kernel/cpu.c

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