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  * @skip_onerr: Do not invoke the functions on error rollback
 106  *              Will go away once the notifiers are gone
 107  * @cant_stop:  Bringup/teardown can't be stopped at this step
 108  */
 109 struct cpuhp_step {
 110         const char              *name;
 111         union {
 112                 int             (*single)(unsigned int cpu);
 113                 int             (*multi)(unsigned int cpu,
 114                                          struct hlist_node *node);
 115         } startup;
 116         union {
 117                 int             (*single)(unsigned int cpu);
 118                 int             (*multi)(unsigned int cpu,
 119                                          struct hlist_node *node);
 120         } teardown;
 121         struct hlist_head       list;
 122         bool                    skip_onerr;
 123         bool                    cant_stop;
 124         bool                    multi_instance;
 125 };
 126
 127 static DEFINE_MUTEX(cpuhp_state_mutex);
 128 static struct cpuhp_step cpuhp_hp_states[];
 129
 130 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 131 {
 132         return cpuhp_hp_states + state;
 133 }
 134
 135 /**
 136  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 137  * @cpu:        The cpu for which the callback should be invoked
 138  * @state:      The state to do callbacks for
 139  * @bringup:    True if the bringup callback should be invoked
 140  * @node:       For multi-instance, do a single entry callback for install/remove
 141  * @lastp:      For multi-instance rollback, remember how far we got
 142  *
 143  * Called from cpu hotplug and from the state register machinery.
 144  */
 145 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 146                                  bool bringup, struct hlist_node *node,
 147                                  struct hlist_node **lastp)
 148 {
 149         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 150         struct cpuhp_step *step = cpuhp_get_step(state);
 151         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 152         int (*cb)(unsigned int cpu);
 153         int ret, cnt;
 154
 155         if (st->fail == state) {
 156                 st->fail = CPUHP_INVALID;
 157
 158                 if (!(bringup ? step->startup.single : step->teardown.single))
 159                         return 0;
 160
 161                 return -EAGAIN;
 162         }
 163
 164         if (!step->multi_instance) {
 165                 WARN_ON_ONCE(lastp && *lastp);
 166                 cb = bringup ? step->startup.single : step->teardown.single;
 167                 if (!cb)
 168                         return 0;
 169                 trace_cpuhp_enter(cpu, st->target, state, cb);
 170                 ret = cb(cpu);
 171                 trace_cpuhp_exit(cpu, st->state, state, ret);
 172                 return ret;
 173         }
 174         cbm = bringup ? step->startup.multi : step->teardown.multi;
 175         if (!cbm)
 176                 return 0;
 177
 178         /* Single invocation for instance add/remove */
 179         if (node) {
 180                 WARN_ON_ONCE(lastp && *lastp);
 181                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 182                 ret = cbm(cpu, node);
 183                 trace_cpuhp_exit(cpu, st->state, state, ret);
 184                 return ret;
 185         }
 186
 187         /* State transition. Invoke on all instances */
 188         cnt = 0;
 189         hlist_for_each(node, &step->list) {
 190                 if (lastp && node == *lastp)
 191                         break;
 192
 193                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 194                 ret = cbm(cpu, node);
 195                 trace_cpuhp_exit(cpu, st->state, state, ret);
 196                 if (ret) {
 197                         if (!lastp)
 198                                 goto err;
 199
 200                         *lastp = node;
 201                         return ret;
 202                 }
 203                 cnt++;
 204         }
 205         if (lastp)
 206                 *lastp = NULL;
 207         return 0;
 208 err:
 209         /* Rollback the instances if one failed */
 210         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 211         if (!cbm)
 212                 return ret;
 213
 214         hlist_for_each(node, &step->list) {
 215                 if (!cnt--)
 216                         break;
 217
 218                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 219                 ret = cbm(cpu, node);
 220                 trace_cpuhp_exit(cpu, st->state, state, ret);
 221                 /*
 222                  * Rollback must not fail,
 223                  */
 224                 WARN_ON_ONCE(ret);
 225         }
 226         return ret;
 227 }
 228
 229 #ifdef CONFIG_SMP
 230 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 231 {
 232         /*
 233          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 234          * purposes as that state is handled explicitly in cpu_down.
 235          */
 236         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 237 }
 238
 239 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 240 {
 241         struct completion *done = bringup ? &st->done_up : &st->done_down;
 242         wait_for_completion(done);
 243 }
 244
 245 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 246 {
 247         struct completion *done = bringup ? &st->done_up : &st->done_down;
 248         complete(done);
 249 }
 250
 251 /*
 252  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 253  */
 254 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 255 {
 256         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 257 }
 258
 259 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 260 static DEFINE_MUTEX(cpu_add_remove_lock);
 261 bool cpuhp_tasks_frozen;
 262 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 263
 264 /*
 265  * The following two APIs (cpu_maps_update_begin/done) must be used when
 266  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 267  */
 268 void cpu_maps_update_begin(void)
 269 {
 270         mutex_lock(&cpu_add_remove_lock);
 271 }
 272
 273 void cpu_maps_update_done(void)
 274 {
 275         mutex_unlock(&cpu_add_remove_lock);
 276 }
 277
 278 /*
 279  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 280  * Should always be manipulated under cpu_add_remove_lock
 281  */
 282 static int cpu_hotplug_disabled;
 283
 284 #ifdef CONFIG_HOTPLUG_CPU
 285
 286 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 287
 288 void cpus_read_lock(void)
 289 {
 290         percpu_down_read(&cpu_hotplug_lock);
 291 }
 292 EXPORT_SYMBOL_GPL(cpus_read_lock);
 293
 294 int cpus_read_trylock(void)
 295 {
 296         return percpu_down_read_trylock(&cpu_hotplug_lock);
 297 }
 298 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 299
 300 void cpus_read_unlock(void)
 301 {
 302         percpu_up_read(&cpu_hotplug_lock);
 303 }
 304 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 305
 306 void cpus_write_lock(void)
 307 {
 308         percpu_down_write(&cpu_hotplug_lock);
 309 }
 310
 311 void cpus_write_unlock(void)
 312 {
 313         percpu_up_write(&cpu_hotplug_lock);
 314 }
 315
 316 void lockdep_assert_cpus_held(void)
 317 {
 318         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 319 }
 320
 321 /*
 322  * Wait for currently running CPU hotplug operations to complete (if any) and
 323  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 324  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 325  * hotplug path before performing hotplug operations. So acquiring that lock
 326  * guarantees mutual exclusion from any currently running hotplug operations.
 327  */
 328 void cpu_hotplug_disable(void)
 329 {
 330         cpu_maps_update_begin();
 331         cpu_hotplug_disabled++;
 332         cpu_maps_update_done();
 333 }
 334 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 335
 336 static void __cpu_hotplug_enable(void)
 337 {
 338         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 339                 return;
 340         cpu_hotplug_disabled--;
 341 }
 342
 343 void cpu_hotplug_enable(void)
 344 {
 345         cpu_maps_update_begin();
 346         __cpu_hotplug_enable();
 347         cpu_maps_update_done();
 348 }
 349 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 350 #endif  /* CONFIG_HOTPLUG_CPU */
 351
 352 #ifdef CONFIG_HOTPLUG_SMT
 353 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 354 EXPORT_SYMBOL_GPL(cpu_smt_control);
 355
 356 static bool cpu_smt_available __read_mostly;
 357
 358 void __init cpu_smt_disable(bool force)
 359 {
 360         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 361                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 362                 return;
 363
 364         if (force) {
 365                 pr_info("SMT: Force disabled\n");
 366                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 367         } else {
 368                 cpu_smt_control = CPU_SMT_DISABLED;
 369         }
 370 }
 371
 372 /*
 373  * The decision whether SMT is supported can only be done after the full
 374  * CPU identification. Called from architecture code before non boot CPUs
 375  * are brought up.
 376  */
 377 void __init cpu_smt_check_topology_early(void)
 378 {
 379         if (!topology_smt_supported())
 380                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 381 }
 382
 383 /*
 384  * If SMT was disabled by BIOS, detect it here, after the CPUs have been
 385  * brought online. This ensures the smt/l1tf sysfs entries are consistent
 386  * with reality. cpu_smt_available is set to true during the bringup of non
 387  * boot CPUs when a SMT sibling is detected. Note, this may overwrite
 388  * cpu_smt_control's previous setting.
 389  */
 390 void __init cpu_smt_check_topology(void)
 391 {
 392         if (!cpu_smt_available)
 393                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 394 }
 395
 396 static int __init smt_cmdline_disable(char *str)
 397 {
 398         cpu_smt_disable(str && !strcmp(str, "force"));
 399         return 0;
 400 }
 401 early_param("nosmt", smt_cmdline_disable);
 402
 403 static inline bool cpu_smt_allowed(unsigned int cpu)
 404 {
 405         if (topology_is_primary_thread(cpu))
 406                 return true;
 407
 408         /*
 409          * If the CPU is not a 'primary' thread and the booted_once bit is
 410          * set then the processor has SMT support. Store this information
 411          * for the late check of SMT support in cpu_smt_check_topology().
 412          */
 413         if (per_cpu(cpuhp_state, cpu).booted_once)
 414                 cpu_smt_available = true;
 415
 416         if (cpu_smt_control == CPU_SMT_ENABLED)
 417                 return true;
 418
 419         /*
 420          * On x86 it's required to boot all logical CPUs at least once so
 421          * that the init code can get a chance to set CR4.MCE on each
 422          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 423          * core will shutdown the machine.
 424          */
 425         return !per_cpu(cpuhp_state, cpu).booted_once;
 426 }
 427 #else
 428 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 429 #endif
 430
 431 static inline enum cpuhp_state
 432 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 433 {
 434         enum cpuhp_state prev_state = st->state;
 435
 436         st->rollback = false;
 437         st->last = NULL;
 438
 439         st->target = target;
 440         st->single = false;
 441         st->bringup = st->state < target;
 442
 443         return prev_state;
 444 }
 445
 446 static inline void
 447 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 448 {
 449         st->rollback = true;
 450
 451         /*
 452          * If we have st->last we need to undo partial multi_instance of this
 453          * state first. Otherwise start undo at the previous state.
 454          */
 455         if (!st->last) {
 456                 if (st->bringup)
 457                         st->state--;
 458                 else
 459                         st->state++;
 460         }
 461
 462         st->target = prev_state;
 463         st->bringup = !st->bringup;
 464 }
 465
 466 /* Regular hotplug invocation of the AP hotplug thread */
 467 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 468 {
 469         if (!st->single && st->state == st->target)
 470                 return;
 471
 472         st->result = 0;
 473         /*
 474          * Make sure the above stores are visible before should_run becomes
 475          * true. Paired with the mb() above in cpuhp_thread_fun()
 476          */
 477         smp_mb();
 478         st->should_run = true;
 479         wake_up_process(st->thread);
 480         wait_for_ap_thread(st, st->bringup);
 481 }
 482
 483 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 484 {
 485         enum cpuhp_state prev_state;
 486         int ret;
 487
 488         prev_state = cpuhp_set_state(st, target);
 489         __cpuhp_kick_ap(st);
 490         if ((ret = st->result)) {
 491                 cpuhp_reset_state(st, prev_state);
 492                 __cpuhp_kick_ap(st);
 493         }
 494
 495         return ret;
 496 }
 497
 498 static int bringup_wait_for_ap(unsigned int cpu)
 499 {
 500         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 501
 502         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 503         wait_for_ap_thread(st, true);
 504         if (WARN_ON_ONCE((!cpu_online(cpu))))
 505                 return -ECANCELED;
 506
 507         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 508         stop_machine_unpark(cpu);
 509         kthread_unpark(st->thread);
 510
 511         /*
 512          * SMT soft disabling on X86 requires to bring the CPU out of the
 513          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 514          * CPU marked itself as booted_once in cpu_notify_starting() so the
 515          * cpu_smt_allowed() check will now return false if this is not the
 516          * primary sibling.
 517          */
 518         if (!cpu_smt_allowed(cpu))
 519                 return -ECANCELED;
 520
 521         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 522                 return 0;
 523
 524         return cpuhp_kick_ap(st, st->target);
 525 }
 526
 527 static int bringup_cpu(unsigned int cpu)
 528 {
 529         struct task_struct *idle = idle_thread_get(cpu);
 530         int ret;
 531
 532         /*
 533          * Some architectures have to walk the irq descriptors to
 534          * setup the vector space for the cpu which comes online.
 535          * Prevent irq alloc/free across the bringup.
 536          */
 537         irq_lock_sparse();
 538
 539         /* Arch-specific enabling code. */
 540         ret = __cpu_up(cpu, idle);
 541         irq_unlock_sparse();
 542         if (ret)
 543                 return ret;
 544         return bringup_wait_for_ap(cpu);
 545 }
 546
 547 /*
 548  * Hotplug state machine related functions
 549  */
 550
 551 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 552 {
 553         for (st->state--; st->state > st->target; st->state--) {
 554                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 555
 556                 if (!step->skip_onerr)
 557                         cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 558         }
 559 }
 560
 561 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 562                               enum cpuhp_state target)
 563 {
 564         enum cpuhp_state prev_state = st->state;
 565         int ret = 0;
 566
 567         while (st->state < target) {
 568                 st->state++;
 569                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 570                 if (ret) {
 571                         st->target = prev_state;
 572                         undo_cpu_up(cpu, st);
 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         /*
 618          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 619          * that if we see ->should_run we also see the rest of the state.
 620          */
 621         smp_mb();
 622
 623         if (WARN_ON_ONCE(!st->should_run))
 624                 return;
 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 (st->rollback) {
 648                 struct cpuhp_step *step = cpuhp_get_step(state);
 649                 if (step->skip_onerr)
 650                         goto next;
 651         }
 652
 653         if (cpuhp_is_atomic_state(state)) {
 654                 local_irq_disable();
 655                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 656                 local_irq_enable();
 657
 658                 /*
 659                  * STARTING/DYING must not fail!
 660                  */
 661                 WARN_ON_ONCE(st->result);
 662         } else {
 663                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 664         }
 665
 666         if (st->result) {
 667                 /*
 668                  * If we fail on a rollback, we're up a creek without no
 669                  * paddle, no way forward, no way back. We loose, thanks for
 670                  * playing.
 671                  */
 672                 WARN_ON_ONCE(st->rollback);
 673                 st->should_run = false;
 674         }
 675
 676 next:
 677         cpuhp_lock_release(bringup);
 678
 679         if (!st->should_run)
 680                 complete_ap_thread(st, bringup);
 681 }
 682
 683 /* Invoke a single callback on a remote cpu */
 684 static int
 685 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 686                          struct hlist_node *node)
 687 {
 688         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 689         int ret;
 690
 691         if (!cpu_online(cpu))
 692                 return 0;
 693
 694         cpuhp_lock_acquire(false);
 695         cpuhp_lock_release(false);
 696
 697         cpuhp_lock_acquire(true);
 698         cpuhp_lock_release(true);
 699
 700         /*
 701          * If we are up and running, use the hotplug thread. For early calls
 702          * we invoke the thread function directly.
 703          */
 704         if (!st->thread)
 705                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 706
 707         st->rollback = false;
 708         st->last = NULL;
 709
 710         st->node = node;
 711         st->bringup = bringup;
 712         st->cb_state = state;
 713         st->single = true;
 714
 715         __cpuhp_kick_ap(st);
 716
 717         /*
 718          * If we failed and did a partial, do a rollback.
 719          */
 720         if ((ret = st->result) && st->last) {
 721                 st->rollback = true;
 722                 st->bringup = !bringup;
 723
 724                 __cpuhp_kick_ap(st);
 725         }
 726
 727         /*
 728          * Clean up the leftovers so the next hotplug operation wont use stale
 729          * data.
 730          */
 731         st->node = st->last = NULL;
 732         return ret;
 733 }
 734
 735 static int cpuhp_kick_ap_work(unsigned int cpu)
 736 {
 737         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 738         enum cpuhp_state prev_state = st->state;
 739         int ret;
 740
 741         cpuhp_lock_acquire(false);
 742         cpuhp_lock_release(false);
 743
 744         cpuhp_lock_acquire(true);
 745         cpuhp_lock_release(true);
 746
 747         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 748         ret = cpuhp_kick_ap(st, st->target);
 749         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 750
 751         return ret;
 752 }
 753
 754 static struct smp_hotplug_thread cpuhp_threads = {
 755         .store                  = &cpuhp_state.thread,
 756         .create                 = &cpuhp_create,
 757         .thread_should_run      = cpuhp_should_run,
 758         .thread_fn              = cpuhp_thread_fun,
 759         .thread_comm            = "cpuhp/%u",
 760         .selfparking            = true,
 761 };
 762
 763 void __init cpuhp_threads_init(void)
 764 {
 765         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 766         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 767 }
 768
 769 #ifdef CONFIG_HOTPLUG_CPU
 770 /**
 771  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 772  * @cpu: a CPU id
 773  *
 774  * This function walks all processes, finds a valid mm struct for each one and
 775  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 776  * trivial, there are various non-obvious corner cases, which this function
 777  * tries to solve in a safe manner.
 778  *
 779  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 780  * be called only for an already offlined CPU.
 781  */
 782 void clear_tasks_mm_cpumask(int cpu)
 783 {
 784         struct task_struct *p;
 785
 786         /*
 787          * This function is called after the cpu is taken down and marked
 788          * offline, so its not like new tasks will ever get this cpu set in
 789          * their mm mask. -- Peter Zijlstra
 790          * Thus, we may use rcu_read_lock() here, instead of grabbing
 791          * full-fledged tasklist_lock.
 792          */
 793         WARN_ON(cpu_online(cpu));
 794         rcu_read_lock();
 795         for_each_process(p) {
 796                 struct task_struct *t;
 797
 798                 /*
 799                  * Main thread might exit, but other threads may still have
 800                  * a valid mm. Find one.
 801                  */
 802                 t = find_lock_task_mm(p);
 803                 if (!t)
 804                         continue;
 805                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 806                 task_unlock(t);
 807         }
 808         rcu_read_unlock();
 809 }
 810
 811 /* Take this CPU down. */
 812 static int take_cpu_down(void *_param)
 813 {
 814         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 815         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 816         int err, cpu = smp_processor_id();
 817         int ret;
 818
 819         /* Ensure this CPU doesn't handle any more interrupts. */
 820         err = __cpu_disable();
 821         if (err < 0)
 822                 return err;
 823
 824         /*
 825          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 826          * do this step again.
 827          */
 828         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 829         st->state--;
 830         /* Invoke the former CPU_DYING callbacks */
 831         for (; st->state > target; st->state--) {
 832                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 833                 /*
 834                  * DYING must not fail!
 835                  */
 836                 WARN_ON_ONCE(ret);
 837         }
 838
 839         /* Give up timekeeping duties */
 840         tick_handover_do_timer();
 841         /* Park the stopper thread */
 842         stop_machine_park(cpu);
 843         return 0;
 844 }
 845
 846 static int takedown_cpu(unsigned int cpu)
 847 {
 848         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 849         int err;
 850
 851         /* Park the smpboot threads */
 852         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 853
 854         /*
 855          * Prevent irq alloc/free while the dying cpu reorganizes the
 856          * interrupt affinities.
 857          */
 858         irq_lock_sparse();
 859
 860         /*
 861          * So now all preempt/rcu users must observe !cpu_active().
 862          */
 863         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 864         if (err) {
 865                 /* CPU refused to die */
 866                 irq_unlock_sparse();
 867                 /* Unpark the hotplug thread so we can rollback there */
 868                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 869                 return err;
 870         }
 871         BUG_ON(cpu_online(cpu));
 872
 873         /*
 874          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 875          * all runnable tasks from the CPU, there's only the idle task left now
 876          * that the migration thread is done doing the stop_machine thing.
 877          *
 878          * Wait for the stop thread to go away.
 879          */
 880         wait_for_ap_thread(st, false);
 881         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 882
 883         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 884         irq_unlock_sparse();
 885
 886         hotplug_cpu__broadcast_tick_pull(cpu);
 887         /* This actually kills the CPU. */
 888         __cpu_die(cpu);
 889
 890         tick_cleanup_dead_cpu(cpu);
 891         rcutree_migrate_callbacks(cpu);
 892         return 0;
 893 }
 894
 895 static void cpuhp_complete_idle_dead(void *arg)
 896 {
 897         struct cpuhp_cpu_state *st = arg;
 898
 899         complete_ap_thread(st, false);
 900 }
 901
 902 void cpuhp_report_idle_dead(void)
 903 {
 904         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 905
 906         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 907         rcu_report_dead(smp_processor_id());
 908         st->state = CPUHP_AP_IDLE_DEAD;
 909         /*
 910          * We cannot call complete after rcu_report_dead() so we delegate it
 911          * to an online cpu.
 912          */
 913         smp_call_function_single(cpumask_first(cpu_online_mask),
 914                                  cpuhp_complete_idle_dead, st, 0);
 915 }
 916
 917 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 918 {
 919         for (st->state++; st->state < st->target; st->state++) {
 920                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 921
 922                 if (!step->skip_onerr)
 923                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 924         }
 925 }
 926
 927 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 928                                 enum cpuhp_state target)
 929 {
 930         enum cpuhp_state prev_state = st->state;
 931         int ret = 0;
 932
 933         for (; st->state > target; st->state--) {
 934                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 935                 if (ret) {
 936                         st->target = prev_state;
 937                         undo_cpu_down(cpu, st);
 938                         break;
 939                 }
 940         }
 941         return ret;
 942 }
 943
 944 /* Requires cpu_add_remove_lock to be held */
 945 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 946                            enum cpuhp_state target)
 947 {
 948         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 949         int prev_state, ret = 0;
 950
 951         if (num_online_cpus() == 1)
 952                 return -EBUSY;
 953
 954         if (!cpu_present(cpu))
 955                 return -EINVAL;
 956
 957         cpus_write_lock();
 958
 959         cpuhp_tasks_frozen = tasks_frozen;
 960
 961         prev_state = cpuhp_set_state(st, target);
 962         /*
 963          * If the current CPU state is in the range of the AP hotplug thread,
 964          * then we need to kick the thread.
 965          */
 966         if (st->state > CPUHP_TEARDOWN_CPU) {
 967                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 968                 ret = cpuhp_kick_ap_work(cpu);
 969                 /*
 970                  * The AP side has done the error rollback already. Just
 971                  * return the error code..
 972                  */
 973                 if (ret)
 974                         goto out;
 975
 976                 /*
 977                  * We might have stopped still in the range of the AP hotplug
 978                  * thread. Nothing to do anymore.
 979                  */
 980                 if (st->state > CPUHP_TEARDOWN_CPU)
 981                         goto out;
 982
 983                 st->target = target;
 984         }
 985         /*
 986          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 987          * to do the further cleanups.
 988          */
 989         ret = cpuhp_down_callbacks(cpu, st, target);
 990         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 991                 cpuhp_reset_state(st, prev_state);
 992                 __cpuhp_kick_ap(st);
 993         }
 994
 995 out:
 996         cpus_write_unlock();
 997         /*
 998          * Do post unplug cleanup. This is still protected against
 999          * concurrent CPU hotplug via cpu_add_remove_lock.
1000          */
1001         lockup_detector_cleanup();
1002         return ret;
1003 }
1004
1005 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1006 {
1007         if (cpu_hotplug_disabled)
1008                 return -EBUSY;
1009         return _cpu_down(cpu, 0, target);
1010 }
1011
1012 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1013 {
1014         int err;
1015
1016         cpu_maps_update_begin();
1017         err = cpu_down_maps_locked(cpu, target);
1018         cpu_maps_update_done();
1019         return err;
1020 }
1021
1022 int cpu_down(unsigned int cpu)
1023 {
1024         return do_cpu_down(cpu, CPUHP_OFFLINE);
1025 }
1026 EXPORT_SYMBOL(cpu_down);
1027
1028 #else
1029 #define takedown_cpu            NULL
1030 #endif /*CONFIG_HOTPLUG_CPU*/
1031
1032 /**
1033  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1034  * @cpu: cpu that just started
1035  *
1036  * It must be called by the arch code on the new cpu, before the new cpu
1037  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1038  */
1039 void notify_cpu_starting(unsigned int cpu)
1040 {
1041         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1042         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1043         int ret;
1044
1045         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1046         st->booted_once = true;
1047         while (st->state < target) {
1048                 st->state++;
1049                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1050                 /*
1051                  * STARTING must not fail!
1052                  */
1053                 WARN_ON_ONCE(ret);
1054         }
1055 }
1056
1057 /*
1058  * Called from the idle task. Wake up the controlling task which brings the
1059  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1060  * the rest of the online bringup to the hotplug thread.
1061  */
1062 void cpuhp_online_idle(enum cpuhp_state state)
1063 {
1064         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1065
1066         /* Happens for the boot cpu */
1067         if (state != CPUHP_AP_ONLINE_IDLE)
1068                 return;
1069
1070         st->state = CPUHP_AP_ONLINE_IDLE;
1071         complete_ap_thread(st, true);
1072 }
1073
1074 /* Requires cpu_add_remove_lock to be held */
1075 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1076 {
1077         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1078         struct task_struct *idle;
1079         int ret = 0;
1080
1081         cpus_write_lock();
1082
1083         if (!cpu_present(cpu)) {
1084                 ret = -EINVAL;
1085                 goto out;
1086         }
1087
1088         /*
1089          * The caller of do_cpu_up might have raced with another
1090          * caller. Ignore it for now.
1091          */
1092         if (st->state >= target)
1093                 goto out;
1094
1095         if (st->state == CPUHP_OFFLINE) {
1096                 /* Let it fail before we try to bring the cpu up */
1097                 idle = idle_thread_get(cpu);
1098                 if (IS_ERR(idle)) {
1099                         ret = PTR_ERR(idle);
1100                         goto out;
1101                 }
1102         }
1103
1104         cpuhp_tasks_frozen = tasks_frozen;
1105
1106         cpuhp_set_state(st, target);
1107         /*
1108          * If the current CPU state is in the range of the AP hotplug thread,
1109          * then we need to kick the thread once more.
1110          */
1111         if (st->state > CPUHP_BRINGUP_CPU) {
1112                 ret = cpuhp_kick_ap_work(cpu);
1113                 /*
1114                  * The AP side has done the error rollback already. Just
1115                  * return the error code..
1116                  */
1117                 if (ret)
1118                         goto out;
1119         }
1120
1121         /*
1122          * Try to reach the target state. We max out on the BP at
1123          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1124          * responsible for bringing it up to the target state.
1125          */
1126         target = min((int)target, CPUHP_BRINGUP_CPU);
1127         ret = cpuhp_up_callbacks(cpu, st, target);
1128 out:
1129         cpus_write_unlock();
1130         return ret;
1131 }
1132
1133 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1134 {
1135         int err = 0;
1136
1137         if (!cpu_possible(cpu)) {
1138                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1139                        cpu);
1140 #if defined(CONFIG_IA64)
1141                 pr_err("please check additional_cpus= boot parameter\n");
1142 #endif
1143                 return -EINVAL;
1144         }
1145
1146         err = try_online_node(cpu_to_node(cpu));
1147         if (err)
1148                 return err;
1149
1150         cpu_maps_update_begin();
1151
1152         if (cpu_hotplug_disabled) {
1153                 err = -EBUSY;
1154                 goto out;
1155         }
1156         if (!cpu_smt_allowed(cpu)) {
1157                 err = -EPERM;
1158                 goto out;
1159         }
1160
1161         err = _cpu_up(cpu, 0, target);
1162 out:
1163         cpu_maps_update_done();
1164         return err;
1165 }
1166
1167 int cpu_up(unsigned int cpu)
1168 {
1169         return do_cpu_up(cpu, CPUHP_ONLINE);
1170 }
1171 EXPORT_SYMBOL_GPL(cpu_up);
1172
1173 #ifdef CONFIG_PM_SLEEP_SMP
1174 static cpumask_var_t frozen_cpus;
1175
1176 int freeze_secondary_cpus(int primary)
1177 {
1178         int cpu, error = 0;
1179
1180         cpu_maps_update_begin();
1181         if (!cpu_online(primary))
1182                 primary = cpumask_first(cpu_online_mask);
1183         /*
1184          * We take down all of the non-boot CPUs in one shot to avoid races
1185          * with the userspace trying to use the CPU hotplug at the same time
1186          */
1187         cpumask_clear(frozen_cpus);
1188
1189         pr_info("Disabling non-boot CPUs ...\n");
1190         for_each_online_cpu(cpu) {
1191                 if (cpu == primary)
1192                         continue;
1193                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1194                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1195                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1196                 if (!error)
1197                         cpumask_set_cpu(cpu, frozen_cpus);
1198                 else {
1199                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1200                         break;
1201                 }
1202         }
1203
1204         if (!error)
1205                 BUG_ON(num_online_cpus() > 1);
1206         else
1207                 pr_err("Non-boot CPUs are not disabled\n");
1208
1209         /*
1210          * Make sure the CPUs won't be enabled by someone else. We need to do
1211          * this even in case of failure as all disable_nonboot_cpus() users are
1212          * supposed to do enable_nonboot_cpus() on the failure path.
1213          */
1214         cpu_hotplug_disabled++;
1215
1216         cpu_maps_update_done();
1217         return error;
1218 }
1219
1220 void __weak arch_enable_nonboot_cpus_begin(void)
1221 {
1222 }
1223
1224 void __weak arch_enable_nonboot_cpus_end(void)
1225 {
1226 }
1227
1228 void enable_nonboot_cpus(void)
1229 {
1230         int cpu, error;
1231
1232         /* Allow everyone to use the CPU hotplug again */
1233         cpu_maps_update_begin();
1234         __cpu_hotplug_enable();
1235         if (cpumask_empty(frozen_cpus))
1236                 goto out;
1237
1238         pr_info("Enabling non-boot CPUs ...\n");
1239
1240         arch_enable_nonboot_cpus_begin();
1241
1242         for_each_cpu(cpu, frozen_cpus) {
1243                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1244                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1245                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1246                 if (!error) {
1247                         pr_info("CPU%d is up\n", cpu);
1248                         continue;
1249                 }
1250                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1251         }
1252
1253         arch_enable_nonboot_cpus_end();
1254
1255         cpumask_clear(frozen_cpus);
1256 out:
1257         cpu_maps_update_done();
1258 }
1259
1260 static int __init alloc_frozen_cpus(void)
1261 {
1262         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1263                 return -ENOMEM;
1264         return 0;
1265 }
1266 core_initcall(alloc_frozen_cpus);
1267
1268 /*
1269  * When callbacks for CPU hotplug notifications are being executed, we must
1270  * ensure that the state of the system with respect to the tasks being frozen
1271  * or not, as reported by the notification, remains unchanged *throughout the
1272  * duration* of the execution of the callbacks.
1273  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1274  *
1275  * This synchronization is implemented by mutually excluding regular CPU
1276  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1277  * Hibernate notifications.
1278  */
1279 static int
1280 cpu_hotplug_pm_callback(struct notifier_block *nb,
1281                         unsigned long action, void *ptr)
1282 {
1283         switch (action) {
1284
1285         case PM_SUSPEND_PREPARE:
1286         case PM_HIBERNATION_PREPARE:
1287                 cpu_hotplug_disable();
1288                 break;
1289
1290         case PM_POST_SUSPEND:
1291         case PM_POST_HIBERNATION:
1292                 cpu_hotplug_enable();
1293                 break;
1294
1295         default:
1296                 return NOTIFY_DONE;
1297         }
1298
1299         return NOTIFY_OK;
1300 }
1301
1302
1303 static int __init cpu_hotplug_pm_sync_init(void)
1304 {
1305         /*
1306          * cpu_hotplug_pm_callback has higher priority than x86
1307          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1308          * to disable cpu hotplug to avoid cpu hotplug race.
1309          */
1310         pm_notifier(cpu_hotplug_pm_callback, 0);
1311         return 0;
1312 }
1313 core_initcall(cpu_hotplug_pm_sync_init);
1314
1315 #endif /* CONFIG_PM_SLEEP_SMP */
1316
1317 int __boot_cpu_id;
1318
1319 #endif /* CONFIG_SMP */
1320
1321 /* Boot processor state steps */
1322 static struct cpuhp_step cpuhp_hp_states[] = {
1323         [CPUHP_OFFLINE] = {
1324                 .name                   = "offline",
1325                 .startup.single         = NULL,
1326                 .teardown.single        = NULL,
1327         },
1328 #ifdef CONFIG_SMP
1329         [CPUHP_CREATE_THREADS]= {
1330                 .name                   = "threads:prepare",
1331                 .startup.single         = smpboot_create_threads,
1332                 .teardown.single        = NULL,
1333                 .cant_stop              = true,
1334         },
1335         [CPUHP_PERF_PREPARE] = {
1336                 .name                   = "perf:prepare",
1337                 .startup.single         = perf_event_init_cpu,
1338                 .teardown.single        = perf_event_exit_cpu,
1339         },
1340         [CPUHP_WORKQUEUE_PREP] = {
1341                 .name                   = "workqueue:prepare",
1342                 .startup.single         = workqueue_prepare_cpu,
1343                 .teardown.single        = NULL,
1344         },
1345         [CPUHP_HRTIMERS_PREPARE] = {
1346                 .name                   = "hrtimers:prepare",
1347                 .startup.single         = hrtimers_prepare_cpu,
1348                 .teardown.single        = hrtimers_dead_cpu,
1349         },
1350         [CPUHP_SMPCFD_PREPARE] = {
1351                 .name                   = "smpcfd:prepare",
1352                 .startup.single         = smpcfd_prepare_cpu,
1353                 .teardown.single        = smpcfd_dead_cpu,
1354         },
1355         [CPUHP_RELAY_PREPARE] = {
1356                 .name                   = "relay:prepare",
1357                 .startup.single         = relay_prepare_cpu,
1358                 .teardown.single        = NULL,
1359         },
1360         [CPUHP_SLAB_PREPARE] = {
1361                 .name                   = "slab:prepare",
1362                 .startup.single         = slab_prepare_cpu,
1363                 .teardown.single        = slab_dead_cpu,
1364         },
1365         [CPUHP_RCUTREE_PREP] = {
1366                 .name                   = "RCU/tree:prepare",
1367                 .startup.single         = rcutree_prepare_cpu,
1368                 .teardown.single        = rcutree_dead_cpu,
1369         },
1370         /*
1371          * On the tear-down path, timers_dead_cpu() must be invoked
1372          * before blk_mq_queue_reinit_notify() from notify_dead(),
1373          * otherwise a RCU stall occurs.
1374          */
1375         [CPUHP_TIMERS_PREPARE] = {
1376                 .name                   = "timers:prepare",
1377                 .startup.single         = timers_prepare_cpu,
1378                 .teardown.single        = timers_dead_cpu,
1379         },
1380         /* Kicks the plugged cpu into life */
1381         [CPUHP_BRINGUP_CPU] = {
1382                 .name                   = "cpu:bringup",
1383                 .startup.single         = bringup_cpu,
1384                 .teardown.single        = NULL,
1385                 .cant_stop              = true,
1386         },
1387         /* Final state before CPU kills itself */
1388         [CPUHP_AP_IDLE_DEAD] = {
1389                 .name                   = "idle:dead",
1390         },
1391         /*
1392          * Last state before CPU enters the idle loop to die. Transient state
1393          * for synchronization.
1394          */
1395         [CPUHP_AP_OFFLINE] = {
1396                 .name                   = "ap:offline",
1397                 .cant_stop              = true,
1398         },
1399         /* First state is scheduler control. Interrupts are disabled */
1400         [CPUHP_AP_SCHED_STARTING] = {
1401                 .name                   = "sched:starting",
1402                 .startup.single         = sched_cpu_starting,
1403                 .teardown.single        = sched_cpu_dying,
1404         },
1405         [CPUHP_AP_RCUTREE_DYING] = {
1406                 .name                   = "RCU/tree:dying",
1407                 .startup.single         = NULL,
1408                 .teardown.single        = rcutree_dying_cpu,
1409         },
1410         [CPUHP_AP_SMPCFD_DYING] = {
1411                 .name                   = "smpcfd:dying",
1412                 .startup.single         = NULL,
1413                 .teardown.single        = smpcfd_dying_cpu,
1414         },
1415         /* Entry state on starting. Interrupts enabled from here on. Transient
1416          * state for synchronsization */
1417         [CPUHP_AP_ONLINE] = {
1418                 .name                   = "ap:online",
1419         },
1420         /*
1421          * Handled on controll processor until the plugged processor manages
1422          * this itself.
1423          */
1424         [CPUHP_TEARDOWN_CPU] = {
1425                 .name                   = "cpu:teardown",
1426                 .startup.single         = NULL,
1427                 .teardown.single        = takedown_cpu,
1428                 .cant_stop              = true,
1429         },
1430         /* Handle smpboot threads park/unpark */
1431         [CPUHP_AP_SMPBOOT_THREADS] = {
1432                 .name                   = "smpboot/threads:online",
1433                 .startup.single         = smpboot_unpark_threads,
1434                 .teardown.single        = smpboot_park_threads,
1435         },
1436         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1437                 .name                   = "irq/affinity:online",
1438                 .startup.single         = irq_affinity_online_cpu,
1439                 .teardown.single        = NULL,
1440         },
1441         [CPUHP_AP_PERF_ONLINE] = {
1442                 .name                   = "perf:online",
1443                 .startup.single         = perf_event_init_cpu,
1444                 .teardown.single        = perf_event_exit_cpu,
1445         },
1446         [CPUHP_AP_WATCHDOG_ONLINE] = {
1447                 .name                   = "lockup_detector:online",
1448                 .startup.single         = lockup_detector_online_cpu,
1449                 .teardown.single        = lockup_detector_offline_cpu,
1450         },
1451         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1452                 .name                   = "workqueue:online",
1453                 .startup.single         = workqueue_online_cpu,
1454                 .teardown.single        = workqueue_offline_cpu,
1455         },
1456         [CPUHP_AP_RCUTREE_ONLINE] = {
1457                 .name                   = "RCU/tree:online",
1458                 .startup.single         = rcutree_online_cpu,
1459                 .teardown.single        = rcutree_offline_cpu,
1460         },
1461 #endif
1462         /*
1463          * The dynamically registered state space is here
1464          */
1465
1466 #ifdef CONFIG_SMP
1467         /* Last state is scheduler control setting the cpu active */
1468         [CPUHP_AP_ACTIVE] = {
1469                 .name                   = "sched:active",
1470                 .startup.single         = sched_cpu_activate,
1471                 .teardown.single        = sched_cpu_deactivate,
1472         },
1473 #endif
1474
1475         /* CPU is fully up and running. */
1476         [CPUHP_ONLINE] = {
1477                 .name                   = "online",
1478                 .startup.single         = NULL,
1479                 .teardown.single        = NULL,
1480         },
1481 };
1482
1483 /* Sanity check for callbacks */
1484 static int cpuhp_cb_check(enum cpuhp_state state)
1485 {
1486         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1487                 return -EINVAL;
1488         return 0;
1489 }
1490
1491 /*
1492  * Returns a free for dynamic slot assignment of the Online state. The states
1493  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1494  * by having no name assigned.
1495  */
1496 static int cpuhp_reserve_state(enum cpuhp_state state)
1497 {
1498         enum cpuhp_state i, end;
1499         struct cpuhp_step *step;
1500
1501         switch (state) {
1502         case CPUHP_AP_ONLINE_DYN:
1503                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1504                 end = CPUHP_AP_ONLINE_DYN_END;
1505                 break;
1506         case CPUHP_BP_PREPARE_DYN:
1507                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1508                 end = CPUHP_BP_PREPARE_DYN_END;
1509                 break;
1510         default:
1511                 return -EINVAL;
1512         }
1513
1514         for (i = state; i <= end; i++, step++) {
1515                 if (!step->name)
1516                         return i;
1517         }
1518         WARN(1, "No more dynamic states available for CPU hotplug\n");
1519         return -ENOSPC;
1520 }
1521
1522 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1523                                  int (*startup)(unsigned int cpu),
1524                                  int (*teardown)(unsigned int cpu),
1525                                  bool multi_instance)
1526 {
1527         /* (Un)Install the callbacks for further cpu hotplug operations */
1528         struct cpuhp_step *sp;
1529         int ret = 0;
1530
1531         /*
1532          * If name is NULL, then the state gets removed.
1533          *
1534          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1535          * the first allocation from these dynamic ranges, so the removal
1536          * would trigger a new allocation and clear the wrong (already
1537          * empty) state, leaving the callbacks of the to be cleared state
1538          * dangling, which causes wreckage on the next hotplug operation.
1539          */
1540         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1541                      state == CPUHP_BP_PREPARE_DYN)) {
1542                 ret = cpuhp_reserve_state(state);
1543                 if (ret < 0)
1544                         return ret;
1545                 state = ret;
1546         }
1547         sp = cpuhp_get_step(state);
1548         if (name && sp->name)
1549                 return -EBUSY;
1550
1551         sp->startup.single = startup;
1552         sp->teardown.single = teardown;
1553         sp->name = name;
1554         sp->multi_instance = multi_instance;
1555         INIT_HLIST_HEAD(&sp->list);
1556         return ret;
1557 }
1558
1559 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1560 {
1561         return cpuhp_get_step(state)->teardown.single;
1562 }
1563
1564 /*
1565  * Call the startup/teardown function for a step either on the AP or
1566  * on the current CPU.
1567  */
1568 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1569                             struct hlist_node *node)
1570 {
1571         struct cpuhp_step *sp = cpuhp_get_step(state);
1572         int ret;
1573
1574         /*
1575          * If there's nothing to do, we done.
1576          * Relies on the union for multi_instance.
1577          */
1578         if ((bringup && !sp->startup.single) ||
1579             (!bringup && !sp->teardown.single))
1580                 return 0;
1581         /*
1582          * The non AP bound callbacks can fail on bringup. On teardown
1583          * e.g. module removal we crash for now.
1584          */
1585 #ifdef CONFIG_SMP
1586         if (cpuhp_is_ap_state(state))
1587                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1588         else
1589                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1590 #else
1591         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1592 #endif
1593         BUG_ON(ret && !bringup);
1594         return ret;
1595 }
1596
1597 /*
1598  * Called from __cpuhp_setup_state on a recoverable failure.
1599  *
1600  * Note: The teardown callbacks for rollback are not allowed to fail!
1601  */
1602 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1603                                    struct hlist_node *node)
1604 {
1605         int cpu;
1606
1607         /* Roll back the already executed steps on the other cpus */
1608         for_each_present_cpu(cpu) {
1609                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1610                 int cpustate = st->state;
1611
1612                 if (cpu >= failedcpu)
1613                         break;
1614
1615                 /* Did we invoke the startup call on that cpu ? */
1616                 if (cpustate >= state)
1617                         cpuhp_issue_call(cpu, state, false, node);
1618         }
1619 }
1620
1621 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1622                                           struct hlist_node *node,
1623                                           bool invoke)
1624 {
1625         struct cpuhp_step *sp;
1626         int cpu;
1627         int ret;
1628
1629         lockdep_assert_cpus_held();
1630
1631         sp = cpuhp_get_step(state);
1632         if (sp->multi_instance == false)
1633                 return -EINVAL;
1634
1635         mutex_lock(&cpuhp_state_mutex);
1636
1637         if (!invoke || !sp->startup.multi)
1638                 goto add_node;
1639
1640         /*
1641          * Try to call the startup callback for each present cpu
1642          * depending on the hotplug state of the cpu.
1643          */
1644         for_each_present_cpu(cpu) {
1645                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1646                 int cpustate = st->state;
1647
1648                 if (cpustate < state)
1649                         continue;
1650
1651                 ret = cpuhp_issue_call(cpu, state, true, node);
1652                 if (ret) {
1653                         if (sp->teardown.multi)
1654                                 cpuhp_rollback_install(cpu, state, node);
1655                         goto unlock;
1656                 }
1657         }
1658 add_node:
1659         ret = 0;
1660         hlist_add_head(node, &sp->list);
1661 unlock:
1662         mutex_unlock(&cpuhp_state_mutex);
1663         return ret;
1664 }
1665
1666 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1667                                bool invoke)
1668 {
1669         int ret;
1670
1671         cpus_read_lock();
1672         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1673         cpus_read_unlock();
1674         return ret;
1675 }
1676 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1677
1678 /**
1679  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1680  * @state:              The state to setup
1681  * @invoke:             If true, the startup function is invoked for cpus where
1682  *                      cpu state >= @state
1683  * @startup:            startup callback function
1684  * @teardown:           teardown callback function
1685  * @multi_instance:     State is set up for multiple instances which get
1686  *                      added afterwards.
1687  *
1688  * The caller needs to hold cpus read locked while calling this function.
1689  * Returns:
1690  *   On success:
1691  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1692  *      0 for all other states
1693  *   On failure: proper (negative) error code
1694  */
1695 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1696                                    const char *name, bool invoke,
1697                                    int (*startup)(unsigned int cpu),
1698                                    int (*teardown)(unsigned int cpu),
1699                                    bool multi_instance)
1700 {
1701         int cpu, ret = 0;
1702         bool dynstate;
1703
1704         lockdep_assert_cpus_held();
1705
1706         if (cpuhp_cb_check(state) || !name)
1707                 return -EINVAL;
1708
1709         mutex_lock(&cpuhp_state_mutex);
1710
1711         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1712                                     multi_instance);
1713
1714         dynstate = state == CPUHP_AP_ONLINE_DYN;
1715         if (ret > 0 && dynstate) {
1716                 state = ret;
1717                 ret = 0;
1718         }
1719
1720         if (ret || !invoke || !startup)
1721                 goto out;
1722
1723         /*
1724          * Try to call the startup callback for each present cpu
1725          * depending on the hotplug state of the cpu.
1726          */
1727         for_each_present_cpu(cpu) {
1728                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1729                 int cpustate = st->state;
1730
1731                 if (cpustate < state)
1732                         continue;
1733
1734                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1735                 if (ret) {
1736                         if (teardown)
1737                                 cpuhp_rollback_install(cpu, state, NULL);
1738                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1739                         goto out;
1740                 }
1741         }
1742 out:
1743         mutex_unlock(&cpuhp_state_mutex);
1744         /*
1745          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1746          * dynamically allocated state in case of success.
1747          */
1748         if (!ret && dynstate)
1749                 return state;
1750         return ret;
1751 }
1752 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1753
1754 int __cpuhp_setup_state(enum cpuhp_state state,
1755                         const char *name, bool invoke,
1756                         int (*startup)(unsigned int cpu),
1757                         int (*teardown)(unsigned int cpu),
1758                         bool multi_instance)
1759 {
1760         int ret;
1761
1762         cpus_read_lock();
1763         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1764                                              teardown, multi_instance);
1765         cpus_read_unlock();
1766         return ret;
1767 }
1768 EXPORT_SYMBOL(__cpuhp_setup_state);
1769
1770 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1771                                   struct hlist_node *node, bool invoke)
1772 {
1773         struct cpuhp_step *sp = cpuhp_get_step(state);
1774         int cpu;
1775
1776         BUG_ON(cpuhp_cb_check(state));
1777
1778         if (!sp->multi_instance)
1779                 return -EINVAL;
1780
1781         cpus_read_lock();
1782         mutex_lock(&cpuhp_state_mutex);
1783
1784         if (!invoke || !cpuhp_get_teardown_cb(state))
1785                 goto remove;
1786         /*
1787          * Call the teardown callback for each present cpu depending
1788          * on the hotplug state of the cpu. This function is not
1789          * allowed to fail currently!
1790          */
1791         for_each_present_cpu(cpu) {
1792                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1793                 int cpustate = st->state;
1794
1795                 if (cpustate >= state)
1796                         cpuhp_issue_call(cpu, state, false, node);
1797         }
1798
1799 remove:
1800         hlist_del(node);
1801         mutex_unlock(&cpuhp_state_mutex);
1802         cpus_read_unlock();
1803
1804         return 0;
1805 }
1806 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1807
1808 /**
1809  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1810  * @state:      The state to remove
1811  * @invoke:     If true, the teardown function is invoked for cpus where
1812  *              cpu state >= @state
1813  *
1814  * The caller needs to hold cpus read locked while calling this function.
1815  * The teardown callback is currently not allowed to fail. Think
1816  * about module removal!
1817  */
1818 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1819 {
1820         struct cpuhp_step *sp = cpuhp_get_step(state);
1821         int cpu;
1822
1823         BUG_ON(cpuhp_cb_check(state));
1824
1825         lockdep_assert_cpus_held();
1826
1827         mutex_lock(&cpuhp_state_mutex);
1828         if (sp->multi_instance) {
1829                 WARN(!hlist_empty(&sp->list),
1830                      "Error: Removing state %d which has instances left.\n",
1831                      state);
1832                 goto remove;
1833         }
1834
1835         if (!invoke || !cpuhp_get_teardown_cb(state))
1836                 goto remove;
1837
1838         /*
1839          * Call the teardown callback for each present cpu depending
1840          * on the hotplug state of the cpu. This function is not
1841          * allowed to fail currently!
1842          */
1843         for_each_present_cpu(cpu) {
1844                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1845                 int cpustate = st->state;
1846
1847                 if (cpustate >= state)
1848                         cpuhp_issue_call(cpu, state, false, NULL);
1849         }
1850 remove:
1851         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1852         mutex_unlock(&cpuhp_state_mutex);
1853 }
1854 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1855
1856 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1857 {
1858         cpus_read_lock();
1859         __cpuhp_remove_state_cpuslocked(state, invoke);
1860         cpus_read_unlock();
1861 }
1862 EXPORT_SYMBOL(__cpuhp_remove_state);
1863
1864 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1865 static ssize_t show_cpuhp_state(struct device *dev,
1866                                 struct device_attribute *attr, char *buf)
1867 {
1868         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1869
1870         return sprintf(buf, "%d\n", st->state);
1871 }
1872 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1873
1874 static ssize_t write_cpuhp_target(struct device *dev,
1875                                   struct device_attribute *attr,
1876                                   const char *buf, size_t count)
1877 {
1878         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1879         struct cpuhp_step *sp;
1880         int target, ret;
1881
1882         ret = kstrtoint(buf, 10, &target);
1883         if (ret)
1884                 return ret;
1885
1886 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1887         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1888                 return -EINVAL;
1889 #else
1890         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1891                 return -EINVAL;
1892 #endif
1893
1894         ret = lock_device_hotplug_sysfs();
1895         if (ret)
1896                 return ret;
1897
1898         mutex_lock(&cpuhp_state_mutex);
1899         sp = cpuhp_get_step(target);
1900         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1901         mutex_unlock(&cpuhp_state_mutex);
1902         if (ret)
1903                 goto out;
1904
1905         if (st->state < target)
1906                 ret = do_cpu_up(dev->id, target);
1907         else
1908                 ret = do_cpu_down(dev->id, target);
1909 out:
1910         unlock_device_hotplug();
1911         return ret ? ret : count;
1912 }
1913
1914 static ssize_t show_cpuhp_target(struct device *dev,
1915                                  struct device_attribute *attr, char *buf)
1916 {
1917         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1918
1919         return sprintf(buf, "%d\n", st->target);
1920 }
1921 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1922
1923
1924 static ssize_t write_cpuhp_fail(struct device *dev,
1925                                 struct device_attribute *attr,
1926                                 const char *buf, size_t count)
1927 {
1928         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1929         struct cpuhp_step *sp;
1930         int fail, ret;
1931
1932         ret = kstrtoint(buf, 10, &fail);
1933         if (ret)
1934                 return ret;
1935
1936         /*
1937          * Cannot fail STARTING/DYING callbacks.
1938          */
1939         if (cpuhp_is_atomic_state(fail))
1940                 return -EINVAL;
1941
1942         /*
1943          * Cannot fail anything that doesn't have callbacks.
1944          */
1945         mutex_lock(&cpuhp_state_mutex);
1946         sp = cpuhp_get_step(fail);
1947         if (!sp->startup.single && !sp->teardown.single)
1948                 ret = -EINVAL;
1949         mutex_unlock(&cpuhp_state_mutex);
1950         if (ret)
1951                 return ret;
1952
1953         st->fail = fail;
1954
1955         return count;
1956 }
1957
1958 static ssize_t show_cpuhp_fail(struct device *dev,
1959                                struct device_attribute *attr, char *buf)
1960 {
1961         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1962
1963         return sprintf(buf, "%d\n", st->fail);
1964 }
1965
1966 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1967
1968 static struct attribute *cpuhp_cpu_attrs[] = {
1969         &dev_attr_state.attr,
1970         &dev_attr_target.attr,
1971         &dev_attr_fail.attr,
1972         NULL
1973 };
1974
1975 static const struct attribute_group cpuhp_cpu_attr_group = {
1976         .attrs = cpuhp_cpu_attrs,
1977         .name = "hotplug",
1978         NULL
1979 };
1980
1981 static ssize_t show_cpuhp_states(struct device *dev,
1982                                  struct device_attribute *attr, char *buf)
1983 {
1984         ssize_t cur, res = 0;
1985         int i;
1986
1987         mutex_lock(&cpuhp_state_mutex);
1988         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1989                 struct cpuhp_step *sp = cpuhp_get_step(i);
1990
1991                 if (sp->name) {
1992                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1993                         buf += cur;
1994                         res += cur;
1995                 }
1996         }
1997         mutex_unlock(&cpuhp_state_mutex);
1998         return res;
1999 }
2000 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2001
2002 static struct attribute *cpuhp_cpu_root_attrs[] = {
2003         &dev_attr_states.attr,
2004         NULL
2005 };
2006
2007 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2008         .attrs = cpuhp_cpu_root_attrs,
2009         .name = "hotplug",
2010         NULL
2011 };
2012
2013 #ifdef CONFIG_HOTPLUG_SMT
2014
2015 static const char *smt_states[] = {
2016         [CPU_SMT_ENABLED]               = "on",
2017         [CPU_SMT_DISABLED]              = "off",
2018         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2019         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2020 };
2021
2022 static ssize_t
2023 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2024 {
2025         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2026 }
2027
2028 static void cpuhp_offline_cpu_device(unsigned int cpu)
2029 {
2030         struct device *dev = get_cpu_device(cpu);
2031
2032         dev->offline = true;
2033         /* Tell user space about the state change */
2034         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2035 }
2036
2037 static void cpuhp_online_cpu_device(unsigned int cpu)
2038 {
2039         struct device *dev = get_cpu_device(cpu);
2040
2041         dev->offline = false;
2042         /* Tell user space about the state change */
2043         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2044 }
2045
2046 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2047 {
2048         int cpu, ret = 0;
2049
2050         cpu_maps_update_begin();
2051         for_each_online_cpu(cpu) {
2052                 if (topology_is_primary_thread(cpu))
2053                         continue;
2054                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2055                 if (ret)
2056                         break;
2057                 /*
2058                  * As this needs to hold the cpu maps lock it's impossible
2059                  * to call device_offline() because that ends up calling
2060                  * cpu_down() which takes cpu maps lock. cpu maps lock
2061                  * needs to be held as this might race against in kernel
2062                  * abusers of the hotplug machinery (thermal management).
2063                  *
2064                  * So nothing would update device:offline state. That would
2065                  * leave the sysfs entry stale and prevent onlining after
2066                  * smt control has been changed to 'off' again. This is
2067                  * called under the sysfs hotplug lock, so it is properly
2068                  * serialized against the regular offline usage.
2069                  */
2070                 cpuhp_offline_cpu_device(cpu);
2071         }
2072         if (!ret)
2073                 cpu_smt_control = ctrlval;
2074         cpu_maps_update_done();
2075         return ret;
2076 }
2077
2078 static int cpuhp_smt_enable(void)
2079 {
2080         int cpu, ret = 0;
2081
2082         cpu_maps_update_begin();
2083         cpu_smt_control = CPU_SMT_ENABLED;
2084         for_each_present_cpu(cpu) {
2085                 /* Skip online CPUs and CPUs on offline nodes */
2086                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2087                         continue;
2088                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2089                 if (ret)
2090                         break;
2091                 /* See comment in cpuhp_smt_disable() */
2092                 cpuhp_online_cpu_device(cpu);
2093         }
2094         cpu_maps_update_done();
2095         return ret;
2096 }
2097
2098 static ssize_t
2099 store_smt_control(struct device *dev, struct device_attribute *attr,
2100                   const char *buf, size_t count)
2101 {
2102         int ctrlval, ret;
2103
2104         if (sysfs_streq(buf, "on"))
2105                 ctrlval = CPU_SMT_ENABLED;
2106         else if (sysfs_streq(buf, "off"))
2107                 ctrlval = CPU_SMT_DISABLED;
2108         else if (sysfs_streq(buf, "forceoff"))
2109                 ctrlval = CPU_SMT_FORCE_DISABLED;
2110         else
2111                 return -EINVAL;
2112
2113         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2114                 return -EPERM;
2115
2116         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2117                 return -ENODEV;
2118
2119         ret = lock_device_hotplug_sysfs();
2120         if (ret)
2121                 return ret;
2122
2123         if (ctrlval != cpu_smt_control) {
2124                 switch (ctrlval) {
2125                 case CPU_SMT_ENABLED:
2126                         ret = cpuhp_smt_enable();
2127                         break;
2128                 case CPU_SMT_DISABLED:
2129                 case CPU_SMT_FORCE_DISABLED:
2130                         ret = cpuhp_smt_disable(ctrlval);
2131                         break;
2132                 }
2133         }
2134
2135         unlock_device_hotplug();
2136         return ret ? ret : count;
2137 }
2138 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2139
2140 static ssize_t
2141 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2142 {
2143         bool active = topology_max_smt_threads() > 1;
2144
2145         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2146 }
2147 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2148
2149 static struct attribute *cpuhp_smt_attrs[] = {
2150         &dev_attr_control.attr,
2151         &dev_attr_active.attr,
2152         NULL
2153 };
2154
2155 static const struct attribute_group cpuhp_smt_attr_group = {
2156         .attrs = cpuhp_smt_attrs,
2157         .name = "smt",
2158         NULL
2159 };
2160
2161 static int __init cpu_smt_state_init(void)
2162 {
2163         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2164                                   &cpuhp_smt_attr_group);
2165 }
2166
2167 #else
2168 static inline int cpu_smt_state_init(void) { return 0; }
2169 #endif
2170
2171 static int __init cpuhp_sysfs_init(void)
2172 {
2173         int cpu, ret;
2174
2175         ret = cpu_smt_state_init();
2176         if (ret)
2177                 return ret;
2178
2179         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2180                                  &cpuhp_cpu_root_attr_group);
2181         if (ret)
2182                 return ret;
2183
2184         for_each_possible_cpu(cpu) {
2185                 struct device *dev = get_cpu_device(cpu);
2186
2187                 if (!dev)
2188                         continue;
2189                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2190                 if (ret)
2191                         return ret;
2192         }
2193         return 0;
2194 }
2195 device_initcall(cpuhp_sysfs_init);
2196 #endif
2197
2198 /*
2199  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2200  * represents all NR_CPUS bits binary values of 1<<nr.
2201  *
2202  * It is used by cpumask_of() to get a constant address to a CPU
2203  * mask value that has a single bit set only.
2204  */
2205
2206 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2207 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2208 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2209 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2210 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2211
2212 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2213
2214         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2215         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2216 #if BITS_PER_LONG > 32
2217         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2218         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2219 #endif
2220 };
2221 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2222
2223 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2224 EXPORT_SYMBOL(cpu_all_bits);
2225
2226 #ifdef CONFIG_INIT_ALL_POSSIBLE
2227 struct cpumask __cpu_possible_mask __read_mostly
2228         = {CPU_BITS_ALL};
2229 #else
2230 struct cpumask __cpu_possible_mask __read_mostly;
2231 #endif
2232 EXPORT_SYMBOL(__cpu_possible_mask);
2233
2234 struct cpumask __cpu_online_mask __read_mostly;
2235 EXPORT_SYMBOL(__cpu_online_mask);
2236
2237 struct cpumask __cpu_present_mask __read_mostly;
2238 EXPORT_SYMBOL(__cpu_present_mask);
2239
2240 struct cpumask __cpu_active_mask __read_mostly;
2241 EXPORT_SYMBOL(__cpu_active_mask);
2242
2243 void init_cpu_present(const struct cpumask *src)
2244 {
2245         cpumask_copy(&__cpu_present_mask, src);
2246 }
2247
2248 void init_cpu_possible(const struct cpumask *src)
2249 {
2250         cpumask_copy(&__cpu_possible_mask, src);
2251 }
2252
2253 void init_cpu_online(const struct cpumask *src)
2254 {
2255         cpumask_copy(&__cpu_online_mask, src);
2256 }
2257
2258 /*
2259  * Activate the first processor.
2260  */
2261 void __init boot_cpu_init(void)
2262 {
2263         int cpu = smp_processor_id();
2264
2265         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2266         set_cpu_online(cpu, true);
2267         set_cpu_active(cpu, true);
2268         set_cpu_present(cpu, true);
2269         set_cpu_possible(cpu, true);
2270
2271 #ifdef CONFIG_SMP
2272         __boot_cpu_id = cpu;
2273 #endif
2274 }
2275
2276 /*
2277  * Must be called _AFTER_ setting up the per_cpu areas
2278  */
2279 void __init boot_cpu_hotplug_init(void)
2280 {
2281 #ifdef CONFIG_SMP
2282         this_cpu_write(cpuhp_state.booted_once, true);
2283 #endif
2284         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2285 }