUHP_BRINGUP_CPU;
590 }
591
592 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
593                               enum cpuhp_state target)
594 {
595         enum cpuhp_state prev_state = st->state;
596         int ret = 0;
597
598         while (st->state < target) {
599                 st->state++;
600                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
601                 if (ret) {
602                         if (can_rollback_cpu(st)) {
603                                 st->target = prev_state;
604                                 undo_cpu_up(cpu, st);
605                         }
606                         break;
607                 }
608         }
609         return ret;
610 }
611
612 /*
613  * The cpu hotplug threads manage the bringup and teardown of the cpus
614  */
615 static void cpuhp_create(unsigned int cpu)
616 {
617         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
618
619         init_completion(&st->done_up);
620         init_completion(&st->done_down);
621 }
622
623 static int cpuhp_should_run(unsigned int cpu)
624 {
625         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
626
627         return st->should_run;
628 }
629
630 /*
631  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
632  * callbacks when a state gets [un]installed at runtime.
633  *
634  * Each invocation of this function by the smpboot thread does a single AP
635  * state callback.
636  *
637  * It has 3 modes of operation:
638  *  - single: runs st->cb_state
639  *  - up:     runs ++st->state, while st->state < st->target
640  *  - down:   runs st->state--, while st->state > st->target
641  *
642  * When complete or on error, should_run is cleared and the completion is fired.
643  */
644 static void cpuhp_thread_fun(unsigned int cpu)
645 {
646         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
647         bool bringup = st->bringup;
648         enum cpuhp_state state;
649
650         if (WARN_ON_ONCE(!st->should_run))
651                 return;
652
653         /*
654          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
655          * that if we see ->should_run we also see the rest of the state.
656          */
657         smp_mb();
658
659         /*
660          * The BP holds the hotplug lock, but we're now running on the AP,
661          * ensure that anybody asserting the lock is held, will actually find
662          * it so.
663          */
664         lockdep_acquire_cpus_lock();
665         cpuhp_lock_acquire(bringup);
666
667         if (st->single) {
668                 state = st->cb_state;
669                 st->should_run = false;
670         } else {
671                 if (bringup) {
672                         st->state++;
673                         state = st->state;
674                         st->should_run = (st->state < st->target);
675                         WARN_ON_ONCE(st->state > st->target);
676                 } else {
677                         state = st->state;
678                         st->state--;
679                         st->should_run = (st->state > st->target);
680                         WARN_ON_ONCE(st->state < st->target);
681                 }
682         }
683
684         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
685
686         if (cpuhp_is_atomic_state(state)) {
687                 local_irq_disable();
688                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
689                 local_irq_enable();
690
691                 /*
692                  * STARTING/DYING must not fail!
693                  */
694                 WARN_ON_ONCE(st->result);
695         } else {
696                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
697         }
698
699         if (st->result) {
700                 /*
701                  * If we fail on a rollback, we're up a creek without no
702                  * paddle, no way forward, no way back. We loose, thanks for
703                  * playing.
704                  */
705                 WARN_ON_ONCE(st->rollback);
706                 st->should_run = false;
707         }
708
709         cpuhp_lock_release(bringup);
710         lockdep_release_cpus_lock();
711
712         if (!st->should_run)
713                 complete_ap_thread(st, bringup);
714 }
715
716 /* Invoke a single callback on a remote cpu */
717 static int
718 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
719                          struct hlist_node *node)
720 {
721         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
722         int ret;
723
724         if (!cpu_online(cpu))
725                 return 0;
726
727         cpuhp_lock_acquire(false);
728         cpuhp_lock_release(false);
729
730         cpuhp_lock_acquire(true);
731         cpuhp_lock_release(true);
732
733         /*
734          * If we are up and running, use the hotplug thread. For early calls
735          * we invoke the thread function directly.
736          */
737         if (!st->thread)
738                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
739
740         st->rollback = false;
741         st->last = NULL;
742
743         st->node = node;
744         st->bringup = bringup;
745         st->cb_state = state;
746         st->single = true;
747
748         __cpuhp_kick_ap(st);
749
750         /*
751          * If we failed and did a partial, do a rollback.
752          */
753         if ((ret = st->result) && st->last) {
754                 st->rollback = true;
755                 st->bringup = !bringup;
756
757                 __cpuhp_kick_ap(st);
758         }
759
760         /*
761          * Clean up the leftovers so the next hotplug operation wont use stale
762          * data.
763          */
764         st->node = st->last = NULL;
765         return ret;
766 }
767
768 static int cpuhp_kick_ap_work(unsigned int cpu)
769 {
770         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
771         enum cpuhp_state prev_state = st->state;
772         int ret;
773
774         cpuhp_lock_acquire(false);
775         cpuhp_lock_release(false);
776
777         cpuhp_lock_acquire(true);
778         cpuhp_lock_release(true);
779
780         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
781         ret = cpuhp_kick_ap(st, st->target);
782         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
783
784         return ret;
785 }
786
787 static struct smp_hotplug_thread cpuhp_threads = {
788         .store                  = &cpuhp_state.thread,
789         .create                 = &cpuhp_create,
790         .thread_should_run      = cpuhp_should_run,
791         .thread_fn              = cpuhp_thread_fun,
792         .thread_comm            = "cpuhp/%u",
793         .selfparking            = true,
794 };
795
796 void __init cpuhp_threads_init(void)
797 {
798         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
799         kthread_unpark(this_cpu_read(cpuhp_state.thread));
800 }
801
802 #ifdef CONFIG_HOTPLUG_CPU
803 /**
804  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
805  * @cpu: a CPU id
806  *
807  * This function walks all processes, finds a valid mm struct for each one and
808  * then clears a corresponding bit in mm's cpumask.  While this all sounds
809  * trivial, there are various non-obvious corner cases, which this function
810  * tries to solve in a safe manner.
811  *
812  * Also note that the function uses a somewhat relaxed locking scheme, so it may
813  * be called only for an already offlined CPU.
814  */
815 void clear_tasks_mm_cpumask(int cpu)
816 {
817         struct task_struct *p;
818
819         /*
820          * This function is called after the cpu is taken down and marked
821          * offline, so its not like new tasks will ever get this cpu set in
822          * their mm mask. -- Peter Zijlstra
823          * Thus, we may use rcu_read_lock() here, instead of grabbing
824          * full-fledged tasklist_lock.
825          */
826         WARN_ON(cpu_online(cpu));
827         rcu_read_lock();
828         for_each_process(p) {
829                 struct task_struct *t;
830
831                 /*
832                  * Main thread might exit, but other threads may still have
833                  * a valid mm. Find one.
834                  */
835                 t = find_lock_task_mm(p);
836                 if (!t)
837                         continue;
838                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
839                 task_unlock(t);
840         }
841         rcu_read_unlock();
842 }
843
844 /* Take this CPU down. */
845 static int take_cpu_down(void *_param)
846 {
847         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
848         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
849         int err, cpu = smp_processor_id();
850         int ret;
851
852         /* Ensure this CPU doesn't handle any more interrupts. */
853         err = __cpu_disable();
854         if (err < 0)
855                 return err;
856
857         /*
858          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
859          * do this step again.
860          */
861         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
862         st->state--;
863         /* Invoke the former CPU_DYING callbacks */
864         for (; st->state > target; st->state--) {
865                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
866                 /*
867                  * DYING must not fail!
868                  */
869                 WARN_ON_ONCE(ret);
870         }
871
872         /* Give up timekeeping duties */
873         tick_handover_do_timer();
874         /* Remove CPU from timer broadcasting */
875         tick_offline_cpu(cpu);
876         /* Park the stopper thread */
877         stop_machine_park(cpu);
878         return 0;
879 }
880
881 static int takedown_cpu(unsigned int cpu)
882 {
883         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
884         int err;
885
886         /* Park the smpboot threads */
887         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
888
889         /*
890          * Prevent irq alloc/free while the dying cpu reorganizes the
891          * interrupt affinities.
892          */
893         irq_lock_sparse();
894
895         /*
896          * So now all preempt/rcu users must observe !cpu_active().
897          */
898         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
899         if (err) {
900                 /* CPU refused to die */
901                 irq_unlock_sparse();
902                 /* Unpark the hotplug thread so we can rollback there */
903                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
904                 return err;
905         }
906         BUG_ON(cpu_online(cpu));
907
908         /*
909          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
910          * all runnable tasks from the CPU, there's only the idle task left now
911          * that the migration thread is done doing the stop_machine thing.
912          *
913          * Wait for the stop thread to go away.
914          */
915         wait_for_ap_thread(st, false);
916         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
917
918         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
919         irq_unlock_sparse();
920
921         hotplug_cpu__broadcast_tick_pull(cpu);
922         /* This actually kills the CPU. */
923         __cpu_die(cpu);
924
925         tick_cleanup_dead_cpu(cpu);
926         rcutree_migrate_callbacks(cpu);
927         return 0;
928 }
929
930 static void cpuhp_complete_idle_dead(void *arg)
931 {
932         struct cpuhp_cpu_state *st = arg;
933
934         complete_ap_thread(st, false);
935 }
936
937 void cpuhp_report_idle_dead(void)
938 {
939         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
940
941         BUG_ON(st->state != CPUHP_AP_OFFLINE);
942         rcu_report_dead(smp_processor_id());
943         st->state = CPUHP_AP_IDLE_DEAD;
944         /*
945          * We cannot call complete after rcu_report_dead() so we delegate it
946          * to an online cpu.
947          */
948         smp_call_function_single(cpumask_first(cpu_online_mask),
949                                  cpuhp_complete_idle_dead, st, 0);
950 }
951
952 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
953 {
954         for (st->state++; st->state < st->target; st->state++)
955                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
956 }
957
958 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
959                                 enum cpuhp_state target)
960 {
961         enum cpuhp_state prev_state = st->state;
962         int ret = 0;
963
964         for (; st->state > target; st->state--) {
965                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
966                 if (ret) {
967                         st->target = prev_state;
968                         if (st->state < prev_state)
969                                 undo_cpu_down(cpu, st);
970                         break;
971                 }
972         }
973         return ret;
974 }
975
976 /* Requires cpu_add_remove_lock to be held */
977 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
978                            enum cpuhp_state target)
979 {
980         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
981         int prev_state, ret = 0;
982
983         if (num_online_cpus() == 1)
984                 return -EBUSY;
985
986         if (!cpu_present(cpu))
987                 return -EINVAL;
988
989         cpus_write_lock();
990
991         cpuhp_tasks_frozen = tasks_frozen;
992
993         prev_state = cpuhp_set_state(st, target);
994         /*
995          * If the current CPU state is in the range of the AP hotplug thread,
996          * then we need to kick the thread.
997          */
998         if (st->state > CPUHP_TEARDOWN_CPU) {
999                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1000                 ret = cpuhp_kick_ap_work(cpu);
1001                 /*
1002                  * The AP side has done the error rollback already. Just
1003                  * return the error code..
1004                  */
1005                 if (ret)
1006                         goto out;
1007
1008                 /*
1009                  * We might have stopped still in the range of the AP hotplug
1010                  * thread. Nothing to do anymore.
1011                  */
1012                 if (st->state > CPUHP_TEARDOWN_CPU)
1013                         goto out;
1014
1015                 st->target = target;
1016         }
1017         /*
1018          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1019          * to do the further cleanups.
1020          */
1021         ret = cpuhp_down_callbacks(cpu, st, target);
1022         if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1023                 cpuhp_reset_state(st, prev_state);
1024                 __cpuhp_kick_ap(st);
1025         }
1026
1027 out:
1028         cpus_write_unlock();
1029         /*
1030          * Do post unplug cleanup. This is still protected against
1031          * concurrent CPU hotplug via cpu_add_remove_lock.
1032          */
1033         lockup_detector_cleanup();
1034         arch_smt_update();
1035         return ret;
1036 }
1037
1038 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1039 {
1040         if (cpu_hotplug_disabled)
1041                 return -EBUSY;
1042         return _cpu_down(cpu, 0, target);
1043 }
1044
1045 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1046 {
1047         int err;
1048
1049         cpu_maps_update_begin();
1050         err = cpu_down_maps_locked(cpu, target);
1051         cpu_maps_update_done();
1052         return err;
1053 }
1054
1055 int cpu_down(unsigned int cpu)
1056 {
1057         return do_cpu_down(cpu, CPUHP_OFFLINE);
1058 }
1059 EXPORT_SYMBOL(cpu_down);
1060
1061 #else
1062 #define takedown_cpu            NULL
1063 #endif /*CONFIG_HOTPLUG_CPU*/
1064
1065 /**
1066  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1067  * @cpu: cpu that just started
1068  *
1069  * It must be called by the arch code on the new cpu, before the new cpu
1070  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1071  */
1072 void notify_cpu_starting(unsigned int cpu)
1073 {
1074         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1075         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1076         int ret;
1077
1078         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1079         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1080         while (st->state < target) {
1081                 st->state++;
1082                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1083                 /*
1084                  * STARTING must not fail!
1085                  */
1086                 WARN_ON_ONCE(ret);
1087         }
1088 }
1089
1090 /*
1091  * Called from the idle task. Wake up the controlling task which brings the
1092  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1093  * the rest of the online bringup to the hotplug thread.
1094  */
1095 void cpuhp_online_idle(enum cpuhp_state state)
1096 {
1097         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1098
1099         /* Happens for the boot cpu */
1100         if (state != CPUHP_AP_ONLINE_IDLE)
1101                 return;
1102
1103         st->state = CPUHP_AP_ONLINE_IDLE;
1104         complete_ap_thread(st, true);
1105 }
1106
1107 /* Requires cpu_add_remove_lock to be held */
1108 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1109 {
1110         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1111         struct task_struct *idle;
1112         int ret = 0;
1113
1114         cpus_write_lock();
1115
1116         if (!cpu_present(cpu)) {
1117                 ret = -EINVAL;
1118                 goto out;
1119         }
1120
1121         /*
1122          * The caller of do_cpu_up might have raced with another
1123          * caller. Ignore it for now.
1124          */
1125         if (st->state >= target)
1126                 goto out;
1127
1128         if (st->state == CPUHP_OFFLINE) {
1129                 /* Let it fail before we try to bring the cpu up */
1130                 idle = idle_thread_get(cpu);
1131                 if (IS_ERR(idle)) {
1132                         ret = PTR_ERR(idle);
1133                         goto out;
1134                 }
1135         }
1136
1137         cpuhp_tasks_frozen = tasks_frozen;
1138
1139         cpuhp_set_state(st, target);
1140         /*
1141          * If the current CPU state is in the range of the AP hotplug thread,
1142          * then we need to kick the thread once more.
1143          */
1144         if (st->state > CPUHP_BRINGUP_CPU) {
1145                 ret = cpuhp_kick_ap_work(cpu);
1146                 /*
1147                  * The AP side has done the error rollback already. Just
1148                  * return the error code..
1149                  */
1150                 if (ret)
1151                         goto out;
1152         }
1153
1154         /*
1155          * Try to reach the target state. We max out on the BP at
1156          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1157          * responsible for bringing it up to the target state.
1158          */
1159         target = min((int)target, CPUHP_BRINGUP_CPU);
1160         ret = cpuhp_up_callbacks(cpu, st, target);
1161 out:
1162         cpus_write_unlock();
1163         arch_smt_update();
1164         return ret;
1165 }
1166
1167 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1168 {
1169         int err = 0;
1170
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",
1173                        cpu);
1174 #if defined(CONFIG_IA64)
1175                 pr_err("please check additional_cpus= boot parameter\n");
1176 #endif
1177                 return -EINVAL;
1178         }
1179
1180         err = try_online_node(cpu_to_node(cpu));
1181         if (err)
1182                 return err;
1183
1184         cpu_maps_update_begin();
1185
1186         if (cpu_hotplug_disabled) {
1187                 err = -EBUSY;
1188                 goto out;
1189         }
1190         if (!cpu_smt_allowed(cpu)) {
1191                 err = -EPERM;
1192                 goto out;
1193         }
1194
1195         err = _cpu_up(cpu, 0, target);
1196 out:
1197         cpu_maps_update_done();
1198         return err;
1199 }
1200
1201 int cpu_up(unsigned int cpu)
1202 {
1203         return do_cpu_up(cpu, CPUHP_ONLINE);
1204 }
1205 EXPORT_SYMBOL_GPL(cpu_up);
1206
1207 #ifdef CONFIG_PM_SLEEP_SMP
1208 static cpumask_var_t frozen_cpus;
1209
1210 int freeze_secondary_cpus(int primary)
1211 {
1212         int cpu, error = 0;
1213
1214         cpu_maps_update_begin();
1215         if (primary == -1) {
1216                 primary = cpumask_first(cpu_online_mask);
1217                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1218                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1219         } else {
1220                 if (!cpu_online(primary))
1221                         primary = cpumask_first(cpu_online_mask);
1222         }
1223
1224         /*
1225          * We take down all of the non-boot CPUs in one shot to avoid races
1226          * with the userspace trying to use the CPU hotplug at the same time
1227          */
1228         cpumask_clear(frozen_cpus);
1229
1230         pr_info("Disabling non-boot CPUs ...\n");
1231         for_each_online_cpu(cpu) {
1232                 if (cpu == primary)
1233                         continue;
1234
1235                 if (pm_wakeup_pending()) {
1236                         pr_info("Wakeup pending. Abort CPU freeze\n");
1237                         error = -EBUSY;
1238                         break;
1239                 }
1240
1241                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1242                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1243                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1244                 if (!error)
1245                         cpumask_set_cpu(cpu, frozen_cpus);
1246                 else {
1247                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1248                         break;
1249                 }
1250         }
1251
1252         if (!error)
1253                 BUG_ON(num_online_cpus() > 1);
1254         else
1255                 pr_err("Non-boot CPUs are not disabled\n");
1256
1257         /*
1258          * Make sure the CPUs won't be enabled by someone else. We need to do
1259          * this even in case of failure as all disable_nonboot_cpus() users are
1260          * supposed to do enable_nonboot_cpus() on the failure path.
1261          */
1262         cpu_hotplug_disabled++;
1263
1264         cpu_maps_update_done();
1265         return error;
1266 }
1267
1268 void __weak arch_enable_nonboot_cpus_begin(void)
1269 {
1270 }
1271
1272 void __weak arch_enable_nonboot_cpus_end(void)
1273 {
1274 }
1275
1276 void enable_nonboot_cpus(void)
1277 {
1278         int cpu, error;
1279
1280         /* Allow everyone to use the CPU hotplug again */
1281         cpu_maps_update_begin();
1282         __cpu_hotplug_enable();
1283         if (cpumask_empty(frozen_cpus))
1284                 goto out;
1285
1286         pr_info("Enabling non-boot CPUs ...\n");
1287
1288         arch_enable_nonboot_cpus_begin();
1289
1290         for_each_cpu(cpu, frozen_cpus) {
1291                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1292                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1293                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1294                 if (!error) {
1295                         pr_info("CPU%d is up\n", cpu);
1296                         continue;
1297                 }
1298                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1299         }
1300
1301         arch_enable_nonboot_cpus_end();
1302
1303         cpumask_clear(frozen_cpus);
1304 out:
1305         cpu_maps_update_done();
1306 }
1307
1308 static int __init alloc_frozen_cpus(void)
1309 {
1310         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1311                 return -ENOMEM;
1312         return 0;
1313 }
1314 core_initcall(alloc_frozen_cpus);
1315
1316 /*
1317  * When callbacks for CPU hotplug notifications are being executed, we must
1318  * ensure that the state of the system with respect to the tasks being frozen
1319  * or not, as reported by the notification, remains unchanged *throughout the
1320  * duration* of the execution of the callbacks.
1321  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1322  *
1323  * This synchronization is implemented by mutually excluding regular CPU
1324  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1325  * Hibernate notifications.
1326  */
1327 static int
1328 cpu_hotplug_pm_callback(struct notifier_block *nb,
1329                         unsigned long action, void *ptr)
1330 {
1331         switch (action) {
1332
1333         case PM_SUSPEND_PREPARE:
1334         case PM_HIBERNATION_PREPARE:
1335                 cpu_hotplug_disable();
1336                 break;
1337
1338         case PM_POST_SUSPEND:
1339         case PM_POST_HIBERNATION:
1340                 cpu_hotplug_enable();
1341                 break;
1342
1343         default:
1344                 return NOTIFY_DONE;
1345         }
1346
1347         return NOTIFY_OK;
1348 }
1349
1350
1351 static int __init cpu_hotplug_pm_sync_init(void)
1352 {
1353         /*
1354          * cpu_hotplug_pm_callback has higher priority than x86
1355          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1356          * to disable cpu hotplug to avoid cpu hotplug race.
1357          */
1358         pm_notifier(cpu_hotplug_pm_callback, 0);
1359         return 0;
1360 }
1361 core_initcall(cpu_hotplug_pm_sync_init);
1362
1363 #endif /* CONFIG_PM_SLEEP_SMP */
1364
1365 int __boot_cpu_id;
1366
1367 #endif /* CONFIG_SMP */
1368
1369 /* Boot processor state steps */
1370 static struct cpuhp_step cpuhp_hp_states[] = {
1371         [CPUHP_OFFLINE] = {
1372                 .name                   = "offline",
1373                 .startup.single         = NULL,
1374                 .teardown.single        = NULL,
1375         },
1376 #ifdef CONFIG_SMP
1377         [CPUHP_CREATE_THREADS]= {
1378                 .name                   = "threads:prepare",
1379                 .startup.single         = smpboot_create_threads,
1380                 .teardown.single        = NULL,
1381                 .cant_stop              = true,
1382         },
1383         [CPUHP_PERF_PREPARE] = {
1384                 .name                   = "perf:prepare",
1385                 .startup.single         = perf_event_init_cpu,
1386                 .teardown.single        = perf_event_exit_cpu,
1387         },
1388         [CPUHP_WORKQUEUE_PREP] = {
1389                 .name                   = "workqueue:prepare",
1390                 .startup.single         = workqueue_prepare_cpu,
1391                 .teardown.single        = NULL,
1392         },
1393         [CPUHP_HRTIMERS_PREPARE] = {
1394                 .name                   = "hrtimers:prepare",
1395                 .startup.single         = hrtimers_prepare_cpu,
1396                 .teardown.single        = hrtimers_dead_cpu,
1397         },
1398         [CPUHP_SMPCFD_PREPARE] = {
1399                 .name                   = "smpcfd:prepare",
1400                 .startup.single         = smpcfd_prepare_cpu,
1401                 .teardown.single        = smpcfd_dead_cpu,
1402         },
1403         [CPUHP_RELAY_PREPARE] = {
1404                 .name                   = "relay:prepare",
1405                 .startup.single         = relay_prepare_cpu,
1406                 .teardown.single        = NULL,
1407         },
1408         [CPUHP_SLAB_PREPARE] = {
1409                 .name                   = "slab:prepare",
1410                 .startup.single         = slab_prepare_cpu,
1411                 .teardown.single        = slab_dead_cpu,
1412         },
1413         [CPUHP_RCUTREE_PREP] = {
1414                 .name                   = "RCU/tree:prepare",
1415                 .startup.single         = rcutree_prepare_cpu,
1416                 .teardown.single        = rcutree_dead_cpu,
1417         },
1418         /*
1419          * On the tear-down path, timers_dead_cpu() must be invoked
1420          * before blk_mq_queue_reinit_notify() from notify_dead(),
1421          * otherwise a RCU stall occurs.
1422          */
1423         [CPUHP_TIMERS_PREPARE] = {
1424                 .name                   = "timers:prepare",
1425                 .startup.single         = timers_prepare_cpu,
1426                 .teardown.single        = timers_dead_cpu,
1427         },
1428         /* Kicks the plugged cpu into life */
1429         [CPUHP_BRINGUP_CPU] = {
1430                 .name                   = "cpu:bringup",
1431                 .startup.single         = bringup_cpu,
1432                 .teardown.single        = NULL,
1433                 .cant_stop              = true,
1434         },
1435         /* Final state before CPU kills itself */
1436         [CPUHP_AP_IDLE_DEAD] = {
1437                 .name                   = "idle:dead",
1438         },
1439         /*
1440          * Last state before CPU enters the idle loop to die. Transient state
1441          * for synchronization.
1442          */
1443         [CPUHP_AP_OFFLINE] = {
1444                 .name                   = "ap:offline",
1445                 .cant_stop              = true,
1446         },
1447         /* First state is scheduler control. Interrupts are disabled */
1448         [CPUHP_AP_SCHED_STARTING] = {
1449                 .name                   = "sched:starting",
1450                 .startup.single         = sched_cpu_starting,
1451                 .teardown.single        = sched_cpu_dying,
1452         },
1453         [CPUHP_AP_RCUTREE_DYING] = {
1454                 .name                   = "RCU/tree:dying",
1455                 .startup.single         = NULL,
1456                 .teardown.single        = rcutree_dying_cpu,
1457         },
1458         [CPUHP_AP_SMPCFD_DYING] = {
1459                 .name                   = "smpcfd:dying",
1460                 .startup.single         = NULL,
1461                 .teardown.single        = smpcfd_dying_cpu,
1462         },
1463         /* Entry state on starting. Interrupts enabled from here on. Transient
1464          * state for synchronsization */
1465         [CPUHP_AP_ONLINE] = {
1466                 .name                   = "ap:online",
1467         },
1468         /*
1469          * Handled on controll processor until the plugged processor manages
1470          * this itself.
1471          */
1472         [CPUHP_TEARDOWN_CPU] = {
1473                 .name                   = "cpu:teardown",
1474                 .startup.single         = NULL,
1475                 .teardown.single        = takedown_cpu,
1476                 .cant_stop              = true,
1477         },
1478         /* Handle smpboot threads park/unpark */
1479         [CPUHP_AP_SMPBOOT_THREADS] = {
1480                 .name                   = "smpboot/threads:online",
1481                 .startup.single         = smpboot_unpark_threads,
1482                 .teardown.single        = smpboot_park_threads,
1483         },
1484         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1485                 .name                   = "irq/affinity:online",
1486                 .startup.single         = irq_affinity_online_cpu,
1487                 .teardown.single        = NULL,
1488         },
1489         [CPUHP_AP_PERF_ONLINE] = {
1490                 .name                   = "perf:online",
1491                 .startup.single         = perf_event_init_cpu,
1492                 .teardown.single        = perf_event_exit_cpu,
1493         },
1494         [CPUHP_AP_WATCHDOG_ONLINE] = {
1495                 .name                   = "lockup_detector:online",
1496                 .startup.single         = lockup_detector_online_cpu,
1497                 .teardown.single        = lockup_detector_offline_cpu,
1498         },
1499         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1500                 .name                   = "workqueue:online",
1501                 .startup.single         = workqueue_online_cpu,
1502                 .teardown.single        = workqueue_offline_cpu,
1503         },
1504         [CPUHP_AP_RCUTREE_ONLINE] = {
1505                 .name                   = "RCU/tree:online",
1506                 .startup.single         = rcutree_online_cpu,
1507                 .teardown.single        = rcutree_offline_cpu,
1508         },
1509 #endif
1510         /*
1511          * The dynamically registered state space is here
1512          */
1513
1514 #ifdef CONFIG_SMP
1515         /* Last state is scheduler control setting the cpu active */
1516         [CPUHP_AP_ACTIVE] = {
1517                 .name                   = "sched:active",
1518                 .startup.single         = sched_cpu_activate,
1519                 .teardown.single        = sched_cpu_deactivate,
1520         },
1521 #endif
1522
1523         /* CPU is fully up and running. */
1524         [CPUHP_ONLINE] = {
1525                 .name                   = "online",
1526                 .startup.single         = NULL,
1527                 .teardown.single        = NULL,
1528         },
1529 };
1530
1531 /* Sanity check for callbacks */
1532 static int cpuhp_cb_check(enum cpuhp_state state)
1533 {
1534         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1535                 return -EINVAL;
1536         return 0;
1537 }
1538
1539 /*
1540  * Returns a free for dynamic slot assignment of the Online state. The states
1541  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1542  * by having no name assigned.
1543  */
1544 static int cpuhp_reserve_state(enum cpuhp_state state)
1545 {
1546         enum cpuhp_state i, end;
1547         struct cpuhp_step *step;
1548
1549         switch (state) {
1550         case CPUHP_AP_ONLINE_DYN:
1551                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1552                 end = CPUHP_AP_ONLINE_DYN_END;
1553                 break;
1554         case CPUHP_BP_PREPARE_DYN:
1555                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1556                 end = CPUHP_BP_PREPARE_DYN_END;
1557                 break;
1558         default:
1559                 return -EINVAL;
1560         }
1561
1562         for (i = state; i <= end; i++, step++) {
1563                 if (!step->name)
1564                         return i;
1565         }
1566         WARN(1, "No more dynamic states available for CPU hotplug\n");
1567         return -ENOSPC;
1568 }
1569
1570 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1571                                  int (*startup)(unsigned int cpu),
1572                                  int (*teardown)(unsigned int cpu),
1573                                  bool multi_instance)
1574 {
1575         /* (Un)Install the callbacks for further cpu hotplug operations */
1576         struct cpuhp_step *sp;
1577         int ret = 0;
1578
1579         /*
1580          * If name is NULL, then the state gets removed.
1581          *
1582          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1583          * the first allocation from these dynamic ranges, so the removal
1584          * would trigger a new allocation and clear the wrong (already
1585          * empty) state, leaving the callbacks of the to be cleared state
1586          * dangling, which causes wreckage on the next hotplug operation.
1587          */
1588         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1589                      state == CPUHP_BP_PREPARE_DYN)) {
1590                 ret = cpuhp_reserve_state(state);
1591                 if (ret < 0)
1592                         return ret;
1593                 state = ret;
1594         }
1595         sp = cpuhp_get_step(state);
1596         if (name && sp->name)
1597                 return -EBUSY;
1598
1599         sp->startup.single = startup;
1600         sp->teardown.single = teardown;
1601         sp->name = name;
1602         sp->multi_instance = multi_instance;
1603         INIT_HLIST_HEAD(&sp->list);
1604         return ret;
1605 }
1606
1607 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1608 {
1609         return cpuhp_get_step(state)->teardown.single;
1610 }
1611
1612 /*
1613  * Call the startup/teardown function for a step either on the AP or
1614  * on the current CPU.
1615  */
1616 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1617                             struct hlist_node *node)
1618 {
1619         struct cpuhp_step *sp = cpuhp_get_step(state);
1620         int ret;
1621
1622         /*
1623          * If there's nothing to do, we done.
1624          * Relies on the union for multi_instance.
1625          */
1626         if ((bringup && !sp->startup.single) ||
1627             (!bringup && !sp->teardown.single))
1628                 return 0;
1629         /*
1630          * The non AP bound callbacks can fail on bringup. On teardown
1631          * e.g. module removal we crash for now.
1632          */
1633 #ifdef CONFIG_SMP
1634         if (cpuhp_is_ap_state(state))
1635                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1636         else
1637                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1638 #else
1639         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1640 #endif
1641         BUG_ON(ret && !bringup);
1642         return ret;
1643 }
1644
1645 /*
1646  * Called from __cpuhp_setup_state on a recoverable failure.
1647  *
1648  * Note: The teardown callbacks for rollback are not allowed to fail!
1649  */
1650 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1651                                    struct hlist_node *node)
1652 {
1653         int cpu;
1654
1655         /* Roll back the already executed steps on the other cpus */
1656         for_each_present_cpu(cpu) {
1657                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1658                 int cpustate = st->state;
1659
1660                 if (cpu >= failedcpu)
1661                         break;
1662
1663                 /* Did we invoke the startup call on that cpu ? */
1664                 if (cpustate >= state)
1665                         cpuhp_issue_call(cpu, state, false, node);
1666         }
1667 }
1668
1669 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1670                                           struct hlist_node *node,
1671                                           bool invoke)
1672 {
1673         struct cpuhp_step *sp;
1674         int cpu;
1675         int ret;
1676
1677         lockdep_assert_cpus_held();
1678
1679         sp = cpuhp_get_step(state);
1680         if (sp->multi_instance == false)
1681                 return -EINVAL;
1682
1683         mutex_lock(&cpuhp_state_mutex);
1684
1685         if (!invoke || !sp->startup.multi)
1686                 goto add_node;
1687
1688         /*
1689          * Try to call the startup callback for each present cpu
1690          * depending on the hotplug state of the cpu.
1691          */
1692         for_each_present_cpu(cpu) {
1693                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1694                 int cpustate = st->state;
1695
1696                 if (cpustate < state)
1697                         continue;
1698
1699                 ret = cpuhp_issue_call(cpu, state, true, node);
1700                 if (ret) {
1701                         if (sp->teardown.multi)
1702                                 cpuhp_rollback_install(cpu, state, node);
1703                         goto unlock;
1704                 }
1705         }
1706 add_node:
1707         ret = 0;
1708         hlist_add_head(node, &sp->list);
1709 unlock:
1710         mutex_unlock(&cpuhp_state_mutex);
1711         return ret;
1712 }
1713
1714 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1715                                bool invoke)
1716 {
1717         int ret;
1718
1719         cpus_read_lock();
1720         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1721         cpus_read_unlock();
1722         return ret;
1723 }
1724 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1725
1726 /**
1727  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1728  * @state:              The state to setup
1729  * @invoke:             If true, the startup function is invoked for cpus where
1730  *                      cpu state >= @state
1731  * @startup:            startup callback function
1732  * @teardown:           teardown callback function
1733  * @multi_instance:     State is set up for multiple instances which get
1734  *                      added afterwards.
1735  *
1736  * The caller needs to hold cpus read locked while calling this function.
1737  * Returns:
1738  *   On success:
1739  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1740  *      0 for all other states
1741  *   On failure: proper (negative) error code
1742  */
1743 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1744                                    const char *name, bool invoke,
1745                                    int (*startup)(unsigned int cpu),
1746                                    int (*teardown)(unsigned int cpu),
1747                                    bool multi_instance)
1748 {
1749         int cpu, ret = 0;
1750         bool dynstate;
1751
1752         lockdep_assert_cpus_held();
1753
1754         if (cpuhp_cb_check(state) || !name)
1755                 return -EINVAL;
1756
1757         mutex_lock(&cpuhp_state_mutex);
1758
1759         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1760                                     multi_instance);
1761
1762         dynstate = state == CPUHP_AP_ONLINE_DYN;
1763         if (ret > 0 && dynstate) {
1764                 state = ret;
1765                 ret = 0;
1766         }
1767
1768         if (ret || !invoke || !startup)
1769                 goto out;
1770
1771         /*
1772          * Try to call the startup callback for each present cpu
1773          * depending on the hotplug state of the cpu.
1774          */
1775         for_each_present_cpu(cpu) {
1776                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1777                 int cpustate = st->state;
1778
1779                 if (cpustate < state)
1780                         continue;
1781
1782                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1783                 if (ret) {
1784                         if (teardown)
1785                                 cpuhp_rollback_install(cpu, state, NULL);
1786                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1787                         goto out;
1788                 }
1789         }
1790 out:
1791         mutex_unlock(&cpuhp_state_mutex);
1792         /*
1793          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1794          * dynamically allocated state in case of success.
1795          */
1796         if (!ret && dynstate)
1797                 return state;
1798         return ret;
1799 }
1800 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1801
1802 int __cpuhp_setup_state(enum cpuhp_state state,
1803                         const char *name, bool invoke,
1804                         int (*startup)(unsigned int cpu),
1805                         int (*teardown)(unsigned int cpu),
1806                         bool multi_instance)
1807 {
1808         int ret;
1809
1810         cpus_read_lock();
1811         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1812                                              teardown, multi_instance);
1813         cpus_read_unlock();
1814         return ret;
1815 }
1816 EXPORT_SYMBOL(__cpuhp_setup_state);
1817
1818 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1819                                   struct hlist_node *node, bool invoke)
1820 {
1821         struct cpuhp_step *sp = cpuhp_get_step(state);
1822         int cpu;
1823
1824         BUG_ON(cpuhp_cb_check(state));
1825
1826         if (!sp->multi_instance)
1827                 return -EINVAL;
1828
1829         cpus_read_lock();
1830         mutex_lock(&cpuhp_state_mutex);
1831
1832         if (!invoke || !cpuhp_get_teardown_cb(state))
1833                 goto remove;
1834         /*
1835          * Call the teardown callback for each present cpu depending
1836          * on the hotplug state of the cpu. This function is not
1837          * allowed to fail currently!
1838          */
1839         for_each_present_cpu(cpu) {
1840                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1841                 int cpustate = st->state;
1842
1843                 if (cpustate >= state)
1844                         cpuhp_issue_call(cpu, state, false, node);
1845         }
1846
1847 remove:
1848         hlist_del(node);
1849         mutex_unlock(&cpuhp_state_mutex);
1850         cpus_read_unlock();
1851
1852         return 0;
1853 }
1854 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1855
1856 /**
1857  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1858  * @state:      The state to remove
1859  * @invoke:     If true, the teardown function is invoked for cpus where
1860  *              cpu state >= @state
1861  *
1862  * The caller needs to hold cpus read locked while calling this function.
1863  * The teardown callback is currently not allowed to fail. Think
1864  * about module removal!
1865  */
1866 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1867 {
1868         struct cpuhp_step *sp = cpuhp_get_step(state);
1869         int cpu;
1870
1871         BUG_ON(cpuhp_cb_check(state));
1872
1873         lockdep_assert_cpus_held();
1874
1875         mutex_lock(&cpuhp_state_mutex);
1876         if (sp->multi_instance) {
1877                 WARN(!hlist_empty(&sp->list),
1878                      "Error: Removing state %d which has instances left.\n",
1879                      state);
1880                 goto remove;
1881         }
1882
1883         if (!invoke || !cpuhp_get_teardown_cb(state))
1884                 goto remove;
1885
1886         /*
1887          * Call the teardown callback for each present cpu depending
1888          * on the hotplug state of the cpu. This function is not
1889          * allowed to fail currently!
1890          */
1891         for_each_present_cpu(cpu) {
1892                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1893                 int cpustate = st->state;
1894
1895                 if (cpustate >= state)
1896                         cpuhp_issue_call(cpu, state, false, NULL);
1897         }
1898 remove:
1899         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1900         mutex_unlock(&cpuhp_state_mutex);
1901 }
1902 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1903
1904 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1905 {
1906         cpus_read_lock();
1907         __cpuhp_remove_state_cpuslocked(state, invoke);
1908         cpus_read_unlock();
1909 }
1910 EXPORT_SYMBOL(__cpuhp_remove_state);
1911
1912 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1913 static ssize_t show_cpuhp_state(struct device *dev,
1914                                 struct device_attribute *attr, char *buf)
1915 {
1916         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1917
1918         return sprintf(buf, "%d\n", st->state);
1919 }
1920 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1921
1922 static ssize_t write_cpuhp_target(struct device *dev,
1923                                   struct device_attribute *attr,
1924                                   const char *buf, size_t count)
1925 {
1926         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1927         struct cpuhp_step *sp;
1928         int target, ret;
1929
1930         ret = kstrtoint(buf, 10, &target);
1931         if (ret)
1932                 return ret;
1933
1934 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1935         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1936                 return -EINVAL;
1937 #else
1938         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1939                 return -EINVAL;
1940 #endif
1941
1942         ret = lock_device_hotplug_sysfs();
1943         if (ret)
1944                 return ret;
1945
1946         mutex_lock(&cpuhp_state_mutex);
1947         sp = cpuhp_get_step(target);
1948         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1949         mutex_unlock(&cpuhp_state_mutex);
1950         if (ret)
1951                 goto out;
1952
1953         if (st->state < target)
1954                 ret = do_cpu_up(dev->id, target);
1955         else
1956                 ret = do_cpu_down(dev->id, target);
1957 out:
1958         unlock_device_hotplug();
1959         return ret ? ret : count;
1960 }
1961
1962 static ssize_t show_cpuhp_target(struct device *dev,
1963                                  struct device_attribute *attr, char *buf)
1964 {
1965         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1966
1967         return sprintf(buf, "%d\n", st->target);
1968 }
1969 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1970
1971
1972 static ssize_t write_cpuhp_fail(struct device *dev,
1973                                 struct device_attribute *attr,
1974                                 const char *buf, size_t count)
1975 {
1976         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1977         struct cpuhp_step *sp;
1978         int fail, ret;
1979
1980         ret = kstrtoint(buf, 10, &fail);
1981         if (ret)
1982                 return ret;
1983
1984         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
1985                 return -EINVAL;
1986
1987         /*
1988          * Cannot fail STARTING/DYING callbacks.
1989          */
1990         if (cpuhp_is_atomic_state(fail))
1991                 return -EINVAL;
1992
1993         /*
1994          * Cannot fail anything that doesn't have callbacks.
1995          */
1996         mutex_lock(&cpuhp_state_mutex);
1997         sp = cpuhp_get_step(fail);
1998         if (!sp->startup.single && !sp->teardown.single)
1999                 ret = -EINVAL;
2000         mutex_unlock(&cpuhp_state_mutex);
2001         if (ret)
2002                 return ret;
2003
2004         st->fail = fail;
2005
2006         return count;
2007 }
2008
2009 static ssize_t show_cpuhp_fail(struct device *dev,
2010                                struct device_attribute *attr, char *buf)
2011 {
2012         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2013
2014         return sprintf(buf, "%d\n", st->fail);
2015 }
2016
2017 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2018
2019 static struct attribute *cpuhp_cpu_attrs[] = {
2020         &dev_attr_state.attr,
2021         &dev_attr_target.attr,
2022         &dev_attr_fail.attr,
2023         NULL
2024 };
2025
2026 static const struct attribute_group cpuhp_cpu_attr_group = {
2027         .attrs = cpuhp_cpu_attrs,
2028         .name = "hotplug",
2029         NULL
2030 };
2031
2032 static ssize_t show_cpuhp_states(struct device *dev,
2033                                  struct device_attribute *attr, char *buf)
2034 {
2035         ssize_t cur, res = 0;
2036         int i;
2037
2038         mutex_lock(&cpuhp_state_mutex);
2039         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2040                 struct cpuhp_step *sp = cpuhp_get_step(i);
2041
2042                 if (sp->name) {
2043                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2044                         buf += cur;
2045                         res += cur;
2046                 }
2047         }
2048         mutex_unlock(&cpuhp_state_mutex);
2049         return res;
2050 }
2051 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2052
2053 static struct attribute *cpuhp_cpu_root_attrs[] = {
2054         &dev_attr_states.attr,
2055         NULL
2056 };
2057
2058 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2059         .attrs = cpuhp_cpu_root_attrs,
2060         .name = "hotplug",
2061         NULL
2062 };
2063
2064 #ifdef CONFIG_HOTPLUG_SMT
2065
2066 static void cpuhp_offline_cpu_device(unsigned int cpu)
2067 {
2068         struct device *dev = get_cpu_device(cpu);
2069
2070         dev->offline = true;
2071         /* Tell user space about the state change */
2072         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2073 }
2074
2075 static void cpuhp_online_cpu_device(unsigned int cpu)
2076 {
2077         struct device *dev = get_cpu_device(cpu);
2078
2079         dev->offline = false;
2080         /* Tell user space about the state change */
2081         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2082 }
2083
2084 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2085 {
2086         int cpu, ret = 0;
2087
2088         cpu_maps_update_begin();
2089         for_each_online_cpu(cpu) {
2090                 if (topology_is_primary_thread(cpu))
2091                         continue;
2092                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2093                 if (ret)
2094                         break;
2095                 /*
2096                  * As this needs to hold the cpu maps lock it's impossible
2097                  * to call device_offline() because that ends up calling
2098                  * cpu_down() which takes cpu maps lock. cpu maps lock
2099                  * needs to be held as this might race against in kernel
2100                  * abusers of the hotplug machinery (thermal management).
2101                  *
2102                  * So nothing would update device:offline state. That would
2103                  * leave the sysfs entry stale and prevent onlining after
2104                  * smt control has been changed to 'off' again. This is
2105                  * called under the sysfs hotplug lock, so it is properly
2106                  * serialized against the regular offline usage.
2107                  */
2108                 cpuhp_offline_cpu_device(cpu);
2109         }
2110         if (!ret)
2111                 cpu_smt_control = ctrlval;
2112         cpu_maps_update_done();
2113         return ret;
2114 }
2115
2116 int cpuhp_smt_enable(void)
2117 {
2118         int cpu, ret = 0;
2119
2120         cpu_maps_update_begin();
2121         cpu_smt_control = CPU_SMT_ENABLED;
2122         for_each_present_cpu(cpu) {
2123                 /* Skip online CPUs and CPUs on offline nodes */
2124                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2125                         continue;
2126                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2127                 if (ret)
2128                         break;
2129                 /* See comment in cpuhp_smt_disable() */
2130                 cpuhp_online_cpu_device(cpu);
2131         }
2132         cpu_maps_update_done();
2133         return ret;
2134 }
2135
2136
2137 static ssize_t
2138 __store_smt_control(struct device *dev, struct device_attribute *attr,
2139                     const char *buf, size_t count)
2140 {
2141         int ctrlval, ret;
2142
2143         if (sysfs_streq(buf, "on"))
2144                 ctrlval = CPU_SMT_ENABLED;
2145         else if (sysfs_streq(buf, "off"))
2146                 ctrlval = CPU_SMT_DISABLED;
2147         else if (sysfs_streq(buf, "forceoff"))
2148                 ctrlval = CPU_SMT_FORCE_DISABLED;
2149         else
2150                 return -EINVAL;
2151
2152         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2153                 return -EPERM;
2154
2155         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2156                 return -ENODEV;
2157
2158         ret = lock_device_hotplug_sysfs();
2159         if (ret)
2160                 return ret;
2161
2162         if (ctrlval != cpu_smt_control) {
2163                 switch (ctrlval) {
2164                 case CPU_SMT_ENABLED:
2165                         ret = cpuhp_smt_enable();
2166                         break;
2167                 case CPU_SMT_DISABLED:
2168                 case CPU_SMT_FORCE_DISABLED:
2169                         ret = cpuhp_smt_disable(ctrlval);
2170                         break;
2171                 }
2172         }
2173
2174         unlock_device_hotplug();
2175         return ret ? ret : count;
2176 }
2177
2178 #else /* !CONFIG_HOTPLUG_SMT */
2179 static ssize_t
2180 __store_smt_control(struct device *dev, struct device_attribute *attr,
2181                     const char *buf, size_t count)
2182 {
2183         return -ENODEV;
2184 }
2185 #endif /* CONFIG_HOTPLUG_SMT */
2186
2187 static const char *smt_states[] = {
2188         [CPU_SMT_ENABLED]               = "on",
2189         [CPU_SMT_DISABLED]              = "off",
2190         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2191         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2192         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2193 };
2194
2195 static ssize_t
2196 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2197 {
2198         const char *state = smt_states[cpu_smt_control];
2199
2200         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2201 }
2202
2203 static ssize_t
2204 store_smt_control(struct device *dev, struct device_attribute *attr,
2205                   const char *buf, size_t count)
2206 {
2207         return __store_smt_control(dev, attr, buf, count);
2208 }
2209 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2210
2211 static ssize_t
2212 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2213 {
2214         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2215 }
2216 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2217
2218 static struct attribute *cpuhp_smt_attrs[] = {
2219         &dev_attr_control.attr,
2220         &dev_attr_active.attr,
2221         NULL
2222 };
2223
2224 static const struct attribute_group cpuhp_smt_attr_group = {
2225         .attrs = cpuhp_smt_attrs,
2226         .name = "smt",
2227         NULL
2228 };
2229
2230 static int __init cpu_smt_sysfs_init(void)
2231 {
2232         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2233                                   &cpuhp_smt_attr_group);
2234 }
2235
2236 static int __init cpuhp_sysfs_init(void)
2237 {
2238         int cpu, ret;
2239
2240         ret = cpu_smt_sysfs_init();
2241         if (ret)
2242                 return ret;
2243
2244         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2245                                  &cpuhp_cpu_root_attr_group);
2246         if (ret)
2247                 return ret;
2248
2249         for_each_possible_cpu(cpu) {
2250                 struct device *dev = get_cpu_device(cpu);
2251
2252                 if (!dev)
2253                         continue;
2254                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2255                 if (ret)
2256                         return ret;
2257         }
2258         return 0;
2259 }
2260 device_initcall(cpuhp_sysfs_init);
2261 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2262
2263 /*
2264  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2265  * represents all NR_CPUS bits binary values of 1<<nr.
2266  *
2267  * It is used by cpumask_of() to get a constant address to a CPU
2268  * mask value that has a single bit set only.
2269  */
2270
2271 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2272 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2273 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2274 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2275 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2276
2277 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2278
2279         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2280         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2281 #if BITS_PER_LONG > 32
2282         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2283         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2284 #endif
2285 };
2286 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2287
2288 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2289 EXPORT_SYMBOL(cpu_all_bits);
2290
2291 #ifdef CONFIG_INIT_ALL_POSSIBLE
2292 struct cpumask __cpu_possible_mask __read_mostly
2293         = {CPU_BITS_ALL};
2294 #else
2295 struct cpumask __cpu_possible_mask __read_mostly;
2296 #endif
2297 EXPORT_SYMBOL(__cpu_possible_mask);
2298
2299 struct cpumask __cpu_online_mask __read_mostly;
2300 EXPORT_SYMBOL(__cpu_online_mask);
2301
2302 struct cpumask __cpu_present_mask __read_mostly;
2303 EXPORT_SYMBOL(__cpu_present_mask);
2304
2305 struct cpumask __cpu_active_mask __read_mostly;
2306 EXPORT_SYMBOL(__cpu_active_mask);
2307
2308 atomic_t __num_online_cpus __read_mostly;
2309 EXPORT_SYMBOL(__num_online_cpus);
2310
2311 void init_cpu_present(const struct cpumask *src)
2312 {
2313         cpumask_copy(&__cpu_present_mask, src);
2314 }
2315
2316 void init_cpu_possible(const struct cpumask *src)
2317 {
2318         cpumask_copy(&__cpu_possible_mask, src);
2319 }
2320
2321 void init_cpu_online(const struct cpumask *src)
2322 {
2323         cpumask_copy(&__cpu_online_mask, src);
2324 }
2325
2326 void set_cpu_online(unsigned int cpu, bool online)
2327 {
2328         /*
2329          * atomic_inc/dec() is required to handle the horrid abuse of this
2330          * function by the reboot and kexec code which invoke it from
2331          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2332          * regular CPU hotplug is properly serialized.
2333          *
2334          * Note, that the fact that __num_online_cpus is of type atomic_t
2335          * does not protect readers which are not serialized against
2336          * concurrent hotplug operations.
2337          */
2338         if (online) {
2339                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2340                         atomic_inc(&__num_online_cpus);
2341         } else {
2342                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2343                         atomic_dec(&__num_online_cpus);
2344         }
2345 }
2346
2347 /*
2348  * Activate the first processor.
2349  */
2350 void __init boot_cpu_init(void)
2351 {
2352         int cpu = smp_processor_id();
2353
2354         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2355         set_cpu_online(cpu, true);
2356         set_cpu_active(cpu, true);
2357         set_cpu_present(cpu, true);
2358         set_cpu_possible(cpu, true);
2359
2360 #ifdef CONFIG_SMP
2361         __boot_cpu_id = cpu;
2362 #endif
2363 }
2364
2365 /*
2366  * Must be called _AFTER_ setting up the per_cpu areas
2367  */
2368 void __init boot_cpu_hotplug_init(void)
2369 {
2370 #ifdef CONFIG_SMP
2371         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2372 #endif
2373         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2374 }
2375
2376 enum cpu_mitigations cpu_mitigations __ro_after_init = CPU_MITIGATIONS_AUTO;
2377
2378 static int __init mitigations_parse_cmdline(char *arg)
2379 {
2380         if (!strcmp(arg, "off"))
2381                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2382         else if (!strcmp(arg, "auto"))
2383                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2384         else if (!strcmp(arg, "auto,nosmt"))
2385                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2386         else
2387                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2388                         arg);
2389
2390         return 0;
2391 }
2392 early_param("mitigations", mitigations_parse_cmdline);