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 void inline cpuhp_lock_acquire(bool bringup)
  85 {
  86         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  87 }
  88
  89 static void inline 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 void inline cpuhp_lock_acquire(bool bringup) { }
  96 static void inline 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_bp_states[];
 129 static struct cpuhp_step cpuhp_ap_states[];
 130
 131 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 132 {
 133         /*
 134          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 135          * purposes as that state is handled explicitly in cpu_down.
 136          */
 137         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 138 }
 139
 140 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 141 {
 142         struct cpuhp_step *sp;
 143
 144         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 145         return sp + state;
 146 }
 147
 148 /**
 149  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 150  * @cpu:        The cpu for which the callback should be invoked
 151  * @state:      The state to do callbacks for
 152  * @bringup:    True if the bringup callback should be invoked
 153  * @node:       For multi-instance, do a single entry callback for install/remove
 154  * @lastp:      For multi-instance rollback, remember how far we got
 155  *
 156  * Called from cpu hotplug and from the state register machinery.
 157  */
 158 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 159                                  bool bringup, struct hlist_node *node,
 160                                  struct hlist_node **lastp)
 161 {
 162         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 163         struct cpuhp_step *step = cpuhp_get_step(state);
 164         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 165         int (*cb)(unsigned int cpu);
 166         int ret, cnt;
 167
 168         if (st->fail == state) {
 169                 st->fail = CPUHP_INVALID;
 170
 171                 if (!(bringup ? step->startup.single : step->teardown.single))
 172                         return 0;
 173
 174                 return -EAGAIN;
 175         }
 176
 177         if (!step->multi_instance) {
 178                 WARN_ON_ONCE(lastp && *lastp);
 179                 cb = bringup ? step->startup.single : step->teardown.single;
 180                 if (!cb)
 181                         return 0;
 182                 trace_cpuhp_enter(cpu, st->target, state, cb);
 183                 ret = cb(cpu);
 184                 trace_cpuhp_exit(cpu, st->state, state, ret);
 185                 return ret;
 186         }
 187         cbm = bringup ? step->startup.multi : step->teardown.multi;
 188         if (!cbm)
 189                 return 0;
 190
 191         /* Single invocation for instance add/remove */
 192         if (node) {
 193                 WARN_ON_ONCE(lastp && *lastp);
 194                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 195                 ret = cbm(cpu, node);
 196                 trace_cpuhp_exit(cpu, st->state, state, ret);
 197                 return ret;
 198         }
 199
 200         /* State transition. Invoke on all instances */
 201         cnt = 0;
 202         hlist_for_each(node, &step->list) {
 203                 if (lastp && node == *lastp)
 204                         break;
 205
 206                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 207                 ret = cbm(cpu, node);
 208                 trace_cpuhp_exit(cpu, st->state, state, ret);
 209                 if (ret) {
 210                         if (!lastp)
 211                                 goto err;
 212
 213                         *lastp = node;
 214                         return ret;
 215                 }
 216                 cnt++;
 217         }
 218         if (lastp)
 219                 *lastp = NULL;
 220         return 0;
 221 err:
 222         /* Rollback the instances if one failed */
 223         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 224         if (!cbm)
 225                 return ret;
 226
 227         hlist_for_each(node, &step->list) {
 228                 if (!cnt--)
 229                         break;
 230
 231                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 232                 ret = cbm(cpu, node);
 233                 trace_cpuhp_exit(cpu, st->state, state, ret);
 234                 /*
 235                  * Rollback must not fail,
 236                  */
 237                 WARN_ON_ONCE(ret);
 238         }
 239         return ret;
 240 }
 241
 242 #ifdef CONFIG_SMP
 243 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 244 {
 245         struct completion *done = bringup ? &st->done_up : &st->done_down;
 246         wait_for_completion(done);
 247 }
 248
 249 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 250 {
 251         struct completion *done = bringup ? &st->done_up : &st->done_down;
 252         complete(done);
 253 }
 254
 255 /*
 256  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 257  */
 258 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 259 {
 260         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 261 }
 262
 263 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 264 static DEFINE_MUTEX(cpu_add_remove_lock);
 265 bool cpuhp_tasks_frozen;
 266 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 267
 268 /*
 269  * The following two APIs (cpu_maps_update_begin/done) must be used when
 270  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 271  */
 272 void cpu_maps_update_begin(void)
 273 {
 274         mutex_lock(&cpu_add_remove_lock);
 275 }
 276
 277 void cpu_maps_update_done(void)
 278 {
 279         mutex_unlock(&cpu_add_remove_lock);
 280 }
 281
 282 /*
 283  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 284  * Should always be manipulated under cpu_add_remove_lock
 285  */
 286 static int cpu_hotplug_disabled;
 287
 288 #ifdef CONFIG_HOTPLUG_CPU
 289
 290 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 291
 292 void cpus_read_lock(void)
 293 {
 294         percpu_down_read(&cpu_hotplug_lock);
 295 }
 296 EXPORT_SYMBOL_GPL(cpus_read_lock);
 297
 298 void cpus_read_unlock(void)
 299 {
 300         percpu_up_read(&cpu_hotplug_lock);
 301 }
 302 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 303
 304 void cpus_write_lock(void)
 305 {
 306         percpu_down_write(&cpu_hotplug_lock);
 307 }
 308
 309 void cpus_write_unlock(void)
 310 {
 311         percpu_up_write(&cpu_hotplug_lock);
 312 }
 313
 314 void lockdep_assert_cpus_held(void)
 315 {
 316         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 317 }
 318
 319 /*
 320  * Wait for currently running CPU hotplug operations to complete (if any) and
 321  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 322  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 323  * hotplug path before performing hotplug operations. So acquiring that lock
 324  * guarantees mutual exclusion from any currently running hotplug operations.
 325  */
 326 void cpu_hotplug_disable(void)
 327 {
 328         cpu_maps_update_begin();
 329         cpu_hotplug_disabled++;
 330         cpu_maps_update_done();
 331 }
 332 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 333
 334 static void __cpu_hotplug_enable(void)
 335 {
 336         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 337                 return;
 338         cpu_hotplug_disabled--;
 339 }
 340
 341 void cpu_hotplug_enable(void)
 342 {
 343         cpu_maps_update_begin();
 344         __cpu_hotplug_enable();
 345         cpu_maps_update_done();
 346 }
 347 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 348 #endif  /* CONFIG_HOTPLUG_CPU */
 349
 350 #ifdef CONFIG_HOTPLUG_SMT
 351 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 352 EXPORT_SYMBOL_GPL(cpu_smt_control);
 353
 354 static bool cpu_smt_available __read_mostly;
 355
 356 void __init cpu_smt_disable(bool force)
 357 {
 358         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 359                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 360                 return;
 361
 362         if (force) {
 363                 pr_info("SMT: Force disabled\n");
 364                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 365         } else {
 366                 cpu_smt_control = CPU_SMT_DISABLED;
 367         }
 368 }
 369
 370 /*
 371  * The decision whether SMT is supported can only be done after the full
 372  * CPU identification. Called from architecture code before non boot CPUs
 373  * are brought up.
 374  */
 375 void __init cpu_smt_check_topology_early(void)
 376 {
 377         if (!topology_smt_supported())
 378                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 379 }
 380
 381 /*
 382  * If SMT was disabled by BIOS, detect it here, after the CPUs have been
 383  * brought online. This ensures the smt/l1tf sysfs entries are consistent
 384  * with reality. cpu_smt_available is set to true during the bringup of non
 385  * boot CPUs when a SMT sibling is detected. Note, this may overwrite
 386  * cpu_smt_control's previous setting.
 387  */
 388 void __init cpu_smt_check_topology(void)
 389 {
 390         if (!cpu_smt_available)
 391                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 392 }
 393
 394 static int __init smt_cmdline_disable(char *str)
 395 {
 396         cpu_smt_disable(str && !strcmp(str, "force"));
 397         return 0;
 398 }
 399 early_param("nosmt", smt_cmdline_disable);
 400
 401 static inline bool cpu_smt_allowed(unsigned int cpu)
 402 {
 403         if (topology_is_primary_thread(cpu))
 404                 return true;
 405
 406         /*
 407          * If the CPU is not a 'primary' thread and the booted_once bit is
 408          * set then the processor has SMT support. Store this information
 409          * for the late check of SMT support in cpu_smt_check_topology().
 410          */
 411         if (per_cpu(cpuhp_state, cpu).booted_once)
 412                 cpu_smt_available = true;
 413
 414         if (cpu_smt_control == CPU_SMT_ENABLED)
 415                 return true;
 416
 417         /*
 418          * On x86 it's required to boot all logical CPUs at least once so
 419          * that the init code can get a chance to set CR4.MCE on each
 420          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 421          * core will shutdown the machine.
 422          */
 423         return !per_cpu(cpuhp_state, cpu).booted_once;
 424 }
 425 #else
 426 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 427 #endif
 428
 429 static inline enum cpuhp_state
 430 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 431 {
 432         enum cpuhp_state prev_state = st->state;
 433
 434         st->rollback = false;
 435         st->last = NULL;
 436
 437         st->target = target;
 438         st->single = false;
 439         st->bringup = st->state < target;
 440
 441         return prev_state;
 442 }
 443
 444 static inline void
 445 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 446 {
 447         st->rollback = true;
 448
 449         /*
 450          * If we have st->last we need to undo partial multi_instance of this
 451          * state first. Otherwise start undo at the previous state.
 452          */
 453         if (!st->last) {
 454                 if (st->bringup)
 455                         st->state--;
 456                 else
 457                         st->state++;
 458         }
 459
 460         st->target = prev_state;
 461         st->bringup = !st->bringup;
 462 }
 463
 464 /* Regular hotplug invocation of the AP hotplug thread */
 465 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 466 {
 467         if (!st->single && st->state == st->target)
 468                 return;
 469
 470         st->result = 0;
 471         /*
 472          * Make sure the above stores are visible before should_run becomes
 473          * true. Paired with the mb() above in cpuhp_thread_fun()
 474          */
 475         smp_mb();
 476         st->should_run = true;
 477         wake_up_process(st->thread);
 478         wait_for_ap_thread(st, st->bringup);
 479 }
 480
 481 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 482 {
 483         enum cpuhp_state prev_state;
 484         int ret;
 485
 486         prev_state = cpuhp_set_state(st, target);
 487         __cpuhp_kick_ap(st);
 488         if ((ret = st->result)) {
 489                 cpuhp_reset_state(st, prev_state);
 490                 __cpuhp_kick_ap(st);
 491         }
 492
 493         return ret;
 494 }
 495
 496 static int bringup_wait_for_ap(unsigned int cpu)
 497 {
 498         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 499
 500         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 501         wait_for_ap_thread(st, true);
 502         if (WARN_ON_ONCE((!cpu_online(cpu))))
 503                 return -ECANCELED;
 504
 505         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 506         stop_machine_unpark(cpu);
 507         kthread_unpark(st->thread);
 508
 509         /*
 510          * SMT soft disabling on X86 requires to bring the CPU out of the
 511          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 512          * CPU marked itself as booted_once in cpu_notify_starting() so the
 513          * cpu_smt_allowed() check will now return false if this is not the
 514          * primary sibling.
 515          */
 516         if (!cpu_smt_allowed(cpu))
 517                 return -ECANCELED;
 518
 519         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 520                 return 0;
 521
 522         return cpuhp_kick_ap(st, st->target);
 523 }
 524
 525 static int bringup_cpu(unsigned int cpu)
 526 {
 527         struct task_struct *idle = idle_thread_get(cpu);
 528         int ret;
 529
 530         /*
 531          * Some architectures have to walk the irq descriptors to
 532          * setup the vector space for the cpu which comes online.
 533          * Prevent irq alloc/free across the bringup.
 534          */
 535         irq_lock_sparse();
 536
 537         /* Arch-specific enabling code. */
 538         ret = __cpu_up(cpu, idle);
 539         irq_unlock_sparse();
 540         if (ret)
 541                 return ret;
 542         return bringup_wait_for_ap(cpu);
 543 }
 544
 545 /*
 546  * Hotplug state machine related functions
 547  */
 548
 549 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 550 {
 551         for (st->state--; st->state > st->target; st->state--) {
 552                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 553
 554                 if (!step->skip_onerr)
 555                         cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 556         }
 557 }
 558
 559 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 560                               enum cpuhp_state target)
 561 {
 562         enum cpuhp_state prev_state = st->state;
 563         int ret = 0;
 564
 565         while (st->state < target) {
 566                 st->state++;
 567                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 568                 if (ret) {
 569                         st->target = prev_state;
 570                         undo_cpu_up(cpu, st);
 571                         break;
 572                 }
 573         }
 574         return ret;
 575 }
 576
 577 /*
 578  * The cpu hotplug threads manage the bringup and teardown of the cpus
 579  */
 580 static void cpuhp_create(unsigned int cpu)
 581 {
 582         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 583
 584         init_completion(&st->done_up);
 585         init_completion(&st->done_down);
 586 }
 587
 588 static int cpuhp_should_run(unsigned int cpu)
 589 {
 590         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 591
 592         return st->should_run;
 593 }
 594
 595 /*
 596  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 597  * callbacks when a state gets [un]installed at runtime.
 598  *
 599  * Each invocation of this function by the smpboot thread does a single AP
 600  * state callback.
 601  *
 602  * It has 3 modes of operation:
 603  *  - single: runs st->cb_state
 604  *  - up:     runs ++st->state, while st->state < st->target
 605  *  - down:   runs st->state--, while st->state > st->target
 606  *
 607  * When complete or on error, should_run is cleared and the completion is fired.
 608  */
 609 static void cpuhp_thread_fun(unsigned int cpu)
 610 {
 611         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 612         bool bringup = st->bringup;
 613         enum cpuhp_state state;
 614
 615         /*
 616          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 617          * that if we see ->should_run we also see the rest of the state.
 618          */
 619         smp_mb();
 620
 621         if (WARN_ON_ONCE(!st->should_run))
 622                 return;
 623
 624         cpuhp_lock_acquire(bringup);
 625
 626         if (st->single) {
 627                 state = st->cb_state;
 628                 st->should_run = false;
 629         } else {
 630                 if (bringup) {
 631                         st->state++;
 632                         state = st->state;
 633                         st->should_run = (st->state < st->target);
 634                         WARN_ON_ONCE(st->state > st->target);
 635                 } else {
 636                         state = st->state;
 637                         st->state--;
 638                         st->should_run = (st->state > st->target);
 639                         WARN_ON_ONCE(st->state < st->target);
 640                 }
 641         }
 642
 643         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 644
 645         if (st->rollback) {
 646                 struct cpuhp_step *step = cpuhp_get_step(state);
 647                 if (step->skip_onerr)
 648                         goto next;
 649         }
 650
 651         if (cpuhp_is_atomic_state(state)) {
 652                 local_irq_disable();
 653                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 654                 local_irq_enable();
 655
 656                 /*
 657                  * STARTING/DYING must not fail!
 658                  */
 659                 WARN_ON_ONCE(st->result);
 660         } else {
 661                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 662         }
 663
 664         if (st->result) {
 665                 /*
 666                  * If we fail on a rollback, we're up a creek without no
 667                  * paddle, no way forward, no way back. We loose, thanks for
 668                  * playing.
 669                  */
 670                 WARN_ON_ONCE(st->rollback);
 671                 st->should_run = false;
 672         }
 673
 674 next:
 675         cpuhp_lock_release(bringup);
 676
 677         if (!st->should_run)
 678                 complete_ap_thread(st, bringup);
 679 }
 680
 681 /* Invoke a single callback on a remote cpu */
 682 static int
 683 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 684                          struct hlist_node *node)
 685 {
 686         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 687         int ret;
 688
 689         if (!cpu_online(cpu))
 690                 return 0;
 691
 692         cpuhp_lock_acquire(false);
 693         cpuhp_lock_release(false);
 694
 695         cpuhp_lock_acquire(true);
 696         cpuhp_lock_release(true);
 697
 698         /*
 699          * If we are up and running, use the hotplug thread. For early calls
 700          * we invoke the thread function directly.
 701          */
 702         if (!st->thread)
 703                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 704
 705         st->rollback = false;
 706         st->last = NULL;
 707
 708         st->node = node;
 709         st->bringup = bringup;
 710         st->cb_state = state;
 711         st->single = true;
 712
 713         __cpuhp_kick_ap(st);
 714
 715         /*
 716          * If we failed and did a partial, do a rollback.
 717          */
 718         if ((ret = st->result) && st->last) {
 719                 st->rollback = true;
 720                 st->bringup = !bringup;
 721
 722                 __cpuhp_kick_ap(st);
 723         }
 724
 725         /*
 726          * Clean up the leftovers so the next hotplug operation wont use stale
 727          * data.
 728          */
 729         st->node = st->last = NULL;
 730         return ret;
 731 }
 732
 733 static int cpuhp_kick_ap_work(unsigned int cpu)
 734 {
 735         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 736         enum cpuhp_state prev_state = st->state;
 737         int ret;
 738
 739         cpuhp_lock_acquire(false);
 740         cpuhp_lock_release(false);
 741
 742         cpuhp_lock_acquire(true);
 743         cpuhp_lock_release(true);
 744
 745         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 746         ret = cpuhp_kick_ap(st, st->target);
 747         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 748
 749         return ret;
 750 }
 751
 752 static struct smp_hotplug_thread cpuhp_threads = {
 753         .store                  = &cpuhp_state.thread,
 754         .create                 = &cpuhp_create,
 755         .thread_should_run      = cpuhp_should_run,
 756         .thread_fn              = cpuhp_thread_fun,
 757         .thread_comm            = "cpuhp/%u",
 758         .selfparking            = true,
 759 };
 760
 761 void __init cpuhp_threads_init(void)
 762 {
 763         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 764         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 765 }
 766
 767 #ifdef CONFIG_HOTPLUG_CPU
 768 /**
 769  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 770  * @cpu: a CPU id
 771  *
 772  * This function walks all processes, finds a valid mm struct for each one and
 773  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 774  * trivial, there are various non-obvious corner cases, which this function
 775  * tries to solve in a safe manner.
 776  *
 777  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 778  * be called only for an already offlined CPU.
 779  */
 780 void clear_tasks_mm_cpumask(int cpu)
 781 {
 782         struct task_struct *p;
 783
 784         /*
 785          * This function is called after the cpu is taken down and marked
 786          * offline, so its not like new tasks will ever get this cpu set in
 787          * their mm mask. -- Peter Zijlstra
 788          * Thus, we may use rcu_read_lock() here, instead of grabbing
 789          * full-fledged tasklist_lock.
 790          */
 791         WARN_ON(cpu_online(cpu));
 792         rcu_read_lock();
 793         for_each_process(p) {
 794                 struct task_struct *t;
 795
 796                 /*
 797                  * Main thread might exit, but other threads may still have
 798                  * a valid mm. Find one.
 799                  */
 800                 t = find_lock_task_mm(p);
 801                 if (!t)
 802                         continue;
 803                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 804                 task_unlock(t);
 805         }
 806         rcu_read_unlock();
 807 }
 808
 809 /* Take this CPU down. */
 810 static int take_cpu_down(void *_param)
 811 {
 812         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 813         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 814         int err, cpu = smp_processor_id();
 815         int ret;
 816
 817         /* Ensure this CPU doesn't handle any more interrupts. */
 818         err = __cpu_disable();
 819         if (err < 0)
 820                 return err;
 821
 822         /*
 823          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 824          * do this step again.
 825          */
 826         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 827         st->state--;
 828         /* Invoke the former CPU_DYING callbacks */
 829         for (; st->state > target; st->state--) {
 830                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 831                 /*
 832                  * DYING must not fail!
 833                  */
 834                 WARN_ON_ONCE(ret);
 835         }
 836
 837         /* Give up timekeeping duties */
 838         tick_handover_do_timer();
 839         /* Park the stopper thread */
 840         stop_machine_park(cpu);
 841         return 0;
 842 }
 843
 844 static int takedown_cpu(unsigned int cpu)
 845 {
 846         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 847         int err;
 848
 849         /* Park the smpboot threads */
 850         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 851
 852         /*
 853          * Prevent irq alloc/free while the dying cpu reorganizes the
 854          * interrupt affinities.
 855          */
 856         irq_lock_sparse();
 857
 858         /*
 859          * So now all preempt/rcu users must observe !cpu_active().
 860          */
 861         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 862         if (err) {
 863                 /* CPU refused to die */
 864                 irq_unlock_sparse();
 865                 /* Unpark the hotplug thread so we can rollback there */
 866                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 867                 return err;
 868         }
 869         BUG_ON(cpu_online(cpu));
 870
 871         /*
 872          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 873          * runnable tasks from the cpu, there's only the idle task left now
 874          * that the migration thread is done doing the stop_machine thing.
 875          *
 876          * Wait for the stop thread to go away.
 877          */
 878         wait_for_ap_thread(st, false);
 879         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 880
 881         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 882         irq_unlock_sparse();
 883
 884         hotplug_cpu__broadcast_tick_pull(cpu);
 885         /* This actually kills the CPU. */
 886         __cpu_die(cpu);
 887
 888         tick_cleanup_dead_cpu(cpu);
 889         rcutree_migrate_callbacks(cpu);
 890         return 0;
 891 }
 892
 893 static void cpuhp_complete_idle_dead(void *arg)
 894 {
 895         struct cpuhp_cpu_state *st = arg;
 896
 897         complete_ap_thread(st, false);
 898 }
 899
 900 void cpuhp_report_idle_dead(void)
 901 {
 902         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 903
 904         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 905         rcu_report_dead(smp_processor_id());
 906         st->state = CPUHP_AP_IDLE_DEAD;
 907         /*
 908          * We cannot call complete after rcu_report_dead() so we delegate it
 909          * to an online cpu.
 910          */
 911         smp_call_function_single(cpumask_first(cpu_online_mask),
 912                                  cpuhp_complete_idle_dead, st, 0);
 913 }
 914
 915 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 916 {
 917         for (st->state++; st->state < st->target; st->state++) {
 918                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 919
 920                 if (!step->skip_onerr)
 921                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 922         }
 923 }
 924
 925 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 926                                 enum cpuhp_state target)
 927 {
 928         enum cpuhp_state prev_state = st->state;
 929         int ret = 0;
 930
 931         for (; st->state > target; st->state--) {
 932                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 933                 if (ret) {
 934                         st->target = prev_state;
 935                         undo_cpu_down(cpu, st);
 936                         break;
 937                 }
 938         }
 939         return ret;
 940 }
 941
 942 /* Requires cpu_add_remove_lock to be held */
 943 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 944                            enum cpuhp_state target)
 945 {
 946         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 947         int prev_state, ret = 0;
 948
 949         if (num_online_cpus() == 1)
 950                 return -EBUSY;
 951
 952         if (!cpu_present(cpu))
 953                 return -EINVAL;
 954
 955         cpus_write_lock();
 956
 957         cpuhp_tasks_frozen = tasks_frozen;
 958
 959         prev_state = cpuhp_set_state(st, target);
 960         /*
 961          * If the current CPU state is in the range of the AP hotplug thread,
 962          * then we need to kick the thread.
 963          */
 964         if (st->state > CPUHP_TEARDOWN_CPU) {
 965                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 966                 ret = cpuhp_kick_ap_work(cpu);
 967                 /*
 968                  * The AP side has done the error rollback already. Just
 969                  * return the error code..
 970                  */
 971                 if (ret)
 972                         goto out;
 973
 974                 /*
 975                  * We might have stopped still in the range of the AP hotplug
 976                  * thread. Nothing to do anymore.
 977                  */
 978                 if (st->state > CPUHP_TEARDOWN_CPU)
 979                         goto out;
 980
 981                 st->target = target;
 982         }
 983         /*
 984          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 985          * to do the further cleanups.
 986          */
 987         ret = cpuhp_down_callbacks(cpu, st, target);
 988         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 989                 cpuhp_reset_state(st, prev_state);
 990                 __cpuhp_kick_ap(st);
 991         }
 992
 993 out:
 994         cpus_write_unlock();
 995         /*
 996          * Do post unplug cleanup. This is still protected against
 997          * concurrent CPU hotplug via cpu_add_remove_lock.
 998          */
 999         lockup_detector_cleanup();
1000         return ret;
1001 }
1002
1003 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1004 {
1005         if (cpu_hotplug_disabled)
1006                 return -EBUSY;
1007         return _cpu_down(cpu, 0, target);
1008 }
1009
1010 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1011 {
1012         int err;
1013
1014         cpu_maps_update_begin();
1015         err = cpu_down_maps_locked(cpu, target);
1016         cpu_maps_update_done();
1017         return err;
1018 }
1019
1020 int cpu_down(unsigned int cpu)
1021 {
1022         return do_cpu_down(cpu, CPUHP_OFFLINE);
1023 }
1024 EXPORT_SYMBOL(cpu_down);
1025
1026 #else
1027 #define takedown_cpu            NULL
1028 #endif /*CONFIG_HOTPLUG_CPU*/
1029
1030 /**
1031  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1032  * @cpu: cpu that just started
1033  *
1034  * It must be called by the arch code on the new cpu, before the new cpu
1035  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1036  */
1037 void notify_cpu_starting(unsigned int cpu)
1038 {
1039         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1040         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1041         int ret;
1042
1043         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1044         st->booted_once = true;
1045         while (st->state < target) {
1046                 st->state++;
1047                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1048                 /*
1049                  * STARTING must not fail!
1050                  */
1051                 WARN_ON_ONCE(ret);
1052         }
1053 }
1054
1055 /*
1056  * Called from the idle task. Wake up the controlling task which brings the
1057  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1058  * the rest of the online bringup to the hotplug thread.
1059  */
1060 void cpuhp_online_idle(enum cpuhp_state state)
1061 {
1062         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1063
1064         /* Happens for the boot cpu */
1065         if (state != CPUHP_AP_ONLINE_IDLE)
1066                 return;
1067
1068         st->state = CPUHP_AP_ONLINE_IDLE;
1069         complete_ap_thread(st, true);
1070 }
1071
1072 /* Requires cpu_add_remove_lock to be held */
1073 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1074 {
1075         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1076         struct task_struct *idle;
1077         int ret = 0;
1078
1079         cpus_write_lock();
1080
1081         if (!cpu_present(cpu)) {
1082                 ret = -EINVAL;
1083                 goto out;
1084         }
1085
1086         /*
1087          * The caller of do_cpu_up might have raced with another
1088          * caller. Ignore it for now.
1089          */
1090         if (st->state >= target)
1091                 goto out;
1092
1093         if (st->state == CPUHP_OFFLINE) {
1094                 /* Let it fail before we try to bring the cpu up */
1095                 idle = idle_thread_get(cpu);
1096                 if (IS_ERR(idle)) {
1097                         ret = PTR_ERR(idle);
1098                         goto out;
1099                 }
1100         }
1101
1102         cpuhp_tasks_frozen = tasks_frozen;
1103
1104         cpuhp_set_state(st, target);
1105         /*
1106          * If the current CPU state is in the range of the AP hotplug thread,
1107          * then we need to kick the thread once more.
1108          */
1109         if (st->state > CPUHP_BRINGUP_CPU) {
1110                 ret = cpuhp_kick_ap_work(cpu);
1111                 /*
1112                  * The AP side has done the error rollback already. Just
1113                  * return the error code..
1114                  */
1115                 if (ret)
1116                         goto out;
1117         }
1118
1119         /*
1120          * Try to reach the target state. We max out on the BP at
1121          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1122          * responsible for bringing it up to the target state.
1123          */
1124         target = min((int)target, CPUHP_BRINGUP_CPU);
1125         ret = cpuhp_up_callbacks(cpu, st, target);
1126 out:
1127         cpus_write_unlock();
1128         return ret;
1129 }
1130
1131 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1132 {
1133         int err = 0;
1134
1135         if (!cpu_possible(cpu)) {
1136                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1137                        cpu);
1138 #if defined(CONFIG_IA64)
1139                 pr_err("please check additional_cpus= boot parameter\n");
1140 #endif
1141                 return -EINVAL;
1142         }
1143
1144         err = try_online_node(cpu_to_node(cpu));
1145         if (err)
1146                 return err;
1147
1148         cpu_maps_update_begin();
1149
1150         if (cpu_hotplug_disabled) {
1151                 err = -EBUSY;
1152                 goto out;
1153         }
1154         if (!cpu_smt_allowed(cpu)) {
1155                 err = -EPERM;
1156                 goto out;
1157         }
1158
1159         err = _cpu_up(cpu, 0, target);
1160 out:
1161         cpu_maps_update_done();
1162         return err;
1163 }
1164
1165 int cpu_up(unsigned int cpu)
1166 {
1167         return do_cpu_up(cpu, CPUHP_ONLINE);
1168 }
1169 EXPORT_SYMBOL_GPL(cpu_up);
1170
1171 #ifdef CONFIG_PM_SLEEP_SMP
1172 static cpumask_var_t frozen_cpus;
1173
1174 int freeze_secondary_cpus(int primary)
1175 {
1176         int cpu, error = 0;
1177
1178         cpu_maps_update_begin();
1179         if (!cpu_online(primary))
1180                 primary = cpumask_first(cpu_online_mask);
1181         /*
1182          * We take down all of the non-boot CPUs in one shot to avoid races
1183          * with the userspace trying to use the CPU hotplug at the same time
1184          */
1185         cpumask_clear(frozen_cpus);
1186
1187         pr_info("Disabling non-boot CPUs ...\n");
1188         for_each_online_cpu(cpu) {
1189                 if (cpu == primary)
1190                         continue;
1191                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1192                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1193                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1194                 if (!error)
1195                         cpumask_set_cpu(cpu, frozen_cpus);
1196                 else {
1197                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1198                         break;
1199                 }
1200         }
1201
1202         if (!error)
1203                 BUG_ON(num_online_cpus() > 1);
1204         else
1205                 pr_err("Non-boot CPUs are not disabled\n");
1206
1207         /*
1208          * Make sure the CPUs won't be enabled by someone else. We need to do
1209          * this even in case of failure as all disable_nonboot_cpus() users are
1210          * supposed to do enable_nonboot_cpus() on the failure path.
1211          */
1212         cpu_hotplug_disabled++;
1213
1214         cpu_maps_update_done();
1215         return error;
1216 }
1217
1218 void __weak arch_enable_nonboot_cpus_begin(void)
1219 {
1220 }
1221
1222 void __weak arch_enable_nonboot_cpus_end(void)
1223 {
1224 }
1225
1226 void enable_nonboot_cpus(void)
1227 {
1228         int cpu, error;
1229
1230         /* Allow everyone to use the CPU hotplug again */
1231         cpu_maps_update_begin();
1232         __cpu_hotplug_enable();
1233         if (cpumask_empty(frozen_cpus))
1234                 goto out;
1235
1236         pr_info("Enabling non-boot CPUs ...\n");
1237
1238         arch_enable_nonboot_cpus_begin();
1239
1240         for_each_cpu(cpu, frozen_cpus) {
1241                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1242                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1243                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1244                 if (!error) {
1245                         pr_info("CPU%d is up\n", cpu);
1246                         continue;
1247                 }
1248                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1249         }
1250
1251         arch_enable_nonboot_cpus_end();
1252
1253         cpumask_clear(frozen_cpus);
1254 out:
1255         cpu_maps_update_done();
1256 }
1257
1258 static int __init alloc_frozen_cpus(void)
1259 {
1260         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1261                 return -ENOMEM;
1262         return 0;
1263 }
1264 core_initcall(alloc_frozen_cpus);
1265
1266 /*
1267  * When callbacks for CPU hotplug notifications are being executed, we must
1268  * ensure that the state of the system with respect to the tasks being frozen
1269  * or not, as reported by the notification, remains unchanged *throughout the
1270  * duration* of the execution of the callbacks.
1271  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1272  *
1273  * This synchronization is implemented by mutually excluding regular CPU
1274  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1275  * Hibernate notifications.
1276  */
1277 static int
1278 cpu_hotplug_pm_callback(struct notifier_block *nb,
1279                         unsigned long action, void *ptr)
1280 {
1281         switch (action) {
1282
1283         case PM_SUSPEND_PREPARE:
1284         case PM_HIBERNATION_PREPARE:
1285                 cpu_hotplug_disable();
1286                 break;
1287
1288         case PM_POST_SUSPEND:
1289         case PM_POST_HIBERNATION:
1290                 cpu_hotplug_enable();
1291                 break;
1292
1293         default:
1294                 return NOTIFY_DONE;
1295         }
1296
1297         return NOTIFY_OK;
1298 }
1299
1300
1301 static int __init cpu_hotplug_pm_sync_init(void)
1302 {
1303         /*
1304          * cpu_hotplug_pm_callback has higher priority than x86
1305          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1306          * to disable cpu hotplug to avoid cpu hotplug race.
1307          */
1308         pm_notifier(cpu_hotplug_pm_callback, 0);
1309         return 0;
1310 }
1311 core_initcall(cpu_hotplug_pm_sync_init);
1312
1313 #endif /* CONFIG_PM_SLEEP_SMP */
1314
1315 int __boot_cpu_id;
1316
1317 #endif /* CONFIG_SMP */
1318
1319 /* Boot processor state steps */
1320 static struct cpuhp_step cpuhp_bp_states[] = {
1321         [CPUHP_OFFLINE] = {
1322                 .name                   = "offline",
1323                 .startup.single         = NULL,
1324                 .teardown.single        = NULL,
1325         },
1326 #ifdef CONFIG_SMP
1327         [CPUHP_CREATE_THREADS]= {
1328                 .name                   = "threads:prepare",
1329                 .startup.single         = smpboot_create_threads,
1330                 .teardown.single        = NULL,
1331                 .cant_stop              = true,
1332         },
1333         [CPUHP_PERF_PREPARE] = {
1334                 .name                   = "perf:prepare",
1335                 .startup.single         = perf_event_init_cpu,
1336                 .teardown.single        = perf_event_exit_cpu,
1337         },
1338         [CPUHP_WORKQUEUE_PREP] = {
1339                 .name                   = "workqueue:prepare",
1340                 .startup.single         = workqueue_prepare_cpu,
1341                 .teardown.single        = NULL,
1342         },
1343         [CPUHP_HRTIMERS_PREPARE] = {
1344                 .name                   = "hrtimers:prepare",
1345                 .startup.single         = hrtimers_prepare_cpu,
1346                 .teardown.single        = hrtimers_dead_cpu,
1347         },
1348         [CPUHP_SMPCFD_PREPARE] = {
1349                 .name                   = "smpcfd:prepare",
1350                 .startup.single         = smpcfd_prepare_cpu,
1351                 .teardown.single        = smpcfd_dead_cpu,
1352         },
1353         [CPUHP_RELAY_PREPARE] = {
1354                 .name                   = "relay:prepare",
1355                 .startup.single         = relay_prepare_cpu,
1356                 .teardown.single        = NULL,
1357         },
1358         [CPUHP_SLAB_PREPARE] = {
1359                 .name                   = "slab:prepare",
1360                 .startup.single         = slab_prepare_cpu,
1361                 .teardown.single        = slab_dead_cpu,
1362         },
1363         [CPUHP_RCUTREE_PREP] = {
1364                 .name                   = "RCU/tree:prepare",
1365                 .startup.single         = rcutree_prepare_cpu,
1366                 .teardown.single        = rcutree_dead_cpu,
1367         },
1368         /*
1369          * On the tear-down path, timers_dead_cpu() must be invoked
1370          * before blk_mq_queue_reinit_notify() from notify_dead(),
1371          * otherwise a RCU stall occurs.
1372          */
1373         [CPUHP_TIMERS_PREPARE] = {
1374                 .name                   = "timers:dead",
1375                 .startup.single         = timers_prepare_cpu,
1376                 .teardown.single        = timers_dead_cpu,
1377         },
1378         /* Kicks the plugged cpu into life */
1379         [CPUHP_BRINGUP_CPU] = {
1380                 .name                   = "cpu:bringup",
1381                 .startup.single         = bringup_cpu,
1382                 .teardown.single        = NULL,
1383                 .cant_stop              = true,
1384         },
1385         /*
1386          * Handled on controll processor until the plugged processor manages
1387          * this itself.
1388          */
1389         [CPUHP_TEARDOWN_CPU] = {
1390                 .name                   = "cpu:teardown",
1391                 .startup.single         = NULL,
1392                 .teardown.single        = takedown_cpu,
1393                 .cant_stop              = true,
1394         },
1395 #else
1396         [CPUHP_BRINGUP_CPU] = { },
1397 #endif
1398 };
1399
1400 /* Application processor state steps */
1401 static struct cpuhp_step cpuhp_ap_states[] = {
1402 #ifdef CONFIG_SMP
1403         /* Final state before CPU kills itself */
1404         [CPUHP_AP_IDLE_DEAD] = {
1405                 .name                   = "idle:dead",
1406         },
1407         /*
1408          * Last state before CPU enters the idle loop to die. Transient state
1409          * for synchronization.
1410          */
1411         [CPUHP_AP_OFFLINE] = {
1412                 .name                   = "ap:offline",
1413                 .cant_stop              = true,
1414         },
1415         /* First state is scheduler control. Interrupts are disabled */
1416         [CPUHP_AP_SCHED_STARTING] = {
1417                 .name                   = "sched:starting",
1418                 .startup.single         = sched_cpu_starting,
1419                 .teardown.single        = sched_cpu_dying,
1420         },
1421         [CPUHP_AP_RCUTREE_DYING] = {
1422                 .name                   = "RCU/tree:dying",
1423                 .startup.single         = NULL,
1424                 .teardown.single        = rcutree_dying_cpu,
1425         },
1426         [CPUHP_AP_SMPCFD_DYING] = {
1427                 .name                   = "smpcfd:dying",
1428                 .startup.single         = NULL,
1429                 .teardown.single        = smpcfd_dying_cpu,
1430         },
1431         /* Entry state on starting. Interrupts enabled from here on. Transient
1432          * state for synchronsization */
1433         [CPUHP_AP_ONLINE] = {
1434                 .name                   = "ap:online",
1435         },
1436         /* Handle smpboot threads park/unpark */
1437         [CPUHP_AP_SMPBOOT_THREADS] = {
1438                 .name                   = "smpboot/threads:online",
1439                 .startup.single         = smpboot_unpark_threads,
1440                 .teardown.single        = smpboot_park_threads,
1441         },
1442         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1443                 .name                   = "irq/affinity:online",
1444                 .startup.single         = irq_affinity_online_cpu,
1445                 .teardown.single        = NULL,
1446         },
1447         [CPUHP_AP_PERF_ONLINE] = {
1448                 .name                   = "perf:online",
1449                 .startup.single         = perf_event_init_cpu,
1450                 .teardown.single        = perf_event_exit_cpu,
1451         },
1452         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1453                 .name                   = "workqueue:online",
1454                 .startup.single         = workqueue_online_cpu,
1455                 .teardown.single        = workqueue_offline_cpu,
1456         },
1457         [CPUHP_AP_RCUTREE_ONLINE] = {
1458                 .name                   = "RCU/tree:online",
1459                 .startup.single         = rcutree_online_cpu,
1460                 .teardown.single        = rcutree_offline_cpu,
1461         },
1462 #endif
1463         /*
1464          * The dynamically registered state space is here
1465          */
1466
1467 #ifdef CONFIG_SMP
1468         /* Last state is scheduler control setting the cpu active */
1469         [CPUHP_AP_ACTIVE] = {
1470                 .name                   = "sched:active",
1471                 .startup.single         = sched_cpu_activate,
1472                 .teardown.single        = sched_cpu_deactivate,
1473         },
1474 #endif
1475
1476         /* CPU is fully up and running. */
1477         [CPUHP_ONLINE] = {
1478                 .name                   = "online",
1479                 .startup.single         = NULL,
1480                 .teardown.single        = NULL,
1481         },
1482 };
1483
1484 /* Sanity check for callbacks */
1485 static int cpuhp_cb_check(enum cpuhp_state state)
1486 {
1487         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1488                 return -EINVAL;
1489         return 0;
1490 }
1491
1492 /*
1493  * Returns a free for dynamic slot assignment of the Online state. The states
1494  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1495  * by having no name assigned.
1496  */
1497 static int cpuhp_reserve_state(enum cpuhp_state state)
1498 {
1499         enum cpuhp_state i, end;
1500         struct cpuhp_step *step;
1501
1502         switch (state) {
1503         case CPUHP_AP_ONLINE_DYN:
1504                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1505                 end = CPUHP_AP_ONLINE_DYN_END;
1506                 break;
1507         case CPUHP_BP_PREPARE_DYN:
1508                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1509                 end = CPUHP_BP_PREPARE_DYN_END;
1510                 break;
1511         default:
1512                 return -EINVAL;
1513         }
1514
1515         for (i = state; i <= end; i++, step++) {
1516                 if (!step->name)
1517                         return i;
1518         }
1519         WARN(1, "No more dynamic states available for CPU hotplug\n");
1520         return -ENOSPC;
1521 }
1522
1523 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1524                                  int (*startup)(unsigned int cpu),
1525                                  int (*teardown)(unsigned int cpu),
1526                                  bool multi_instance)
1527 {
1528         /* (Un)Install the callbacks for further cpu hotplug operations */
1529         struct cpuhp_step *sp;
1530         int ret = 0;
1531
1532         /*
1533          * If name is NULL, then the state gets removed.
1534          *
1535          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1536          * the first allocation from these dynamic ranges, so the removal
1537          * would trigger a new allocation and clear the wrong (already
1538          * empty) state, leaving the callbacks of the to be cleared state
1539          * dangling, which causes wreckage on the next hotplug operation.
1540          */
1541         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1542                      state == CPUHP_BP_PREPARE_DYN)) {
1543                 ret = cpuhp_reserve_state(state);
1544                 if (ret < 0)
1545                         return ret;
1546                 state = ret;
1547         }
1548         sp = cpuhp_get_step(state);
1549         if (name && sp->name)
1550                 return -EBUSY;
1551
1552         sp->startup.single = startup;
1553         sp->teardown.single = teardown;
1554         sp->name = name;
1555         sp->multi_instance = multi_instance;
1556         INIT_HLIST_HEAD(&sp->list);
1557         return ret;
1558 }
1559
1560 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1561 {
1562         return cpuhp_get_step(state)->teardown.single;
1563 }
1564
1565 /*
1566  * Call the startup/teardown function for a step either on the AP or
1567  * on the current CPU.
1568  */
1569 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1570                             struct hlist_node *node)
1571 {
1572         struct cpuhp_step *sp = cpuhp_get_step(state);
1573         int ret;
1574
1575         /*
1576          * If there's nothing to do, we done.
1577          * Relies on the union for multi_instance.
1578          */
1579         if ((bringup && !sp->startup.single) ||
1580             (!bringup && !sp->teardown.single))
1581                 return 0;
1582         /*
1583          * The non AP bound callbacks can fail on bringup. On teardown
1584          * e.g. module removal we crash for now.
1585          */
1586 #ifdef CONFIG_SMP
1587         if (cpuhp_is_ap_state(state))
1588                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1589         else
1590                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1591 #else
1592         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1593 #endif
1594         BUG_ON(ret && !bringup);
1595         return ret;
1596 }
1597
1598 /*
1599  * Called from __cpuhp_setup_state on a recoverable failure.
1600  *
1601  * Note: The teardown callbacks for rollback are not allowed to fail!
1602  */
1603 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1604                                    struct hlist_node *node)
1605 {
1606         int cpu;
1607
1608         /* Roll back the already executed steps on the other cpus */
1609         for_each_present_cpu(cpu) {
1610                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1611                 int cpustate = st->state;
1612
1613                 if (cpu >= failedcpu)
1614                         break;
1615
1616                 /* Did we invoke the startup call on that cpu ? */
1617                 if (cpustate >= state)
1618                         cpuhp_issue_call(cpu, state, false, node);
1619         }
1620 }
1621
1622 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1623                                           struct hlist_node *node,
1624                                           bool invoke)
1625 {
1626         struct cpuhp_step *sp;
1627         int cpu;
1628         int ret;
1629
1630         lockdep_assert_cpus_held();
1631
1632         sp = cpuhp_get_step(state);
1633         if (sp->multi_instance == false)
1634                 return -EINVAL;
1635
1636         mutex_lock(&cpuhp_state_mutex);
1637
1638         if (!invoke || !sp->startup.multi)
1639                 goto add_node;
1640
1641         /*
1642          * Try to call the startup callback for each present cpu
1643          * depending on the hotplug state of the cpu.
1644          */
1645         for_each_present_cpu(cpu) {
1646                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1647                 int cpustate = st->state;
1648
1649                 if (cpustate < state)
1650                         continue;
1651
1652                 ret = cpuhp_issue_call(cpu, state, true, node);
1653                 if (ret) {
1654                         if (sp->teardown.multi)
1655                                 cpuhp_rollback_install(cpu, state, node);
1656                         goto unlock;
1657                 }
1658         }
1659 add_node:
1660         ret = 0;
1661         hlist_add_head(node, &sp->list);
1662 unlock:
1663         mutex_unlock(&cpuhp_state_mutex);
1664         return ret;
1665 }
1666
1667 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1668                                bool invoke)
1669 {
1670         int ret;
1671
1672         cpus_read_lock();
1673         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1674         cpus_read_unlock();
1675         return ret;
1676 }
1677 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1678
1679 /**
1680  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1681  * @state:              The state to setup
1682  * @invoke:             If true, the startup function is invoked for cpus where
1683  *                      cpu state >= @state
1684  * @startup:            startup callback function
1685  * @teardown:           teardown callback function
1686  * @multi_instance:     State is set up for multiple instances which get
1687  *                      added afterwards.
1688  *
1689  * The caller needs to hold cpus read locked while calling this function.
1690  * Returns:
1691  *   On success:
1692  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1693  *      0 for all other states
1694  *   On failure: proper (negative) error code
1695  */
1696 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1697                                    const char *name, bool invoke,
1698                                    int (*startup)(unsigned int cpu),
1699                                    int (*teardown)(unsigned int cpu),
1700                                    bool multi_instance)
1701 {
1702         int cpu, ret = 0;
1703         bool dynstate;
1704
1705         lockdep_assert_cpus_held();
1706
1707         if (cpuhp_cb_check(state) || !name)
1708                 return -EINVAL;
1709
1710         mutex_lock(&cpuhp_state_mutex);
1711
1712         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1713                                     multi_instance);
1714
1715         dynstate = state == CPUHP_AP_ONLINE_DYN;
1716         if (ret > 0 && dynstate) {
1717                 state = ret;
1718                 ret = 0;
1719         }
1720
1721         if (ret || !invoke || !startup)
1722                 goto out;
1723
1724         /*
1725          * Try to call the startup callback for each present cpu
1726          * depending on the hotplug state of the cpu.
1727          */
1728         for_each_present_cpu(cpu) {
1729                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1730                 int cpustate = st->state;
1731
1732                 if (cpustate < state)
1733                         continue;
1734
1735                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1736                 if (ret) {
1737                         if (teardown)
1738                                 cpuhp_rollback_install(cpu, state, NULL);
1739                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1740                         goto out;
1741                 }
1742         }
1743 out:
1744         mutex_unlock(&cpuhp_state_mutex);
1745         /*
1746          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1747          * dynamically allocated state in case of success.
1748          */
1749         if (!ret && dynstate)
1750                 return state;
1751         return ret;
1752 }
1753 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1754
1755 int __cpuhp_setup_state(enum cpuhp_state state,
1756                         const char *name, bool invoke,
1757                         int (*startup)(unsigned int cpu),
1758                         int (*teardown)(unsigned int cpu),
1759                         bool multi_instance)
1760 {
1761         int ret;
1762
1763         cpus_read_lock();
1764         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1765                                              teardown, multi_instance);
1766         cpus_read_unlock();
1767         return ret;
1768 }
1769 EXPORT_SYMBOL(__cpuhp_setup_state);
1770
1771 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1772                                   struct hlist_node *node, bool invoke)
1773 {
1774         struct cpuhp_step *sp = cpuhp_get_step(state);
1775         int cpu;
1776
1777         BUG_ON(cpuhp_cb_check(state));
1778
1779         if (!sp->multi_instance)
1780                 return -EINVAL;
1781
1782         cpus_read_lock();
1783         mutex_lock(&cpuhp_state_mutex);
1784
1785         if (!invoke || !cpuhp_get_teardown_cb(state))
1786                 goto remove;
1787         /*
1788          * Call the teardown callback for each present cpu depending
1789          * on the hotplug state of the cpu. This function is not
1790          * allowed to fail currently!
1791          */
1792         for_each_present_cpu(cpu) {
1793                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1794                 int cpustate = st->state;
1795
1796                 if (cpustate >= state)
1797                         cpuhp_issue_call(cpu, state, false, node);
1798         }
1799
1800 remove:
1801         hlist_del(node);
1802         mutex_unlock(&cpuhp_state_mutex);
1803         cpus_read_unlock();
1804
1805         return 0;
1806 }
1807 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1808
1809 /**
1810  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1811  * @state:      The state to remove
1812  * @invoke:     If true, the teardown function is invoked for cpus where
1813  *              cpu state >= @state
1814  *
1815  * The caller needs to hold cpus read locked while calling this function.
1816  * The teardown callback is currently not allowed to fail. Think
1817  * about module removal!
1818  */
1819 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1820 {
1821         struct cpuhp_step *sp = cpuhp_get_step(state);
1822         int cpu;
1823
1824         BUG_ON(cpuhp_cb_check(state));
1825
1826         lockdep_assert_cpus_held();
1827
1828         mutex_lock(&cpuhp_state_mutex);
1829         if (sp->multi_instance) {
1830                 WARN(!hlist_empty(&sp->list),
1831                      "Error: Removing state %d which has instances left.\n",
1832                      state);
1833                 goto remove;
1834         }
1835
1836         if (!invoke || !cpuhp_get_teardown_cb(state))
1837                 goto remove;
1838
1839         /*
1840          * Call the teardown callback for each present cpu depending
1841          * on the hotplug state of the cpu. This function is not
1842          * allowed to fail currently!
1843          */
1844         for_each_present_cpu(cpu) {
1845                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1846                 int cpustate = st->state;
1847
1848                 if (cpustate >= state)
1849                         cpuhp_issue_call(cpu, state, false, NULL);
1850         }
1851 remove:
1852         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1853         mutex_unlock(&cpuhp_state_mutex);
1854 }
1855 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1856
1857 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1858 {
1859         cpus_read_lock();
1860         __cpuhp_remove_state_cpuslocked(state, invoke);
1861         cpus_read_unlock();
1862 }
1863 EXPORT_SYMBOL(__cpuhp_remove_state);
1864
1865 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1866 static ssize_t show_cpuhp_state(struct device *dev,
1867                                 struct device_attribute *attr, char *buf)
1868 {
1869         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1870
1871         return sprintf(buf, "%d\n", st->state);
1872 }
1873 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1874
1875 static ssize_t write_cpuhp_target(struct device *dev,
1876                                   struct device_attribute *attr,
1877                                   const char *buf, size_t count)
1878 {
1879         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1880         struct cpuhp_step *sp;
1881         int target, ret;
1882
1883         ret = kstrtoint(buf, 10, &target);
1884         if (ret)
1885                 return ret;
1886
1887 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1888         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1889                 return -EINVAL;
1890 #else
1891         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1892                 return -EINVAL;
1893 #endif
1894
1895         ret = lock_device_hotplug_sysfs();
1896         if (ret)
1897                 return ret;
1898
1899         mutex_lock(&cpuhp_state_mutex);
1900         sp = cpuhp_get_step(target);
1901         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1902         mutex_unlock(&cpuhp_state_mutex);
1903         if (ret)
1904                 goto out;
1905
1906         if (st->state < target)
1907                 ret = do_cpu_up(dev->id, target);
1908         else
1909                 ret = do_cpu_down(dev->id, target);
1910 out:
1911         unlock_device_hotplug();
1912         return ret ? ret : count;
1913 }
1914
1915 static ssize_t show_cpuhp_target(struct device *dev,
1916                                  struct device_attribute *attr, char *buf)
1917 {
1918         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1919
1920         return sprintf(buf, "%d\n", st->target);
1921 }
1922 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1923
1924
1925 static ssize_t write_cpuhp_fail(struct device *dev,
1926                                 struct device_attribute *attr,
1927                                 const char *buf, size_t count)
1928 {
1929         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1930         struct cpuhp_step *sp;
1931         int fail, ret;
1932
1933         ret = kstrtoint(buf, 10, &fail);
1934         if (ret)
1935                 return ret;
1936
1937         /*
1938          * Cannot fail STARTING/DYING callbacks.
1939          */
1940         if (cpuhp_is_atomic_state(fail))
1941                 return -EINVAL;
1942
1943         /*
1944          * Cannot fail anything that doesn't have callbacks.
1945          */
1946         mutex_lock(&cpuhp_state_mutex);
1947         sp = cpuhp_get_step(fail);
1948         if (!sp->startup.single && !sp->teardown.single)
1949                 ret = -EINVAL;
1950         mutex_unlock(&cpuhp_state_mutex);
1951         if (ret)
1952                 return ret;
1953
1954         st->fail = fail;
1955
1956         return count;
1957 }
1958
1959 static ssize_t show_cpuhp_fail(struct device *dev,
1960                                struct device_attribute *attr, char *buf)
1961 {
1962         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1963
1964         return sprintf(buf, "%d\n", st->fail);
1965 }
1966
1967 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1968
1969 static struct attribute *cpuhp_cpu_attrs[] = {
1970         &dev_attr_state.attr,
1971         &dev_attr_target.attr,
1972         &dev_attr_fail.attr,
1973         NULL
1974 };
1975
1976 static const struct attribute_group cpuhp_cpu_attr_group = {
1977         .attrs = cpuhp_cpu_attrs,
1978         .name = "hotplug",
1979         NULL
1980 };
1981
1982 static ssize_t show_cpuhp_states(struct device *dev,
1983                                  struct device_attribute *attr, char *buf)
1984 {
1985         ssize_t cur, res = 0;
1986         int i;
1987
1988         mutex_lock(&cpuhp_state_mutex);
1989         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1990                 struct cpuhp_step *sp = cpuhp_get_step(i);
1991
1992                 if (sp->name) {
1993                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1994                         buf += cur;
1995                         res += cur;
1996                 }
1997         }
1998         mutex_unlock(&cpuhp_state_mutex);
1999         return res;
2000 }
2001 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2002
2003 static struct attribute *cpuhp_cpu_root_attrs[] = {
2004         &dev_attr_states.attr,
2005         NULL
2006 };
2007
2008 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2009         .attrs = cpuhp_cpu_root_attrs,
2010         .name = "hotplug",
2011         NULL
2012 };
2013
2014 #ifdef CONFIG_HOTPLUG_SMT
2015
2016 static const char *smt_states[] = {
2017         [CPU_SMT_ENABLED]               = "on",
2018         [CPU_SMT_DISABLED]              = "off",
2019         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2020         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2021 };
2022
2023 static ssize_t
2024 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2025 {
2026         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2027 }
2028
2029 static void cpuhp_offline_cpu_device(unsigned int cpu)
2030 {
2031         struct device *dev = get_cpu_device(cpu);
2032
2033         dev->offline = true;
2034         /* Tell user space about the state change */
2035         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2036 }
2037
2038 static void cpuhp_online_cpu_device(unsigned int cpu)
2039 {
2040         struct device *dev = get_cpu_device(cpu);
2041
2042         dev->offline = false;
2043         /* Tell user space about the state change */
2044         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2045 }
2046
2047 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2048 {
2049         int cpu, ret = 0;
2050
2051         cpu_maps_update_begin();
2052         for_each_online_cpu(cpu) {
2053                 if (topology_is_primary_thread(cpu))
2054                         continue;
2055                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2056                 if (ret)
2057                         break;
2058                 /*
2059                  * As this needs to hold the cpu maps lock it's impossible
2060                  * to call device_offline() because that ends up calling
2061                  * cpu_down() which takes cpu maps lock. cpu maps lock
2062                  * needs to be held as this might race against in kernel
2063                  * abusers of the hotplug machinery (thermal management).
2064                  *
2065                  * So nothing would update device:offline state. That would
2066                  * leave the sysfs entry stale and prevent onlining after
2067                  * smt control has been changed to 'off' again. This is
2068                  * called under the sysfs hotplug lock, so it is properly
2069                  * serialized against the regular offline usage.
2070                  */
2071                 cpuhp_offline_cpu_device(cpu);
2072         }
2073         if (!ret)
2074                 cpu_smt_control = ctrlval;
2075         cpu_maps_update_done();
2076         return ret;
2077 }
2078
2079 static int cpuhp_smt_enable(void)
2080 {
2081         int cpu, ret = 0;
2082
2083         cpu_maps_update_begin();
2084         cpu_smt_control = CPU_SMT_ENABLED;
2085         for_each_present_cpu(cpu) {
2086                 /* Skip online CPUs and CPUs on offline nodes */
2087                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2088                         continue;
2089                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2090                 if (ret)
2091                         break;
2092                 /* See comment in cpuhp_smt_disable() */
2093                 cpuhp_online_cpu_device(cpu);
2094         }
2095         cpu_maps_update_done();
2096         return ret;
2097 }
2098
2099 static ssize_t
2100 store_smt_control(struct device *dev, struct device_attribute *attr,
2101                   const char *buf, size_t count)
2102 {
2103         int ctrlval, ret;
2104
2105         if (sysfs_streq(buf, "on"))
2106                 ctrlval = CPU_SMT_ENABLED;
2107         else if (sysfs_streq(buf, "off"))
2108                 ctrlval = CPU_SMT_DISABLED;
2109         else if (sysfs_streq(buf, "forceoff"))
2110                 ctrlval = CPU_SMT_FORCE_DISABLED;
2111         else
2112                 return -EINVAL;
2113
2114         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2115                 return -EPERM;
2116
2117         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2118                 return -ENODEV;
2119
2120         ret = lock_device_hotplug_sysfs();
2121         if (ret)
2122                 return ret;
2123
2124         if (ctrlval != cpu_smt_control) {
2125                 switch (ctrlval) {
2126                 case CPU_SMT_ENABLED:
2127                         ret = cpuhp_smt_enable();
2128                         break;
2129                 case CPU_SMT_DISABLED:
2130                 case CPU_SMT_FORCE_DISABLED:
2131                         ret = cpuhp_smt_disable(ctrlval);
2132                         break;
2133                 }
2134         }
2135
2136         unlock_device_hotplug();
2137         return ret ? ret : count;
2138 }
2139 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2140
2141 static ssize_t
2142 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2143 {
2144         bool active = topology_max_smt_threads() > 1;
2145
2146         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2147 }
2148 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2149
2150 static struct attribute *cpuhp_smt_attrs[] = {
2151         &dev_attr_control.attr,
2152         &dev_attr_active.attr,
2153         NULL
2154 };
2155
2156 static const struct attribute_group cpuhp_smt_attr_group = {
2157         .attrs = cpuhp_smt_attrs,
2158         .name = "smt",
2159         NULL
2160 };
2161
2162 static int __init cpu_smt_state_init(void)
2163 {
2164         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2165                                   &cpuhp_smt_attr_group);
2166 }
2167
2168 #else
2169 static inline int cpu_smt_state_init(void) { return 0; }
2170 #endif
2171
2172 static int __init cpuhp_sysfs_init(void)
2173 {
2174         int cpu, ret;
2175
2176         ret = cpu_smt_state_init();
2177         if (ret)
2178                 return ret;
2179
2180         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2181                                  &cpuhp_cpu_root_attr_group);
2182         if (ret)
2183                 return ret;
2184
2185         for_each_possible_cpu(cpu) {
2186                 struct device *dev = get_cpu_device(cpu);
2187
2188                 if (!dev)
2189                         continue;
2190                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2191                 if (ret)
2192                         return ret;
2193         }
2194         return 0;
2195 }
2196 device_initcall(cpuhp_sysfs_init);
2197 #endif
2198
2199 /*
2200  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2201  * represents all NR_CPUS bits binary values of 1<<nr.
2202  *
2203  * It is used by cpumask_of() to get a constant address to a CPU
2204  * mask value that has a single bit set only.
2205  */
2206
2207 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2208 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2209 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2210 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2211 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2212
2213 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2214
2215         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2216         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2217 #if BITS_PER_LONG > 32
2218         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2219         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2220 #endif
2221 };
2222 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2223
2224 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2225 EXPORT_SYMBOL(cpu_all_bits);
2226
2227 #ifdef CONFIG_INIT_ALL_POSSIBLE
2228 struct cpumask __cpu_possible_mask __read_mostly
2229         = {CPU_BITS_ALL};
2230 #else
2231 struct cpumask __cpu_possible_mask __read_mostly;
2232 #endif
2233 EXPORT_SYMBOL(__cpu_possible_mask);
2234
2235 struct cpumask __cpu_online_mask __read_mostly;
2236 EXPORT_SYMBOL(__cpu_online_mask);
2237
2238 struct cpumask __cpu_present_mask __read_mostly;
2239 EXPORT_SYMBOL(__cpu_present_mask);
2240
2241 struct cpumask __cpu_active_mask __read_mostly;
2242 EXPORT_SYMBOL(__cpu_active_mask);
2243
2244 void init_cpu_present(const struct cpumask *src)
2245 {
2246         cpumask_copy(&__cpu_present_mask, src);
2247 }
2248
2249 void init_cpu_possible(const struct cpumask *src)
2250 {
2251         cpumask_copy(&__cpu_possible_mask, src);
2252 }
2253
2254 void init_cpu_online(const struct cpumask *src)
2255 {
2256         cpumask_copy(&__cpu_online_mask, src);
2257 }
2258
2259 /*
2260  * Activate the first processor.
2261  */
2262 void __init boot_cpu_init(void)
2263 {
2264         int cpu = smp_processor_id();
2265
2266         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2267         set_cpu_online(cpu, true);
2268         set_cpu_active(cpu, true);
2269         set_cpu_present(cpu, true);
2270         set_cpu_possible(cpu, true);
2271
2272 #ifdef CONFIG_SMP
2273         __boot_cpu_id = cpu;
2274 #endif
2275 }
2276
2277 /*
2278  * Must be called _AFTER_ setting up the per_cpu areas
2279  */
2280 void __init boot_cpu_hotplug_init(void)
2281 {
2282 #ifdef CONFIG_SMP
2283         this_cpu_write(cpuhp_state.booted_once, true);
2284 #endif
2285         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2286 }