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