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         rcutree_migrate_callbacks(cpu);
 654         return 0;
 655 }
 656
 657 static void cpuhp_complete_idle_dead(void *arg)
 658 {
 659         struct cpuhp_cpu_state *st = arg;
 660
 661         complete(&st->done);
 662 }
 663
 664 void cpuhp_report_idle_dead(void)
 665 {
 666         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 667
 668         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 669         rcu_report_dead(smp_processor_id());
 670         st->state = CPUHP_AP_IDLE_DEAD;
 671         /*
 672          * We cannot call complete after rcu_report_dead() so we delegate it
 673          * to an online cpu.
 674          */
 675         smp_call_function_single(cpumask_first(cpu_online_mask),
 676                                  cpuhp_complete_idle_dead, st, 0);
 677 }
 678
 679 #else
 680 #define takedown_cpu            NULL
 681 #endif
 682
 683 #ifdef CONFIG_HOTPLUG_CPU
 684
 685 /* Requires cpu_add_remove_lock to be held */
 686 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 687                            enum cpuhp_state target)
 688 {
 689         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 690         int prev_state, ret = 0;
 691
 692         if (num_online_cpus() == 1)
 693                 return -EBUSY;
 694
 695         if (!cpu_present(cpu))
 696                 return -EINVAL;
 697
 698         cpus_write_lock();
 699
 700         cpuhp_tasks_frozen = tasks_frozen;
 701
 702         prev_state = st->state;
 703         st->target = target;
 704         /*
 705          * If the current CPU state is in the range of the AP hotplug thread,
 706          * then we need to kick the thread.
 707          */
 708         if (st->state > CPUHP_TEARDOWN_CPU) {
 709                 ret = cpuhp_kick_ap_work(cpu);
 710                 /*
 711                  * The AP side has done the error rollback already. Just
 712                  * return the error code..
 713                  */
 714                 if (ret)
 715                         goto out;
 716
 717                 /*
 718                  * We might have stopped still in the range of the AP hotplug
 719                  * thread. Nothing to do anymore.
 720                  */
 721                 if (st->state > CPUHP_TEARDOWN_CPU)
 722                         goto out;
 723         }
 724         /*
 725          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 726          * to do the further cleanups.
 727          */
 728         ret = cpuhp_down_callbacks(cpu, st, target);
 729         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 730                 st->target = prev_state;
 731                 st->rollback = true;
 732                 cpuhp_kick_ap_work(cpu);
 733         }
 734
 735 out:
 736         cpus_write_unlock();
 737         return ret;
 738 }
 739
 740 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 741 {
 742         int err;
 743
 744         cpu_maps_update_begin();
 745
 746         if (cpu_hotplug_disabled) {
 747                 err = -EBUSY;
 748                 goto out;
 749         }
 750
 751         err = _cpu_down(cpu, 0, target);
 752
 753 out:
 754         cpu_maps_update_done();
 755         return err;
 756 }
 757 int cpu_down(unsigned int cpu)
 758 {
 759         return do_cpu_down(cpu, CPUHP_OFFLINE);
 760 }
 761 EXPORT_SYMBOL(cpu_down);
 762 #endif /*CONFIG_HOTPLUG_CPU*/
 763
 764 /**
 765  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
 766  * @cpu: cpu that just started
 767  *
 768  * It must be called by the arch code on the new cpu, before the new cpu
 769  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 770  */
 771 void notify_cpu_starting(unsigned int cpu)
 772 {
 773         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 774         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 775
 776         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 777         while (st->state < target) {
 778                 st->state++;
 779                 cpuhp_invoke_callback(cpu, st->state, true, NULL);
 780         }
 781 }
 782
 783 /*
 784  * Called from the idle task. Wake up the controlling task which brings the
 785  * stopper and the hotplug thread of the upcoming CPU up and then delegates
 786  * the rest of the online bringup to the hotplug thread.
 787  */
 788 void cpuhp_online_idle(enum cpuhp_state state)
 789 {
 790         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 791
 792         /* Happens for the boot cpu */
 793         if (state != CPUHP_AP_ONLINE_IDLE)
 794                 return;
 795
 796         st->state = CPUHP_AP_ONLINE_IDLE;
 797         complete(&st->done);
 798 }
 799
 800 /* Requires cpu_add_remove_lock to be held */
 801 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 802 {
 803         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 804         struct task_struct *idle;
 805         int ret = 0;
 806
 807         cpus_write_lock();
 808
 809         if (!cpu_present(cpu)) {
 810                 ret = -EINVAL;
 811                 goto out;
 812         }
 813
 814         /*
 815          * The caller of do_cpu_up might have raced with another
 816          * caller. Ignore it for now.
 817          */
 818         if (st->state >= target)
 819                 goto out;
 820
 821         if (st->state == CPUHP_OFFLINE) {
 822                 /* Let it fail before we try to bring the cpu up */
 823                 idle = idle_thread_get(cpu);
 824                 if (IS_ERR(idle)) {
 825                         ret = PTR_ERR(idle);
 826                         goto out;
 827                 }
 828         }
 829
 830         cpuhp_tasks_frozen = tasks_frozen;
 831
 832         st->target = target;
 833         /*
 834          * If the current CPU state is in the range of the AP hotplug thread,
 835          * then we need to kick the thread once more.
 836          */
 837         if (st->state > CPUHP_BRINGUP_CPU) {
 838                 ret = cpuhp_kick_ap_work(cpu);
 839                 /*
 840                  * The AP side has done the error rollback already. Just
 841                  * return the error code..
 842                  */
 843                 if (ret)
 844                         goto out;
 845         }
 846
 847         /*
 848          * Try to reach the target state. We max out on the BP at
 849          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
 850          * responsible for bringing it up to the target state.
 851          */
 852         target = min((int)target, CPUHP_BRINGUP_CPU);
 853         ret = cpuhp_up_callbacks(cpu, st, target);
 854 out:
 855         cpus_write_unlock();
 856         return ret;
 857 }
 858
 859 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 860 {
 861         int err = 0;
 862
 863         if (!cpu_possible(cpu)) {
 864                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
 865                        cpu);
 866 #if defined(CONFIG_IA64)
 867                 pr_err("please check additional_cpus= boot parameter\n");
 868 #endif
 869                 return -EINVAL;
 870         }
 871
 872         err = try_online_node(cpu_to_node(cpu));
 873         if (err)
 874                 return err;
 875
 876         cpu_maps_update_begin();
 877
 878         if (cpu_hotplug_disabled) {
 879                 err = -EBUSY;
 880                 goto out;
 881         }
 882
 883         err = _cpu_up(cpu, 0, target);
 884 out:
 885         cpu_maps_update_done();
 886         return err;
 887 }
 888
 889 int cpu_up(unsigned int cpu)
 890 {
 891         return do_cpu_up(cpu, CPUHP_ONLINE);
 892 }
 893 EXPORT_SYMBOL_GPL(cpu_up);
 894
 895 #ifdef CONFIG_PM_SLEEP_SMP
 896 static cpumask_var_t frozen_cpus;
 897
 898 int freeze_secondary_cpus(int primary)
 899 {
 900         int cpu, error = 0;
 901
 902         cpu_maps_update_begin();
 903         if (!cpu_online(primary))
 904                 primary = cpumask_first(cpu_online_mask);
 905         /*
 906          * We take down all of the non-boot CPUs in one shot to avoid races
 907          * with the userspace trying to use the CPU hotplug at the same time
 908          */
 909         cpumask_clear(frozen_cpus);
 910
 911         pr_info("Disabling non-boot CPUs ...\n");
 912         for_each_online_cpu(cpu) {
 913                 if (cpu == primary)
 914                         continue;
 915                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
 916                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
 917                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
 918                 if (!error)
 919                         cpumask_set_cpu(cpu, frozen_cpus);
 920                 else {
 921                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
 922                         break;
 923                 }
 924         }
 925
 926         if (!error)
 927                 BUG_ON(num_online_cpus() > 1);
 928         else
 929                 pr_err("Non-boot CPUs are not disabled\n");
 930
 931         /*
 932          * Make sure the CPUs won't be enabled by someone else. We need to do
 933          * this even in case of failure as all disable_nonboot_cpus() users are
 934          * supposed to do enable_nonboot_cpus() on the failure path.
 935          */
 936         cpu_hotplug_disabled++;
 937
 938         cpu_maps_update_done();
 939         return error;
 940 }
 941
 942 void __weak arch_enable_nonboot_cpus_begin(void)
 943 {
 944 }
 945
 946 void __weak arch_enable_nonboot_cpus_end(void)
 947 {
 948 }
 949
 950 void enable_nonboot_cpus(void)
 951 {
 952         int cpu, error;
 953
 954         /* Allow everyone to use the CPU hotplug again */
 955         cpu_maps_update_begin();
 956         __cpu_hotplug_enable();
 957         if (cpumask_empty(frozen_cpus))
 958                 goto out;
 959
 960         pr_info("Enabling non-boot CPUs ...\n");
 961
 962         arch_enable_nonboot_cpus_begin();
 963
 964         for_each_cpu(cpu, frozen_cpus) {
 965                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
 966                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
 967                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
 968                 if (!error) {
 969                         pr_info("CPU%d is up\n", cpu);
 970                         continue;
 971                 }
 972                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
 973         }
 974
 975         arch_enable_nonboot_cpus_end();
 976
 977         cpumask_clear(frozen_cpus);
 978 out:
 979         cpu_maps_update_done();
 980 }
 981
 982 static int __init alloc_frozen_cpus(void)
 983 {
 984         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 985                 return -ENOMEM;
 986         return 0;
 987 }
 988 core_initcall(alloc_frozen_cpus);
 989
 990 /*
 991  * When callbacks for CPU hotplug notifications are being executed, we must
 992  * ensure that the state of the system with respect to the tasks being frozen
 993  * or not, as reported by the notification, remains unchanged *throughout the
 994  * duration* of the execution of the callbacks.
 995  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
 996  *
 997  * This synchronization is implemented by mutually excluding regular CPU
 998  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
 999  * Hibernate notifications.
1000  */
1001 static int
1002 cpu_hotplug_pm_callback(struct notifier_block *nb,
1003                         unsigned long action, void *ptr)
1004 {
1005         switch (action) {
1006
1007         case PM_SUSPEND_PREPARE:
1008         case PM_HIBERNATION_PREPARE:
1009                 cpu_hotplug_disable();
1010                 break;
1011
1012         case PM_POST_SUSPEND:
1013         case PM_POST_HIBERNATION:
1014                 cpu_hotplug_enable();
1015                 break;
1016
1017         default:
1018                 return NOTIFY_DONE;
1019         }
1020
1021         return NOTIFY_OK;
1022 }
1023
1024
1025 static int __init cpu_hotplug_pm_sync_init(void)
1026 {
1027         /*
1028          * cpu_hotplug_pm_callback has higher priority than x86
1029          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1030          * to disable cpu hotplug to avoid cpu hotplug race.
1031          */
1032         pm_notifier(cpu_hotplug_pm_callback, 0);
1033         return 0;
1034 }
1035 core_initcall(cpu_hotplug_pm_sync_init);
1036
1037 #endif /* CONFIG_PM_SLEEP_SMP */
1038
1039 int __boot_cpu_id;
1040
1041 #endif /* CONFIG_SMP */
1042
1043 /* Boot processor state steps */
1044 static struct cpuhp_step cpuhp_bp_states[] = {
1045         [CPUHP_OFFLINE] = {
1046                 .name                   = "offline",
1047                 .startup.single         = NULL,
1048                 .teardown.single        = NULL,
1049         },
1050 #ifdef CONFIG_SMP
1051         [CPUHP_CREATE_THREADS]= {
1052                 .name                   = "threads:prepare",
1053                 .startup.single         = smpboot_create_threads,
1054                 .teardown.single        = NULL,
1055                 .cant_stop              = true,
1056         },
1057         [CPUHP_PERF_PREPARE] = {
1058                 .name                   = "perf:prepare",
1059                 .startup.single         = perf_event_init_cpu,
1060                 .teardown.single        = perf_event_exit_cpu,
1061         },
1062         [CPUHP_WORKQUEUE_PREP] = {
1063                 .name                   = "workqueue:prepare",
1064                 .startup.single         = workqueue_prepare_cpu,
1065                 .teardown.single        = NULL,
1066         },
1067         [CPUHP_HRTIMERS_PREPARE] = {
1068                 .name                   = "hrtimers:prepare",
1069                 .startup.single         = hrtimers_prepare_cpu,
1070                 .teardown.single        = hrtimers_dead_cpu,
1071         },
1072         [CPUHP_SMPCFD_PREPARE] = {
1073                 .name                   = "smpcfd:prepare",
1074                 .startup.single         = smpcfd_prepare_cpu,
1075                 .teardown.single        = smpcfd_dead_cpu,
1076         },
1077         [CPUHP_RELAY_PREPARE] = {
1078                 .name                   = "relay:prepare",
1079                 .startup.single         = relay_prepare_cpu,
1080                 .teardown.single        = NULL,
1081         },
1082         [CPUHP_SLAB_PREPARE] = {
1083                 .name                   = "slab:prepare",
1084                 .startup.single         = slab_prepare_cpu,
1085                 .teardown.single        = slab_dead_cpu,
1086         },
1087         [CPUHP_RCUTREE_PREP] = {
1088                 .name                   = "RCU/tree:prepare",
1089                 .startup.single         = rcutree_prepare_cpu,
1090                 .teardown.single        = rcutree_dead_cpu,
1091         },
1092         /*
1093          * On the tear-down path, timers_dead_cpu() must be invoked
1094          * before blk_mq_queue_reinit_notify() from notify_dead(),
1095          * otherwise a RCU stall occurs.
1096          */
1097         [CPUHP_TIMERS_DEAD] = {
1098                 .name                   = "timers:dead",
1099                 .startup.single         = NULL,
1100                 .teardown.single        = timers_dead_cpu,
1101         },
1102         /* Kicks the plugged cpu into life */
1103         [CPUHP_BRINGUP_CPU] = {
1104                 .name                   = "cpu:bringup",
1105                 .startup.single         = bringup_cpu,
1106                 .teardown.single        = NULL,
1107                 .cant_stop              = true,
1108         },
1109         [CPUHP_AP_SMPCFD_DYING] = {
1110                 .name                   = "smpcfd:dying",
1111                 .startup.single         = NULL,
1112                 .teardown.single        = smpcfd_dying_cpu,
1113         },
1114         /*
1115          * Handled on controll processor until the plugged processor manages
1116          * this itself.
1117          */
1118         [CPUHP_TEARDOWN_CPU] = {
1119                 .name                   = "cpu:teardown",
1120                 .startup.single         = NULL,
1121                 .teardown.single        = takedown_cpu,
1122                 .cant_stop              = true,
1123         },
1124 #else
1125         [CPUHP_BRINGUP_CPU] = { },
1126 #endif
1127 };
1128
1129 /* Application processor state steps */
1130 static struct cpuhp_step cpuhp_ap_states[] = {
1131 #ifdef CONFIG_SMP
1132         /* Final state before CPU kills itself */
1133         [CPUHP_AP_IDLE_DEAD] = {
1134                 .name                   = "idle:dead",
1135         },
1136         /*
1137          * Last state before CPU enters the idle loop to die. Transient state
1138          * for synchronization.
1139          */
1140         [CPUHP_AP_OFFLINE] = {
1141                 .name                   = "ap:offline",
1142                 .cant_stop              = true,
1143         },
1144         /* First state is scheduler control. Interrupts are disabled */
1145         [CPUHP_AP_SCHED_STARTING] = {
1146                 .name                   = "sched:starting",
1147                 .startup.single         = sched_cpu_starting,
1148                 .teardown.single        = sched_cpu_dying,
1149         },
1150         [CPUHP_AP_RCUTREE_DYING] = {
1151                 .name                   = "RCU/tree:dying",
1152                 .startup.single         = NULL,
1153                 .teardown.single        = rcutree_dying_cpu,
1154         },
1155         /* Entry state on starting. Interrupts enabled from here on. Transient
1156          * state for synchronsization */
1157         [CPUHP_AP_ONLINE] = {
1158                 .name                   = "ap:online",
1159         },
1160         /* Handle smpboot threads park/unpark */
1161         [CPUHP_AP_SMPBOOT_THREADS] = {
1162                 .name                   = "smpboot/threads:online",
1163                 .startup.single         = smpboot_unpark_threads,
1164                 .teardown.single        = NULL,
1165         },
1166         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1167                 .name                   = "irq/affinity:online",
1168                 .startup.single         = irq_affinity_online_cpu,
1169                 .teardown.single        = NULL,
1170         },
1171         [CPUHP_AP_PERF_ONLINE] = {
1172                 .name                   = "perf:online",
1173                 .startup.single         = perf_event_init_cpu,
1174                 .teardown.single        = perf_event_exit_cpu,
1175         },
1176         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1177                 .name                   = "workqueue:online",
1178                 .startup.single         = workqueue_online_cpu,
1179                 .teardown.single        = workqueue_offline_cpu,
1180         },
1181         [CPUHP_AP_RCUTREE_ONLINE] = {
1182                 .name                   = "RCU/tree:online",
1183                 .startup.single         = rcutree_online_cpu,
1184                 .teardown.single        = rcutree_offline_cpu,
1185         },
1186 #endif
1187         /*
1188          * The dynamically registered state space is here
1189          */
1190
1191 #ifdef CONFIG_SMP
1192         /* Last state is scheduler control setting the cpu active */
1193         [CPUHP_AP_ACTIVE] = {
1194                 .name                   = "sched:active",
1195                 .startup.single         = sched_cpu_activate,
1196                 .teardown.single        = sched_cpu_deactivate,
1197         },
1198 #endif
1199
1200         /* CPU is fully up and running. */
1201         [CPUHP_ONLINE] = {
1202                 .name                   = "online",
1203                 .startup.single         = NULL,
1204                 .teardown.single        = NULL,
1205         },
1206 };
1207
1208 /* Sanity check for callbacks */
1209 static int cpuhp_cb_check(enum cpuhp_state state)
1210 {
1211         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1212                 return -EINVAL;
1213         return 0;
1214 }
1215
1216 /*
1217  * Returns a free for dynamic slot assignment of the Online state. The states
1218  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1219  * by having no name assigned.
1220  */
1221 static int cpuhp_reserve_state(enum cpuhp_state state)
1222 {
1223         enum cpuhp_state i, end;
1224         struct cpuhp_step *step;
1225
1226         switch (state) {
1227         case CPUHP_AP_ONLINE_DYN:
1228                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1229                 end = CPUHP_AP_ONLINE_DYN_END;
1230                 break;
1231         case CPUHP_BP_PREPARE_DYN:
1232                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1233                 end = CPUHP_BP_PREPARE_DYN_END;
1234                 break;
1235         default:
1236                 return -EINVAL;
1237         }
1238
1239         for (i = state; i <= end; i++, step++) {
1240                 if (!step->name)
1241                         return i;
1242         }
1243         WARN(1, "No more dynamic states available for CPU hotplug\n");
1244         return -ENOSPC;
1245 }
1246
1247 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1248                                  int (*startup)(unsigned int cpu),
1249                                  int (*teardown)(unsigned int cpu),
1250                                  bool multi_instance)
1251 {
1252         /* (Un)Install the callbacks for further cpu hotplug operations */
1253         struct cpuhp_step *sp;
1254         int ret = 0;
1255
1256         /*
1257          * If name is NULL, then the state gets removed.
1258          *
1259          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1260          * the first allocation from these dynamic ranges, so the removal
1261          * would trigger a new allocation and clear the wrong (already
1262          * empty) state, leaving the callbacks of the to be cleared state
1263          * dangling, which causes wreckage on the next hotplug operation.
1264          */
1265         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1266                      state == CPUHP_BP_PREPARE_DYN)) {
1267                 ret = cpuhp_reserve_state(state);
1268                 if (ret < 0)
1269                         return ret;
1270                 state = ret;
1271         }
1272         sp = cpuhp_get_step(state);
1273         if (name && sp->name)
1274                 return -EBUSY;
1275
1276         sp->startup.single = startup;
1277         sp->teardown.single = teardown;
1278         sp->name = name;
1279         sp->multi_instance = multi_instance;
1280         INIT_HLIST_HEAD(&sp->list);
1281         return ret;
1282 }
1283
1284 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1285 {
1286         return cpuhp_get_step(state)->teardown.single;
1287 }
1288
1289 /*
1290  * Call the startup/teardown function for a step either on the AP or
1291  * on the current CPU.
1292  */
1293 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1294                             struct hlist_node *node)
1295 {
1296         struct cpuhp_step *sp = cpuhp_get_step(state);
1297         int ret;
1298
1299         if ((bringup && !sp->startup.single) ||
1300             (!bringup && !sp->teardown.single))
1301                 return 0;
1302         /*
1303          * The non AP bound callbacks can fail on bringup. On teardown
1304          * e.g. module removal we crash for now.
1305          */
1306 #ifdef CONFIG_SMP
1307         if (cpuhp_is_ap_state(state))
1308                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1309         else
1310                 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1311 #else
1312         ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1313 #endif
1314         BUG_ON(ret && !bringup);
1315         return ret;
1316 }
1317
1318 /*
1319  * Called from __cpuhp_setup_state on a recoverable failure.
1320  *
1321  * Note: The teardown callbacks for rollback are not allowed to fail!
1322  */
1323 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1324                                    struct hlist_node *node)
1325 {
1326         int cpu;
1327
1328         /* Roll back the already executed steps on the other cpus */
1329         for_each_present_cpu(cpu) {
1330                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1331                 int cpustate = st->state;
1332
1333                 if (cpu >= failedcpu)
1334                         break;
1335
1336                 /* Did we invoke the startup call on that cpu ? */
1337                 if (cpustate >= state)
1338                         cpuhp_issue_call(cpu, state, false, node);
1339         }
1340 }
1341
1342 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1343                                           struct hlist_node *node,
1344                                           bool invoke)
1345 {
1346         struct cpuhp_step *sp;
1347         int cpu;
1348         int ret;
1349
1350         lockdep_assert_cpus_held();
1351
1352         sp = cpuhp_get_step(state);
1353         if (sp->multi_instance == false)
1354                 return -EINVAL;
1355
1356         mutex_lock(&cpuhp_state_mutex);
1357
1358         if (!invoke || !sp->startup.multi)
1359                 goto add_node;
1360
1361         /*
1362          * Try to call the startup callback for each present cpu
1363          * depending on the hotplug state of the cpu.
1364          */
1365         for_each_present_cpu(cpu) {
1366                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1367                 int cpustate = st->state;
1368
1369                 if (cpustate < state)
1370                         continue;
1371
1372                 ret = cpuhp_issue_call(cpu, state, true, node);
1373                 if (ret) {
1374                         if (sp->teardown.multi)
1375                                 cpuhp_rollback_install(cpu, state, node);
1376                         goto unlock;
1377                 }
1378         }
1379 add_node:
1380         ret = 0;
1381         hlist_add_head(node, &sp->list);
1382 unlock:
1383         mutex_unlock(&cpuhp_state_mutex);
1384         return ret;
1385 }
1386
1387 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1388                                bool invoke)
1389 {
1390         int ret;
1391
1392         cpus_read_lock();
1393         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1394         cpus_read_unlock();
1395         return ret;
1396 }
1397 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1398
1399 /**
1400  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1401  * @state:              The state to setup
1402  * @invoke:             If true, the startup function is invoked for cpus where
1403  *                      cpu state >= @state
1404  * @startup:            startup callback function
1405  * @teardown:           teardown callback function
1406  * @multi_instance:     State is set up for multiple instances which get
1407  *                      added afterwards.
1408  *
1409  * The caller needs to hold cpus read locked while calling this function.
1410  * Returns:
1411  *   On success:
1412  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1413  *      0 for all other states
1414  *   On failure: proper (negative) error code
1415  */
1416 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1417                                    const char *name, bool invoke,
1418                                    int (*startup)(unsigned int cpu),
1419                                    int (*teardown)(unsigned int cpu),
1420                                    bool multi_instance)
1421 {
1422         int cpu, ret = 0;
1423         bool dynstate;
1424
1425         lockdep_assert_cpus_held();
1426
1427         if (cpuhp_cb_check(state) || !name)
1428                 return -EINVAL;
1429
1430         mutex_lock(&cpuhp_state_mutex);
1431
1432         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1433                                     multi_instance);
1434
1435         dynstate = state == CPUHP_AP_ONLINE_DYN;
1436         if (ret > 0 && dynstate) {
1437                 state = ret;
1438                 ret = 0;
1439         }
1440
1441         if (ret || !invoke || !startup)
1442                 goto out;
1443
1444         /*
1445          * Try to call the startup callback for each present cpu
1446          * depending on the hotplug state of the cpu.
1447          */
1448         for_each_present_cpu(cpu) {
1449                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1450                 int cpustate = st->state;
1451
1452                 if (cpustate < state)
1453                         continue;
1454
1455                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1456                 if (ret) {
1457                         if (teardown)
1458                                 cpuhp_rollback_install(cpu, state, NULL);
1459                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1460                         goto out;
1461                 }
1462         }
1463 out:
1464         mutex_unlock(&cpuhp_state_mutex);
1465         /*
1466          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1467          * dynamically allocated state in case of success.
1468          */
1469         if (!ret && dynstate)
1470                 return state;
1471         return ret;
1472 }
1473 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1474
1475 int __cpuhp_setup_state(enum cpuhp_state state,
1476                         const char *name, bool invoke,
1477                         int (*startup)(unsigned int cpu),
1478                         int (*teardown)(unsigned int cpu),
1479                         bool multi_instance)
1480 {
1481         int ret;
1482
1483         cpus_read_lock();
1484         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1485                                              teardown, multi_instance);
1486         cpus_read_unlock();
1487         return ret;
1488 }
1489 EXPORT_SYMBOL(__cpuhp_setup_state);
1490
1491 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1492                                   struct hlist_node *node, bool invoke)
1493 {
1494         struct cpuhp_step *sp = cpuhp_get_step(state);
1495         int cpu;
1496
1497         BUG_ON(cpuhp_cb_check(state));
1498
1499         if (!sp->multi_instance)
1500                 return -EINVAL;
1501
1502         cpus_read_lock();
1503         mutex_lock(&cpuhp_state_mutex);
1504
1505         if (!invoke || !cpuhp_get_teardown_cb(state))
1506                 goto remove;
1507         /*
1508          * Call the teardown callback for each present cpu depending
1509          * on the hotplug state of the cpu. This function is not
1510          * allowed to fail currently!
1511          */
1512         for_each_present_cpu(cpu) {
1513                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1514                 int cpustate = st->state;
1515
1516                 if (cpustate >= state)
1517                         cpuhp_issue_call(cpu, state, false, node);
1518         }
1519
1520 remove:
1521         hlist_del(node);
1522         mutex_unlock(&cpuhp_state_mutex);
1523         cpus_read_unlock();
1524
1525         return 0;
1526 }
1527 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1528
1529 /**
1530  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1531  * @state:      The state to remove
1532  * @invoke:     If true, the teardown function is invoked for cpus where
1533  *              cpu state >= @state
1534  *
1535  * The caller needs to hold cpus read locked while calling this function.
1536  * The teardown callback is currently not allowed to fail. Think
1537  * about module removal!
1538  */
1539 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1540 {
1541         struct cpuhp_step *sp = cpuhp_get_step(state);
1542         int cpu;
1543
1544         BUG_ON(cpuhp_cb_check(state));
1545
1546         lockdep_assert_cpus_held();
1547
1548         mutex_lock(&cpuhp_state_mutex);
1549         if (sp->multi_instance) {
1550                 WARN(!hlist_empty(&sp->list),
1551                      "Error: Removing state %d which has instances left.\n",
1552                      state);
1553                 goto remove;
1554         }
1555
1556         if (!invoke || !cpuhp_get_teardown_cb(state))
1557                 goto remove;
1558
1559         /*
1560          * Call the teardown callback for each present cpu depending
1561          * on the hotplug state of the cpu. This function is not
1562          * allowed to fail currently!
1563          */
1564         for_each_present_cpu(cpu) {
1565                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1566                 int cpustate = st->state;
1567
1568                 if (cpustate >= state)
1569                         cpuhp_issue_call(cpu, state, false, NULL);
1570         }
1571 remove:
1572         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1573         mutex_unlock(&cpuhp_state_mutex);
1574 }
1575 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1576
1577 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1578 {
1579         cpus_read_lock();
1580         __cpuhp_remove_state_cpuslocked(state, invoke);
1581         cpus_read_unlock();
1582 }
1583 EXPORT_SYMBOL(__cpuhp_remove_state);
1584
1585 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1586 static ssize_t show_cpuhp_state(struct device *dev,
1587                                 struct device_attribute *attr, char *buf)
1588 {
1589         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1590
1591         return sprintf(buf, "%d\n", st->state);
1592 }
1593 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1594
1595 static ssize_t write_cpuhp_target(struct device *dev,
1596                                   struct device_attribute *attr,
1597                                   const char *buf, size_t count)
1598 {
1599         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1600         struct cpuhp_step *sp;
1601         int target, ret;
1602
1603         ret = kstrtoint(buf, 10, &target);
1604         if (ret)
1605                 return ret;
1606
1607 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1608         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1609                 return -EINVAL;
1610 #else
1611         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1612                 return -EINVAL;
1613 #endif
1614
1615         ret = lock_device_hotplug_sysfs();
1616         if (ret)
1617                 return ret;
1618
1619         mutex_lock(&cpuhp_state_mutex);
1620         sp = cpuhp_get_step(target);
1621         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1622         mutex_unlock(&cpuhp_state_mutex);
1623         if (ret)
1624                 goto out;
1625
1626         if (st->state < target)
1627                 ret = do_cpu_up(dev->id, target);
1628         else
1629                 ret = do_cpu_down(dev->id, target);
1630 out:
1631         unlock_device_hotplug();
1632         return ret ? ret : count;
1633 }
1634
1635 static ssize_t show_cpuhp_target(struct device *dev,
1636                                  struct device_attribute *attr, char *buf)
1637 {
1638         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1639
1640         return sprintf(buf, "%d\n", st->target);
1641 }
1642 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1643
1644 static struct attribute *cpuhp_cpu_attrs[] = {
1645         &dev_attr_state.attr,
1646         &dev_attr_target.attr,
1647         NULL
1648 };
1649
1650 static const struct attribute_group cpuhp_cpu_attr_group = {
1651         .attrs = cpuhp_cpu_attrs,
1652         .name = "hotplug",
1653         NULL
1654 };
1655
1656 static ssize_t show_cpuhp_states(struct device *dev,
1657                                  struct device_attribute *attr, char *buf)
1658 {
1659         ssize_t cur, res = 0;
1660         int i;
1661
1662         mutex_lock(&cpuhp_state_mutex);
1663         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1664                 struct cpuhp_step *sp = cpuhp_get_step(i);
1665
1666                 if (sp->name) {
1667                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1668                         buf += cur;
1669                         res += cur;
1670                 }
1671         }
1672         mutex_unlock(&cpuhp_state_mutex);
1673         return res;
1674 }
1675 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1676
1677 static struct attribute *cpuhp_cpu_root_attrs[] = {
1678         &dev_attr_states.attr,
1679         NULL
1680 };
1681
1682 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1683         .attrs = cpuhp_cpu_root_attrs,
1684         .name = "hotplug",
1685         NULL
1686 };
1687
1688 static int __init cpuhp_sysfs_init(void)
1689 {
1690         int cpu, ret;
1691
1692         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1693                                  &cpuhp_cpu_root_attr_group);
1694         if (ret)
1695                 return ret;
1696
1697         for_each_possible_cpu(cpu) {
1698                 struct device *dev = get_cpu_device(cpu);
1699
1700                 if (!dev)
1701                         continue;
1702                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1703                 if (ret)
1704                         return ret;
1705         }
1706         return 0;
1707 }
1708 device_initcall(cpuhp_sysfs_init);
1709 #endif
1710
1711 /*
1712  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1713  * represents all NR_CPUS bits binary values of 1<<nr.
1714  *
1715  * It is used by cpumask_of() to get a constant address to a CPU
1716  * mask value that has a single bit set only.
1717  */
1718
1719 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1720 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1721 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1722 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1723 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1724
1725 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1726
1727         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1728         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1729 #if BITS_PER_LONG > 32
1730         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1731         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1732 #endif
1733 };
1734 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1735
1736 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1737 EXPORT_SYMBOL(cpu_all_bits);
1738
1739 #ifdef CONFIG_INIT_ALL_POSSIBLE
1740 struct cpumask __cpu_possible_mask __read_mostly
1741         = {CPU_BITS_ALL};
1742 #else
1743 struct cpumask __cpu_possible_mask __read_mostly;
1744 #endif
1745 EXPORT_SYMBOL(__cpu_possible_mask);
1746
1747 struct cpumask __cpu_online_mask __read_mostly;
1748 EXPORT_SYMBOL(__cpu_online_mask);
1749
1750 struct cpumask __cpu_present_mask __read_mostly;
1751 EXPORT_SYMBOL(__cpu_present_mask);
1752
1753 struct cpumask __cpu_active_mask __read_mostly;
1754 EXPORT_SYMBOL(__cpu_active_mask);
1755
1756 void init_cpu_present(const struct cpumask *src)
1757 {
1758         cpumask_copy(&__cpu_present_mask, src);
1759 }
1760
1761 void init_cpu_possible(const struct cpumask *src)
1762 {
1763         cpumask_copy(&__cpu_possible_mask, src);
1764 }
1765
1766 void init_cpu_online(const struct cpumask *src)
1767 {
1768         cpumask_copy(&__cpu_online_mask, src);
1769 }
1770
1771 /*
1772  * Activate the first processor.
1773  */
1774 void __init boot_cpu_init(void)
1775 {
1776         int cpu = smp_processor_id();
1777
1778         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1779         set_cpu_online(cpu, true);
1780         set_cpu_active(cpu, true);
1781         set_cpu_present(cpu, true);
1782         set_cpu_possible(cpu, true);
1783
1784 #ifdef CONFIG_SMP
1785         __boot_cpu_id = cpu;
1786 #endif
1787 }
1788
1789 /*
1790  * Must be called _AFTER_ setting up the per_cpu areas
1791  */
1792 void __init boot_cpu_state_init(void)
1793 {
1794         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1795 }