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