kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/proc_fs.h>
   7 #include <linux/smp.h>
   8 #include <linux/init.h>
   9 #include <linux/notifier.h>
  10 #include <linux/sched/signal.h>
  11 #include <linux/sched/hotplug.h>
  12 #include <linux/sched/isolation.h>
  13 #include <linux/sched/task.h>
  14 #include <linux/sched/smt.h>
  15 #include <linux/unistd.h>
  16 #include <linux/cpu.h>
  17 #include <linux/oom.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/export.h>
  20 #include <linux/bug.h>
  21 #include <linux/kthread.h>
  22 #include <linux/stop_machine.h>
  23 #include <linux/mutex.h>
  24 #include <linux/gfp.h>
  25 #include <linux/suspend.h>
  26 #include <linux/lockdep.h>
  27 #include <linux/tick.h>
  28 #include <linux/irq.h>
  29 #include <linux/nmi.h>
  30 #include <linux/smpboot.h>
  31 #include <linux/relay.h>
  32 #include <linux/slab.h>
  33 #include <linux/percpu-rwsem.h>
  34
  35 #include <trace/events/power.h>
  36 #define CREATE_TRACE_POINTS
  37 #include <trace/events/cpuhp.h>
  38
  39 #include "smpboot.h"
  40
  41 /**
  42  * cpuhp_cpu_state - Per cpu hotplug state storage
  43  * @state:      The current cpu state
  44  * @target:     The target state
  45  * @thread:     Pointer to the hotplug thread
  46  * @should_run: Thread should execute
  47  * @rollback:   Perform a rollback
  48  * @single:     Single callback invocation
  49  * @bringup:    Single callback bringup or teardown selector
  50  * @cb_state:   The state for a single callback (install/uninstall)
  51  * @result:     Result of the operation
  52  * @done_up:    Signal completion to the issuer of the task for cpu-up
  53  * @done_down:  Signal completion to the issuer of the task for cpu-down
  54  */
  55 struct cpuhp_cpu_state {
  56         enum cpuhp_state        state;
  57         enum cpuhp_state        target;
  58         enum cpuhp_state        fail;
  59 #ifdef CONFIG_SMP
  60         struct task_struct      *thread;
  61         bool                    should_run;
  62         bool                    rollback;
  63         bool                    single;
  64         bool                    bringup;
  65         struct hlist_node       *node;
  66         struct hlist_node       *last;
  67         enum cpuhp_state        cb_state;
  68         int                     result;
  69         struct completion       done_up;
  70         struct completion       done_down;
  71 #endif
  72 };
  73
  74 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  75         .fail = CPUHP_INVALID,
  76 };
  77
  78 #ifdef CONFIG_SMP
  79 cpumask_t cpus_booted_once_mask;
  80 #endif
  81
  82 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  83 static struct lockdep_map cpuhp_state_up_map =
  84         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  85 static struct lockdep_map cpuhp_state_down_map =
  86         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  87
  88
  89 static inline void cpuhp_lock_acquire(bool bringup)
  90 {
  91         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  92 }
  93
  94 static inline void cpuhp_lock_release(bool bringup)
  95 {
  96         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  97 }
  98 #else
  99
 100 static inline void cpuhp_lock_acquire(bool bringup) { }
 101 static inline void cpuhp_lock_release(bool bringup) { }
 102
 103 #endif
 104
 105 /**
 106  * cpuhp_step - Hotplug state machine step
 107  * @name:       Name of the step
 108  * @startup:    Startup function of the step
 109  * @teardown:   Teardown function of the step
 110  * @cant_stop:  Bringup/teardown can't be stopped at this step
 111  */
 112 struct cpuhp_step {
 113         const char              *name;
 114         union {
 115                 int             (*single)(unsigned int cpu);
 116                 int             (*multi)(unsigned int cpu,
 117                                          struct hlist_node *node);
 118         } startup;
 119         union {
 120                 int             (*single)(unsigned int cpu);
 121                 int             (*multi)(unsigned int cpu,
 122                                          struct hlist_node *node);
 123         } teardown;
 124         struct hlist_head       list;
 125         bool                    cant_stop;
 126         bool                    multi_instance;
 127 };
 128
 129 static DEFINE_MUTEX(cpuhp_state_mutex);
 130 static struct cpuhp_step cpuhp_hp_states[];
 131
 132 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 133 {
 134         return cpuhp_hp_states + state;
 135 }
 136
 137 /**
 138  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 139  * @cpu:        The cpu for which the callback should be invoked
 140  * @state:      The state to do callbacks for
 141  * @bringup:    True if the bringup callback should be invoked
 142  * @node:       For multi-instance, do a single entry callback for install/remove
 143  * @lastp:      For multi-instance rollback, remember how far we got
 144  *
 145  * Called from cpu hotplug and from the state register machinery.
 146  */
 147 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 148                                  bool bringup, struct hlist_node *node,
 149                                  struct hlist_node **lastp)
 150 {
 151         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 152         struct cpuhp_step *step = cpuhp_get_step(state);
 153         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 154         int (*cb)(unsigned int cpu);
 155         int ret, cnt;
 156
 157         if (st->fail == state) {
 158                 st->fail = CPUHP_INVALID;
 159
 160                 if (!(bringup ? step->startup.single : step->teardown.single))
 161                         return 0;
 162
 163                 return -EAGAIN;
 164         }
 165
 166         if (!step->multi_instance) {
 167                 WARN_ON_ONCE(lastp && *lastp);
 168                 cb = bringup ? step->startup.single : step->teardown.single;
 169                 if (!cb)
 170                         return 0;
 171                 trace_cpuhp_enter(cpu, st->target, state, cb);
 172                 ret = cb(cpu);
 173                 trace_cpuhp_exit(cpu, st->state, state, ret);
 174                 return ret;
 175         }
 176         cbm = bringup ? step->startup.multi : step->teardown.multi;
 177         if (!cbm)
 178                 return 0;
 179
 180         /* Single invocation for instance add/remove */
 181         if (node) {
 182                 WARN_ON_ONCE(lastp && *lastp);
 183                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 184                 ret = cbm(cpu, node);
 185                 trace_cpuhp_exit(cpu, st->state, state, ret);
 186                 return ret;
 187         }
 188
 189         /* State transition. Invoke on all instances */
 190         cnt = 0;
 191         hlist_for_each(node, &step->list) {
 192                 if (lastp && node == *lastp)
 193                         break;
 194
 195                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 196                 ret = cbm(cpu, node);
 197                 trace_cpuhp_exit(cpu, st->state, state, ret);
 198                 if (ret) {
 199                         if (!lastp)
 200                                 goto err;
 201
 202                         *lastp = node;
 203                         return ret;
 204                 }
 205                 cnt++;
 206         }
 207         if (lastp)
 208                 *lastp = NULL;
 209         return 0;
 210 err:
 211         /* Rollback the instances if one failed */
 212         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 213         if (!cbm)
 214                 return ret;
 215
 216         hlist_for_each(node, &step->list) {
 217                 if (!cnt--)
 218                         break;
 219
 220                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 221                 ret = cbm(cpu, node);
 222                 trace_cpuhp_exit(cpu, st->state, state, ret);
 223                 /*
 224                  * Rollback must not fail,
 225                  */
 226                 WARN_ON_ONCE(ret);
 227         }
 228         return ret;
 229 }
 230
 231 #ifdef CONFIG_SMP
 232 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 233 {
 234         /*
 235          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 236          * purposes as that state is handled explicitly in cpu_down.
 237          */
 238         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 239 }
 240
 241 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 242 {
 243         struct completion *done = bringup ? &st->done_up : &st->done_down;
 244         wait_for_completion(done);
 245 }
 246
 247 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 248 {
 249         struct completion *done = bringup ? &st->done_up : &st->done_down;
 250         complete(done);
 251 }
 252
 253 /*
 254  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 255  */
 256 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 257 {
 258         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 259 }
 260
 261 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 262 static DEFINE_MUTEX(cpu_add_remove_lock);
 263 bool cpuhp_tasks_frozen;
 264 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 265
 266 /*
 267  * The following two APIs (cpu_maps_update_begin/done) must be used when
 268  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 269  */
 270 void cpu_maps_update_begin(void)
 271 {
 272         mutex_lock(&cpu_add_remove_lock);
 273 }
 274
 275 void cpu_maps_update_done(void)
 276 {
 277         mutex_unlock(&cpu_add_remove_lock);
 278 }
 279
 280 /*
 281  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 282  * Should always be manipulated under cpu_add_remove_lock
 283  */
 284 static int cpu_hotplug_disabled;
 285
 286 #ifdef CONFIG_HOTPLUG_CPU
 287
 288 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 289
 290 void cpus_read_lock(void)
 291 {
 292         percpu_down_read(&cpu_hotplug_lock);
 293 }
 294 EXPORT_SYMBOL_GPL(cpus_read_lock);
 295
 296 int cpus_read_trylock(void)
 297 {
 298         return percpu_down_read_trylock(&cpu_hotplug_lock);
 299 }
 300 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 301
 302 void cpus_read_unlock(void)
 303 {
 304         percpu_up_read(&cpu_hotplug_lock);
 305 }
 306 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 307
 308 void cpus_write_lock(void)
 309 {
 310         percpu_down_write(&cpu_hotplug_lock);
 311 }
 312
 313 void cpus_write_unlock(void)
 314 {
 315         percpu_up_write(&cpu_hotplug_lock);
 316 }
 317
 318 void lockdep_assert_cpus_held(void)
 319 {
 320         /*
 321          * We can't have hotplug operations before userspace starts running,
 322          * and some init codepaths will knowingly not take the hotplug lock.
 323          * This is all valid, so mute lockdep until it makes sense to report
 324          * unheld locks.
 325          */
 326         if (system_state < SYSTEM_RUNNING)
 327                 return;
 328
 329         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 330 }
 331
 332 static void lockdep_acquire_cpus_lock(void)
 333 {
 334         rwsem_acquire(&cpu_hotplug_lock.rw_sem.dep_map, 0, 0, _THIS_IP_);
 335 }
 336
 337 static void lockdep_release_cpus_lock(void)
 338 {
 339         rwsem_release(&cpu_hotplug_lock.rw_sem.dep_map, _THIS_IP_);
 340 }
 341
 342 /*
 343  * Wait for currently running CPU hotplug operations to complete (if any) and
 344  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 345  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 346  * hotplug path before performing hotplug operations. So acquiring that lock
 347  * guarantees mutual exclusion from any currently running hotplug operations.
 348  */
 349 void cpu_hotplug_disable(void)
 350 {
 351         cpu_maps_update_begin();
 352         cpu_hotplug_disabled++;
 353         cpu_maps_update_done();
 354 }
 355 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 356
 357 static void __cpu_hotplug_enable(void)
 358 {
 359         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 360                 return;
 361         cpu_hotplug_disabled--;
 362 }
 363
 364 void cpu_hotplug_enable(void)
 365 {
 366         cpu_maps_update_begin();
 367         __cpu_hotplug_enable();
 368         cpu_maps_update_done();
 369 }
 370 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 371
 372 #else
 373
 374 static void lockdep_acquire_cpus_lock(void)
 375 {
 376 }
 377
 378 static void lockdep_release_cpus_lock(void)
 379 {
 380 }
 381
 382 #endif  /* CONFIG_HOTPLUG_CPU */
 383
 384 /*
 385  * Architectures that need SMT-specific errata handling during SMT hotplug
 386  * should override this.
 387  */
 388 void __weak arch_smt_update(void) { }
 389
 390 #ifdef CONFIG_HOTPLUG_SMT
 391 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 392
 393 void __init cpu_smt_disable(bool force)
 394 {
 395         if (!cpu_smt_possible())
 396                 return;
 397
 398         if (force) {
 399                 pr_info("SMT: Force disabled\n");
 400                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 401         } else {
 402                 pr_info("SMT: disabled\n");
 403                 cpu_smt_control = CPU_SMT_DISABLED;
 404         }
 405 }
 406
 407 /*
 408  * The decision whether SMT is supported can only be done after the full
 409  * CPU identification. Called from architecture code.
 410  */
 411 void __init cpu_smt_check_topology(void)
 412 {
 413         if (!topology_smt_supported())
 414                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 415 }
 416
 417 static int __init smt_cmdline_disable(char *str)
 418 {
 419         cpu_smt_disable(str && !strcmp(str, "force"));
 420         return 0;
 421 }
 422 early_param("nosmt", smt_cmdline_disable);
 423
 424 static inline bool cpu_smt_allowed(unsigned int cpu)
 425 {
 426         if (cpu_smt_control == CPU_SMT_ENABLED)
 427                 return true;
 428
 429         if (topology_is_primary_thread(cpu))
 430                 return true;
 431
 432         /*
 433          * On x86 it's required to boot all logical CPUs at least once so
 434          * that the init code can get a chance to set CR4.MCE on each
 435          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 436          * core will shutdown the machine.
 437          */
 438         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
 439 }
 440
 441 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
 442 bool cpu_smt_possible(void)
 443 {
 444         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
 445                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
 446 }
 447 EXPORT_SYMBOL_GPL(cpu_smt_possible);
 448 #else
 449 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 450 #endif
 451
 452 static inline enum cpuhp_state
 453 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 454 {
 455         enum cpuhp_state prev_state = st->state;
 456
 457         st->rollback = false;
 458         st->last = NULL;
 459
 460         st->target = target;
 461         st->single = false;
 462         st->bringup = st->state < target;
 463
 464         return prev_state;
 465 }
 466
 467 static inline void
 468 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 469 {
 470         st->rollback = true;
 471
 472         /*
 473          * If we have st->last we need to undo partial multi_instance of this
 474          * state first. Otherwise start undo at the previous state.
 475          */
 476         if (!st->last) {
 477                 if (st->bringup)
 478                         st->state--;
 479                 else
 480                         st->state++;
 481         }
 482
 483         st->target = prev_state;
 484         st->bringup = !st->bringup;
 485 }
 486
 487 /* Regular hotplug invocation of the AP hotplug thread */
 488 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 489 {
 490         if (!st->single && st->state == st->target)
 491                 return;
 492
 493         st->result = 0;
 494         /*
 495          * Make sure the above stores are visible before should_run becomes
 496          * true. Paired with the mb() above in cpuhp_thread_fun()
 497          */
 498         smp_mb();
 499         st->should_run = true;
 500         wake_up_process(st->thread);
 501         wait_for_ap_thread(st, st->bringup);
 502 }
 503
 504 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 505 {
 506         enum cpuhp_state prev_state;
 507         int ret;
 508
 509         prev_state = cpuhp_set_state(st, target);
 510         __cpuhp_kick_ap(st);
 511         if ((ret = st->result)) {
 512                 cpuhp_reset_state(st, prev_state);
 513                 __cpuhp_kick_ap(st);
 514         }
 515
 516         return ret;
 517 }
 518
 519 static int bringup_wait_for_ap(unsigned int cpu)
 520 {
 521         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 522
 523         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 524         wait_for_ap_thread(st, true);
 525         if (WARN_ON_ONCE((!cpu_online(cpu))))
 526                 return -ECANCELED;
 527
 528         /* Unpark the hotplug thread of the target cpu */
 529         kthread_unpark(st->thread);
 530
 531         /*
 532          * SMT soft disabling on X86 requires to bring the CPU out of the
 533          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 534          * CPU marked itself as booted_once in notify_cpu_starting() so the
 535          * cpu_smt_allowed() check will now return false if this is not the
 536          * primary sibling.
 537          */
 538         if (!cpu_smt_allowed(cpu))
 539                 return -ECANCELED;
 540
 541         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 542                 return 0;
 543
 544         return cpuhp_kick_ap(st, st->target);
 545 }
 546
 547 static int bringup_cpu(unsigned int cpu)
 548 {
 549         struct task_struct *idle = idle_thread_get(cpu);
 550         int ret;
 551
 552         /*
 553          * Some architectures have to walk the irq descriptors to
 554          * setup the vector space for the cpu which comes online.
 555          * Prevent irq alloc/free across the bringup.
 556          */
 557         irq_lock_sparse();
 558
 559         /* Arch-specific enabling code. */
 560         ret = __cpu_up(cpu, idle);
 561         irq_unlock_sparse();
 562         if (ret)
 563                 return ret;
 564         return bringup_wait_for_ap(cpu);
 565 }
 566
 567 /*
 568  * Hotplug state machine related functions
 569  */
 570
 571 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 572 {
 573         for (st->state--; st->state > st->target; st->state--)
 574                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 575 }
 576
 577 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 578 {
 579         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 580                 return true;
 581         /*
 582          * When CPU hotplug is disabled, then taking the CPU down is not
 583          * possible because takedown_cpu() and the architecture and
 584          * subsystem specific mechanisms are not available. So the CPU
 585          * which would be completely unplugged again needs to stay around
 586          * in the current state.
 587          */
 588         return st->state <= CPUHP_BRINGUP_CPU;
 589 }
 590
 591 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 592                               enum cpuhp_state target)
 593 {
 594         enum cpuhp_state prev_state = st->state;
 595         int ret = 0;
 596
 597         while (st->state < target) {
 598                 st->state++;
 599                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 600                 if (ret) {
 601                         if (can_rollback_cpu(st)) {
 602                                 st->target = prev_state;
 603                                 undo_cpu_up(cpu, st);
 604                         }
 605                         break;
 606                 }
 607         }
 608         return ret;
 609 }
 610
 611 /*
 612  * The cpu hotplug threads manage the bringup and teardown of the cpus
 613  */
 614 static void cpuhp_create(unsigned int cpu)
 615 {
 616         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 617
 618         init_completion(&st->done_up);
 619         init_completion(&st->done_down);
 620 }
 621
 622 static int cpuhp_should_run(unsigned int cpu)
 623 {
 624         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 625
 626         return st->should_run;
 627 }
 628
 629 /*
 630  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 631  * callbacks when a state gets [un]installed at runtime.
 632  *
 633  * Each invocation of this function by the smpboot thread does a single AP
 634  * state callback.
 635  *
 636  * It has 3 modes of operation:
 637  *  - single: runs st->cb_state
 638  *  - up:     runs ++st->state, while st->state < st->target
 639  *  - down:   runs st->state--, while st->state > st->target
 640  *
 641  * When complete or on error, should_run is cleared and the completion is fired.
 642  */
 643 static void cpuhp_thread_fun(unsigned int cpu)
 644 {
 645         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 646         bool bringup = st->bringup;
 647         enum cpuhp_state state;
 648
 649         if (WARN_ON_ONCE(!st->should_run))
 650                 return;
 651
 652         /*
 653          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 654          * that if we see ->should_run we also see the rest of the state.
 655          */
 656         smp_mb();
 657
 658         /*
 659          * The BP holds the hotplug lock, but we're now running on the AP,
 660          * ensure that anybody asserting the lock is held, will actually find
 661          * it so.
 662          */
 663         lockdep_acquire_cpus_lock();
 664         cpuhp_lock_acquire(bringup);
 665
 666         if (st->single) {
 667                 state = st->cb_state;
 668                 st->should_run = false;
 669         } else {
 670                 if (bringup) {
 671                         st->state++;
 672                         state = st->state;
 673                         st->should_run = (st->state < st->target);
 674                         WARN_ON_ONCE(st->state > st->target);
 675                 } else {
 676                         state = st->state;
 677                         st->state--;
 678                         st->should_run = (st->state > st->target);
 679                         WARN_ON_ONCE(st->state < st->target);
 680                 }
 681         }
 682
 683         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 684
 685         if (cpuhp_is_atomic_state(state)) {
 686                 local_irq_disable();
 687                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 688                 local_irq_enable();
 689
 690                 /*
 691                  * STARTING/DYING must not fail!
 692                  */
 693                 WARN_ON_ONCE(st->result);
 694         } else {
 695                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 696         }
 697
 698         if (st->result) {
 699                 /*
 700                  * If we fail on a rollback, we're up a creek without no
 701                  * paddle, no way forward, no way back. We loose, thanks for
 702                  * playing.
 703                  */
 704                 WARN_ON_ONCE(st->rollback);
 705                 st->should_run = false;
 706         }
 707
 708         cpuhp_lock_release(bringup);
 709         lockdep_release_cpus_lock();
 710
 711         if (!st->should_run)
 712                 complete_ap_thread(st, bringup);
 713 }
 714
 715 /* Invoke a single callback on a remote cpu */
 716 static int
 717 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 718                          struct hlist_node *node)
 719 {
 720         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 721         int ret;
 722
 723         if (!cpu_online(cpu))
 724                 return 0;
 725
 726         cpuhp_lock_acquire(false);
 727         cpuhp_lock_release(false);
 728
 729         cpuhp_lock_acquire(true);
 730         cpuhp_lock_release(true);
 731
 732         /*
 733          * If we are up and running, use the hotplug thread. For early calls
 734          * we invoke the thread function directly.
 735          */
 736         if (!st->thread)
 737                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 738
 739         st->rollback = false;
 740         st->last = NULL;
 741
 742         st->node = node;
 743         st->bringup = bringup;
 744         st->cb_state = state;
 745         st->single = true;
 746
 747         __cpuhp_kick_ap(st);
 748
 749         /*
 750          * If we failed and did a partial, do a rollback.
 751          */
 752         if ((ret = st->result) && st->last) {
 753                 st->rollback = true;
 754                 st->bringup = !bringup;
 755
 756                 __cpuhp_kick_ap(st);
 757         }
 758
 759         /*
 760          * Clean up the leftovers so the next hotplug operation wont use stale
 761          * data.
 762          */
 763         st->node = st->last = NULL;
 764         return ret;
 765 }
 766
 767 static int cpuhp_kick_ap_work(unsigned int cpu)
 768 {
 769         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 770         enum cpuhp_state prev_state = st->state;
 771         int ret;
 772
 773         cpuhp_lock_acquire(false);
 774         cpuhp_lock_release(false);
 775
 776         cpuhp_lock_acquire(true);
 777         cpuhp_lock_release(true);
 778
 779         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 780         ret = cpuhp_kick_ap(st, st->target);
 781         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 782
 783         return ret;
 784 }
 785
 786 static struct smp_hotplug_thread cpuhp_threads = {
 787         .store                  = &cpuhp_state.thread,
 788         .create                 = &cpuhp_create,
 789         .thread_should_run      = cpuhp_should_run,
 790         .thread_fn              = cpuhp_thread_fun,
 791         .thread_comm            = "cpuhp/%u",
 792         .selfparking            = true,
 793 };
 794
 795 void __init cpuhp_threads_init(void)
 796 {
 797         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 798         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 799 }
 800
 801 #ifdef CONFIG_HOTPLUG_CPU
 802 /**
 803  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 804  * @cpu: a CPU id
 805  *
 806  * This function walks all processes, finds a valid mm struct for each one and
 807  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 808  * trivial, there are various non-obvious corner cases, which this function
 809  * tries to solve in a safe manner.
 810  *
 811  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 812  * be called only for an already offlined CPU.
 813  */
 814 void clear_tasks_mm_cpumask(int cpu)
 815 {
 816         struct task_struct *p;
 817
 818         /*
 819          * This function is called after the cpu is taken down and marked
 820          * offline, so its not like new tasks will ever get this cpu set in
 821          * their mm mask. -- Peter Zijlstra
 822          * Thus, we may use rcu_read_lock() here, instead of grabbing
 823          * full-fledged tasklist_lock.
 824          */
 825         WARN_ON(cpu_online(cpu));
 826         rcu_read_lock();
 827         for_each_process(p) {
 828                 struct task_struct *t;
 829
 830                 /*
 831                  * Main thread might exit, but other threads may still have
 832                  * a valid mm. Find one.
 833                  */
 834                 t = find_lock_task_mm(p);
 835                 if (!t)
 836                         continue;
 837                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 838                 task_unlock(t);
 839         }
 840         rcu_read_unlock();
 841 }
 842
 843 /* Take this CPU down. */
 844 static int take_cpu_down(void *_param)
 845 {
 846         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 847         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 848         int err, cpu = smp_processor_id();
 849         int ret;
 850
 851         /* Ensure this CPU doesn't handle any more interrupts. */
 852         err = __cpu_disable();
 853         if (err < 0)
 854                 return err;
 855
 856         /*
 857          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 858          * do this step again.
 859          */
 860         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 861         st->state--;
 862         /* Invoke the former CPU_DYING callbacks */
 863         for (; st->state > target; st->state--) {
 864                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 865                 /*
 866                  * DYING must not fail!
 867                  */
 868                 WARN_ON_ONCE(ret);
 869         }
 870
 871         /* Give up timekeeping duties */
 872         tick_handover_do_timer();
 873         /* Remove CPU from timer broadcasting */
 874         tick_offline_cpu(cpu);
 875         /* Park the stopper thread */
 876         stop_machine_park(cpu);
 877         return 0;
 878 }
 879
 880 static int takedown_cpu(unsigned int cpu)
 881 {
 882         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 883         int err;
 884
 885         /* Park the smpboot threads */
 886         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 887
 888         /*
 889          * Prevent irq alloc/free while the dying cpu reorganizes the
 890          * interrupt affinities.
 891          */
 892         irq_lock_sparse();
 893
 894         /*
 895          * So now all preempt/rcu users must observe !cpu_active().
 896          */
 897         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 898         if (err) {
 899                 /* CPU refused to die */
 900                 irq_unlock_sparse();
 901                 /* Unpark the hotplug thread so we can rollback there */
 902                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 903                 return err;
 904         }
 905         BUG_ON(cpu_online(cpu));
 906
 907         /*
 908          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 909          * all runnable tasks from the CPU, there's only the idle task left now
 910          * that the migration thread is done doing the stop_machine thing.
 911          *
 912          * Wait for the stop thread to go away.
 913          */
 914         wait_for_ap_thread(st, false);
 915         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 916
 917         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 918         irq_unlock_sparse();
 919
 920         hotplug_cpu__broadcast_tick_pull(cpu);
 921         /* This actually kills the CPU. */
 922         __cpu_die(cpu);
 923
 924         tick_cleanup_dead_cpu(cpu);
 925         rcutree_migrate_callbacks(cpu);
 926         return 0;
 927 }
 928
 929 static void cpuhp_complete_idle_dead(void *arg)
 930 {
 931         struct cpuhp_cpu_state *st = arg;
 932
 933         complete_ap_thread(st, false);
 934 }
 935
 936 void cpuhp_report_idle_dead(void)
 937 {
 938         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 939
 940         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 941         rcu_report_dead(smp_processor_id());
 942         st->state = CPUHP_AP_IDLE_DEAD;
 943         /*
 944          * We cannot call complete after rcu_report_dead() so we delegate it
 945          * to an online cpu.
 946          */
 947         smp_call_function_single(cpumask_first(cpu_online_mask),
 948                                  cpuhp_complete_idle_dead, st, 0);
 949 }
 950
 951 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 952 {
 953         for (st->state++; st->state < st->target; st->state++)
 954                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 955 }
 956
 957 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 958                                 enum cpuhp_state target)
 959 {
 960         enum cpuhp_state prev_state = st->state;
 961         int ret = 0;
 962
 963         for (; st->state > target; st->state--) {
 964                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 965                 if (ret) {
 966                         st->target = prev_state;
 967                         if (st->state < prev_state)
 968                                 undo_cpu_down(cpu, st);
 969                         break;
 970                 }
 971         }
 972         return ret;
 973 }
 974
 975 /* Requires cpu_add_remove_lock to be held */
 976 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 977                            enum cpuhp_state target)
 978 {
 979         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 980         int prev_state, ret = 0;
 981
 982         if (num_online_cpus() == 1)
 983                 return -EBUSY;
 984
 985         if (!cpu_present(cpu))
 986                 return -EINVAL;
 987
 988         cpus_write_lock();
 989
 990         cpuhp_tasks_frozen = tasks_frozen;
 991
 992         prev_state = cpuhp_set_state(st, target);
 993         /*
 994          * If the current CPU state is in the range of the AP hotplug thread,
 995          * then we need to kick the thread.
 996          */
 997         if (st->state > CPUHP_TEARDOWN_CPU) {
 998                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 999                 ret = cpuhp_kick_ap_work(cpu);
1000                 /*
1001                  * The AP side has done the error rollback already. Just
1002                  * return the error code..
1003                  */
1004                 if (ret)
1005                         goto out;
1006
1007                 /*
1008                  * We might have stopped still in the range of the AP hotplug
1009                  * thread. Nothing to do anymore.
1010                  */
1011                 if (st->state > CPUHP_TEARDOWN_CPU)
1012                         goto out;
1013
1014                 st->target = target;
1015         }
1016         /*
1017          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1018          * to do the further cleanups.
1019          */
1020         ret = cpuhp_down_callbacks(cpu, st, target);
1021         if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1022                 cpuhp_reset_state(st, prev_state);
1023                 __cpuhp_kick_ap(st);
1024         }
1025
1026 out:
1027         cpus_write_unlock();
1028         /*
1029          * Do post unplug cleanup. This is still protected against
1030          * concurrent CPU hotplug via cpu_add_remove_lock.
1031          */
1032         lockup_detector_cleanup();
1033         arch_smt_update();
1034         return ret;
1035 }
1036
1037 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1038 {
1039         if (cpu_hotplug_disabled)
1040                 return -EBUSY;
1041         return _cpu_down(cpu, 0, target);
1042 }
1043
1044 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1045 {
1046         int err;
1047
1048         cpu_maps_update_begin();
1049         err = cpu_down_maps_locked(cpu, target);
1050         cpu_maps_update_done();
1051         return err;
1052 }
1053
1054 int cpu_down(unsigned int cpu)
1055 {
1056         return do_cpu_down(cpu, CPUHP_OFFLINE);
1057 }
1058 EXPORT_SYMBOL(cpu_down);
1059
1060 #else
1061 #define takedown_cpu            NULL
1062 #endif /*CONFIG_HOTPLUG_CPU*/
1063
1064 /**
1065  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1066  * @cpu: cpu that just started
1067  *
1068  * It must be called by the arch code on the new cpu, before the new cpu
1069  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1070  */
1071 void notify_cpu_starting(unsigned int cpu)
1072 {
1073         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1074         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1075         int ret;
1076
1077         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1078         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1079         while (st->state < target) {
1080                 st->state++;
1081                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1082                 /*
1083                  * STARTING must not fail!
1084                  */
1085                 WARN_ON_ONCE(ret);
1086         }
1087 }
1088
1089 /*
1090  * Called from the idle task. Wake up the controlling task which brings the
1091  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1092  * online bringup to the hotplug thread.
1093  */
1094 void cpuhp_online_idle(enum cpuhp_state state)
1095 {
1096         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1097
1098         /* Happens for the boot cpu */
1099         if (state != CPUHP_AP_ONLINE_IDLE)
1100                 return;
1101
1102         /*
1103          * Unpart the stopper thread before we start the idle loop (and start
1104          * scheduling); this ensures the stopper task is always available.
1105          */
1106         stop_machine_unpark(smp_processor_id());
1107
1108         st->state = CPUHP_AP_ONLINE_IDLE;
1109         complete_ap_thread(st, true);
1110 }
1111
1112 /* Requires cpu_add_remove_lock to be held */
1113 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1114 {
1115         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1116         struct task_struct *idle;
1117         int ret = 0;
1118
1119         cpus_write_lock();
1120
1121         if (!cpu_present(cpu)) {
1122                 ret = -EINVAL;
1123                 goto out;
1124         }
1125
1126         /*
1127          * The caller of do_cpu_up might have raced with another
1128          * caller. Ignore it for now.
1129          */
1130         if (st->state >= target)
1131                 goto out;
1132
1133         if (st->state == CPUHP_OFFLINE) {
1134                 /* Let it fail before we try to bring the cpu up */
1135                 idle = idle_thread_get(cpu);
1136                 if (IS_ERR(idle)) {
1137                         ret = PTR_ERR(idle);
1138                         goto out;
1139                 }
1140         }
1141
1142         cpuhp_tasks_frozen = tasks_frozen;
1143
1144         cpuhp_set_state(st, target);
1145         /*
1146          * If the current CPU state is in the range of the AP hotplug thread,
1147          * then we need to kick the thread once more.
1148          */
1149         if (st->state > CPUHP_BRINGUP_CPU) {
1150                 ret = cpuhp_kick_ap_work(cpu);
1151                 /*
1152                  * The AP side has done the error rollback already. Just
1153                  * return the error code..
1154                  */
1155                 if (ret)
1156                         goto out;
1157         }
1158
1159         /*
1160          * Try to reach the target state. We max out on the BP at
1161          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1162          * responsible for bringing it up to the target state.
1163          */
1164         target = min((int)target, CPUHP_BRINGUP_CPU);
1165         ret = cpuhp_up_callbacks(cpu, st, target);
1166 out:
1167         cpus_write_unlock();
1168         arch_smt_update();
1169         return ret;
1170 }
1171
1172 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1173 {
1174         int err = 0;
1175
1176         if (!cpu_possible(cpu)) {
1177                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1178                        cpu);
1179 #if defined(CONFIG_IA64)
1180                 pr_err("please check additional_cpus= boot parameter\n");
1181 #endif
1182                 return -EINVAL;
1183         }
1184
1185         err = try_online_node(cpu_to_node(cpu));
1186         if (err)
1187                 return err;
1188
1189         cpu_maps_update_begin();
1190
1191         if (cpu_hotplug_disabled) {
1192                 err = -EBUSY;
1193                 goto out;
1194         }
1195         if (!cpu_smt_allowed(cpu)) {
1196                 err = -EPERM;
1197                 goto out;
1198         }
1199
1200         err = _cpu_up(cpu, 0, target);
1201 out:
1202         cpu_maps_update_done();
1203         return err;
1204 }
1205
1206 int cpu_up(unsigned int cpu)
1207 {
1208         return do_cpu_up(cpu, CPUHP_ONLINE);
1209 }
1210 EXPORT_SYMBOL_GPL(cpu_up);
1211
1212 #ifdef CONFIG_PM_SLEEP_SMP
1213 static cpumask_var_t frozen_cpus;
1214
1215 int freeze_secondary_cpus(int primary)
1216 {
1217         int cpu, error = 0;
1218
1219         cpu_maps_update_begin();
1220         if (primary == -1) {
1221                 primary = cpumask_first(cpu_online_mask);
1222                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1223                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1224         } else {
1225                 if (!cpu_online(primary))
1226                         primary = cpumask_first(cpu_online_mask);
1227         }
1228
1229         /*
1230          * We take down all of the non-boot CPUs in one shot to avoid races
1231          * with the userspace trying to use the CPU hotplug at the same time
1232          */
1233         cpumask_clear(frozen_cpus);
1234
1235         pr_info("Disabling non-boot CPUs ...\n");
1236         for_each_online_cpu(cpu) {
1237                 if (cpu == primary)
1238                         continue;
1239
1240                 if (pm_wakeup_pending()) {
1241                         pr_info("Wakeup pending. Abort CPU freeze\n");
1242                         error = -EBUSY;
1243                         break;
1244                 }
1245
1246                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1247                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1248                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1249                 if (!error)
1250                         cpumask_set_cpu(cpu, frozen_cpus);
1251                 else {
1252                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1253                         break;
1254                 }
1255         }
1256
1257         if (!error)
1258                 BUG_ON(num_online_cpus() > 1);
1259         else
1260                 pr_err("Non-boot CPUs are not disabled\n");
1261
1262         /*
1263          * Make sure the CPUs won't be enabled by someone else. We need to do
1264          * this even in case of failure as all disable_nonboot_cpus() users are
1265          * supposed to do enable_nonboot_cpus() on the failure path.
1266          */
1267         cpu_hotplug_disabled++;
1268
1269         cpu_maps_update_done();
1270         return error;
1271 }
1272
1273 void __weak arch_enable_nonboot_cpus_begin(void)
1274 {
1275 }
1276
1277 void __weak arch_enable_nonboot_cpus_end(void)
1278 {
1279 }
1280
1281 void enable_nonboot_cpus(void)
1282 {
1283         int cpu, error;
1284
1285         /* Allow everyone to use the CPU hotplug again */
1286         cpu_maps_update_begin();
1287         __cpu_hotplug_enable();
1288         if (cpumask_empty(frozen_cpus))
1289                 goto out;
1290
1291         pr_info("Enabling non-boot CPUs ...\n");
1292
1293         arch_enable_nonboot_cpus_begin();
1294
1295         for_each_cpu(cpu, frozen_cpus) {
1296                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1297                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1298                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1299                 if (!error) {
1300                         pr_info("CPU%d is up\n", cpu);
1301                         continue;
1302                 }
1303                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1304         }
1305
1306         arch_enable_nonboot_cpus_end();
1307
1308         cpumask_clear(frozen_cpus);
1309 out:
1310         cpu_maps_update_done();
1311 }
1312
1313 static int __init alloc_frozen_cpus(void)
1314 {
1315         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1316                 return -ENOMEM;
1317         return 0;
1318 }
1319 core_initcall(alloc_frozen_cpus);
1320
1321 /*
1322  * When callbacks for CPU hotplug notifications are being executed, we must
1323  * ensure that the state of the system with respect to the tasks being frozen
1324  * or not, as reported by the notification, remains unchanged *throughout the
1325  * duration* of the execution of the callbacks.
1326  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1327  *
1328  * This synchronization is implemented by mutually excluding regular CPU
1329  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1330  * Hibernate notifications.
1331  */
1332 static int
1333 cpu_hotplug_pm_callback(struct notifier_block *nb,
1334                         unsigned long action, void *ptr)
1335 {
1336         switch (action) {
1337
1338         case PM_SUSPEND_PREPARE:
1339         case PM_HIBERNATION_PREPARE:
1340                 cpu_hotplug_disable();
1341                 break;
1342
1343         case PM_POST_SUSPEND:
1344         case PM_POST_HIBERNATION:
1345                 cpu_hotplug_enable();
1346                 break;
1347
1348         default:
1349                 return NOTIFY_DONE;
1350         }
1351
1352         return NOTIFY_OK;
1353 }
1354
1355
1356 static int __init cpu_hotplug_pm_sync_init(void)
1357 {
1358         /*
1359          * cpu_hotplug_pm_callback has higher priority than x86
1360          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1361          * to disable cpu hotplug to avoid cpu hotplug race.
1362          */
1363         pm_notifier(cpu_hotplug_pm_callback, 0);
1364         return 0;
1365 }
1366 core_initcall(cpu_hotplug_pm_sync_init);
1367
1368 #endif /* CONFIG_PM_SLEEP_SMP */
1369
1370 int __boot_cpu_id;
1371
1372 #endif /* CONFIG_SMP */
1373
1374 /* Boot processor state steps */
1375 static struct cpuhp_step cpuhp_hp_states[] = {
1376         [CPUHP_OFFLINE] = {
1377                 .name                   = "offline",
1378                 .startup.single         = NULL,
1379                 .teardown.single        = NULL,
1380         },
1381 #ifdef CONFIG_SMP
1382         [CPUHP_CREATE_THREADS]= {
1383                 .name                   = "threads:prepare",
1384                 .startup.single         = smpboot_create_threads,
1385                 .teardown.single        = NULL,
1386                 .cant_stop              = true,
1387         },
1388         [CPUHP_PERF_PREPARE] = {
1389                 .name                   = "perf:prepare",
1390                 .startup.single         = perf_event_init_cpu,
1391                 .teardown.single        = perf_event_exit_cpu,
1392         },
1393         [CPUHP_WORKQUEUE_PREP] = {
1394                 .name                   = "workqueue:prepare",
1395                 .startup.single         = workqueue_prepare_cpu,
1396                 .teardown.single        = NULL,
1397         },
1398         [CPUHP_HRTIMERS_PREPARE] = {
1399                 .name                   = "hrtimers:prepare",
1400                 .startup.single         = hrtimers_prepare_cpu,
1401                 .teardown.single        = hrtimers_dead_cpu,
1402         },
1403         [CPUHP_SMPCFD_PREPARE] = {
1404                 .name                   = "smpcfd:prepare",
1405                 .startup.single         = smpcfd_prepare_cpu,
1406                 .teardown.single        = smpcfd_dead_cpu,
1407         },
1408         [CPUHP_RELAY_PREPARE] = {
1409                 .name                   = "relay:prepare",
1410                 .startup.single         = relay_prepare_cpu,
1411                 .teardown.single        = NULL,
1412         },
1413         [CPUHP_SLAB_PREPARE] = {
1414                 .name                   = "slab:prepare",
1415                 .startup.single         = slab_prepare_cpu,
1416                 .teardown.single        = slab_dead_cpu,
1417         },
1418         [CPUHP_RCUTREE_PREP] = {
1419                 .name                   = "RCU/tree:prepare",
1420                 .startup.single         = rcutree_prepare_cpu,
1421                 .teardown.single        = rcutree_dead_cpu,
1422         },
1423         /*
1424          * On the tear-down path, timers_dead_cpu() must be invoked
1425          * before blk_mq_queue_reinit_notify() from notify_dead(),
1426          * otherwise a RCU stall occurs.
1427          */
1428         [CPUHP_TIMERS_PREPARE] = {
1429                 .name                   = "timers:prepare",
1430                 .startup.single         = timers_prepare_cpu,
1431                 .teardown.single        = timers_dead_cpu,
1432         },
1433         /* Kicks the plugged cpu into life */
1434         [CPUHP_BRINGUP_CPU] = {
1435                 .name                   = "cpu:bringup",
1436                 .startup.single         = bringup_cpu,
1437                 .teardown.single        = NULL,
1438                 .cant_stop              = true,
1439         },
1440         /* Final state before CPU kills itself */
1441         [CPUHP_AP_IDLE_DEAD] = {
1442                 .name                   = "idle:dead",
1443         },
1444         /*
1445          * Last state before CPU enters the idle loop to die. Transient state
1446          * for synchronization.
1447          */
1448         [CPUHP_AP_OFFLINE] = {
1449                 .name                   = "ap:offline",
1450                 .cant_stop              = true,
1451         },
1452         /* First state is scheduler control. Interrupts are disabled */
1453         [CPUHP_AP_SCHED_STARTING] = {
1454                 .name                   = "sched:starting",
1455                 .startup.single         = sched_cpu_starting,
1456                 .teardown.single        = sched_cpu_dying,
1457         },
1458         [CPUHP_AP_RCUTREE_DYING] = {
1459                 .name                   = "RCU/tree:dying",
1460                 .startup.single         = NULL,
1461                 .teardown.single        = rcutree_dying_cpu,
1462         },
1463         [CPUHP_AP_SMPCFD_DYING] = {
1464                 .name                   = "smpcfd:dying",
1465                 .startup.single         = NULL,
1466                 .teardown.single        = smpcfd_dying_cpu,
1467         },
1468         /* Entry state on starting. Interrupts enabled from here on. Transient
1469          * state for synchronsization */
1470         [CPUHP_AP_ONLINE] = {
1471                 .name                   = "ap:online",
1472         },
1473         /*
1474          * Handled on controll processor until the plugged processor manages
1475          * this itself.
1476          */
1477         [CPUHP_TEARDOWN_CPU] = {
1478                 .name                   = "cpu:teardown",
1479                 .startup.single         = NULL,
1480                 .teardown.single        = takedown_cpu,
1481                 .cant_stop              = true,
1482         },
1483         /* Handle smpboot threads park/unpark */
1484         [CPUHP_AP_SMPBOOT_THREADS] = {
1485                 .name                   = "smpboot/threads:online",
1486                 .startup.single         = smpboot_unpark_threads,
1487                 .teardown.single        = smpboot_park_threads,
1488         },
1489         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1490                 .name                   = "irq/affinity:online",
1491                 .startup.single         = irq_affinity_online_cpu,
1492                 .teardown.single        = NULL,
1493         },
1494         [CPUHP_AP_PERF_ONLINE] = {
1495                 .name                   = "perf:online",
1496                 .startup.single         = perf_event_init_cpu,
1497                 .teardown.single        = perf_event_exit_cpu,
1498         },
1499         [CPUHP_AP_WATCHDOG_ONLINE] = {
1500                 .name                   = "lockup_detector:online",
1501                 .startup.single         = lockup_detector_online_cpu,
1502                 .teardown.single        = lockup_detector_offline_cpu,
1503         },
1504         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1505                 .name                   = "workqueue:online",
1506                 .startup.single         = workqueue_online_cpu,
1507                 .teardown.single        = workqueue_offline_cpu,
1508         },
1509         [CPUHP_AP_RCUTREE_ONLINE] = {
1510                 .name                   = "RCU/tree:online",
1511                 .startup.single         = rcutree_online_cpu,
1512                 .teardown.single        = rcutree_offline_cpu,
1513         },
1514 #endif
1515         /*
1516          * The dynamically registered state space is here
1517          */
1518
1519 #ifdef CONFIG_SMP
1520         /* Last state is scheduler control setting the cpu active */
1521         [CPUHP_AP_ACTIVE] = {
1522                 .name                   = "sched:active",
1523                 .startup.single         = sched_cpu_activate,
1524                 .teardown.single        = sched_cpu_deactivate,
1525         },
1526 #endif
1527
1528         /* CPU is fully up and running. */
1529         [CPUHP_ONLINE] = {
1530                 .name                   = "online",
1531                 .startup.single         = NULL,
1532                 .teardown.single        = NULL,
1533         },
1534 };
1535
1536 /* Sanity check for callbacks */
1537 static int cpuhp_cb_check(enum cpuhp_state state)
1538 {
1539         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1540                 return -EINVAL;
1541         return 0;
1542 }
1543
1544 /*
1545  * Returns a free for dynamic slot assignment of the Online state. The states
1546  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1547  * by having no name assigned.
1548  */
1549 static int cpuhp_reserve_state(enum cpuhp_state state)
1550 {
1551         enum cpuhp_state i, end;
1552         struct cpuhp_step *step;
1553
1554         switch (state) {
1555         case CPUHP_AP_ONLINE_DYN:
1556                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1557                 end = CPUHP_AP_ONLINE_DYN_END;
1558                 break;
1559         case CPUHP_BP_PREPARE_DYN:
1560                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1561                 end = CPUHP_BP_PREPARE_DYN_END;
1562                 break;
1563         default:
1564                 return -EINVAL;
1565         }
1566
1567         for (i = state; i <= end; i++, step++) {
1568                 if (!step->name)
1569                         return i;
1570         }
1571         WARN(1, "No more dynamic states available for CPU hotplug\n");
1572         return -ENOSPC;
1573 }
1574
1575 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1576                                  int (*startup)(unsigned int cpu),
1577                                  int (*teardown)(unsigned int cpu),
1578                                  bool multi_instance)
1579 {
1580         /* (Un)Install the callbacks for further cpu hotplug operations */
1581         struct cpuhp_step *sp;
1582         int ret = 0;
1583
1584         /*
1585          * If name is NULL, then the state gets removed.
1586          *
1587          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1588          * the first allocation from these dynamic ranges, so the removal
1589          * would trigger a new allocation and clear the wrong (already
1590          * empty) state, leaving the callbacks of the to be cleared state
1591          * dangling, which causes wreckage on the next hotplug operation.
1592          */
1593         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1594                      state == CPUHP_BP_PREPARE_DYN)) {
1595                 ret = cpuhp_reserve_state(state);
1596                 if (ret < 0)
1597                         return ret;
1598                 state = ret;
1599         }
1600         sp = cpuhp_get_step(state);
1601         if (name && sp->name)
1602                 return -EBUSY;
1603
1604         sp->startup.single = startup;
1605         sp->teardown.single = teardown;
1606         sp->name = name;
1607         sp->multi_instance = multi_instance;
1608         INIT_HLIST_HEAD(&sp->list);
1609         return ret;
1610 }
1611
1612 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1613 {
1614         return cpuhp_get_step(state)->teardown.single;
1615 }
1616
1617 /*
1618  * Call the startup/teardown function for a step either on the AP or
1619  * on the current CPU.
1620  */
1621 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1622                             struct hlist_node *node)
1623 {
1624         struct cpuhp_step *sp = cpuhp_get_step(state);
1625         int ret;
1626
1627         /*
1628          * If there's nothing to do, we done.
1629          * Relies on the union for multi_instance.
1630          */
1631         if ((bringup && !sp->startup.single) ||
1632             (!bringup && !sp->teardown.single))
1633                 return 0;
1634         /*
1635          * The non AP bound callbacks can fail on bringup. On teardown
1636          * e.g. module removal we crash for now.
1637          */
1638 #ifdef CONFIG_SMP
1639         if (cpuhp_is_ap_state(state))
1640                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1641         else
1642                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1643 #else
1644         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1645 #endif
1646         BUG_ON(ret && !bringup);
1647         return ret;
1648 }
1649
1650 /*
1651  * Called from __cpuhp_setup_state on a recoverable failure.
1652  *
1653  * Note: The teardown callbacks for rollback are not allowed to fail!
1654  */
1655 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1656                                    struct hlist_node *node)
1657 {
1658         int cpu;
1659
1660         /* Roll back the already executed steps on the other cpus */
1661         for_each_present_cpu(cpu) {
1662                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1663                 int cpustate = st->state;
1664
1665                 if (cpu >= failedcpu)
1666                         break;
1667
1668                 /* Did we invoke the startup call on that cpu ? */
1669                 if (cpustate >= state)
1670                         cpuhp_issue_call(cpu, state, false, node);
1671         }
1672 }
1673
1674 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1675                                           struct hlist_node *node,
1676                                           bool invoke)
1677 {
1678         struct cpuhp_step *sp;
1679         int cpu;
1680         int ret;
1681
1682         lockdep_assert_cpus_held();
1683
1684         sp = cpuhp_get_step(state);
1685         if (sp->multi_instance == false)
1686                 return -EINVAL;
1687
1688         mutex_lock(&cpuhp_state_mutex);
1689
1690         if (!invoke || !sp->startup.multi)
1691                 goto add_node;
1692
1693         /*
1694          * Try to call the startup callback for each present cpu
1695          * depending on the hotplug state of the cpu.
1696          */
1697         for_each_present_cpu(cpu) {
1698                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1699                 int cpustate = st->state;
1700
1701                 if (cpustate < state)
1702                         continue;
1703
1704                 ret = cpuhp_issue_call(cpu, state, true, node);
1705                 if (ret) {
1706                         if (sp->teardown.multi)
1707                                 cpuhp_rollback_install(cpu, state, node);
1708                         goto unlock;
1709                 }
1710         }
1711 add_node:
1712         ret = 0;
1713         hlist_add_head(node, &sp->list);
1714 unlock:
1715         mutex_unlock(&cpuhp_state_mutex);
1716         return ret;
1717 }
1718
1719 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1720                                bool invoke)
1721 {
1722         int ret;
1723
1724         cpus_read_lock();
1725         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1726         cpus_read_unlock();
1727         return ret;
1728 }
1729 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1730
1731 /**
1732  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1733  * @state:              The state to setup
1734  * @invoke:             If true, the startup function is invoked for cpus where
1735  *                      cpu state >= @state
1736  * @startup:            startup callback function
1737  * @teardown:           teardown callback function
1738  * @multi_instance:     State is set up for multiple instances which get
1739  *                      added afterwards.
1740  *
1741  * The caller needs to hold cpus read locked while calling this function.
1742  * Returns:
1743  *   On success:
1744  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1745  *      0 for all other states
1746  *   On failure: proper (negative) error code
1747  */
1748 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1749                                    const char *name, bool invoke,
1750                                    int (*startup)(unsigned int cpu),
1751                                    int (*teardown)(unsigned int cpu),
1752                                    bool multi_instance)
1753 {
1754         int cpu, ret = 0;
1755         bool dynstate;
1756
1757         lockdep_assert_cpus_held();
1758
1759         if (cpuhp_cb_check(state) || !name)
1760                 return -EINVAL;
1761
1762         mutex_lock(&cpuhp_state_mutex);
1763
1764         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1765                                     multi_instance);
1766
1767         dynstate = state == CPUHP_AP_ONLINE_DYN;
1768         if (ret > 0 && dynstate) {
1769                 state = ret;
1770                 ret = 0;
1771         }
1772
1773         if (ret || !invoke || !startup)
1774                 goto out;
1775
1776         /*
1777          * Try to call the startup callback for each present cpu
1778          * depending on the hotplug state of the cpu.
1779          */
1780         for_each_present_cpu(cpu) {
1781                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1782                 int cpustate = st->state;
1783
1784                 if (cpustate < state)
1785                         continue;
1786
1787                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1788                 if (ret) {
1789                         if (teardown)
1790                                 cpuhp_rollback_install(cpu, state, NULL);
1791                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1792                         goto out;
1793                 }
1794         }
1795 out:
1796         mutex_unlock(&cpuhp_state_mutex);
1797         /*
1798          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1799          * dynamically allocated state in case of success.
1800          */
1801         if (!ret && dynstate)
1802                 return state;
1803         return ret;
1804 }
1805 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1806
1807 int __cpuhp_setup_state(enum cpuhp_state state,
1808                         const char *name, bool invoke,
1809                         int (*startup)(unsigned int cpu),
1810                         int (*teardown)(unsigned int cpu),
1811                         bool multi_instance)
1812 {
1813         int ret;
1814
1815         cpus_read_lock();
1816         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1817                                              teardown, multi_instance);
1818         cpus_read_unlock();
1819         return ret;
1820 }
1821 EXPORT_SYMBOL(__cpuhp_setup_state);
1822
1823 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1824                                   struct hlist_node *node, bool invoke)
1825 {
1826         struct cpuhp_step *sp = cpuhp_get_step(state);
1827         int cpu;
1828
1829         BUG_ON(cpuhp_cb_check(state));
1830
1831         if (!sp->multi_instance)
1832                 return -EINVAL;
1833
1834         cpus_read_lock();
1835         mutex_lock(&cpuhp_state_mutex);
1836
1837         if (!invoke || !cpuhp_get_teardown_cb(state))
1838                 goto remove;
1839         /*
1840          * Call the teardown callback for each present cpu depending
1841          * on the hotplug state of the cpu. This function is not
1842          * allowed to fail currently!
1843          */
1844         for_each_present_cpu(cpu) {
1845                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1846                 int cpustate = st->state;
1847
1848                 if (cpustate >= state)
1849                         cpuhp_issue_call(cpu, state, false, node);
1850         }
1851
1852 remove:
1853         hlist_del(node);
1854         mutex_unlock(&cpuhp_state_mutex);
1855         cpus_read_unlock();
1856
1857         return 0;
1858 }
1859 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1860
1861 /**
1862  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1863  * @state:      The state to remove
1864  * @invoke:     If true, the teardown function is invoked for cpus where
1865  *              cpu state >= @state
1866  *
1867  * The caller needs to hold cpus read locked while calling this function.
1868  * The teardown callback is currently not allowed to fail. Think
1869  * about module removal!
1870  */
1871 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1872 {
1873         struct cpuhp_step *sp = cpuhp_get_step(state);
1874         int cpu;
1875
1876         BUG_ON(cpuhp_cb_check(state));
1877
1878         lockdep_assert_cpus_held();
1879
1880         mutex_lock(&cpuhp_state_mutex);
1881         if (sp->multi_instance) {
1882                 WARN(!hlist_empty(&sp->list),
1883                      "Error: Removing state %d which has instances left.\n",
1884                      state);
1885                 goto remove;
1886         }
1887
1888         if (!invoke || !cpuhp_get_teardown_cb(state))
1889                 goto remove;
1890
1891         /*
1892          * Call the teardown callback for each present cpu depending
1893          * on the hotplug state of the cpu. This function is not
1894          * allowed to fail currently!
1895          */
1896         for_each_present_cpu(cpu) {
1897                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1898                 int cpustate = st->state;
1899
1900                 if (cpustate >= state)
1901                         cpuhp_issue_call(cpu, state, false, NULL);
1902         }
1903 remove:
1904         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1905         mutex_unlock(&cpuhp_state_mutex);
1906 }
1907 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1908
1909 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1910 {
1911         cpus_read_lock();
1912         __cpuhp_remove_state_cpuslocked(state, invoke);
1913         cpus_read_unlock();
1914 }
1915 EXPORT_SYMBOL(__cpuhp_remove_state);
1916
1917 #ifdef CONFIG_HOTPLUG_SMT
1918 static void cpuhp_offline_cpu_device(unsigned int cpu)
1919 {
1920         struct device *dev = get_cpu_device(cpu);
1921
1922         dev->offline = true;
1923         /* Tell user space about the state change */
1924         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
1925 }
1926
1927 static void cpuhp_online_cpu_device(unsigned int cpu)
1928 {
1929         struct device *dev = get_cpu_device(cpu);
1930
1931         dev->offline = false;
1932         /* Tell user space about the state change */
1933         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
1934 }
1935
1936 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
1937 {
1938         int cpu, ret = 0;
1939
1940         cpu_maps_update_begin();
1941         for_each_online_cpu(cpu) {
1942                 if (topology_is_primary_thread(cpu))
1943                         continue;
1944                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1945                 if (ret)
1946                         break;
1947                 /*
1948                  * As this needs to hold the cpu maps lock it's impossible
1949                  * to call device_offline() because that ends up calling
1950                  * cpu_down() which takes cpu maps lock. cpu maps lock
1951                  * needs to be held as this might race against in kernel
1952                  * abusers of the hotplug machinery (thermal management).
1953                  *
1954                  * So nothing would update device:offline state. That would
1955                  * leave the sysfs entry stale and prevent onlining after
1956                  * smt control has been changed to 'off' again. This is
1957                  * called under the sysfs hotplug lock, so it is properly
1958                  * serialized against the regular offline usage.
1959                  */
1960                 cpuhp_offline_cpu_device(cpu);
1961         }
1962         if (!ret)
1963                 cpu_smt_control = ctrlval;
1964         cpu_maps_update_done();
1965         return ret;
1966 }
1967
1968 int cpuhp_smt_enable(void)
1969 {
1970         int cpu, ret = 0;
1971
1972         cpu_maps_update_begin();
1973         cpu_smt_control = CPU_SMT_ENABLED;
1974         for_each_present_cpu(cpu) {
1975                 /* Skip online CPUs and CPUs on offline nodes */
1976                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
1977                         continue;
1978                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
1979                 if (ret)
1980                         break;
1981                 /* See comment in cpuhp_smt_disable() */
1982                 cpuhp_online_cpu_device(cpu);
1983         }
1984         cpu_maps_update_done();
1985         return ret;
1986 }
1987 #endif
1988
1989 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1990 static ssize_t show_cpuhp_state(struct device *dev,
1991                                 struct device_attribute *attr, char *buf)
1992 {
1993         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1994
1995         return sprintf(buf, "%d\n", st->state);
1996 }
1997 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1998
1999 static ssize_t write_cpuhp_target(struct device *dev,
2000                                   struct device_attribute *attr,
2001                                   const char *buf, size_t count)
2002 {
2003         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2004         struct cpuhp_step *sp;
2005         int target, ret;
2006
2007         ret = kstrtoint(buf, 10, &target);
2008         if (ret)
2009                 return ret;
2010
2011 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2012         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2013                 return -EINVAL;
2014 #else
2015         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2016                 return -EINVAL;
2017 #endif
2018
2019         ret = lock_device_hotplug_sysfs();
2020         if (ret)
2021                 return ret;
2022
2023         mutex_lock(&cpuhp_state_mutex);
2024         sp = cpuhp_get_step(target);
2025         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2026         mutex_unlock(&cpuhp_state_mutex);
2027         if (ret)
2028                 goto out;
2029
2030         if (st->state < target)
2031                 ret = do_cpu_up(dev->id, target);
2032         else
2033                 ret = do_cpu_down(dev->id, target);
2034 out:
2035         unlock_device_hotplug();
2036         return ret ? ret : count;
2037 }
2038
2039 static ssize_t show_cpuhp_target(struct device *dev,
2040                                  struct device_attribute *attr, char *buf)
2041 {
2042         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2043
2044         return sprintf(buf, "%d\n", st->target);
2045 }
2046 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
2047
2048
2049 static ssize_t write_cpuhp_fail(struct device *dev,
2050                                 struct device_attribute *attr,
2051                                 const char *buf, size_t count)
2052 {
2053         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2054         struct cpuhp_step *sp;
2055         int fail, ret;
2056
2057         ret = kstrtoint(buf, 10, &fail);
2058         if (ret)
2059                 return ret;
2060
2061         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2062                 return -EINVAL;
2063
2064         /*
2065          * Cannot fail STARTING/DYING callbacks.
2066          */
2067         if (cpuhp_is_atomic_state(fail))
2068                 return -EINVAL;
2069
2070         /*
2071          * Cannot fail anything that doesn't have callbacks.
2072          */
2073         mutex_lock(&cpuhp_state_mutex);
2074         sp = cpuhp_get_step(fail);
2075         if (!sp->startup.single && !sp->teardown.single)
2076                 ret = -EINVAL;
2077         mutex_unlock(&cpuhp_state_mutex);
2078         if (ret)
2079                 return ret;
2080
2081         st->fail = fail;
2082
2083         return count;
2084 }
2085
2086 static ssize_t show_cpuhp_fail(struct device *dev,
2087                                struct device_attribute *attr, char *buf)
2088 {
2089         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2090
2091         return sprintf(buf, "%d\n", st->fail);
2092 }
2093
2094 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2095
2096 static struct attribute *cpuhp_cpu_attrs[] = {
2097         &dev_attr_state.attr,
2098         &dev_attr_target.attr,
2099         &dev_attr_fail.attr,
2100         NULL
2101 };
2102
2103 static const struct attribute_group cpuhp_cpu_attr_group = {
2104         .attrs = cpuhp_cpu_attrs,
2105         .name = "hotplug",
2106         NULL
2107 };
2108
2109 static ssize_t show_cpuhp_states(struct device *dev,
2110                                  struct device_attribute *attr, char *buf)
2111 {
2112         ssize_t cur, res = 0;
2113         int i;
2114
2115         mutex_lock(&cpuhp_state_mutex);
2116         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2117                 struct cpuhp_step *sp = cpuhp_get_step(i);
2118
2119                 if (sp->name) {
2120                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2121                         buf += cur;
2122                         res += cur;
2123                 }
2124         }
2125         mutex_unlock(&cpuhp_state_mutex);
2126         return res;
2127 }
2128 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2129
2130 static struct attribute *cpuhp_cpu_root_attrs[] = {
2131         &dev_attr_states.attr,
2132         NULL
2133 };
2134
2135 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2136         .attrs = cpuhp_cpu_root_attrs,
2137         .name = "hotplug",
2138         NULL
2139 };
2140
2141 #ifdef CONFIG_HOTPLUG_SMT
2142
2143 static ssize_t
2144 __store_smt_control(struct device *dev, struct device_attribute *attr,
2145                     const char *buf, size_t count)
2146 {
2147         int ctrlval, ret;
2148
2149         if (sysfs_streq(buf, "on"))
2150                 ctrlval = CPU_SMT_ENABLED;
2151         else if (sysfs_streq(buf, "off"))
2152                 ctrlval = CPU_SMT_DISABLED;
2153         else if (sysfs_streq(buf, "forceoff"))
2154                 ctrlval = CPU_SMT_FORCE_DISABLED;
2155         else
2156                 return -EINVAL;
2157
2158         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2159                 return -EPERM;
2160
2161         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2162                 return -ENODEV;
2163
2164         ret = lock_device_hotplug_sysfs();
2165         if (ret)
2166                 return ret;
2167
2168         if (ctrlval != cpu_smt_control) {
2169                 switch (ctrlval) {
2170                 case CPU_SMT_ENABLED:
2171                         ret = cpuhp_smt_enable();
2172                         break;
2173                 case CPU_SMT_DISABLED:
2174                 case CPU_SMT_FORCE_DISABLED:
2175                         ret = cpuhp_smt_disable(ctrlval);
2176                         break;
2177                 }
2178         }
2179
2180         unlock_device_hotplug();
2181         return ret ? ret : count;
2182 }
2183
2184 #else /* !CONFIG_HOTPLUG_SMT */
2185 static ssize_t
2186 __store_smt_control(struct device *dev, struct device_attribute *attr,
2187                     const char *buf, size_t count)
2188 {
2189         return -ENODEV;
2190 }
2191 #endif /* CONFIG_HOTPLUG_SMT */
2192
2193 static const char *smt_states[] = {
2194         [CPU_SMT_ENABLED]               = "on",
2195         [CPU_SMT_DISABLED]              = "off",
2196         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2197         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2198         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2199 };
2200
2201 static ssize_t
2202 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2203 {
2204         const char *state = smt_states[cpu_smt_control];
2205
2206         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2207 }
2208
2209 static ssize_t
2210 store_smt_control(struct device *dev, struct device_attribute *attr,
2211                   const char *buf, size_t count)
2212 {
2213         return __store_smt_control(dev, attr, buf, count);
2214 }
2215 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2216
2217 static ssize_t
2218 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2219 {
2220         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2221 }
2222 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2223
2224 static struct attribute *cpuhp_smt_attrs[] = {
2225         &dev_attr_control.attr,
2226         &dev_attr_active.attr,
2227         NULL
2228 };
2229
2230 static const struct attribute_group cpuhp_smt_attr_group = {
2231         .attrs = cpuhp_smt_attrs,
2232         .name = "smt",
2233         NULL
2234 };
2235
2236 static int __init cpu_smt_sysfs_init(void)
2237 {
2238         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2239                                   &cpuhp_smt_attr_group);
2240 }
2241
2242 static int __init cpuhp_sysfs_init(void)
2243 {
2244         int cpu, ret;
2245
2246         ret = cpu_smt_sysfs_init();
2247         if (ret)
2248                 return ret;
2249
2250         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2251                                  &cpuhp_cpu_root_attr_group);
2252         if (ret)
2253                 return ret;
2254
2255         for_each_possible_cpu(cpu) {
2256                 struct device *dev = get_cpu_device(cpu);
2257
2258                 if (!dev)
2259                         continue;
2260                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2261                 if (ret)
2262                         return ret;
2263         }
2264         return 0;
2265 }
2266 device_initcall(cpuhp_sysfs_init);
2267 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2268
2269 /*
2270  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2271  * represents all NR_CPUS bits binary values of 1<<nr.
2272  *
2273  * It is used by cpumask_of() to get a constant address to a CPU
2274  * mask value that has a single bit set only.
2275  */
2276
2277 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2278 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2279 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2280 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2281 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2282
2283 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2284
2285         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2286         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2287 #if BITS_PER_LONG > 32
2288         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2289         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2290 #endif
2291 };
2292 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2293
2294 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2295 EXPORT_SYMBOL(cpu_all_bits);
2296
2297 #ifdef CONFIG_INIT_ALL_POSSIBLE
2298 struct cpumask __cpu_possible_mask __read_mostly
2299         = {CPU_BITS_ALL};
2300 #else
2301 struct cpumask __cpu_possible_mask __read_mostly;
2302 #endif
2303 EXPORT_SYMBOL(__cpu_possible_mask);
2304
2305 struct cpumask __cpu_online_mask __read_mostly;
2306 EXPORT_SYMBOL(__cpu_online_mask);
2307
2308 struct cpumask __cpu_present_mask __read_mostly;
2309 EXPORT_SYMBOL(__cpu_present_mask);
2310
2311 struct cpumask __cpu_active_mask __read_mostly;
2312 EXPORT_SYMBOL(__cpu_active_mask);
2313
2314 atomic_t __num_online_cpus __read_mostly;
2315 EXPORT_SYMBOL(__num_online_cpus);
2316
2317 void init_cpu_present(const struct cpumask *src)
2318 {
2319         cpumask_copy(&__cpu_present_mask, src);
2320 }
2321
2322 void init_cpu_possible(const struct cpumask *src)
2323 {
2324         cpumask_copy(&__cpu_possible_mask, src);
2325 }
2326
2327 void init_cpu_online(const struct cpumask *src)
2328 {
2329         cpumask_copy(&__cpu_online_mask, src);
2330 }
2331
2332 void set_cpu_online(unsigned int cpu, bool online)
2333 {
2334         /*
2335          * atomic_inc/dec() is required to handle the horrid abuse of this
2336          * function by the reboot and kexec code which invoke it from
2337          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2338          * regular CPU hotplug is properly serialized.
2339          *
2340          * Note, that the fact that __num_online_cpus is of type atomic_t
2341          * does not protect readers which are not serialized against
2342          * concurrent hotplug operations.
2343          */
2344         if (online) {
2345                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2346                         atomic_inc(&__num_online_cpus);
2347         } else {
2348                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2349                         atomic_dec(&__num_online_cpus);
2350         }
2351 }
2352
2353 /*
2354  * Activate the first processor.
2355  */
2356 void __init boot_cpu_init(void)
2357 {
2358         int cpu = smp_processor_id();
2359
2360         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2361         set_cpu_online(cpu, true);
2362         set_cpu_active(cpu, true);
2363         set_cpu_present(cpu, true);
2364         set_cpu_possible(cpu, true);
2365
2366 #ifdef CONFIG_SMP
2367         __boot_cpu_id = cpu;
2368 #endif
2369 }
2370
2371 /*
2372  * Must be called _AFTER_ setting up the per_cpu areas
2373  */
2374 void __init boot_cpu_hotplug_init(void)
2375 {
2376 #ifdef CONFIG_SMP
2377         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2378 #endif
2379         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2380 }
2381
2382 /*
2383  * These are used for a global "mitigations=" cmdline option for toggling
2384  * optional CPU mitigations.
2385  */
2386 enum cpu_mitigations {
2387         CPU_MITIGATIONS_OFF,
2388         CPU_MITIGATIONS_AUTO,
2389         CPU_MITIGATIONS_AUTO_NOSMT,
2390 };
2391
2392 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2393         CPU_MITIGATIONS_AUTO;
2394
2395 static int __init mitigations_parse_cmdline(char *arg)
2396 {
2397         if (!strcmp(arg, "off"))
2398                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2399         else if (!strcmp(arg, "auto"))
2400                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2401         else if (!strcmp(arg, "auto,nosmt"))
2402                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2403         else
2404                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2405                         arg);
2406
2407         return 0;
2408 }
2409 early_param("mitigations", mitigations_parse_cmdline);
2410
2411 /* mitigations=off */
2412 bool cpu_mitigations_off(void)
2413 {
2414         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2415 }
2416 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2417
2418 /* mitigations=auto,nosmt */
2419 bool cpu_mitigations_auto_nosmt(void)
2420 {
2421         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2422 }
2423 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);