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.h>
  11 #include <linux/unistd.h>
  12 #include <linux/cpu.h>
  13 #include <linux/oom.h>
  14 #include <linux/rcupdate.h>
  15 #include <linux/export.h>
  16 #include <linux/bug.h>
  17 #include <linux/kthread.h>
  18 #include <linux/stop_machine.h>
  19 #include <linux/mutex.h>
  20 #include <linux/gfp.h>
  21 #include <linux/suspend.h>
  22 #include <linux/lockdep.h>
  23 #include <linux/tick.h>
  24 #include <linux/irq.h>
  25 #include <linux/smpboot.h>
  26 #include <linux/relay.h>
  27 #include <linux/slab.h>
  28
  29 #include <trace/events/power.h>
  30 #define CREATE_TRACE_POINTS
  31 #include <trace/events/cpuhp.h>
  32
  33 #include "smpboot.h"
  34
  35 /**
  36  * cpuhp_cpu_state - Per cpu hotplug state storage
  37  * @state:      The current cpu state
  38  * @target:     The target state
  39  * @thread:     Pointer to the hotplug thread
  40  * @should_run: Thread should execute
  41  * @rollback:   Perform a rollback
  42  * @single:     Single callback invocation
  43  * @bringup:    Single callback bringup or teardown selector
  44  * @cb_state:   The state for a single callback (install/uninstall)
  45  * @result:     Result of the operation
  46  * @done:       Signal completion to the issuer of the task
  47  */
  48 struct cpuhp_cpu_state {
  49         enum cpuhp_state        state;
  50         enum cpuhp_state        target;
  51 #ifdef CONFIG_SMP
  52         struct task_struct      *thread;
  53         bool                    should_run;
  54         bool                    rollback;
  55         bool                    single;
  56         bool                    bringup;
  57         struct hlist_node       *node;
  58         enum cpuhp_state        cb_state;
  59         int                     result;
  60         struct completion       done;
  61 #endif
  62 };
  63
  64 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  65
  66 /**
  67  * cpuhp_step - Hotplug state machine step
  68  * @name:       Name of the step
  69  * @startup:    Startup function of the step
  70  * @teardown:   Teardown function of the step
  71  * @skip_onerr: Do not invoke the functions on error rollback
  72  *              Will go away once the notifiers are gone
  73  * @cant_stop:  Bringup/teardown can't be stopped at this step
  74  */
  75 struct cpuhp_step {
  76         const char              *name;
  77         union {
  78                 int             (*single)(unsigned int cpu);
  79                 int             (*multi)(unsigned int cpu,
  80                                          struct hlist_node *node);
  81         } startup;
  82         union {
  83                 int             (*single)(unsigned int cpu);
  84                 int             (*multi)(unsigned int cpu,
  85                                          struct hlist_node *node);
  86         } teardown;
  87         struct hlist_head       list;
  88         bool                    skip_onerr;
  89         bool                    cant_stop;
  90         bool                    multi_instance;
  91 };
  92
  93 static DEFINE_MUTEX(cpuhp_state_mutex);
  94 static struct cpuhp_step cpuhp_bp_states[];
  95 static struct cpuhp_step cpuhp_ap_states[];
  96
  97 static bool cpuhp_is_ap_state(enum cpuhp_state state)
  98 {
  99         /*
 100          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 101          * purposes as that state is handled explicitly in cpu_down.
 102          */
 103         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 104 }
 105
 106 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 107 {
 108         struct cpuhp_step *sp;
 109
 110         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 111         return sp + state;
 112 }
 113
 114 /**
 115  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 116  * @cpu:        The cpu for which the callback should be invoked
 117  * @step:       The step in the state machine
 118  * @bringup:    True if the bringup callback should be invoked
 119  *
 120  * Called from cpu hotplug and from the state register machinery.
 121  */
 122 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 123                                  bool bringup, struct hlist_node *node)
 124 {
 125         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 126         struct cpuhp_step *step = cpuhp_get_step(state);
 127         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 128         int (*cb)(unsigned int cpu);
 129         int ret, cnt;
 130
 131         if (!step->multi_instance) {
 132                 cb = bringup ? step->startup.single : step->teardown.single;
 133                 if (!cb)
 134                         return 0;
 135                 trace_cpuhp_enter(cpu, st->target, state, cb);
 136                 ret = cb(cpu);
 137                 trace_cpuhp_exit(cpu, st->state, state, ret);
 138                 return ret;
 139         }
 140         cbm = bringup ? step->startup.multi : step->teardown.multi;
 141         if (!cbm)
 142                 return 0;
 143
 144         /* Single invocation for instance add/remove */
 145         if (node) {
 146                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 147                 ret = cbm(cpu, node);
 148                 trace_cpuhp_exit(cpu, st->state, state, ret);
 149                 return ret;
 150         }
 151
 152         /* State transition. Invoke on all instances */
 153         cnt = 0;
 154         hlist_for_each(node, &step->list) {
 155                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 156                 ret = cbm(cpu, node);
 157                 trace_cpuhp_exit(cpu, st->state, state, ret);
 158                 if (ret)
 159                         goto err;
 160                 cnt++;
 161         }
 162         return 0;
 163 err:
 164         /* Rollback the instances if one failed */
 165         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 166         if (!cbm)
 167                 return ret;
 168
 169         hlist_for_each(node, &step->list) {
 170                 if (!cnt--)
 171                         break;
 172                 cbm(cpu, node);
 173         }
 174         return ret;
 175 }
 176
 177 #ifdef CONFIG_SMP
 178 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 179 static DEFINE_MUTEX(cpu_add_remove_lock);
 180 bool cpuhp_tasks_frozen;
 181 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 182
 183 /*
 184  * The following two APIs (cpu_maps_update_begin/done) must be used when
 185  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 186  */
 187 void cpu_maps_update_begin(void)
 188 {
 189         mutex_lock(&cpu_add_remove_lock);
 190 }
 191
 192 void cpu_maps_update_done(void)
 193 {
 194         mutex_unlock(&cpu_add_remove_lock);
 195 }
 196
 197 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 198  * Should always be manipulated under cpu_add_remove_lock
 199  */
 200 static int cpu_hotplug_disabled;
 201
 202 #ifdef CONFIG_HOTPLUG_CPU
 203
 204 static struct {
 205         struct task_struct *active_writer;
 206         /* wait queue to wake up the active_writer */
 207         wait_queue_head_t wq;
 208         /* verifies that no writer will get active while readers are active */
 209         struct mutex lock;
 210         /*
 211          * Also blocks the new readers during
 212          * an ongoing cpu hotplug operation.
 213          */
 214         atomic_t refcount;
 215
 216 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 217         struct lockdep_map dep_map;
 218 #endif
 219 } cpu_hotplug = {
 220         .active_writer = NULL,
 221         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 222         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 223 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 224         .dep_map = STATIC_LOCKDEP_MAP_INIT("cpu_hotplug.dep_map", &cpu_hotplug.dep_map),
 225 #endif
 226 };
 227
 228 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 229 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 230 #define cpuhp_lock_acquire_tryread() \
 231                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 232 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 233 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 234
 235
 236 void get_online_cpus(void)
 237 {
 238         might_sleep();
 239         if (cpu_hotplug.active_writer == current)
 240                 return;
 241         cpuhp_lock_acquire_read();
 242         mutex_lock(&cpu_hotplug.lock);
 243         atomic_inc(&cpu_hotplug.refcount);
 244         mutex_unlock(&cpu_hotplug.lock);
 245 }
 246 EXPORT_SYMBOL_GPL(get_online_cpus);
 247
 248 void put_online_cpus(void)
 249 {
 250         int refcount;
 251
 252         if (cpu_hotplug.active_writer == current)
 253                 return;
 254
 255         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 256         if (WARN_ON(refcount < 0)) /* try to fix things up */
 257                 atomic_inc(&cpu_hotplug.refcount);
 258
 259         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 260                 wake_up(&cpu_hotplug.wq);
 261
 262         cpuhp_lock_release();
 263
 264 }
 265 EXPORT_SYMBOL_GPL(put_online_cpus);
 266
 267 /*
 268  * This ensures that the hotplug operation can begin only when the
 269  * refcount goes to zero.
 270  *
 271  * Note that during a cpu-hotplug operation, the new readers, if any,
 272  * will be blocked by the cpu_hotplug.lock
 273  *
 274  * Since cpu_hotplug_begin() is always called after invoking
 275  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 276  *
 277  * Note that theoretically, there is a possibility of a livelock:
 278  * - Refcount goes to zero, last reader wakes up the sleeping
 279  *   writer.
 280  * - Last reader unlocks the cpu_hotplug.lock.
 281  * - A new reader arrives at this moment, bumps up the refcount.
 282  * - The writer acquires the cpu_hotplug.lock finds the refcount
 283  *   non zero and goes to sleep again.
 284  *
 285  * However, this is very difficult to achieve in practice since
 286  * get_online_cpus() not an api which is called all that often.
 287  *
 288  */
 289 void cpu_hotplug_begin(void)
 290 {
 291         DEFINE_WAIT(wait);
 292
 293         cpu_hotplug.active_writer = current;
 294         cpuhp_lock_acquire();
 295
 296         for (;;) {
 297                 mutex_lock(&cpu_hotplug.lock);
 298                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 299                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 300                                 break;
 301                 mutex_unlock(&cpu_hotplug.lock);
 302                 schedule();
 303         }
 304         finish_wait(&cpu_hotplug.wq, &wait);
 305 }
 306
 307 void cpu_hotplug_done(void)
 308 {
 309         cpu_hotplug.active_writer = NULL;
 310         mutex_unlock(&cpu_hotplug.lock);
 311         cpuhp_lock_release();
 312 }
 313
 314 /*
 315  * Wait for currently running CPU hotplug operations to complete (if any) and
 316  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 317  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 318  * hotplug path before performing hotplug operations. So acquiring that lock
 319  * guarantees mutual exclusion from any currently running hotplug operations.
 320  */
 321 void cpu_hotplug_disable(void)
 322 {
 323         cpu_maps_update_begin();
 324         cpu_hotplug_disabled++;
 325         cpu_maps_update_done();
 326 }
 327 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 328
 329 static void __cpu_hotplug_enable(void)
 330 {
 331         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 332                 return;
 333         cpu_hotplug_disabled--;
 334 }
 335
 336 void cpu_hotplug_enable(void)
 337 {
 338         cpu_maps_update_begin();
 339         __cpu_hotplug_enable();
 340         cpu_maps_update_done();
 341 }
 342 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 343 #endif  /* CONFIG_HOTPLUG_CPU */
 344
 345 /* Notifier wrappers for transitioning to state machine */
 346
 347 static int bringup_wait_for_ap(unsigned int cpu)
 348 {
 349         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 350
 351         wait_for_completion(&st->done);
 352         return st->result;
 353 }
 354
 355 static int bringup_cpu(unsigned int cpu)
 356 {
 357         struct task_struct *idle = idle_thread_get(cpu);
 358         int ret;
 359
 360         /*
 361          * Some architectures have to walk the irq descriptors to
 362          * setup the vector space for the cpu which comes online.
 363          * Prevent irq alloc/free across the bringup.
 364          */
 365         irq_lock_sparse();
 366
 367         /* Arch-specific enabling code. */
 368         ret = __cpu_up(cpu, idle);
 369         irq_unlock_sparse();
 370         if (ret)
 371                 return ret;
 372         ret = bringup_wait_for_ap(cpu);
 373         BUG_ON(!cpu_online(cpu));
 374         return ret;
 375 }
 376
 377 /*
 378  * Hotplug state machine related functions
 379  */
 380 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 381 {
 382         for (st->state++; st->state < st->target; st->state++) {
 383                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 384
 385                 if (!step->skip_onerr)
 386                         cpuhp_invoke_callback(cpu, st->state, true, NULL);
 387         }
 388 }
 389
 390 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 391                                 enum cpuhp_state target)
 392 {
 393         enum cpuhp_state prev_state = st->state;
 394         int ret = 0;
 395
 396         for (; st->state > target; st->state--) {
 397                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
 398                 if (ret) {
 399                         st->target = prev_state;
 400                         undo_cpu_down(cpu, st);
 401                         break;
 402                 }
 403         }
 404         return ret;
 405 }
 406
 407 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 408 {
 409         for (st->state--; st->state > st->target; st->state--) {
 410                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 411
 412                 if (!step->skip_onerr)
 413                         cpuhp_invoke_callback(cpu, st->state, false, NULL);
 414         }
 415 }
 416
 417 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 418                               enum cpuhp_state target)
 419 {
 420         enum cpuhp_state prev_state = st->state;
 421         int ret = 0;
 422
 423         while (st->state < target) {
 424                 st->state++;
 425                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
 426                 if (ret) {
 427                         st->target = prev_state;
 428                         undo_cpu_up(cpu, st);
 429                         break;
 430                 }
 431         }
 432         return ret;
 433 }
 434
 435 /*
 436  * The cpu hotplug threads manage the bringup and teardown of the cpus
 437  */
 438 static void cpuhp_create(unsigned int cpu)
 439 {
 440         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 441
 442         init_completion(&st->done);
 443 }
 444
 445 static int cpuhp_should_run(unsigned int cpu)
 446 {
 447         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 448
 449         return st->should_run;
 450 }
 451
 452 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 453 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 454 {
 455         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 456
 457         return cpuhp_down_callbacks(cpu, st, target);
 458 }
 459
 460 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 461 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 462 {
 463         return cpuhp_up_callbacks(cpu, st, st->target);
 464 }
 465
 466 /*
 467  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 468  * callbacks when a state gets [un]installed at runtime.
 469  */
 470 static void cpuhp_thread_fun(unsigned int cpu)
 471 {
 472         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 473         int ret = 0;
 474
 475         /*
 476          * Paired with the mb() in cpuhp_kick_ap_work and
 477          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 478          */
 479         smp_mb();
 480         if (!st->should_run)
 481                 return;
 482
 483         st->should_run = false;
 484
 485         /* Single callback invocation for [un]install ? */
 486         if (st->single) {
 487                 if (st->cb_state < CPUHP_AP_ONLINE) {
 488                         local_irq_disable();
 489                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 490                                                     st->bringup, st->node);
 491                         local_irq_enable();
 492                 } else {
 493                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 494                                                     st->bringup, st->node);
 495                 }
 496         } else if (st->rollback) {
 497                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 498
 499                 undo_cpu_down(cpu, st);
 500                 st->rollback = false;
 501         } else {
 502                 /* Cannot happen .... */
 503                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 504
 505                 /* Regular hotplug work */
 506                 if (st->state < st->target)
 507                         ret = cpuhp_ap_online(cpu, st);
 508                 else if (st->state > st->target)
 509                         ret = cpuhp_ap_offline(cpu, st);
 510         }
 511         st->result = ret;
 512         complete(&st->done);
 513 }
 514
 515 /* Invoke a single callback on a remote cpu */
 516 static int
 517 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 518                          struct hlist_node *node)
 519 {
 520         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 521
 522         if (!cpu_online(cpu))
 523                 return 0;
 524
 525         /*
 526          * If we are up and running, use the hotplug thread. For early calls
 527          * we invoke the thread function directly.
 528          */
 529         if (!st->thread)
 530                 return cpuhp_invoke_callback(cpu, state, bringup, node);
 531
 532         st->cb_state = state;
 533         st->single = true;
 534         st->bringup = bringup;
 535         st->node = node;
 536
 537         /*
 538          * Make sure the above stores are visible before should_run becomes
 539          * true. Paired with the mb() above in cpuhp_thread_fun()
 540          */
 541         smp_mb();
 542         st->should_run = true;
 543         wake_up_process(st->thread);
 544         wait_for_completion(&st->done);
 545         return st->result;
 546 }
 547
 548 /* Regular hotplug invocation of the AP hotplug thread */
 549 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 550 {
 551         st->result = 0;
 552         st->single = false;
 553         /*
 554          * Make sure the above stores are visible before should_run becomes
 555          * true. Paired with the mb() above in cpuhp_thread_fun()
 556          */
 557         smp_mb();
 558         st->should_run = true;
 559         wake_up_process(st->thread);
 560 }
 561
 562 static int cpuhp_kick_ap_work(unsigned int cpu)
 563 {
 564         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 565         enum cpuhp_state state = st->state;
 566
 567         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 568         __cpuhp_kick_ap_work(st);
 569         wait_for_completion(&st->done);
 570         trace_cpuhp_exit(cpu, st->state, state, st->result);
 571         return st->result;
 572 }
 573
 574 static struct smp_hotplug_thread cpuhp_threads = {
 575         .store                  = &cpuhp_state.thread,
 576         .create                 = &cpuhp_create,
 577         .thread_should_run      = cpuhp_should_run,
 578         .thread_fn              = cpuhp_thread_fun,
 579         .thread_comm            = "cpuhp/%u",
 580         .selfparking            = true,
 581 };
 582
 583 void __init cpuhp_threads_init(void)
 584 {
 585         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 586         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 587 }
 588
 589 #ifdef CONFIG_HOTPLUG_CPU
 590 /**
 591  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 592  * @cpu: a CPU id
 593  *
 594  * This function walks all processes, finds a valid mm struct for each one and
 595  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 596  * trivial, there are various non-obvious corner cases, which this function
 597  * tries to solve in a safe manner.
 598  *
 599  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 600  * be called only for an already offlined CPU.
 601  */
 602 void clear_tasks_mm_cpumask(int cpu)
 603 {
 604         struct task_struct *p;
 605
 606         /*
 607          * This function is called after the cpu is taken down and marked
 608          * offline, so its not like new tasks will ever get this cpu set in
 609          * their mm mask. -- Peter Zijlstra
 610          * Thus, we may use rcu_read_lock() here, instead of grabbing
 611          * full-fledged tasklist_lock.
 612          */
 613         WARN_ON(cpu_online(cpu));
 614         rcu_read_lock();
 615         for_each_process(p) {
 616                 struct task_struct *t;
 617
 618                 /*
 619                  * Main thread might exit, but other threads may still have
 620                  * a valid mm. Find one.
 621                  */
 622                 t = find_lock_task_mm(p);
 623                 if (!t)
 624                         continue;
 625                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 626                 task_unlock(t);
 627         }
 628         rcu_read_unlock();
 629 }
 630
 631 static inline void check_for_tasks(int dead_cpu)
 632 {
 633         struct task_struct *g, *p;
 634
 635         read_lock(&tasklist_lock);
 636         for_each_process_thread(g, p) {
 637                 if (!p->on_rq)
 638                         continue;
 639                 /*
 640                  * We do the check with unlocked task_rq(p)->lock.
 641                  * Order the reading to do not warn about a task,
 642                  * which was running on this cpu in the past, and
 643                  * it's just been woken on another cpu.
 644                  */
 645                 rmb();
 646                 if (task_cpu(p) != dead_cpu)
 647                         continue;
 648
 649                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 650                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 651         }
 652         read_unlock(&tasklist_lock);
 653 }
 654
 655 /* Take this CPU down. */
 656 static int take_cpu_down(void *_param)
 657 {
 658         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 659         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 660         int err, cpu = smp_processor_id();
 661
 662         /* Ensure this CPU doesn't handle any more interrupts. */
 663         err = __cpu_disable();
 664         if (err < 0)
 665                 return err;
 666
 667         /*
 668          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 669          * do this step again.
 670          */
 671         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 672         st->state--;
 673         /* Invoke the former CPU_DYING callbacks */
 674         for (; st->state > target; st->state--)
 675                 cpuhp_invoke_callback(cpu, st->state, false, NULL);
 676
 677         /* Give up timekeeping duties */
 678         tick_handover_do_timer();
 679         /* Park the stopper thread */
 680         stop_machine_park(cpu);
 681         return 0;
 682 }
 683
 684 static int takedown_cpu(unsigned int cpu)
 685 {
 686         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 687         int err;
 688
 689         /* Park the smpboot threads */
 690         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 691         smpboot_park_threads(cpu);
 692
 693         /*
 694          * Prevent irq alloc/free while the dying cpu reorganizes the
 695          * interrupt affinities.
 696          */
 697         irq_lock_sparse();
 698
 699         /*
 700          * So now all preempt/rcu users must observe !cpu_active().
 701          */
 702         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 703         if (err) {
 704                 /* CPU refused to die */
 705                 irq_unlock_sparse();
 706                 /* Unpark the hotplug thread so we can rollback there */
 707                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 708                 return err;
 709         }
 710         BUG_ON(cpu_online(cpu));
 711
 712         /*
 713          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 714          * runnable tasks from the cpu, there's only the idle task left now
 715          * that the migration thread is done doing the stop_machine thing.
 716          *
 717          * Wait for the stop thread to go away.
 718          */
 719         wait_for_completion(&st->done);
 720         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 721
 722         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 723         irq_unlock_sparse();
 724
 725         hotplug_cpu__broadcast_tick_pull(cpu);
 726         /* This actually kills the CPU. */
 727         __cpu_die(cpu);
 728
 729         tick_cleanup_dead_cpu(cpu);
 730         return 0;
 731 }
 732
 733 static void cpuhp_complete_idle_dead(void *arg)
 734 {
 735         struct cpuhp_cpu_state *st = arg;
 736
 737         complete(&st->done);
 738 }
 739
 740 void cpuhp_report_idle_dead(void)
 741 {
 742         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 743
 744         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 745         rcu_report_dead(smp_processor_id());
 746         st->state = CPUHP_AP_IDLE_DEAD;
 747         /*
 748          * We cannot call complete after rcu_report_dead() so we delegate it
 749          * to an online cpu.
 750          */
 751         smp_call_function_single(cpumask_first(cpu_online_mask),
 752                                  cpuhp_complete_idle_dead, st, 0);
 753 }
 754
 755 #else
 756 #define takedown_cpu            NULL
 757 #endif
 758
 759 #ifdef CONFIG_HOTPLUG_CPU
 760
 761 /* Requires cpu_add_remove_lock to be held */
 762 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 763                            enum cpuhp_state target)
 764 {
 765         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 766         int prev_state, ret = 0;
 767
 768         if (num_online_cpus() == 1)
 769                 return -EBUSY;
 770
 771         if (!cpu_present(cpu))
 772                 return -EINVAL;
 773
 774         cpu_hotplug_begin();
 775
 776         cpuhp_tasks_frozen = tasks_frozen;
 777
 778         prev_state = st->state;
 779         st->target = target;
 780         /*
 781          * If the current CPU state is in the range of the AP hotplug thread,
 782          * then we need to kick the thread.
 783          */
 784         if (st->state > CPUHP_TEARDOWN_CPU) {
 785                 ret = cpuhp_kick_ap_work(cpu);
 786                 /*
 787                  * The AP side has done the error rollback already. Just
 788                  * return the error code..
 789                  */
 790                 if (ret)
 791                         goto out;
 792
 793                 /*
 794                  * We might have stopped still in the range of the AP hotplug
 795                  * thread. Nothing to do anymore.
 796                  */
 797                 if (st->state > CPUHP_TEARDOWN_CPU)
 798                         goto out;
 799         }
 800         /*
 801          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 802          * to do the further cleanups.
 803          */
 804         ret = cpuhp_down_callbacks(cpu, st, target);
 805         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 806                 st->target = prev_state;
 807                 st->rollback = true;
 808                 cpuhp_kick_ap_work(cpu);
 809         }
 810
 811 out:
 812         cpu_hotplug_done();
 813         return ret;
 814 }
 815
 816 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 817 {
 818         int err;
 819
 820         cpu_maps_update_begin();
 821
 822         if (cpu_hotplug_disabled) {
 823                 err = -EBUSY;
 824                 goto out;
 825         }
 826
 827         err = _cpu_down(cpu, 0, target);
 828
 829 out:
 830         cpu_maps_update_done();
 831         return err;
 832 }
 833 int cpu_down(unsigned int cpu)
 834 {
 835         return do_cpu_down(cpu, CPUHP_OFFLINE);
 836 }
 837 EXPORT_SYMBOL(cpu_down);
 838 #endif /*CONFIG_HOTPLUG_CPU*/
 839
 840 /**
 841  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
 842  * @cpu: cpu that just started
 843  *
 844  * It must be called by the arch code on the new cpu, before the new cpu
 845  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 846  */
 847 void notify_cpu_starting(unsigned int cpu)
 848 {
 849         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 850         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 851
 852         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 853         while (st->state < target) {
 854                 st->state++;
 855                 cpuhp_invoke_callback(cpu, st->state, true, NULL);
 856         }
 857 }
 858
 859 /*
 860  * Called from the idle task. We need to set active here, so we can kick off
 861  * the stopper thread and unpark the smpboot threads. If the target state is
 862  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 863  * cpu further.
 864  */
 865 void cpuhp_online_idle(enum cpuhp_state state)
 866 {
 867         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 868         unsigned int cpu = smp_processor_id();
 869
 870         /* Happens for the boot cpu */
 871         if (state != CPUHP_AP_ONLINE_IDLE)
 872                 return;
 873
 874         st->state = CPUHP_AP_ONLINE_IDLE;
 875
 876         /* Unpark the stopper thread and the hotplug thread of this cpu */
 877         stop_machine_unpark(cpu);
 878         kthread_unpark(st->thread);
 879
 880         /* Should we go further up ? */
 881         if (st->target > CPUHP_AP_ONLINE_IDLE)
 882                 __cpuhp_kick_ap_work(st);
 883         else
 884                 complete(&st->done);
 885 }
 886
 887 /* Requires cpu_add_remove_lock to be held */
 888 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 889 {
 890         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 891         struct task_struct *idle;
 892         int ret = 0;
 893
 894         cpu_hotplug_begin();
 895
 896         if (!cpu_present(cpu)) {
 897                 ret = -EINVAL;
 898                 goto out;
 899         }
 900
 901         /*
 902          * The caller of do_cpu_up might have raced with another
 903          * caller. Ignore it for now.
 904          */
 905         if (st->state >= target)
 906                 goto out;
 907
 908         if (st->state == CPUHP_OFFLINE) {
 909                 /* Let it fail before we try to bring the cpu up */
 910                 idle = idle_thread_get(cpu);
 911                 if (IS_ERR(idle)) {
 912                         ret = PTR_ERR(idle);
 913                         goto out;
 914                 }
 915         }
 916
 917         cpuhp_tasks_frozen = tasks_frozen;
 918
 919         st->target = target;
 920         /*
 921          * If the current CPU state is in the range of the AP hotplug thread,
 922          * then we need to kick the thread once more.
 923          */
 924         if (st->state > CPUHP_BRINGUP_CPU) {
 925                 ret = cpuhp_kick_ap_work(cpu);
 926                 /*
 927                  * The AP side has done the error rollback already. Just
 928                  * return the error code..
 929                  */
 930                 if (ret)
 931                         goto out;
 932         }
 933
 934         /*
 935          * Try to reach the target state. We max out on the BP at
 936          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
 937          * responsible for bringing it up to the target state.
 938          */
 939         target = min((int)target, CPUHP_BRINGUP_CPU);
 940         ret = cpuhp_up_callbacks(cpu, st, target);
 941 out:
 942         cpu_hotplug_done();
 943         return ret;
 944 }
 945
 946 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 947 {
 948         int err = 0;
 949
 950         if (!cpu_possible(cpu)) {
 951                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
 952                        cpu);
 953 #if defined(CONFIG_IA64)
 954                 pr_err("please check additional_cpus= boot parameter\n");
 955 #endif
 956                 return -EINVAL;
 957         }
 958
 959         err = try_online_node(cpu_to_node(cpu));
 960         if (err)
 961                 return err;
 962
 963         cpu_maps_update_begin();
 964
 965         if (cpu_hotplug_disabled) {
 966                 err = -EBUSY;
 967                 goto out;
 968         }
 969
 970         err = _cpu_up(cpu, 0, target);
 971 out:
 972         cpu_maps_update_done();
 973         return err;
 974 }
 975
 976 int cpu_up(unsigned int cpu)
 977 {
 978         return do_cpu_up(cpu, CPUHP_ONLINE);
 979 }
 980 EXPORT_SYMBOL_GPL(cpu_up);
 981
 982 #ifdef CONFIG_PM_SLEEP_SMP
 983 static cpumask_var_t frozen_cpus;
 984
 985 int freeze_secondary_cpus(int primary)
 986 {
 987         int cpu, error = 0;
 988
 989         cpu_maps_update_begin();
 990         if (!cpu_online(primary))
 991                 primary = cpumask_first(cpu_online_mask);
 992         /*
 993          * We take down all of the non-boot CPUs in one shot to avoid races
 994          * with the userspace trying to use the CPU hotplug at the same time
 995          */
 996         cpumask_clear(frozen_cpus);
 997
 998         pr_info("Disabling non-boot CPUs ...\n");
 999         for_each_online_cpu(cpu) {
1000                 if (cpu == primary)
1001                         continue;
1002                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1003                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1004                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1005                 if (!error)
1006                         cpumask_set_cpu(cpu, frozen_cpus);
1007                 else {
1008                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1009                         break;
1010                 }
1011         }
1012
1013         if (!error)
1014                 BUG_ON(num_online_cpus() > 1);
1015         else
1016                 pr_err("Non-boot CPUs are not disabled\n");
1017
1018         /*
1019          * Make sure the CPUs won't be enabled by someone else. We need to do
1020          * this even in case of failure as all disable_nonboot_cpus() users are
1021          * supposed to do enable_nonboot_cpus() on the failure path.
1022          */
1023         cpu_hotplug_disabled++;
1024
1025         cpu_maps_update_done();
1026         return error;
1027 }
1028
1029 void __weak arch_enable_nonboot_cpus_begin(void)
1030 {
1031 }
1032
1033 void __weak arch_enable_nonboot_cpus_end(void)
1034 {
1035 }
1036
1037 void enable_nonboot_cpus(void)
1038 {
1039         int cpu, error;
1040
1041         /* Allow everyone to use the CPU hotplug again */
1042         cpu_maps_update_begin();
1043         __cpu_hotplug_enable();
1044         if (cpumask_empty(frozen_cpus))
1045                 goto out;
1046
1047         pr_info("Enabling non-boot CPUs ...\n");
1048
1049         arch_enable_nonboot_cpus_begin();
1050
1051         for_each_cpu(cpu, frozen_cpus) {
1052                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1053                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1054                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1055                 if (!error) {
1056                         pr_info("CPU%d is up\n", cpu);
1057                         continue;
1058                 }
1059                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1060         }
1061
1062         arch_enable_nonboot_cpus_end();
1063
1064         cpumask_clear(frozen_cpus);
1065 out:
1066         cpu_maps_update_done();
1067 }
1068
1069 static int __init alloc_frozen_cpus(void)
1070 {
1071         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1072                 return -ENOMEM;
1073         return 0;
1074 }
1075 core_initcall(alloc_frozen_cpus);
1076
1077 /*
1078  * When callbacks for CPU hotplug notifications are being executed, we must
1079  * ensure that the state of the system with respect to the tasks being frozen
1080  * or not, as reported by the notification, remains unchanged *throughout the
1081  * duration* of the execution of the callbacks.
1082  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1083  *
1084  * This synchronization is implemented by mutually excluding regular CPU
1085  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1086  * Hibernate notifications.
1087  */
1088 static int
1089 cpu_hotplug_pm_callback(struct notifier_block *nb,
1090                         unsigned long action, void *ptr)
1091 {
1092         switch (action) {
1093
1094         case PM_SUSPEND_PREPARE:
1095         case PM_HIBERNATION_PREPARE:
1096                 cpu_hotplug_disable();
1097                 break;
1098
1099         case PM_POST_SUSPEND:
1100         case PM_POST_HIBERNATION:
1101                 cpu_hotplug_enable();
1102                 break;
1103
1104         default:
1105                 return NOTIFY_DONE;
1106         }
1107
1108         return NOTIFY_OK;
1109 }
1110
1111
1112 static int __init cpu_hotplug_pm_sync_init(void)
1113 {
1114         /*
1115          * cpu_hotplug_pm_callback has higher priority than x86
1116          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1117          * to disable cpu hotplug to avoid cpu hotplug race.
1118          */
1119         pm_notifier(cpu_hotplug_pm_callback, 0);
1120         return 0;
1121 }
1122 core_initcall(cpu_hotplug_pm_sync_init);
1123
1124 #endif /* CONFIG_PM_SLEEP_SMP */
1125
1126 #endif /* CONFIG_SMP */
1127
1128 /* Boot processor state steps */
1129 static struct cpuhp_step cpuhp_bp_states[] = {
1130         [CPUHP_OFFLINE] = {
1131                 .name                   = "offline",
1132                 .startup.single         = NULL,
1133                 .teardown.single        = NULL,
1134         },
1135 #ifdef CONFIG_SMP
1136         [CPUHP_CREATE_THREADS]= {
1137                 .name                   = "threads:prepare",
1138                 .startup.single         = smpboot_create_threads,
1139                 .teardown.single        = NULL,
1140                 .cant_stop              = true,
1141         },
1142         [CPUHP_PERF_PREPARE] = {
1143                 .name                   = "perf:prepare",
1144                 .startup.single         = perf_event_init_cpu,
1145                 .teardown.single        = perf_event_exit_cpu,
1146         },
1147         [CPUHP_WORKQUEUE_PREP] = {
1148                 .name                   = "workqueue:prepare",
1149                 .startup.single         = workqueue_prepare_cpu,
1150                 .teardown.single        = NULL,
1151         },
1152         [CPUHP_HRTIMERS_PREPARE] = {
1153                 .name                   = "hrtimers:prepare",
1154                 .startup.single         = hrtimers_prepare_cpu,
1155                 .teardown.single        = hrtimers_dead_cpu,
1156         },
1157         [CPUHP_SMPCFD_PREPARE] = {
1158                 .name                   = "smpcfd:prepare",
1159                 .startup.single         = smpcfd_prepare_cpu,
1160                 .teardown.single        = smpcfd_dead_cpu,
1161         },
1162         [CPUHP_RELAY_PREPARE] = {
1163                 .name                   = "relay:prepare",
1164                 .startup.single         = relay_prepare_cpu,
1165                 .teardown.single        = NULL,
1166         },
1167         [CPUHP_SLAB_PREPARE] = {
1168                 .name                   = "slab:prepare",
1169                 .startup.single         = slab_prepare_cpu,
1170                 .teardown.single        = slab_dead_cpu,
1171         },
1172         [CPUHP_RCUTREE_PREP] = {
1173                 .name                   = "RCU/tree:prepare",
1174                 .startup.single         = rcutree_prepare_cpu,
1175                 .teardown.single        = rcutree_dead_cpu,
1176         },
1177         /*
1178          * On the tear-down path, timers_dead_cpu() must be invoked
1179          * before blk_mq_queue_reinit_notify() from notify_dead(),
1180          * otherwise a RCU stall occurs.
1181          */
1182         [CPUHP_TIMERS_DEAD] = {
1183                 .name                   = "timers:dead",
1184                 .startup.single         = NULL,
1185                 .teardown.single        = timers_dead_cpu,
1186         },
1187         /* Kicks the plugged cpu into life */
1188         [CPUHP_BRINGUP_CPU] = {
1189                 .name                   = "cpu:bringup",
1190                 .startup.single         = bringup_cpu,
1191                 .teardown.single        = NULL,
1192                 .cant_stop              = true,
1193         },
1194         [CPUHP_AP_SMPCFD_DYING] = {
1195                 .name                   = "smpcfd:dying",
1196                 .startup.single         = NULL,
1197                 .teardown.single        = smpcfd_dying_cpu,
1198         },
1199         /*
1200          * Handled on controll processor until the plugged processor manages
1201          * this itself.
1202          */
1203         [CPUHP_TEARDOWN_CPU] = {
1204                 .name                   = "cpu:teardown",
1205                 .startup.single         = NULL,
1206                 .teardown.single        = takedown_cpu,
1207                 .cant_stop              = true,
1208         },
1209 #else
1210         [CPUHP_BRINGUP_CPU] = { },
1211 #endif
1212 };
1213
1214 /* Application processor state steps */
1215 static struct cpuhp_step cpuhp_ap_states[] = {
1216 #ifdef CONFIG_SMP
1217         /* Final state before CPU kills itself */
1218         [CPUHP_AP_IDLE_DEAD] = {
1219                 .name                   = "idle:dead",
1220         },
1221         /*
1222          * Last state before CPU enters the idle loop to die. Transient state
1223          * for synchronization.
1224          */
1225         [CPUHP_AP_OFFLINE] = {
1226                 .name                   = "ap:offline",
1227                 .cant_stop              = true,
1228         },
1229         /* First state is scheduler control. Interrupts are disabled */
1230         [CPUHP_AP_SCHED_STARTING] = {
1231                 .name                   = "sched:starting",
1232                 .startup.single         = sched_cpu_starting,
1233                 .teardown.single        = sched_cpu_dying,
1234         },
1235         [CPUHP_AP_RCUTREE_DYING] = {
1236                 .name                   = "RCU/tree:dying",
1237                 .startup.single         = NULL,
1238                 .teardown.single        = rcutree_dying_cpu,
1239         },
1240         /* Entry state on starting. Interrupts enabled from here on. Transient
1241          * state for synchronsization */
1242         [CPUHP_AP_ONLINE] = {
1243                 .name                   = "ap:online",
1244         },
1245         /* Handle smpboot threads park/unpark */
1246         [CPUHP_AP_SMPBOOT_THREADS] = {
1247                 .name                   = "smpboot/threads:online",
1248                 .startup.single         = smpboot_unpark_threads,
1249                 .teardown.single        = NULL,
1250         },
1251         [CPUHP_AP_PERF_ONLINE] = {
1252                 .name                   = "perf:online",
1253                 .startup.single         = perf_event_init_cpu,
1254                 .teardown.single        = perf_event_exit_cpu,
1255         },
1256         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1257                 .name                   = "workqueue:online",
1258                 .startup.single         = workqueue_online_cpu,
1259                 .teardown.single        = workqueue_offline_cpu,
1260         },
1261         [CPUHP_AP_RCUTREE_ONLINE] = {
1262                 .name                   = "RCU/tree:online",
1263                 .startup.single         = rcutree_online_cpu,
1264                 .teardown.single        = rcutree_offline_cpu,
1265         },
1266 #endif
1267         /*
1268          * The dynamically registered state space is here
1269          */
1270
1271 #ifdef CONFIG_SMP
1272         /* Last state is scheduler control setting the cpu active */
1273         [CPUHP_AP_ACTIVE] = {
1274                 .name                   = "sched:active",
1275                 .startup.single         = sched_cpu_activate,
1276                 .teardown.single        = sched_cpu_deactivate,
1277         },
1278 #endif
1279
1280         /* CPU is fully up and running. */
1281         [CPUHP_ONLINE] = {
1282                 .name                   = "online",
1283                 .startup.single         = NULL,
1284                 .teardown.single        = NULL,
1285         },
1286 };
1287
1288 /* Sanity check for callbacks */
1289 static int cpuhp_cb_check(enum cpuhp_state state)
1290 {
1291         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1292                 return -EINVAL;
1293         return 0;
1294 }
1295
1296 /*
1297  * Returns a free for dynamic slot assignment of the Online state. The states
1298  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1299  * by having no name assigned.
1300  */
1301 static int cpuhp_reserve_state(enum cpuhp_state state)
1302 {
1303         enum cpuhp_state i, end;
1304         struct cpuhp_step *step;
1305
1306         switch (state) {
1307         case CPUHP_AP_ONLINE_DYN:
1308                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1309                 end = CPUHP_AP_ONLINE_DYN_END;
1310                 break;
1311         case CPUHP_BP_PREPARE_DYN:
1312                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1313                 end = CPUHP_BP_PREPARE_DYN_END;
1314                 break;
1315         default:
1316                 return -EINVAL;
1317         }
1318
1319         for (i = state; i <= end; i++, step++) {
1320                 if (!step->name)
1321                         return i;
1322         }
1323         WARN(1, "No more dynamic states available for CPU hotplug\n");
1324         return -ENOSPC;
1325 }
1326
1327 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1328                                  int (*startup)(unsigned int cpu),
1329                                  int (*teardown)(unsigned int cpu),
1330                                  bool multi_instance)
1331 {
1332         /* (Un)Install the callbacks for further cpu hotplug operations */
1333         struct cpuhp_step *sp;
1334         int ret = 0;
1335
1336         mutex_lock(&cpuhp_state_mutex);
1337
1338         if (state == CPUHP_AP_ONLINE_DYN || state == CPUHP_BP_PREPARE_DYN) {
1339                 ret = cpuhp_reserve_state(state);
1340                 if (ret < 0)
1341                         goto out;
1342                 state = ret;
1343         }
1344         sp = cpuhp_get_step(state);
1345         if (name && sp->name) {
1346                 ret = -EBUSY;
1347                 goto out;
1348         }
1349         sp->startup.single = startup;
1350         sp->teardown.single = teardown;
1351         sp->name = name;
1352         sp->multi_instance = multi_instance;
1353         INIT_HLIST_HEAD(&sp->list);
1354 out:
1355         mutex_unlock(&cpuhp_state_mutex);
1356         return ret;
1357 }
1358
1359 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1360 {
1361         return cpuhp_get_step(state)->teardown.single;
1362 }
1363
1364 /*
1365  * Call the startup/teardown function for a step either on the AP or
1366  * on the current CPU.
1367  */
1368 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1369                             struct hlist_node *node)
1370 {
1371         struct cpuhp_step *sp = cpuhp_get_step(state);
1372         int ret;
1373
1374         if ((bringup && !sp->startup.single) ||
1375             (!bringup && !sp->teardown.single))
1376                 return 0;
1377         /*
1378          * The non AP bound callbacks can fail on bringup. On teardown
1379          * e.g. module removal we crash for now.
1380          */
1381 #ifdef CONFIG_SMP
1382         if (cpuhp_is_ap_state(state))
1383                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1384         else
1385                 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1386 #else
1387         ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1388 #endif
1389         BUG_ON(ret && !bringup);
1390         return ret;
1391 }
1392
1393 /*
1394  * Called from __cpuhp_setup_state on a recoverable failure.
1395  *
1396  * Note: The teardown callbacks for rollback are not allowed to fail!
1397  */
1398 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1399                                    struct hlist_node *node)
1400 {
1401         int cpu;
1402
1403         /* Roll back the already executed steps on the other cpus */
1404         for_each_present_cpu(cpu) {
1405                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1406                 int cpustate = st->state;
1407
1408                 if (cpu >= failedcpu)
1409                         break;
1410
1411                 /* Did we invoke the startup call on that cpu ? */
1412                 if (cpustate >= state)
1413                         cpuhp_issue_call(cpu, state, false, node);
1414         }
1415 }
1416
1417 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1418                                bool invoke)
1419 {
1420         struct cpuhp_step *sp;
1421         int cpu;
1422         int ret;
1423
1424         sp = cpuhp_get_step(state);
1425         if (sp->multi_instance == false)
1426                 return -EINVAL;
1427
1428         get_online_cpus();
1429
1430         if (!invoke || !sp->startup.multi)
1431                 goto add_node;
1432
1433         /*
1434          * Try to call the startup callback for each present cpu
1435          * depending on the hotplug state of the cpu.
1436          */
1437         for_each_present_cpu(cpu) {
1438                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1439                 int cpustate = st->state;
1440
1441                 if (cpustate < state)
1442                         continue;
1443
1444                 ret = cpuhp_issue_call(cpu, state, true, node);
1445                 if (ret) {
1446                         if (sp->teardown.multi)
1447                                 cpuhp_rollback_install(cpu, state, node);
1448                         goto err;
1449                 }
1450         }
1451 add_node:
1452         ret = 0;
1453         mutex_lock(&cpuhp_state_mutex);
1454         hlist_add_head(node, &sp->list);
1455         mutex_unlock(&cpuhp_state_mutex);
1456
1457 err:
1458         put_online_cpus();
1459         return ret;
1460 }
1461 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1462
1463 /**
1464  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1465  * @state:              The state to setup
1466  * @invoke:             If true, the startup function is invoked for cpus where
1467  *                      cpu state >= @state
1468  * @startup:            startup callback function
1469  * @teardown:           teardown callback function
1470  * @multi_instance:     State is set up for multiple instances which get
1471  *                      added afterwards.
1472  *
1473  * Returns:
1474  *   On success:
1475  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1476  *      0 for all other states
1477  *   On failure: proper (negative) error code
1478  */
1479 int __cpuhp_setup_state(enum cpuhp_state state,
1480                         const char *name, bool invoke,
1481                         int (*startup)(unsigned int cpu),
1482                         int (*teardown)(unsigned int cpu),
1483                         bool multi_instance)
1484 {
1485         int cpu, ret = 0;
1486         bool dynstate;
1487
1488         if (cpuhp_cb_check(state) || !name)
1489                 return -EINVAL;
1490
1491         get_online_cpus();
1492
1493         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1494                                     multi_instance);
1495
1496         dynstate = state == CPUHP_AP_ONLINE_DYN;
1497         if (ret > 0 && dynstate) {
1498                 state = ret;
1499                 ret = 0;
1500         }
1501
1502         if (ret || !invoke || !startup)
1503                 goto out;
1504
1505         /*
1506          * Try to call the startup callback for each present cpu
1507          * depending on the hotplug state of the cpu.
1508          */
1509         for_each_present_cpu(cpu) {
1510                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1511                 int cpustate = st->state;
1512
1513                 if (cpustate < state)
1514                         continue;
1515
1516                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1517                 if (ret) {
1518                         if (teardown)
1519                                 cpuhp_rollback_install(cpu, state, NULL);
1520                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1521                         goto out;
1522                 }
1523         }
1524 out:
1525         put_online_cpus();
1526         /*
1527          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1528          * dynamically allocated state in case of success.
1529          */
1530         if (!ret && dynstate)
1531                 return state;
1532         return ret;
1533 }
1534 EXPORT_SYMBOL(__cpuhp_setup_state);
1535
1536 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1537                                   struct hlist_node *node, bool invoke)
1538 {
1539         struct cpuhp_step *sp = cpuhp_get_step(state);
1540         int cpu;
1541
1542         BUG_ON(cpuhp_cb_check(state));
1543
1544         if (!sp->multi_instance)
1545                 return -EINVAL;
1546
1547         get_online_cpus();
1548         if (!invoke || !cpuhp_get_teardown_cb(state))
1549                 goto remove;
1550         /*
1551          * Call the teardown callback for each present cpu depending
1552          * on the hotplug state of the cpu. This function is not
1553          * allowed to fail currently!
1554          */
1555         for_each_present_cpu(cpu) {
1556                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1557                 int cpustate = st->state;
1558
1559                 if (cpustate >= state)
1560                         cpuhp_issue_call(cpu, state, false, node);
1561         }
1562
1563 remove:
1564         mutex_lock(&cpuhp_state_mutex);
1565         hlist_del(node);
1566         mutex_unlock(&cpuhp_state_mutex);
1567         put_online_cpus();
1568
1569         return 0;
1570 }
1571 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1572 /**
1573  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1574  * @state:      The state to remove
1575  * @invoke:     If true, the teardown function is invoked for cpus where
1576  *              cpu state >= @state
1577  *
1578  * The teardown callback is currently not allowed to fail. Think
1579  * about module removal!
1580  */
1581 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1582 {
1583         struct cpuhp_step *sp = cpuhp_get_step(state);
1584         int cpu;
1585
1586         BUG_ON(cpuhp_cb_check(state));
1587
1588         get_online_cpus();
1589
1590         if (sp->multi_instance) {
1591                 WARN(!hlist_empty(&sp->list),
1592                      "Error: Removing state %d which has instances left.\n",
1593                      state);
1594                 goto remove;
1595         }
1596
1597         if (!invoke || !cpuhp_get_teardown_cb(state))
1598                 goto remove;
1599
1600         /*
1601          * Call the teardown callback for each present cpu depending
1602          * on the hotplug state of the cpu. This function is not
1603          * allowed to fail currently!
1604          */
1605         for_each_present_cpu(cpu) {
1606                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1607                 int cpustate = st->state;
1608
1609                 if (cpustate >= state)
1610                         cpuhp_issue_call(cpu, state, false, NULL);
1611         }
1612 remove:
1613         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1614         put_online_cpus();
1615 }
1616 EXPORT_SYMBOL(__cpuhp_remove_state);
1617
1618 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1619 static ssize_t show_cpuhp_state(struct device *dev,
1620                                 struct device_attribute *attr, char *buf)
1621 {
1622         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1623
1624         return sprintf(buf, "%d\n", st->state);
1625 }
1626 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1627
1628 static ssize_t write_cpuhp_target(struct device *dev,
1629                                   struct device_attribute *attr,
1630                                   const char *buf, size_t count)
1631 {
1632         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1633         struct cpuhp_step *sp;
1634         int target, ret;
1635
1636         ret = kstrtoint(buf, 10, &target);
1637         if (ret)
1638                 return ret;
1639
1640 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1641         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1642                 return -EINVAL;
1643 #else
1644         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1645                 return -EINVAL;
1646 #endif
1647
1648         ret = lock_device_hotplug_sysfs();
1649         if (ret)
1650                 return ret;
1651
1652         mutex_lock(&cpuhp_state_mutex);
1653         sp = cpuhp_get_step(target);
1654         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1655         mutex_unlock(&cpuhp_state_mutex);
1656         if (ret)
1657                 return ret;
1658
1659         if (st->state < target)
1660                 ret = do_cpu_up(dev->id, target);
1661         else
1662                 ret = do_cpu_down(dev->id, target);
1663
1664         unlock_device_hotplug();
1665         return ret ? ret : count;
1666 }
1667
1668 static ssize_t show_cpuhp_target(struct device *dev,
1669                                  struct device_attribute *attr, char *buf)
1670 {
1671         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1672
1673         return sprintf(buf, "%d\n", st->target);
1674 }
1675 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1676
1677 static struct attribute *cpuhp_cpu_attrs[] = {
1678         &dev_attr_state.attr,
1679         &dev_attr_target.attr,
1680         NULL
1681 };
1682
1683 static struct attribute_group cpuhp_cpu_attr_group = {
1684         .attrs = cpuhp_cpu_attrs,
1685         .name = "hotplug",
1686         NULL
1687 };
1688
1689 static ssize_t show_cpuhp_states(struct device *dev,
1690                                  struct device_attribute *attr, char *buf)
1691 {
1692         ssize_t cur, res = 0;
1693         int i;
1694
1695         mutex_lock(&cpuhp_state_mutex);
1696         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1697                 struct cpuhp_step *sp = cpuhp_get_step(i);
1698
1699                 if (sp->name) {
1700                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1701                         buf += cur;
1702                         res += cur;
1703                 }
1704         }
1705         mutex_unlock(&cpuhp_state_mutex);
1706         return res;
1707 }
1708 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1709
1710 static struct attribute *cpuhp_cpu_root_attrs[] = {
1711         &dev_attr_states.attr,
1712         NULL
1713 };
1714
1715 static struct attribute_group cpuhp_cpu_root_attr_group = {
1716         .attrs = cpuhp_cpu_root_attrs,
1717         .name = "hotplug",
1718         NULL
1719 };
1720
1721 static int __init cpuhp_sysfs_init(void)
1722 {
1723         int cpu, ret;
1724
1725         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1726                                  &cpuhp_cpu_root_attr_group);
1727         if (ret)
1728                 return ret;
1729
1730         for_each_possible_cpu(cpu) {
1731                 struct device *dev = get_cpu_device(cpu);
1732
1733                 if (!dev)
1734                         continue;
1735                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1736                 if (ret)
1737                         return ret;
1738         }
1739         return 0;
1740 }
1741 device_initcall(cpuhp_sysfs_init);
1742 #endif
1743
1744 /*
1745  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1746  * represents all NR_CPUS bits binary values of 1<<nr.
1747  *
1748  * It is used by cpumask_of() to get a constant address to a CPU
1749  * mask value that has a single bit set only.
1750  */
1751
1752 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1753 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1754 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1755 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1756 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1757
1758 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1759
1760         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1761         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1762 #if BITS_PER_LONG > 32
1763         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1764         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1765 #endif
1766 };
1767 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1768
1769 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1770 EXPORT_SYMBOL(cpu_all_bits);
1771
1772 #ifdef CONFIG_INIT_ALL_POSSIBLE
1773 struct cpumask __cpu_possible_mask __read_mostly
1774         = {CPU_BITS_ALL};
1775 #else
1776 struct cpumask __cpu_possible_mask __read_mostly;
1777 #endif
1778 EXPORT_SYMBOL(__cpu_possible_mask);
1779
1780 struct cpumask __cpu_online_mask __read_mostly;
1781 EXPORT_SYMBOL(__cpu_online_mask);
1782
1783 struct cpumask __cpu_present_mask __read_mostly;
1784 EXPORT_SYMBOL(__cpu_present_mask);
1785
1786 struct cpumask __cpu_active_mask __read_mostly;
1787 EXPORT_SYMBOL(__cpu_active_mask);
1788
1789 void init_cpu_present(const struct cpumask *src)
1790 {
1791         cpumask_copy(&__cpu_present_mask, src);
1792 }
1793
1794 void init_cpu_possible(const struct cpumask *src)
1795 {
1796         cpumask_copy(&__cpu_possible_mask, src);
1797 }
1798
1799 void init_cpu_online(const struct cpumask *src)
1800 {
1801         cpumask_copy(&__cpu_online_mask, src);
1802 }
1803
1804 /*
1805  * Activate the first processor.
1806  */
1807 void __init boot_cpu_init(void)
1808 {
1809         int cpu = smp_processor_id();
1810
1811         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1812         set_cpu_online(cpu, true);
1813         set_cpu_active(cpu, true);
1814         set_cpu_present(cpu, true);
1815         set_cpu_possible(cpu, true);
1816 }
1817
1818 /*
1819  * Must be called _AFTER_ setting up the per_cpu areas
1820  */
1821 void __init boot_cpu_state_init(void)
1822 {
1823         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1824 }