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