kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/proc_fs.h>
   7 #include <linux/smp.h>
   8 #include <linux/init.h>
   9 #include <linux/notifier.h>
  10 #include <linux/sched/signal.h>
  11 #include <linux/sched/hotplug.h>
  12 #include <linux/sched/task.h>
  13 #include <linux/unistd.h>
  14 #include <linux/cpu.h>
  15 #include <linux/oom.h>
  16 #include <linux/rcupdate.h>
  17 #include <linux/export.h>
  18 #include <linux/bug.h>
  19 #include <linux/kthread.h>
  20 #include <linux/stop_machine.h>
  21 #include <linux/mutex.h>
  22 #include <linux/gfp.h>
  23 #include <linux/suspend.h>
  24 #include <linux/lockdep.h>
  25 #include <linux/tick.h>
  26 #include <linux/irq.h>
  27 #include <linux/nmi.h>
  28 #include <linux/smpboot.h>
  29 #include <linux/relay.h>
  30 #include <linux/slab.h>
  31 #include <linux/percpu-rwsem.h>
  32
  33 #include <trace/events/power.h>
  34 #define CREATE_TRACE_POINTS
  35 #include <trace/events/cpuhp.h>
  36
  37 #include "smpboot.h"
  38
  39 /**
  40  * cpuhp_cpu_state - Per cpu hotplug state storage
  41  * @state:      The current cpu state
  42  * @target:     The target state
  43  * @thread:     Pointer to the hotplug thread
  44  * @should_run: Thread should execute
  45  * @rollback:   Perform a rollback
  46  * @single:     Single callback invocation
  47  * @bringup:    Single callback bringup or teardown selector
  48  * @cb_state:   The state for a single callback (install/uninstall)
  49  * @result:     Result of the operation
  50  * @done_up:    Signal completion to the issuer of the task for cpu-up
  51  * @done_down:  Signal completion to the issuer of the task for cpu-down
  52  */
  53 struct cpuhp_cpu_state {
  54         enum cpuhp_state        state;
  55         enum cpuhp_state        target;
  56         enum cpuhp_state        fail;
  57 #ifdef CONFIG_SMP
  58         struct task_struct      *thread;
  59         bool                    should_run;
  60         bool                    rollback;
  61         bool                    single;
  62         bool                    bringup;
  63         bool                    booted_once;
  64         struct hlist_node       *node;
  65         struct hlist_node       *last;
  66         enum cpuhp_state        cb_state;
  67         int                     result;
  68         struct completion       done_up;
  69         struct completion       done_down;
  70 #endif
  71 };
  72
  73 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  74         .fail = CPUHP_INVALID,
  75 };
  76
  77 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  78 static struct lockdep_map cpuhp_state_up_map =
  79         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  80 static struct lockdep_map cpuhp_state_down_map =
  81         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  82
  83
  84 static inline void cpuhp_lock_acquire(bool bringup)
  85 {
  86         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  87 }
  88
  89 static inline void cpuhp_lock_release(bool bringup)
  90 {
  91         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  92 }
  93 #else
  94
  95 static inline void cpuhp_lock_acquire(bool bringup) { }
  96 static inline void cpuhp_lock_release(bool bringup) { }
  97
  98 #endif
  99
 100 /**
 101  * cpuhp_step - Hotplug state machine step
 102  * @name:       Name of the step
 103  * @startup:    Startup function of the step
 104  * @teardown:   Teardown function of the step
 105  * @cant_stop:  Bringup/teardown can't be stopped at this step
 106  */
 107 struct cpuhp_step {
 108         const char              *name;
 109         union {
 110                 int             (*single)(unsigned int cpu);
 111                 int             (*multi)(unsigned int cpu,
 112                                          struct hlist_node *node);
 113         } startup;
 114         union {
 115                 int             (*single)(unsigned int cpu);
 116                 int             (*multi)(unsigned int cpu,
 117                                          struct hlist_node *node);
 118         } teardown;
 119         struct hlist_head       list;
 120         bool                    cant_stop;
 121         bool                    multi_instance;
 122 };
 123
 124 static DEFINE_MUTEX(cpuhp_state_mutex);
 125 static struct cpuhp_step cpuhp_hp_states[];
 126
 127 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 128 {
 129         return cpuhp_hp_states + state;
 130 }
 131
 132 /**
 133  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 134  * @cpu:        The cpu for which the callback should be invoked
 135  * @state:      The state to do callbacks for
 136  * @bringup:    True if the bringup callback should be invoked
 137  * @node:       For multi-instance, do a single entry callback for install/remove
 138  * @lastp:      For multi-instance rollback, remember how far we got
 139  *
 140  * Called from cpu hotplug and from the state register machinery.
 141  */
 142 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 143                                  bool bringup, struct hlist_node *node,
 144                                  struct hlist_node **lastp)
 145 {
 146         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 147         struct cpuhp_step *step = cpuhp_get_step(state);
 148         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 149         int (*cb)(unsigned int cpu);
 150         int ret, cnt;
 151
 152         if (st->fail == state) {
 153                 st->fail = CPUHP_INVALID;
 154
 155                 if (!(bringup ? step->startup.single : step->teardown.single))
 156                         return 0;
 157
 158                 return -EAGAIN;
 159         }
 160
 161         if (!step->multi_instance) {
 162                 WARN_ON_ONCE(lastp && *lastp);
 163                 cb = bringup ? step->startup.single : step->teardown.single;
 164                 if (!cb)
 165                         return 0;
 166                 trace_cpuhp_enter(cpu, st->target, state, cb);
 167                 ret = cb(cpu);
 168                 trace_cpuhp_exit(cpu, st->state, state, ret);
 169                 return ret;
 170         }
 171         cbm = bringup ? step->startup.multi : step->teardown.multi;
 172         if (!cbm)
 173                 return 0;
 174
 175         /* Single invocation for instance add/remove */
 176         if (node) {
 177                 WARN_ON_ONCE(lastp && *lastp);
 178                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 179                 ret = cbm(cpu, node);
 180                 trace_cpuhp_exit(cpu, st->state, state, ret);
 181                 return ret;
 182         }
 183
 184         /* State transition. Invoke on all instances */
 185         cnt = 0;
 186         hlist_for_each(node, &step->list) {
 187                 if (lastp && node == *lastp)
 188                         break;
 189
 190                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 191                 ret = cbm(cpu, node);
 192                 trace_cpuhp_exit(cpu, st->state, state, ret);
 193                 if (ret) {
 194                         if (!lastp)
 195                                 goto err;
 196
 197                         *lastp = node;
 198                         return ret;
 199                 }
 200                 cnt++;
 201         }
 202         if (lastp)
 203                 *lastp = NULL;
 204         return 0;
 205 err:
 206         /* Rollback the instances if one failed */
 207         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 208         if (!cbm)
 209                 return ret;
 210
 211         hlist_for_each(node, &step->list) {
 212                 if (!cnt--)
 213                         break;
 214
 215                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 216                 ret = cbm(cpu, node);
 217                 trace_cpuhp_exit(cpu, st->state, state, ret);
 218                 /*
 219                  * Rollback must not fail,
 220                  */
 221                 WARN_ON_ONCE(ret);
 222         }
 223         return ret;
 224 }
 225
 226 #ifdef CONFIG_SMP
 227 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 228 {
 229         /*
 230          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 231          * purposes as that state is handled explicitly in cpu_down.
 232          */
 233         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 234 }
 235
 236 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 237 {
 238         struct completion *done = bringup ? &st->done_up : &st->done_down;
 239         wait_for_completion(done);
 240 }
 241
 242 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 243 {
 244         struct completion *done = bringup ? &st->done_up : &st->done_down;
 245         complete(done);
 246 }
 247
 248 /*
 249  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 250  */
 251 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 252 {
 253         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 254 }
 255
 256 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 257 static DEFINE_MUTEX(cpu_add_remove_lock);
 258 bool cpuhp_tasks_frozen;
 259 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 260
 261 /*
 262  * The following two APIs (cpu_maps_update_begin/done) must be used when
 263  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 264  */
 265 void cpu_maps_update_begin(void)
 266 {
 267         mutex_lock(&cpu_add_remove_lock);
 268 }
 269
 270 void cpu_maps_update_done(void)
 271 {
 272         mutex_unlock(&cpu_add_remove_lock);
 273 }
 274
 275 /*
 276  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 277  * Should always be manipulated under cpu_add_remove_lock
 278  */
 279 static int cpu_hotplug_disabled;
 280
 281 #ifdef CONFIG_HOTPLUG_CPU
 282
 283 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 284
 285 void cpus_read_lock(void)
 286 {
 287         percpu_down_read(&cpu_hotplug_lock);
 288 }
 289 EXPORT_SYMBOL_GPL(cpus_read_lock);
 290
 291 int cpus_read_trylock(void)
 292 {
 293         return percpu_down_read_trylock(&cpu_hotplug_lock);
 294 }
 295 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 296
 297 void cpus_read_unlock(void)
 298 {
 299         percpu_up_read(&cpu_hotplug_lock);
 300 }
 301 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 302
 303 void cpus_write_lock(void)
 304 {
 305         percpu_down_write(&cpu_hotplug_lock);
 306 }
 307
 308 void cpus_write_unlock(void)
 309 {
 310         percpu_up_write(&cpu_hotplug_lock);
 311 }
 312
 313 void lockdep_assert_cpus_held(void)
 314 {
 315         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 316 }
 317
 318 /*
 319  * Wait for currently running CPU hotplug operations to complete (if any) and
 320  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 321  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 322  * hotplug path before performing hotplug operations. So acquiring that lock
 323  * guarantees mutual exclusion from any currently running hotplug operations.
 324  */
 325 void cpu_hotplug_disable(void)
 326 {
 327         cpu_maps_update_begin();
 328         cpu_hotplug_disabled++;
 329         cpu_maps_update_done();
 330 }
 331 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 332
 333 static void __cpu_hotplug_enable(void)
 334 {
 335         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 336                 return;
 337         cpu_hotplug_disabled--;
 338 }
 339
 340 void cpu_hotplug_enable(void)
 341 {
 342         cpu_maps_update_begin();
 343         __cpu_hotplug_enable();
 344         cpu_maps_update_done();
 345 }
 346 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 347 #endif  /* CONFIG_HOTPLUG_CPU */
 348
 349 #ifdef CONFIG_HOTPLUG_SMT
 350 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 351 EXPORT_SYMBOL_GPL(cpu_smt_control);
 352
 353 static bool cpu_smt_available __read_mostly;
 354
 355 void __init cpu_smt_disable(bool force)
 356 {
 357         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 358                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 359                 return;
 360
 361         if (force) {
 362                 pr_info("SMT: Force disabled\n");
 363                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 364         } else {
 365                 cpu_smt_control = CPU_SMT_DISABLED;
 366         }
 367 }
 368
 369 /*
 370  * The decision whether SMT is supported can only be done after the full
 371  * CPU identification. Called from architecture code before non boot CPUs
 372  * are brought up.
 373  */
 374 void __init cpu_smt_check_topology_early(void)
 375 {
 376         if (!topology_smt_supported())
 377                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 378 }
 379
 380 /*
 381  * If SMT was disabled by BIOS, detect it here, after the CPUs have been
 382  * brought online. This ensures the smt/l1tf sysfs entries are consistent
 383  * with reality. cpu_smt_available is set to true during the bringup of non
 384  * boot CPUs when a SMT sibling is detected. Note, this may overwrite
 385  * cpu_smt_control's previous setting.
 386  */
 387 void __init cpu_smt_check_topology(void)
 388 {
 389         if (!cpu_smt_available)
 390                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 391 }
 392
 393 static int __init smt_cmdline_disable(char *str)
 394 {
 395         cpu_smt_disable(str && !strcmp(str, "force"));
 396         return 0;
 397 }
 398 early_param("nosmt", smt_cmdline_disable);
 399
 400 static inline bool cpu_smt_allowed(unsigned int cpu)
 401 {
 402         if (topology_is_primary_thread(cpu))
 403                 return true;
 404
 405         /*
 406          * If the CPU is not a 'primary' thread and the booted_once bit is
 407          * set then the processor has SMT support. Store this information
 408          * for the late check of SMT support in cpu_smt_check_topology().
 409          */
 410         if (per_cpu(cpuhp_state, cpu).booted_once)
 411                 cpu_smt_available = true;
 412
 413         if (cpu_smt_control == CPU_SMT_ENABLED)
 414                 return true;
 415
 416         /*
 417          * On x86 it's required to boot all logical CPUs at least once so
 418          * that the init code can get a chance to set CR4.MCE on each
 419          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 420          * core will shutdown the machine.
 421          */
 422         return !per_cpu(cpuhp_state, cpu).booted_once;
 423 }
 424 #else
 425 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 426 #endif
 427
 428 static inline enum cpuhp_state
 429 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 430 {
 431         enum cpuhp_state prev_state = st->state;
 432
 433         st->rollback = false;
 434         st->last = NULL;
 435
 436         st->target = target;
 437         st->single = false;
 438         st->bringup = st->state < target;
 439
 440         return prev_state;
 441 }
 442
 443 static inline void
 444 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 445 {
 446         st->rollback = true;
 447
 448         /*
 449          * If we have st->last we need to undo partial multi_instance of this
 450          * state first. Otherwise start undo at the previous state.
 451          */
 452         if (!st->last) {
 453                 if (st->bringup)
 454                         st->state--;
 455                 else
 456                         st->state++;
 457         }
 458
 459         st->target = prev_state;
 460         st->bringup = !st->bringup;
 461 }
 462
 463 /* Regular hotplug invocation of the AP hotplug thread */
 464 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 465 {
 466         if (!st->single && st->state == st->target)
 467                 return;
 468
 469         st->result = 0;
 470         /*
 471          * Make sure the above stores are visible before should_run becomes
 472          * true. Paired with the mb() above in cpuhp_thread_fun()
 473          */
 474         smp_mb();
 475         st->should_run = true;
 476         wake_up_process(st->thread);
 477         wait_for_ap_thread(st, st->bringup);
 478 }
 479
 480 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 481 {
 482         enum cpuhp_state prev_state;
 483         int ret;
 484
 485         prev_state = cpuhp_set_state(st, target);
 486         __cpuhp_kick_ap(st);
 487         if ((ret = st->result)) {
 488                 cpuhp_reset_state(st, prev_state);
 489                 __cpuhp_kick_ap(st);
 490         }
 491
 492         return ret;
 493 }
 494
 495 static int bringup_wait_for_ap(unsigned int cpu)
 496 {
 497         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 498
 499         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 500         wait_for_ap_thread(st, true);
 501         if (WARN_ON_ONCE((!cpu_online(cpu))))
 502                 return -ECANCELED;
 503
 504         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 505         stop_machine_unpark(cpu);
 506         kthread_unpark(st->thread);
 507
 508         /*
 509          * SMT soft disabling on X86 requires to bring the CPU out of the
 510          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 511          * CPU marked itself as booted_once in cpu_notify_starting() so the
 512          * cpu_smt_allowed() check will now return false if this is not the
 513          * primary sibling.
 514          */
 515         if (!cpu_smt_allowed(cpu))
 516                 return -ECANCELED;
 517
 518         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 519                 return 0;
 520
 521         return cpuhp_kick_ap(st, st->target);
 522 }
 523
 524 static int bringup_cpu(unsigned int cpu)
 525 {
 526         struct task_struct *idle = idle_thread_get(cpu);
 527         int ret;
 528
 529         /*
 530          * Some architectures have to walk the irq descriptors to
 531          * setup the vector space for the cpu which comes online.
 532          * Prevent irq alloc/free across the bringup.
 533          */
 534         irq_lock_sparse();
 535
 536         /* Arch-specific enabling code. */
 537         ret = __cpu_up(cpu, idle);
 538         irq_unlock_sparse();
 539         if (ret)
 540                 return ret;
 541         return bringup_wait_for_ap(cpu);
 542 }
 543
 544 /*
 545  * Hotplug state machine related functions
 546  */
 547
 548 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 549 {
 550         for (st->state--; st->state > st->target; st->state--)
 551                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 552 }
 553
 554 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 555                               enum cpuhp_state target)
 556 {
 557         enum cpuhp_state prev_state = st->state;
 558         int ret = 0;
 559
 560         while (st->state < target) {
 561                 st->state++;
 562                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 563                 if (ret) {
 564                         st->target = prev_state;
 565                         undo_cpu_up(cpu, st);
 566                         break;
 567                 }
 568         }
 569         return ret;
 570 }
 571
 572 /*
 573  * The cpu hotplug threads manage the bringup and teardown of the cpus
 574  */
 575 static void cpuhp_create(unsigned int cpu)
 576 {
 577         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 578
 579         init_completion(&st->done_up);
 580         init_completion(&st->done_down);
 581 }
 582
 583 static int cpuhp_should_run(unsigned int cpu)
 584 {
 585         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 586
 587         return st->should_run;
 588 }
 589
 590 /*
 591  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 592  * callbacks when a state gets [un]installed at runtime.
 593  *
 594  * Each invocation of this function by the smpboot thread does a single AP
 595  * state callback.
 596  *
 597  * It has 3 modes of operation:
 598  *  - single: runs st->cb_state
 599  *  - up:     runs ++st->state, while st->state < st->target
 600  *  - down:   runs st->state--, while st->state > st->target
 601  *
 602  * When complete or on error, should_run is cleared and the completion is fired.
 603  */
 604 static void cpuhp_thread_fun(unsigned int cpu)
 605 {
 606         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 607         bool bringup = st->bringup;
 608         enum cpuhp_state state;
 609
 610         /*
 611          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 612          * that if we see ->should_run we also see the rest of the state.
 613          */
 614         smp_mb();
 615
 616         if (WARN_ON_ONCE(!st->should_run))
 617                 return;
 618
 619         cpuhp_lock_acquire(bringup);
 620
 621         if (st->single) {
 622                 state = st->cb_state;
 623                 st->should_run = false;
 624         } else {
 625                 if (bringup) {
 626                         st->state++;
 627                         state = st->state;
 628                         st->should_run = (st->state < st->target);
 629                         WARN_ON_ONCE(st->state > st->target);
 630                 } else {
 631                         state = st->state;
 632                         st->state--;
 633                         st->should_run = (st->state > st->target);
 634                         WARN_ON_ONCE(st->state < st->target);
 635                 }
 636         }
 637
 638         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 639
 640         if (cpuhp_is_atomic_state(state)) {
 641                 local_irq_disable();
 642                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 643                 local_irq_enable();
 644
 645                 /*
 646                  * STARTING/DYING must not fail!
 647                  */
 648                 WARN_ON_ONCE(st->result);
 649         } else {
 650                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 651         }
 652
 653         if (st->result) {
 654                 /*
 655                  * If we fail on a rollback, we're up a creek without no
 656                  * paddle, no way forward, no way back. We loose, thanks for
 657                  * playing.
 658                  */
 659                 WARN_ON_ONCE(st->rollback);
 660                 st->should_run = false;
 661         }
 662
 663         cpuhp_lock_release(bringup);
 664
 665         if (!st->should_run)
 666                 complete_ap_thread(st, bringup);
 667 }
 668
 669 /* Invoke a single callback on a remote cpu */
 670 static int
 671 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 672                          struct hlist_node *node)
 673 {
 674         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 675         int ret;
 676
 677         if (!cpu_online(cpu))
 678                 return 0;
 679
 680         cpuhp_lock_acquire(false);
 681         cpuhp_lock_release(false);
 682
 683         cpuhp_lock_acquire(true);
 684         cpuhp_lock_release(true);
 685
 686         /*
 687          * If we are up and running, use the hotplug thread. For early calls
 688          * we invoke the thread function directly.
 689          */
 690         if (!st->thread)
 691                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 692
 693         st->rollback = false;
 694         st->last = NULL;
 695
 696         st->node = node;
 697         st->bringup = bringup;
 698         st->cb_state = state;
 699         st->single = true;
 700
 701         __cpuhp_kick_ap(st);
 702
 703         /*
 704          * If we failed and did a partial, do a rollback.
 705          */
 706         if ((ret = st->result) && st->last) {
 707                 st->rollback = true;
 708                 st->bringup = !bringup;
 709
 710                 __cpuhp_kick_ap(st);
 711         }
 712
 713         /*
 714          * Clean up the leftovers so the next hotplug operation wont use stale
 715          * data.
 716          */
 717         st->node = st->last = NULL;
 718         return ret;
 719 }
 720
 721 static int cpuhp_kick_ap_work(unsigned int cpu)
 722 {
 723         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 724         enum cpuhp_state prev_state = st->state;
 725         int ret;
 726
 727         cpuhp_lock_acquire(false);
 728         cpuhp_lock_release(false);
 729
 730         cpuhp_lock_acquire(true);
 731         cpuhp_lock_release(true);
 732
 733         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 734         ret = cpuhp_kick_ap(st, st->target);
 735         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 736
 737         return ret;
 738 }
 739
 740 static struct smp_hotplug_thread cpuhp_threads = {
 741         .store                  = &cpuhp_state.thread,
 742         .create                 = &cpuhp_create,
 743         .thread_should_run      = cpuhp_should_run,
 744         .thread_fn              = cpuhp_thread_fun,
 745         .thread_comm            = "cpuhp/%u",
 746         .selfparking            = true,
 747 };
 748
 749 void __init cpuhp_threads_init(void)
 750 {
 751         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 752         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 753 }
 754
 755 #ifdef CONFIG_HOTPLUG_CPU
 756 /**
 757  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 758  * @cpu: a CPU id
 759  *
 760  * This function walks all processes, finds a valid mm struct for each one and
 761  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 762  * trivial, there are various non-obvious corner cases, which this function
 763  * tries to solve in a safe manner.
 764  *
 765  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 766  * be called only for an already offlined CPU.
 767  */
 768 void clear_tasks_mm_cpumask(int cpu)
 769 {
 770         struct task_struct *p;
 771
 772         /*
 773          * This function is called after the cpu is taken down and marked
 774          * offline, so its not like new tasks will ever get this cpu set in
 775          * their mm mask. -- Peter Zijlstra
 776          * Thus, we may use rcu_read_lock() here, instead of grabbing
 777          * full-fledged tasklist_lock.
 778          */
 779         WARN_ON(cpu_online(cpu));
 780         rcu_read_lock();
 781         for_each_process(p) {
 782                 struct task_struct *t;
 783
 784                 /*
 785                  * Main thread might exit, but other threads may still have
 786                  * a valid mm. Find one.
 787                  */
 788                 t = find_lock_task_mm(p);
 789                 if (!t)
 790                         continue;
 791                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 792                 task_unlock(t);
 793         }
 794         rcu_read_unlock();
 795 }
 796
 797 /* Take this CPU down. */
 798 static int take_cpu_down(void *_param)
 799 {
 800         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 801         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 802         int err, cpu = smp_processor_id();
 803         int ret;
 804
 805         /* Ensure this CPU doesn't handle any more interrupts. */
 806         err = __cpu_disable();
 807         if (err < 0)
 808                 return err;
 809
 810         /*
 811          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 812          * do this step again.
 813          */
 814         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 815         st->state--;
 816         /* Invoke the former CPU_DYING callbacks */
 817         for (; st->state > target; st->state--) {
 818                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 819                 /*
 820                  * DYING must not fail!
 821                  */
 822                 WARN_ON_ONCE(ret);
 823         }
 824
 825         /* Give up timekeeping duties */
 826         tick_handover_do_timer();
 827         /* Park the stopper thread */
 828         stop_machine_park(cpu);
 829         return 0;
 830 }
 831
 832 static int takedown_cpu(unsigned int cpu)
 833 {
 834         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 835         int err;
 836
 837         /* Park the smpboot threads */
 838         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 839
 840         /*
 841          * Prevent irq alloc/free while the dying cpu reorganizes the
 842          * interrupt affinities.
 843          */
 844         irq_lock_sparse();
 845
 846         /*
 847          * So now all preempt/rcu users must observe !cpu_active().
 848          */
 849         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 850         if (err) {
 851                 /* CPU refused to die */
 852                 irq_unlock_sparse();
 853                 /* Unpark the hotplug thread so we can rollback there */
 854                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 855                 return err;
 856         }
 857         BUG_ON(cpu_online(cpu));
 858
 859         /*
 860          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 861          * all runnable tasks from the CPU, there's only the idle task left now
 862          * that the migration thread is done doing the stop_machine thing.
 863          *
 864          * Wait for the stop thread to go away.
 865          */
 866         wait_for_ap_thread(st, false);
 867         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 868
 869         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 870         irq_unlock_sparse();
 871
 872         hotplug_cpu__broadcast_tick_pull(cpu);
 873         /* This actually kills the CPU. */
 874         __cpu_die(cpu);
 875
 876         tick_cleanup_dead_cpu(cpu);
 877         rcutree_migrate_callbacks(cpu);
 878         return 0;
 879 }
 880
 881 static void cpuhp_complete_idle_dead(void *arg)
 882 {
 883         struct cpuhp_cpu_state *st = arg;
 884
 885         complete_ap_thread(st, false);
 886 }
 887
 888 void cpuhp_report_idle_dead(void)
 889 {
 890         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 891
 892         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 893         rcu_report_dead(smp_processor_id());
 894         st->state = CPUHP_AP_IDLE_DEAD;
 895         /*
 896          * We cannot call complete after rcu_report_dead() so we delegate it
 897          * to an online cpu.
 898          */
 899         smp_call_function_single(cpumask_first(cpu_online_mask),
 900                                  cpuhp_complete_idle_dead, st, 0);
 901 }
 902
 903 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 904 {
 905         for (st->state++; st->state < st->target; st->state++)
 906                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 907 }
 908
 909 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 910                                 enum cpuhp_state target)
 911 {
 912         enum cpuhp_state prev_state = st->state;
 913         int ret = 0;
 914
 915         for (; st->state > target; st->state--) {
 916                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 917                 if (ret) {
 918                         st->target = prev_state;
 919                         undo_cpu_down(cpu, st);
 920                         break;
 921                 }
 922         }
 923         return ret;
 924 }
 925
 926 /* Requires cpu_add_remove_lock to be held */
 927 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 928                            enum cpuhp_state target)
 929 {
 930         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 931         int prev_state, ret = 0;
 932
 933         if (num_online_cpus() == 1)
 934                 return -EBUSY;
 935
 936         if (!cpu_present(cpu))
 937                 return -EINVAL;
 938
 939         cpus_write_lock();
 940
 941         cpuhp_tasks_frozen = tasks_frozen;
 942
 943         prev_state = cpuhp_set_state(st, target);
 944         /*
 945          * If the current CPU state is in the range of the AP hotplug thread,
 946          * then we need to kick the thread.
 947          */
 948         if (st->state > CPUHP_TEARDOWN_CPU) {
 949                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 950                 ret = cpuhp_kick_ap_work(cpu);
 951                 /*
 952                  * The AP side has done the error rollback already. Just
 953                  * return the error code..
 954                  */
 955                 if (ret)
 956                         goto out;
 957
 958                 /*
 959                  * We might have stopped still in the range of the AP hotplug
 960                  * thread. Nothing to do anymore.
 961                  */
 962                 if (st->state > CPUHP_TEARDOWN_CPU)
 963                         goto out;
 964
 965                 st->target = target;
 966         }
 967         /*
 968          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 969          * to do the further cleanups.
 970          */
 971         ret = cpuhp_down_callbacks(cpu, st, target);
 972         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 973                 cpuhp_reset_state(st, prev_state);
 974                 __cpuhp_kick_ap(st);
 975         }
 976
 977 out:
 978         cpus_write_unlock();
 979         /*
 980          * Do post unplug cleanup. This is still protected against
 981          * concurrent CPU hotplug via cpu_add_remove_lock.
 982          */
 983         lockup_detector_cleanup();
 984         return ret;
 985 }
 986
 987 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
 988 {
 989         if (cpu_hotplug_disabled)
 990                 return -EBUSY;
 991         return _cpu_down(cpu, 0, target);
 992 }
 993
 994 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 995 {
 996         int err;
 997
 998         cpu_maps_update_begin();
 999         err = cpu_down_maps_locked(cpu, target);
1000         cpu_maps_update_done();
1001         return err;
1002 }
1003
1004 int cpu_down(unsigned int cpu)
1005 {
1006         return do_cpu_down(cpu, CPUHP_OFFLINE);
1007 }
1008 EXPORT_SYMBOL(cpu_down);
1009
1010 #else
1011 #define takedown_cpu            NULL
1012 #endif /*CONFIG_HOTPLUG_CPU*/
1013
1014 /**
1015  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1016  * @cpu: cpu that just started
1017  *
1018  * It must be called by the arch code on the new cpu, before the new cpu
1019  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1020  */
1021 void notify_cpu_starting(unsigned int cpu)
1022 {
1023         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1024         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1025         int ret;
1026
1027         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1028         st->booted_once = true;
1029         while (st->state < target) {
1030                 st->state++;
1031                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1032                 /*
1033                  * STARTING must not fail!
1034                  */
1035                 WARN_ON_ONCE(ret);
1036         }
1037 }
1038
1039 /*
1040  * Called from the idle task. Wake up the controlling task which brings the
1041  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1042  * the rest of the online bringup to the hotplug thread.
1043  */
1044 void cpuhp_online_idle(enum cpuhp_state state)
1045 {
1046         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1047
1048         /* Happens for the boot cpu */
1049         if (state != CPUHP_AP_ONLINE_IDLE)
1050                 return;
1051
1052         st->state = CPUHP_AP_ONLINE_IDLE;
1053         complete_ap_thread(st, true);
1054 }
1055
1056 /* Requires cpu_add_remove_lock to be held */
1057 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1058 {
1059         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1060         struct task_struct *idle;
1061         int ret = 0;
1062
1063         cpus_write_lock();
1064
1065         if (!cpu_present(cpu)) {
1066                 ret = -EINVAL;
1067                 goto out;
1068         }
1069
1070         /*
1071          * The caller of do_cpu_up might have raced with another
1072          * caller. Ignore it for now.
1073          */
1074         if (st->state >= target)
1075                 goto out;
1076
1077         if (st->state == CPUHP_OFFLINE) {
1078                 /* Let it fail before we try to bring the cpu up */
1079                 idle = idle_thread_get(cpu);
1080                 if (IS_ERR(idle)) {
1081                         ret = PTR_ERR(idle);
1082                         goto out;
1083                 }
1084         }
1085
1086         cpuhp_tasks_frozen = tasks_frozen;
1087
1088         cpuhp_set_state(st, target);
1089         /*
1090          * If the current CPU state is in the range of the AP hotplug thread,
1091          * then we need to kick the thread once more.
1092          */
1093         if (st->state > CPUHP_BRINGUP_CPU) {
1094                 ret = cpuhp_kick_ap_work(cpu);
1095                 /*
1096                  * The AP side has done the error rollback already. Just
1097                  * return the error code..
1098                  */
1099                 if (ret)
1100                         goto out;
1101         }
1102
1103         /*
1104          * Try to reach the target state. We max out on the BP at
1105          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1106          * responsible for bringing it up to the target state.
1107          */
1108         target = min((int)target, CPUHP_BRINGUP_CPU);
1109         ret = cpuhp_up_callbacks(cpu, st, target);
1110 out:
1111         cpus_write_unlock();
1112         return ret;
1113 }
1114
1115 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1116 {
1117         int err = 0;
1118
1119         if (!cpu_possible(cpu)) {
1120                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1121                        cpu);
1122 #if defined(CONFIG_IA64)
1123                 pr_err("please check additional_cpus= boot parameter\n");
1124 #endif
1125                 return -EINVAL;
1126         }
1127
1128         err = try_online_node(cpu_to_node(cpu));
1129         if (err)
1130                 return err;
1131
1132         cpu_maps_update_begin();
1133
1134         if (cpu_hotplug_disabled) {
1135                 err = -EBUSY;
1136                 goto out;
1137         }
1138         if (!cpu_smt_allowed(cpu)) {
1139                 err = -EPERM;
1140                 goto out;
1141         }
1142
1143         err = _cpu_up(cpu, 0, target);
1144 out:
1145         cpu_maps_update_done();
1146         return err;
1147 }
1148
1149 int cpu_up(unsigned int cpu)
1150 {
1151         return do_cpu_up(cpu, CPUHP_ONLINE);
1152 }
1153 EXPORT_SYMBOL_GPL(cpu_up);
1154
1155 #ifdef CONFIG_PM_SLEEP_SMP
1156 static cpumask_var_t frozen_cpus;
1157
1158 int freeze_secondary_cpus(int primary)
1159 {
1160         int cpu, error = 0;
1161
1162         cpu_maps_update_begin();
1163         if (!cpu_online(primary))
1164                 primary = cpumask_first(cpu_online_mask);
1165         /*
1166          * We take down all of the non-boot CPUs in one shot to avoid races
1167          * with the userspace trying to use the CPU hotplug at the same time
1168          */
1169         cpumask_clear(frozen_cpus);
1170
1171         pr_info("Disabling non-boot CPUs ...\n");
1172         for_each_online_cpu(cpu) {
1173                 if (cpu == primary)
1174                         continue;
1175                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1176                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1177                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1178                 if (!error)
1179                         cpumask_set_cpu(cpu, frozen_cpus);
1180                 else {
1181                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1182                         break;
1183                 }
1184         }
1185
1186         if (!error)
1187                 BUG_ON(num_online_cpus() > 1);
1188         else
1189                 pr_err("Non-boot CPUs are not disabled\n");
1190
1191         /*
1192          * Make sure the CPUs won't be enabled by someone else. We need to do
1193          * this even in case of failure as all disable_nonboot_cpus() users are
1194          * supposed to do enable_nonboot_cpus() on the failure path.
1195          */
1196         cpu_hotplug_disabled++;
1197
1198         cpu_maps_update_done();
1199         return error;
1200 }
1201
1202 void __weak arch_enable_nonboot_cpus_begin(void)
1203 {
1204 }
1205
1206 void __weak arch_enable_nonboot_cpus_end(void)
1207 {
1208 }
1209
1210 void enable_nonboot_cpus(void)
1211 {
1212         int cpu, error;
1213
1214         /* Allow everyone to use the CPU hotplug again */
1215         cpu_maps_update_begin();
1216         __cpu_hotplug_enable();
1217         if (cpumask_empty(frozen_cpus))
1218                 goto out;
1219
1220         pr_info("Enabling non-boot CPUs ...\n");
1221
1222         arch_enable_nonboot_cpus_begin();
1223
1224         for_each_cpu(cpu, frozen_cpus) {
1225                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1226                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1227                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1228                 if (!error) {
1229                         pr_info("CPU%d is up\n", cpu);
1230                         continue;
1231                 }
1232                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1233         }
1234
1235         arch_enable_nonboot_cpus_end();
1236
1237         cpumask_clear(frozen_cpus);
1238 out:
1239         cpu_maps_update_done();
1240 }
1241
1242 static int __init alloc_frozen_cpus(void)
1243 {
1244         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1245                 return -ENOMEM;
1246         return 0;
1247 }
1248 core_initcall(alloc_frozen_cpus);
1249
1250 /*
1251  * When callbacks for CPU hotplug notifications are being executed, we must
1252  * ensure that the state of the system with respect to the tasks being frozen
1253  * or not, as reported by the notification, remains unchanged *throughout the
1254  * duration* of the execution of the callbacks.
1255  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1256  *
1257  * This synchronization is implemented by mutually excluding regular CPU
1258  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1259  * Hibernate notifications.
1260  */
1261 static int
1262 cpu_hotplug_pm_callback(struct notifier_block *nb,
1263                         unsigned long action, void *ptr)
1264 {
1265         switch (action) {
1266
1267         case PM_SUSPEND_PREPARE:
1268         case PM_HIBERNATION_PREPARE:
1269                 cpu_hotplug_disable();
1270                 break;
1271
1272         case PM_POST_SUSPEND:
1273         case PM_POST_HIBERNATION:
1274                 cpu_hotplug_enable();
1275                 break;
1276
1277         default:
1278                 return NOTIFY_DONE;
1279         }
1280
1281         return NOTIFY_OK;
1282 }
1283
1284
1285 static int __init cpu_hotplug_pm_sync_init(void)
1286 {
1287         /*
1288          * cpu_hotplug_pm_callback has higher priority than x86
1289          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1290          * to disable cpu hotplug to avoid cpu hotplug race.
1291          */
1292         pm_notifier(cpu_hotplug_pm_callback, 0);
1293         return 0;
1294 }
1295 core_initcall(cpu_hotplug_pm_sync_init);
1296
1297 #endif /* CONFIG_PM_SLEEP_SMP */
1298
1299 int __boot_cpu_id;
1300
1301 #endif /* CONFIG_SMP */
1302
1303 /* Boot processor state steps */
1304 static struct cpuhp_step cpuhp_hp_states[] = {
1305         [CPUHP_OFFLINE] = {
1306                 .name                   = "offline",
1307                 .startup.single         = NULL,
1308                 .teardown.single        = NULL,
1309         },
1310 #ifdef CONFIG_SMP
1311         [CPUHP_CREATE_THREADS]= {
1312                 .name                   = "threads:prepare",
1313                 .startup.single         = smpboot_create_threads,
1314                 .teardown.single        = NULL,
1315                 .cant_stop              = true,
1316         },
1317         [CPUHP_PERF_PREPARE] = {
1318                 .name                   = "perf:prepare",
1319                 .startup.single         = perf_event_init_cpu,
1320                 .teardown.single        = perf_event_exit_cpu,
1321         },
1322         [CPUHP_WORKQUEUE_PREP] = {
1323                 .name                   = "workqueue:prepare",
1324                 .startup.single         = workqueue_prepare_cpu,
1325                 .teardown.single        = NULL,
1326         },
1327         [CPUHP_HRTIMERS_PREPARE] = {
1328                 .name                   = "hrtimers:prepare",
1329                 .startup.single         = hrtimers_prepare_cpu,
1330                 .teardown.single        = hrtimers_dead_cpu,
1331         },
1332         [CPUHP_SMPCFD_PREPARE] = {
1333                 .name                   = "smpcfd:prepare",
1334                 .startup.single         = smpcfd_prepare_cpu,
1335                 .teardown.single        = smpcfd_dead_cpu,
1336         },
1337         [CPUHP_RELAY_PREPARE] = {
1338                 .name                   = "relay:prepare",
1339                 .startup.single         = relay_prepare_cpu,
1340                 .teardown.single        = NULL,
1341         },
1342         [CPUHP_SLAB_PREPARE] = {
1343                 .name                   = "slab:prepare",
1344                 .startup.single         = slab_prepare_cpu,
1345                 .teardown.single        = slab_dead_cpu,
1346         },
1347         [CPUHP_RCUTREE_PREP] = {
1348                 .name                   = "RCU/tree:prepare",
1349                 .startup.single         = rcutree_prepare_cpu,
1350                 .teardown.single        = rcutree_dead_cpu,
1351         },
1352         /*
1353          * On the tear-down path, timers_dead_cpu() must be invoked
1354          * before blk_mq_queue_reinit_notify() from notify_dead(),
1355          * otherwise a RCU stall occurs.
1356          */
1357         [CPUHP_TIMERS_PREPARE] = {
1358                 .name                   = "timers:prepare",
1359                 .startup.single         = timers_prepare_cpu,
1360                 .teardown.single        = timers_dead_cpu,
1361         },
1362         /* Kicks the plugged cpu into life */
1363         [CPUHP_BRINGUP_CPU] = {
1364                 .name                   = "cpu:bringup",
1365                 .startup.single         = bringup_cpu,
1366                 .teardown.single        = NULL,
1367                 .cant_stop              = true,
1368         },
1369         /* Final state before CPU kills itself */
1370         [CPUHP_AP_IDLE_DEAD] = {
1371                 .name                   = "idle:dead",
1372         },
1373         /*
1374          * Last state before CPU enters the idle loop to die. Transient state
1375          * for synchronization.
1376          */
1377         [CPUHP_AP_OFFLINE] = {
1378                 .name                   = "ap:offline",
1379                 .cant_stop              = true,
1380         },
1381         /* First state is scheduler control. Interrupts are disabled */
1382         [CPUHP_AP_SCHED_STARTING] = {
1383                 .name                   = "sched:starting",
1384                 .startup.single         = sched_cpu_starting,
1385                 .teardown.single        = sched_cpu_dying,
1386         },
1387         [CPUHP_AP_RCUTREE_DYING] = {
1388                 .name                   = "RCU/tree:dying",
1389                 .startup.single         = NULL,
1390                 .teardown.single        = rcutree_dying_cpu,
1391         },
1392         [CPUHP_AP_SMPCFD_DYING] = {
1393                 .name                   = "smpcfd:dying",
1394                 .startup.single         = NULL,
1395                 .teardown.single        = smpcfd_dying_cpu,
1396         },
1397         /* Entry state on starting. Interrupts enabled from here on. Transient
1398          * state for synchronsization */
1399         [CPUHP_AP_ONLINE] = {
1400                 .name                   = "ap:online",
1401         },
1402         /*
1403          * Handled on controll processor until the plugged processor manages
1404          * this itself.
1405          */
1406         [CPUHP_TEARDOWN_CPU] = {
1407                 .name                   = "cpu:teardown",
1408                 .startup.single         = NULL,
1409                 .teardown.single        = takedown_cpu,
1410                 .cant_stop              = true,
1411         },
1412         /* Handle smpboot threads park/unpark */
1413         [CPUHP_AP_SMPBOOT_THREADS] = {
1414                 .name                   = "smpboot/threads:online",
1415                 .startup.single         = smpboot_unpark_threads,
1416                 .teardown.single        = smpboot_park_threads,
1417         },
1418         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1419                 .name                   = "irq/affinity:online",
1420                 .startup.single         = irq_affinity_online_cpu,
1421                 .teardown.single        = NULL,
1422         },
1423         [CPUHP_AP_PERF_ONLINE] = {
1424                 .name                   = "perf:online",
1425                 .startup.single         = perf_event_init_cpu,
1426                 .teardown.single        = perf_event_exit_cpu,
1427         },
1428         [CPUHP_AP_WATCHDOG_ONLINE] = {
1429                 .name                   = "lockup_detector:online",
1430                 .startup.single         = lockup_detector_online_cpu,
1431                 .teardown.single        = lockup_detector_offline_cpu,
1432         },
1433         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1434                 .name                   = "workqueue:online",
1435                 .startup.single         = workqueue_online_cpu,
1436                 .teardown.single        = workqueue_offline_cpu,
1437         },
1438         [CPUHP_AP_RCUTREE_ONLINE] = {
1439                 .name                   = "RCU/tree:online",
1440                 .startup.single         = rcutree_online_cpu,
1441                 .teardown.single        = rcutree_offline_cpu,
1442         },
1443 #endif
1444         /*
1445          * The dynamically registered state space is here
1446          */
1447
1448 #ifdef CONFIG_SMP
1449         /* Last state is scheduler control setting the cpu active */
1450         [CPUHP_AP_ACTIVE] = {
1451                 .name                   = "sched:active",
1452                 .startup.single         = sched_cpu_activate,
1453                 .teardown.single        = sched_cpu_deactivate,
1454         },
1455 #endif
1456
1457         /* CPU is fully up and running. */
1458         [CPUHP_ONLINE] = {
1459                 .name                   = "online",
1460                 .startup.single         = NULL,
1461                 .teardown.single        = NULL,
1462         },
1463 };
1464
1465 /* Sanity check for callbacks */
1466 static int cpuhp_cb_check(enum cpuhp_state state)
1467 {
1468         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1469                 return -EINVAL;
1470         return 0;
1471 }
1472
1473 /*
1474  * Returns a free for dynamic slot assignment of the Online state. The states
1475  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1476  * by having no name assigned.
1477  */
1478 static int cpuhp_reserve_state(enum cpuhp_state state)
1479 {
1480         enum cpuhp_state i, end;
1481         struct cpuhp_step *step;
1482
1483         switch (state) {
1484         case CPUHP_AP_ONLINE_DYN:
1485                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1486                 end = CPUHP_AP_ONLINE_DYN_END;
1487                 break;
1488         case CPUHP_BP_PREPARE_DYN:
1489                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1490                 end = CPUHP_BP_PREPARE_DYN_END;
1491                 break;
1492         default:
1493                 return -EINVAL;
1494         }
1495
1496         for (i = state; i <= end; i++, step++) {
1497                 if (!step->name)
1498                         return i;
1499         }
1500         WARN(1, "No more dynamic states available for CPU hotplug\n");
1501         return -ENOSPC;
1502 }
1503
1504 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1505                                  int (*startup)(unsigned int cpu),
1506                                  int (*teardown)(unsigned int cpu),
1507                                  bool multi_instance)
1508 {
1509         /* (Un)Install the callbacks for further cpu hotplug operations */
1510         struct cpuhp_step *sp;
1511         int ret = 0;
1512
1513         /*
1514          * If name is NULL, then the state gets removed.
1515          *
1516          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1517          * the first allocation from these dynamic ranges, so the removal
1518          * would trigger a new allocation and clear the wrong (already
1519          * empty) state, leaving the callbacks of the to be cleared state
1520          * dangling, which causes wreckage on the next hotplug operation.
1521          */
1522         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1523                      state == CPUHP_BP_PREPARE_DYN)) {
1524                 ret = cpuhp_reserve_state(state);
1525                 if (ret < 0)
1526                         return ret;
1527                 state = ret;
1528         }
1529         sp = cpuhp_get_step(state);
1530         if (name && sp->name)
1531                 return -EBUSY;
1532
1533         sp->startup.single = startup;
1534         sp->teardown.single = teardown;
1535         sp->name = name;
1536         sp->multi_instance = multi_instance;
1537         INIT_HLIST_HEAD(&sp->list);
1538         return ret;
1539 }
1540
1541 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1542 {
1543         return cpuhp_get_step(state)->teardown.single;
1544 }
1545
1546 /*
1547  * Call the startup/teardown function for a step either on the AP or
1548  * on the current CPU.
1549  */
1550 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1551                             struct hlist_node *node)
1552 {
1553         struct cpuhp_step *sp = cpuhp_get_step(state);
1554         int ret;
1555
1556         /*
1557          * If there's nothing to do, we done.
1558          * Relies on the union for multi_instance.
1559          */
1560         if ((bringup && !sp->startup.single) ||
1561             (!bringup && !sp->teardown.single))
1562                 return 0;
1563         /*
1564          * The non AP bound callbacks can fail on bringup. On teardown
1565          * e.g. module removal we crash for now.
1566          */
1567 #ifdef CONFIG_SMP
1568         if (cpuhp_is_ap_state(state))
1569                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1570         else
1571                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1572 #else
1573         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1574 #endif
1575         BUG_ON(ret && !bringup);
1576         return ret;
1577 }
1578
1579 /*
1580  * Called from __cpuhp_setup_state on a recoverable failure.
1581  *
1582  * Note: The teardown callbacks for rollback are not allowed to fail!
1583  */
1584 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1585                                    struct hlist_node *node)
1586 {
1587         int cpu;
1588
1589         /* Roll back the already executed steps on the other cpus */
1590         for_each_present_cpu(cpu) {
1591                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1592                 int cpustate = st->state;
1593
1594                 if (cpu >= failedcpu)
1595                         break;
1596
1597                 /* Did we invoke the startup call on that cpu ? */
1598                 if (cpustate >= state)
1599                         cpuhp_issue_call(cpu, state, false, node);
1600         }
1601 }
1602
1603 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1604                                           struct hlist_node *node,
1605                                           bool invoke)
1606 {
1607         struct cpuhp_step *sp;
1608         int cpu;
1609         int ret;
1610
1611         lockdep_assert_cpus_held();
1612
1613         sp = cpuhp_get_step(state);
1614         if (sp->multi_instance == false)
1615                 return -EINVAL;
1616
1617         mutex_lock(&cpuhp_state_mutex);
1618
1619         if (!invoke || !sp->startup.multi)
1620                 goto add_node;
1621
1622         /*
1623          * Try to call the startup callback for each present cpu
1624          * depending on the hotplug state of the cpu.
1625          */
1626         for_each_present_cpu(cpu) {
1627                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1628                 int cpustate = st->state;
1629
1630                 if (cpustate < state)
1631                         continue;
1632
1633                 ret = cpuhp_issue_call(cpu, state, true, node);
1634                 if (ret) {
1635                         if (sp->teardown.multi)
1636                                 cpuhp_rollback_install(cpu, state, node);
1637                         goto unlock;
1638                 }
1639         }
1640 add_node:
1641         ret = 0;
1642         hlist_add_head(node, &sp->list);
1643 unlock:
1644         mutex_unlock(&cpuhp_state_mutex);
1645         return ret;
1646 }
1647
1648 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1649                                bool invoke)
1650 {
1651         int ret;
1652
1653         cpus_read_lock();
1654         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1655         cpus_read_unlock();
1656         return ret;
1657 }
1658 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1659
1660 /**
1661  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1662  * @state:              The state to setup
1663  * @invoke:             If true, the startup function is invoked for cpus where
1664  *                      cpu state >= @state
1665  * @startup:            startup callback function
1666  * @teardown:           teardown callback function
1667  * @multi_instance:     State is set up for multiple instances which get
1668  *                      added afterwards.
1669  *
1670  * The caller needs to hold cpus read locked while calling this function.
1671  * Returns:
1672  *   On success:
1673  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1674  *      0 for all other states
1675  *   On failure: proper (negative) error code
1676  */
1677 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1678                                    const char *name, bool invoke,
1679                                    int (*startup)(unsigned int cpu),
1680                                    int (*teardown)(unsigned int cpu),
1681                                    bool multi_instance)
1682 {
1683         int cpu, ret = 0;
1684         bool dynstate;
1685
1686         lockdep_assert_cpus_held();
1687
1688         if (cpuhp_cb_check(state) || !name)
1689                 return -EINVAL;
1690
1691         mutex_lock(&cpuhp_state_mutex);
1692
1693         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1694                                     multi_instance);
1695
1696         dynstate = state == CPUHP_AP_ONLINE_DYN;
1697         if (ret > 0 && dynstate) {
1698                 state = ret;
1699                 ret = 0;
1700         }
1701
1702         if (ret || !invoke || !startup)
1703                 goto out;
1704
1705         /*
1706          * Try to call the startup callback for each present cpu
1707          * depending on the hotplug state of the cpu.
1708          */
1709         for_each_present_cpu(cpu) {
1710                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1711                 int cpustate = st->state;
1712
1713                 if (cpustate < state)
1714                         continue;
1715
1716                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1717                 if (ret) {
1718                         if (teardown)
1719                                 cpuhp_rollback_install(cpu, state, NULL);
1720                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1721                         goto out;
1722                 }
1723         }
1724 out:
1725         mutex_unlock(&cpuhp_state_mutex);
1726         /*
1727          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1728          * dynamically allocated state in case of success.
1729          */
1730         if (!ret && dynstate)
1731                 return state;
1732         return ret;
1733 }
1734 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1735
1736 int __cpuhp_setup_state(enum cpuhp_state state,
1737                         const char *name, bool invoke,
1738                         int (*startup)(unsigned int cpu),
1739                         int (*teardown)(unsigned int cpu),
1740                         bool multi_instance)
1741 {
1742         int ret;
1743
1744         cpus_read_lock();
1745         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1746                                              teardown, multi_instance);
1747         cpus_read_unlock();
1748         return ret;
1749 }
1750 EXPORT_SYMBOL(__cpuhp_setup_state);
1751
1752 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1753                                   struct hlist_node *node, bool invoke)
1754 {
1755         struct cpuhp_step *sp = cpuhp_get_step(state);
1756         int cpu;
1757
1758         BUG_ON(cpuhp_cb_check(state));
1759
1760         if (!sp->multi_instance)
1761                 return -EINVAL;
1762
1763         cpus_read_lock();
1764         mutex_lock(&cpuhp_state_mutex);
1765
1766         if (!invoke || !cpuhp_get_teardown_cb(state))
1767                 goto remove;
1768         /*
1769          * Call the teardown callback for each present cpu depending
1770          * on the hotplug state of the cpu. This function is not
1771          * allowed to fail currently!
1772          */
1773         for_each_present_cpu(cpu) {
1774                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1775                 int cpustate = st->state;
1776
1777                 if (cpustate >= state)
1778                         cpuhp_issue_call(cpu, state, false, node);
1779         }
1780
1781 remove:
1782         hlist_del(node);
1783         mutex_unlock(&cpuhp_state_mutex);
1784         cpus_read_unlock();
1785
1786         return 0;
1787 }
1788 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1789
1790 /**
1791  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1792  * @state:      The state to remove
1793  * @invoke:     If true, the teardown function is invoked for cpus where
1794  *              cpu state >= @state
1795  *
1796  * The caller needs to hold cpus read locked while calling this function.
1797  * The teardown callback is currently not allowed to fail. Think
1798  * about module removal!
1799  */
1800 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1801 {
1802         struct cpuhp_step *sp = cpuhp_get_step(state);
1803         int cpu;
1804
1805         BUG_ON(cpuhp_cb_check(state));
1806
1807         lockdep_assert_cpus_held();
1808
1809         mutex_lock(&cpuhp_state_mutex);
1810         if (sp->multi_instance) {
1811                 WARN(!hlist_empty(&sp->list),
1812                      "Error: Removing state %d which has instances left.\n",
1813                      state);
1814                 goto remove;
1815         }
1816
1817         if (!invoke || !cpuhp_get_teardown_cb(state))
1818                 goto remove;
1819
1820         /*
1821          * Call the teardown callback for each present cpu depending
1822          * on the hotplug state of the cpu. This function is not
1823          * allowed to fail currently!
1824          */
1825         for_each_present_cpu(cpu) {
1826                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1827                 int cpustate = st->state;
1828
1829                 if (cpustate >= state)
1830                         cpuhp_issue_call(cpu, state, false, NULL);
1831         }
1832 remove:
1833         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1834         mutex_unlock(&cpuhp_state_mutex);
1835 }
1836 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1837
1838 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1839 {
1840         cpus_read_lock();
1841         __cpuhp_remove_state_cpuslocked(state, invoke);
1842         cpus_read_unlock();
1843 }
1844 EXPORT_SYMBOL(__cpuhp_remove_state);
1845
1846 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1847 static ssize_t show_cpuhp_state(struct device *dev,
1848                                 struct device_attribute *attr, char *buf)
1849 {
1850         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1851
1852         return sprintf(buf, "%d\n", st->state);
1853 }
1854 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1855
1856 static ssize_t write_cpuhp_target(struct device *dev,
1857                                   struct device_attribute *attr,
1858                                   const char *buf, size_t count)
1859 {
1860         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1861         struct cpuhp_step *sp;
1862         int target, ret;
1863
1864         ret = kstrtoint(buf, 10, &target);
1865         if (ret)
1866                 return ret;
1867
1868 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1869         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1870                 return -EINVAL;
1871 #else
1872         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1873                 return -EINVAL;
1874 #endif
1875
1876         ret = lock_device_hotplug_sysfs();
1877         if (ret)
1878                 return ret;
1879
1880         mutex_lock(&cpuhp_state_mutex);
1881         sp = cpuhp_get_step(target);
1882         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1883         mutex_unlock(&cpuhp_state_mutex);
1884         if (ret)
1885                 goto out;
1886
1887         if (st->state < target)
1888                 ret = do_cpu_up(dev->id, target);
1889         else
1890                 ret = do_cpu_down(dev->id, target);
1891 out:
1892         unlock_device_hotplug();
1893         return ret ? ret : count;
1894 }
1895
1896 static ssize_t show_cpuhp_target(struct device *dev,
1897                                  struct device_attribute *attr, char *buf)
1898 {
1899         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1900
1901         return sprintf(buf, "%d\n", st->target);
1902 }
1903 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1904
1905
1906 static ssize_t write_cpuhp_fail(struct device *dev,
1907                                 struct device_attribute *attr,
1908                                 const char *buf, size_t count)
1909 {
1910         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1911         struct cpuhp_step *sp;
1912         int fail, ret;
1913
1914         ret = kstrtoint(buf, 10, &fail);
1915         if (ret)
1916                 return ret;
1917
1918         /*
1919          * Cannot fail STARTING/DYING callbacks.
1920          */
1921         if (cpuhp_is_atomic_state(fail))
1922                 return -EINVAL;
1923
1924         /*
1925          * Cannot fail anything that doesn't have callbacks.
1926          */
1927         mutex_lock(&cpuhp_state_mutex);
1928         sp = cpuhp_get_step(fail);
1929         if (!sp->startup.single && !sp->teardown.single)
1930                 ret = -EINVAL;
1931         mutex_unlock(&cpuhp_state_mutex);
1932         if (ret)
1933                 return ret;
1934
1935         st->fail = fail;
1936
1937         return count;
1938 }
1939
1940 static ssize_t show_cpuhp_fail(struct device *dev,
1941                                struct device_attribute *attr, char *buf)
1942 {
1943         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1944
1945         return sprintf(buf, "%d\n", st->fail);
1946 }
1947
1948 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1949
1950 static struct attribute *cpuhp_cpu_attrs[] = {
1951         &dev_attr_state.attr,
1952         &dev_attr_target.attr,
1953         &dev_attr_fail.attr,
1954         NULL
1955 };
1956
1957 static const struct attribute_group cpuhp_cpu_attr_group = {
1958         .attrs = cpuhp_cpu_attrs,
1959         .name = "hotplug",
1960         NULL
1961 };
1962
1963 static ssize_t show_cpuhp_states(struct device *dev,
1964                                  struct device_attribute *attr, char *buf)
1965 {
1966         ssize_t cur, res = 0;
1967         int i;
1968
1969         mutex_lock(&cpuhp_state_mutex);
1970         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1971                 struct cpuhp_step *sp = cpuhp_get_step(i);
1972
1973                 if (sp->name) {
1974                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1975                         buf += cur;
1976                         res += cur;
1977                 }
1978         }
1979         mutex_unlock(&cpuhp_state_mutex);
1980         return res;
1981 }
1982 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1983
1984 static struct attribute *cpuhp_cpu_root_attrs[] = {
1985         &dev_attr_states.attr,
1986         NULL
1987 };
1988
1989 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1990         .attrs = cpuhp_cpu_root_attrs,
1991         .name = "hotplug",
1992         NULL
1993 };
1994
1995 #ifdef CONFIG_HOTPLUG_SMT
1996
1997 static const char *smt_states[] = {
1998         [CPU_SMT_ENABLED]               = "on",
1999         [CPU_SMT_DISABLED]              = "off",
2000         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2001         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2002 };
2003
2004 static ssize_t
2005 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2006 {
2007         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2008 }
2009
2010 static void cpuhp_offline_cpu_device(unsigned int cpu)
2011 {
2012         struct device *dev = get_cpu_device(cpu);
2013
2014         dev->offline = true;
2015         /* Tell user space about the state change */
2016         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2017 }
2018
2019 static void cpuhp_online_cpu_device(unsigned int cpu)
2020 {
2021         struct device *dev = get_cpu_device(cpu);
2022
2023         dev->offline = false;
2024         /* Tell user space about the state change */
2025         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2026 }
2027
2028 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2029 {
2030         int cpu, ret = 0;
2031
2032         cpu_maps_update_begin();
2033         for_each_online_cpu(cpu) {
2034                 if (topology_is_primary_thread(cpu))
2035                         continue;
2036                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2037                 if (ret)
2038                         break;
2039                 /*
2040                  * As this needs to hold the cpu maps lock it's impossible
2041                  * to call device_offline() because that ends up calling
2042                  * cpu_down() which takes cpu maps lock. cpu maps lock
2043                  * needs to be held as this might race against in kernel
2044                  * abusers of the hotplug machinery (thermal management).
2045                  *
2046                  * So nothing would update device:offline state. That would
2047                  * leave the sysfs entry stale and prevent onlining after
2048                  * smt control has been changed to 'off' again. This is
2049                  * called under the sysfs hotplug lock, so it is properly
2050                  * serialized against the regular offline usage.
2051                  */
2052                 cpuhp_offline_cpu_device(cpu);
2053         }
2054         if (!ret)
2055                 cpu_smt_control = ctrlval;
2056         cpu_maps_update_done();
2057         return ret;
2058 }
2059
2060 static int cpuhp_smt_enable(void)
2061 {
2062         int cpu, ret = 0;
2063
2064         cpu_maps_update_begin();
2065         cpu_smt_control = CPU_SMT_ENABLED;
2066         for_each_present_cpu(cpu) {
2067                 /* Skip online CPUs and CPUs on offline nodes */
2068                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2069                         continue;
2070                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2071                 if (ret)
2072                         break;
2073                 /* See comment in cpuhp_smt_disable() */
2074                 cpuhp_online_cpu_device(cpu);
2075         }
2076         cpu_maps_update_done();
2077         return ret;
2078 }
2079
2080 static ssize_t
2081 store_smt_control(struct device *dev, struct device_attribute *attr,
2082                   const char *buf, size_t count)
2083 {
2084         int ctrlval, ret;
2085
2086         if (sysfs_streq(buf, "on"))
2087                 ctrlval = CPU_SMT_ENABLED;
2088         else if (sysfs_streq(buf, "off"))
2089                 ctrlval = CPU_SMT_DISABLED;
2090         else if (sysfs_streq(buf, "forceoff"))
2091                 ctrlval = CPU_SMT_FORCE_DISABLED;
2092         else
2093                 return -EINVAL;
2094
2095         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2096                 return -EPERM;
2097
2098         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2099                 return -ENODEV;
2100
2101         ret = lock_device_hotplug_sysfs();
2102         if (ret)
2103                 return ret;
2104
2105         if (ctrlval != cpu_smt_control) {
2106                 switch (ctrlval) {
2107                 case CPU_SMT_ENABLED:
2108                         ret = cpuhp_smt_enable();
2109                         break;
2110                 case CPU_SMT_DISABLED:
2111                 case CPU_SMT_FORCE_DISABLED:
2112                         ret = cpuhp_smt_disable(ctrlval);
2113                         break;
2114                 }
2115         }
2116
2117         unlock_device_hotplug();
2118         return ret ? ret : count;
2119 }
2120 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2121
2122 static ssize_t
2123 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2124 {
2125         bool active = topology_max_smt_threads() > 1;
2126
2127         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2128 }
2129 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2130
2131 static struct attribute *cpuhp_smt_attrs[] = {
2132         &dev_attr_control.attr,
2133         &dev_attr_active.attr,
2134         NULL
2135 };
2136
2137 static const struct attribute_group cpuhp_smt_attr_group = {
2138         .attrs = cpuhp_smt_attrs,
2139         .name = "smt",
2140         NULL
2141 };
2142
2143 static int __init cpu_smt_state_init(void)
2144 {
2145         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2146                                   &cpuhp_smt_attr_group);
2147 }
2148
2149 #else
2150 static inline int cpu_smt_state_init(void) { return 0; }
2151 #endif
2152
2153 static int __init cpuhp_sysfs_init(void)
2154 {
2155         int cpu, ret;
2156
2157         ret = cpu_smt_state_init();
2158         if (ret)
2159                 return ret;
2160
2161         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2162                                  &cpuhp_cpu_root_attr_group);
2163         if (ret)
2164                 return ret;
2165
2166         for_each_possible_cpu(cpu) {
2167                 struct device *dev = get_cpu_device(cpu);
2168
2169                 if (!dev)
2170                         continue;
2171                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2172                 if (ret)
2173                         return ret;
2174         }
2175         return 0;
2176 }
2177 device_initcall(cpuhp_sysfs_init);
2178 #endif
2179
2180 /*
2181  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2182  * represents all NR_CPUS bits binary values of 1<<nr.
2183  *
2184  * It is used by cpumask_of() to get a constant address to a CPU
2185  * mask value that has a single bit set only.
2186  */
2187
2188 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2189 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2190 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2191 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2192 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2193
2194 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2195
2196         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2197         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2198 #if BITS_PER_LONG > 32
2199         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2200         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2201 #endif
2202 };
2203 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2204
2205 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2206 EXPORT_SYMBOL(cpu_all_bits);
2207
2208 #ifdef CONFIG_INIT_ALL_POSSIBLE
2209 struct cpumask __cpu_possible_mask __read_mostly
2210         = {CPU_BITS_ALL};
2211 #else
2212 struct cpumask __cpu_possible_mask __read_mostly;
2213 #endif
2214 EXPORT_SYMBOL(__cpu_possible_mask);
2215
2216 struct cpumask __cpu_online_mask __read_mostly;
2217 EXPORT_SYMBOL(__cpu_online_mask);
2218
2219 struct cpumask __cpu_present_mask __read_mostly;
2220 EXPORT_SYMBOL(__cpu_present_mask);
2221
2222 struct cpumask __cpu_active_mask __read_mostly;
2223 EXPORT_SYMBOL(__cpu_active_mask);
2224
2225 void init_cpu_present(const struct cpumask *src)
2226 {
2227         cpumask_copy(&__cpu_present_mask, src);
2228 }
2229
2230 void init_cpu_possible(const struct cpumask *src)
2231 {
2232         cpumask_copy(&__cpu_possible_mask, src);
2233 }
2234
2235 void init_cpu_online(const struct cpumask *src)
2236 {
2237         cpumask_copy(&__cpu_online_mask, src);
2238 }
2239
2240 /*
2241  * Activate the first processor.
2242  */
2243 void __init boot_cpu_init(void)
2244 {
2245         int cpu = smp_processor_id();
2246
2247         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2248         set_cpu_online(cpu, true);
2249         set_cpu_active(cpu, true);
2250         set_cpu_present(cpu, true);
2251         set_cpu_possible(cpu, true);
2252
2253 #ifdef CONFIG_SMP
2254         __boot_cpu_id = cpu;
2255 #endif
2256 }
2257
2258 /*
2259  * Must be called _AFTER_ setting up the per_cpu areas
2260  */
2261 void __init boot_cpu_hotplug_init(void)
2262 {
2263 #ifdef CONFIG_SMP
2264         this_cpu_write(cpuhp_state.booted_once, true);
2265 #endif
2266         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2267 }