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/sched/mm.h>
   7 #include <linux/proc_fs.h>
   8 #include <linux/smp.h>
   9 #include <linux/init.h>
  10 #include <linux/notifier.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/sched/hotplug.h>
  13 #include <linux/sched/isolation.h>
  14 #include <linux/sched/task.h>
  15 #include <linux/sched/smt.h>
  16 #include <linux/unistd.h>
  17 #include <linux/cpu.h>
  18 #include <linux/oom.h>
  19 #include <linux/rcupdate.h>
  20 #include <linux/export.h>
  21 #include <linux/bug.h>
  22 #include <linux/kthread.h>
  23 #include <linux/stop_machine.h>
  24 #include <linux/mutex.h>
  25 #include <linux/gfp.h>
  26 #include <linux/suspend.h>
  27 #include <linux/lockdep.h>
  28 #include <linux/tick.h>
  29 #include <linux/irq.h>
  30 #include <linux/nmi.h>
  31 #include <linux/smpboot.h>
  32 #include <linux/relay.h>
  33 #include <linux/slab.h>
  34 #include <linux/percpu-rwsem.h>
  35 #include <linux/cpuset.h>
  36
  37 #include <trace/events/power.h>
  38 #define CREATE_TRACE_POINTS
  39 #include <trace/events/cpuhp.h>
  40
  41 #include "smpboot.h"
  42
  43 /**
  44  * cpuhp_cpu_state - Per cpu hotplug state storage
  45  * @state:      The current cpu state
  46  * @target:     The target state
  47  * @thread:     Pointer to the hotplug thread
  48  * @should_run: Thread should execute
  49  * @rollback:   Perform a rollback
  50  * @single:     Single callback invocation
  51  * @bringup:    Single callback bringup or teardown selector
  52  * @cb_state:   The state for a single callback (install/uninstall)
  53  * @result:     Result of the operation
  54  * @done_up:    Signal completion to the issuer of the task for cpu-up
  55  * @done_down:  Signal completion to the issuer of the task for cpu-down
  56  */
  57 struct cpuhp_cpu_state {
  58         enum cpuhp_state        state;
  59         enum cpuhp_state        target;
  60         enum cpuhp_state        fail;
  61 #ifdef CONFIG_SMP
  62         struct task_struct      *thread;
  63         bool                    should_run;
  64         bool                    rollback;
  65         bool                    single;
  66         bool                    bringup;
  67         struct hlist_node       *node;
  68         struct hlist_node       *last;
  69         enum cpuhp_state        cb_state;
  70         int                     result;
  71         struct completion       done_up;
  72         struct completion       done_down;
  73 #endif
  74 };
  75
  76 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  77         .fail = CPUHP_INVALID,
  78 };
  79
  80 #ifdef CONFIG_SMP
  81 cpumask_t cpus_booted_once_mask;
  82 #endif
  83
  84 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  85 static struct lockdep_map cpuhp_state_up_map =
  86         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  87 static struct lockdep_map cpuhp_state_down_map =
  88         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  89
  90
  91 static inline void cpuhp_lock_acquire(bool bringup)
  92 {
  93         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  94 }
  95
  96 static inline void cpuhp_lock_release(bool bringup)
  97 {
  98         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  99 }
 100 #else
 101
 102 static inline void cpuhp_lock_acquire(bool bringup) { }
 103 static inline void cpuhp_lock_release(bool bringup) { }
 104
 105 #endif
 106
 107 /**
 108  * cpuhp_step - Hotplug state machine step
 109  * @name:       Name of the step
 110  * @startup:    Startup function of the step
 111  * @teardown:   Teardown function of the step
 112  * @cant_stop:  Bringup/teardown can't be stopped at this step
 113  */
 114 struct cpuhp_step {
 115         const char              *name;
 116         union {
 117                 int             (*single)(unsigned int cpu);
 118                 int             (*multi)(unsigned int cpu,
 119                                          struct hlist_node *node);
 120         } startup;
 121         union {
 122                 int             (*single)(unsigned int cpu);
 123                 int             (*multi)(unsigned int cpu,
 124                                          struct hlist_node *node);
 125         } teardown;
 126         struct hlist_head       list;
 127         bool                    cant_stop;
 128         bool                    multi_instance;
 129 };
 130
 131 static DEFINE_MUTEX(cpuhp_state_mutex);
 132 static struct cpuhp_step cpuhp_hp_states[];
 133
 134 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 135 {
 136         return cpuhp_hp_states + state;
 137 }
 138
 139 /**
 140  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 141  * @cpu:        The cpu for which the callback should be invoked
 142  * @state:      The state to do callbacks for
 143  * @bringup:    True if the bringup callback should be invoked
 144  * @node:       For multi-instance, do a single entry callback for install/remove
 145  * @lastp:      For multi-instance rollback, remember how far we got
 146  *
 147  * Called from cpu hotplug and from the state register machinery.
 148  */
 149 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 150                                  bool bringup, struct hlist_node *node,
 151                                  struct hlist_node **lastp)
 152 {
 153         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 154         struct cpuhp_step *step = cpuhp_get_step(state);
 155         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 156         int (*cb)(unsigned int cpu);
 157         int ret, cnt;
 158
 159         if (st->fail == state) {
 160                 st->fail = CPUHP_INVALID;
 161
 162                 if (!(bringup ? step->startup.single : step->teardown.single))
 163                         return 0;
 164
 165                 return -EAGAIN;
 166         }
 167
 168         if (!step->multi_instance) {
 169                 WARN_ON_ONCE(lastp && *lastp);
 170                 cb = bringup ? step->startup.single : step->teardown.single;
 171                 if (!cb)
 172                         return 0;
 173                 trace_cpuhp_enter(cpu, st->target, state, cb);
 174                 ret = cb(cpu);
 175                 trace_cpuhp_exit(cpu, st->state, state, ret);
 176                 return ret;
 177         }
 178         cbm = bringup ? step->startup.multi : step->teardown.multi;
 179         if (!cbm)
 180                 return 0;
 181
 182         /* Single invocation for instance add/remove */
 183         if (node) {
 184                 WARN_ON_ONCE(lastp && *lastp);
 185                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 186                 ret = cbm(cpu, node);
 187                 trace_cpuhp_exit(cpu, st->state, state, ret);
 188                 return ret;
 189         }
 190
 191         /* State transition. Invoke on all instances */
 192         cnt = 0;
 193         hlist_for_each(node, &step->list) {
 194                 if (lastp && node == *lastp)
 195                         break;
 196
 197                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 198                 ret = cbm(cpu, node);
 199                 trace_cpuhp_exit(cpu, st->state, state, ret);
 200                 if (ret) {
 201                         if (!lastp)
 202                                 goto err;
 203
 204                         *lastp = node;
 205                         return ret;
 206                 }
 207                 cnt++;
 208         }
 209         if (lastp)
 210                 *lastp = NULL;
 211         return 0;
 212 err:
 213         /* Rollback the instances if one failed */
 214         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 215         if (!cbm)
 216                 return ret;
 217
 218         hlist_for_each(node, &step->list) {
 219                 if (!cnt--)
 220                         break;
 221
 222                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 223                 ret = cbm(cpu, node);
 224                 trace_cpuhp_exit(cpu, st->state, state, ret);
 225                 /*
 226                  * Rollback must not fail,
 227                  */
 228                 WARN_ON_ONCE(ret);
 229         }
 230         return ret;
 231 }
 232
 233 #ifdef CONFIG_SMP
 234 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 235 {
 236         /*
 237          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 238          * purposes as that state is handled explicitly in cpu_down.
 239          */
 240         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 241 }
 242
 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 int cpus_read_trylock(void)
 299 {
 300         return percpu_down_read_trylock(&cpu_hotplug_lock);
 301 }
 302 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 303
 304 void cpus_read_unlock(void)
 305 {
 306         percpu_up_read(&cpu_hotplug_lock);
 307 }
 308 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 309
 310 void cpus_write_lock(void)
 311 {
 312         percpu_down_write(&cpu_hotplug_lock);
 313 }
 314
 315 void cpus_write_unlock(void)
 316 {
 317         percpu_up_write(&cpu_hotplug_lock);
 318 }
 319
 320 void lockdep_assert_cpus_held(void)
 321 {
 322         /*
 323          * We can't have hotplug operations before userspace starts running,
 324          * and some init codepaths will knowingly not take the hotplug lock.
 325          * This is all valid, so mute lockdep until it makes sense to report
 326          * unheld locks.
 327          */
 328         if (system_state < SYSTEM_RUNNING)
 329                 return;
 330
 331         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 332 }
 333
 334 static void lockdep_acquire_cpus_lock(void)
 335 {
 336         rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
 337 }
 338
 339 static void lockdep_release_cpus_lock(void)
 340 {
 341         rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
 342 }
 343
 344 /*
 345  * Wait for currently running CPU hotplug operations to complete (if any) and
 346  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 347  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 348  * hotplug path before performing hotplug operations. So acquiring that lock
 349  * guarantees mutual exclusion from any currently running hotplug operations.
 350  */
 351 void cpu_hotplug_disable(void)
 352 {
 353         cpu_maps_update_begin();
 354         cpu_hotplug_disabled++;
 355         cpu_maps_update_done();
 356 }
 357 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 358
 359 static void __cpu_hotplug_enable(void)
 360 {
 361         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 362                 return;
 363         cpu_hotplug_disabled--;
 364 }
 365
 366 void cpu_hotplug_enable(void)
 367 {
 368         cpu_maps_update_begin();
 369         __cpu_hotplug_enable();
 370         cpu_maps_update_done();
 371 }
 372 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 373
 374 #else
 375
 376 static void lockdep_acquire_cpus_lock(void)
 377 {
 378 }
 379
 380 static void lockdep_release_cpus_lock(void)
 381 {
 382 }
 383
 384 #endif  /* CONFIG_HOTPLUG_CPU */
 385
 386 /*
 387  * Architectures that need SMT-specific errata handling during SMT hotplug
 388  * should override this.
 389  */
 390 void __weak arch_smt_update(void) { }
 391
 392 #ifdef CONFIG_HOTPLUG_SMT
 393 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 394
 395 void __init cpu_smt_disable(bool force)
 396 {
 397         if (!cpu_smt_possible())
 398                 return;
 399
 400         if (force) {
 401                 pr_info("SMT: Force disabled\n");
 402                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 403         } else {
 404                 pr_info("SMT: disabled\n");
 405                 cpu_smt_control = CPU_SMT_DISABLED;
 406         }
 407 }
 408
 409 /*
 410  * The decision whether SMT is supported can only be done after the full
 411  * CPU identification. Called from architecture code.
 412  */
 413 void __init cpu_smt_check_topology(void)
 414 {
 415         if (!topology_smt_supported())
 416                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 417 }
 418
 419 static int __init smt_cmdline_disable(char *str)
 420 {
 421         cpu_smt_disable(str && !strcmp(str, "force"));
 422         return 0;
 423 }
 424 early_param("nosmt", smt_cmdline_disable);
 425
 426 static inline bool cpu_smt_allowed(unsigned int cpu)
 427 {
 428         if (cpu_smt_control == CPU_SMT_ENABLED)
 429                 return true;
 430
 431         if (topology_is_primary_thread(cpu))
 432                 return true;
 433
 434         /*
 435          * On x86 it's required to boot all logical CPUs at least once so
 436          * that the init code can get a chance to set CR4.MCE on each
 437          * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
 438          * core will shutdown the machine.
 439          */
 440         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
 441 }
 442
 443 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
 444 bool cpu_smt_possible(void)
 445 {
 446         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
 447                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
 448 }
 449 EXPORT_SYMBOL_GPL(cpu_smt_possible);
 450 #else
 451 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 452 #endif
 453
 454 static inline enum cpuhp_state
 455 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 456 {
 457         enum cpuhp_state prev_state = st->state;
 458
 459         st->rollback = false;
 460         st->last = NULL;
 461
 462         st->target = target;
 463         st->single = false;
 464         st->bringup = st->state < target;
 465
 466         return prev_state;
 467 }
 468
 469 static inline void
 470 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 471 {
 472         st->rollback = true;
 473
 474         /*
 475          * If we have st->last we need to undo partial multi_instance of this
 476          * state first. Otherwise start undo at the previous state.
 477          */
 478         if (!st->last) {
 479                 if (st->bringup)
 480                         st->state--;
 481                 else
 482                         st->state++;
 483         }
 484
 485         st->target = prev_state;
 486         st->bringup = !st->bringup;
 487 }
 488
 489 /* Regular hotplug invocation of the AP hotplug thread */
 490 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 491 {
 492         if (!st->single && st->state == st->target)
 493                 return;
 494
 495         st->result = 0;
 496         /*
 497          * Make sure the above stores are visible before should_run becomes
 498          * true. Paired with the mb() above in cpuhp_thread_fun()
 499          */
 500         smp_mb();
 501         st->should_run = true;
 502         wake_up_process(st->thread);
 503         wait_for_ap_thread(st, st->bringup);
 504 }
 505
 506 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 507 {
 508         enum cpuhp_state prev_state;
 509         int ret;
 510
 511         prev_state = cpuhp_set_state(st, target);
 512         __cpuhp_kick_ap(st);
 513         if ((ret = st->result)) {
 514                 cpuhp_reset_state(st, prev_state);
 515                 __cpuhp_kick_ap(st);
 516         }
 517
 518         return ret;
 519 }
 520
 521 static int bringup_wait_for_ap(unsigned int cpu)
 522 {
 523         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 524
 525         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 526         wait_for_ap_thread(st, true);
 527         if (WARN_ON_ONCE((!cpu_online(cpu))))
 528                 return -ECANCELED;
 529
 530         /* Unpark the hotplug thread of the target cpu */
 531         kthread_unpark(st->thread);
 532
 533         /*
 534          * SMT soft disabling on X86 requires to bring the CPU out of the
 535          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 536          * CPU marked itself as booted_once in notify_cpu_starting() so the
 537          * cpu_smt_allowed() check will now return false if this is not the
 538          * primary sibling.
 539          */
 540         if (!cpu_smt_allowed(cpu))
 541                 return -ECANCELED;
 542
 543         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 544                 return 0;
 545
 546         return cpuhp_kick_ap(st, st->target);
 547 }
 548
 549 static int bringup_cpu(unsigned int cpu)
 550 {
 551         struct task_struct *idle = idle_thread_get(cpu);
 552         int ret;
 553
 554         /*
 555          * Some architectures have to walk the irq descriptors to
 556          * setup the vector space for the cpu which comes online.
 557          * Prevent irq alloc/free across the bringup.
 558          */
 559         irq_lock_sparse();
 560
 561         /* Arch-specific enabling code. */
 562         ret = __cpu_up(cpu, idle);
 563         irq_unlock_sparse();
 564         if (ret)
 565                 return ret;
 566         return bringup_wait_for_ap(cpu);
 567 }
 568
 569 static int finish_cpu(unsigned int cpu)
 570 {
 571         struct task_struct *idle = idle_thread_get(cpu);
 572         struct mm_struct *mm = idle->active_mm;
 573
 574         /*
 575          * idle_task_exit() will have switched to &init_mm, now
 576          * clean up any remaining active_mm state.
 577          */
 578         if (mm != &init_mm)
 579                 idle->active_mm = &init_mm;
 580         mmdrop(mm);
 581         return 0;
 582 }
 583
 584 /*
 585  * Hotplug state machine related functions
 586  */
 587
 588 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 589 {
 590         for (st->state--; st->state > st->target; st->state--)
 591                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 592 }
 593
 594 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 595 {
 596         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 597                 return true;
 598         /*
 599          * When CPU hotplug is disabled, then taking the CPU down is not
 600          * possible because takedown_cpu() and the architecture and
 601          * subsystem specific mechanisms are not available. So the CPU
 602          * which would be completely unplugged again needs to stay around
 603          * in the current state.
 604          */
 605         return st->state <= CPUHP_BRINGUP_CPU;
 606 }
 607
 608 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 609                               enum cpuhp_state target)
 610 {
 611         enum cpuhp_state prev_state = st->state;
 612         int ret = 0;
 613
 614         while (st->state < target) {
 615                 st->state++;
 616                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 617                 if (ret) {
 618                         if (can_rollback_cpu(st)) {
 619                                 st->target = prev_state;
 620                                 undo_cpu_up(cpu, st);
 621                         }
 622                         break;
 623                 }
 624         }
 625         return ret;
 626 }
 627
 628 /*
 629  * The cpu hotplug threads manage the bringup and teardown of the cpus
 630  */
 631 static void cpuhp_create(unsigned int cpu)
 632 {
 633         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 634
 635         init_completion(&st->done_up);
 636         init_completion(&st->done_down);
 637 }
 638
 639 static int cpuhp_should_run(unsigned int cpu)
 640 {
 641         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 642
 643         return st->should_run;
 644 }
 645
 646 /*
 647  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 648  * callbacks when a state gets [un]installed at runtime.
 649  *
 650  * Each invocation of this function by the smpboot thread does a single AP
 651  * state callback.
 652  *
 653  * It has 3 modes of operation:
 654  *  - single: runs st->cb_state
 655  *  - up:     runs ++st->state, while st->state < st->target
 656  *  - down:   runs st->state--, while st->state > st->target
 657  *
 658  * When complete or on error, should_run is cleared and the completion is fired.
 659  */
 660 static void cpuhp_thread_fun(unsigned int cpu)
 661 {
 662         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 663         bool bringup = st->bringup;
 664         enum cpuhp_state state;
 665
 666         if (WARN_ON_ONCE(!st->should_run))
 667                 return;
 668
 669         /*
 670          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 671          * that if we see ->should_run we also see the rest of the state.
 672          */
 673         smp_mb();
 674
 675         /*
 676          * The BP holds the hotplug lock, but we're now running on the AP,
 677          * ensure that anybody asserting the lock is held, will actually find
 678          * it so.
 679          */
 680         lockdep_acquire_cpus_lock();
 681         cpuhp_lock_acquire(bringup);
 682
 683         if (st->single) {
 684                 state = st->cb_state;
 685                 st->should_run = false;
 686         } else {
 687                 if (bringup) {
 688                         st->state++;
 689                         state = st->state;
 690                         st->should_run = (st->state < st->target);
 691                         WARN_ON_ONCE(st->state > st->target);
 692                 } else {
 693                         state = st->state;
 694                         st->state--;
 695                         st->should_run = (st->state > st->target);
 696                         WARN_ON_ONCE(st->state < st->target);
 697                 }
 698         }
 699
 700         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 701
 702         if (cpuhp_is_atomic_state(state)) {
 703                 local_irq_disable();
 704                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 705                 local_irq_enable();
 706
 707                 /*
 708                  * STARTING/DYING must not fail!
 709                  */
 710                 WARN_ON_ONCE(st->result);
 711         } else {
 712                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 713         }
 714
 715         if (st->result) {
 716                 /*
 717                  * If we fail on a rollback, we're up a creek without no
 718                  * paddle, no way forward, no way back. We loose, thanks for
 719                  * playing.
 720                  */
 721                 WARN_ON_ONCE(st->rollback);
 722                 st->should_run = false;
 723         }
 724
 725         cpuhp_lock_release(bringup);
 726         lockdep_release_cpus_lock();
 727
 728         if (!st->should_run)
 729                 complete_ap_thread(st, bringup);
 730 }
 731
 732 /* Invoke a single callback on a remote cpu */
 733 static int
 734 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 735                          struct hlist_node *node)
 736 {
 737         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 738         int ret;
 739
 740         if (!cpu_online(cpu))
 741                 return 0;
 742
 743         cpuhp_lock_acquire(false);
 744         cpuhp_lock_release(false);
 745
 746         cpuhp_lock_acquire(true);
 747         cpuhp_lock_release(true);
 748
 749         /*
 750          * If we are up and running, use the hotplug thread. For early calls
 751          * we invoke the thread function directly.
 752          */
 753         if (!st->thread)
 754                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 755
 756         st->rollback = false;
 757         st->last = NULL;
 758
 759         st->node = node;
 760         st->bringup = bringup;
 761         st->cb_state = state;
 762         st->single = true;
 763
 764         __cpuhp_kick_ap(st);
 765
 766         /*
 767          * If we failed and did a partial, do a rollback.
 768          */
 769         if ((ret = st->result) && st->last) {
 770                 st->rollback = true;
 771                 st->bringup = !bringup;
 772
 773                 __cpuhp_kick_ap(st);
 774         }
 775
 776         /*
 777          * Clean up the leftovers so the next hotplug operation wont use stale
 778          * data.
 779          */
 780         st->node = st->last = NULL;
 781         return ret;
 782 }
 783
 784 static int cpuhp_kick_ap_work(unsigned int cpu)
 785 {
 786         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 787         enum cpuhp_state prev_state = st->state;
 788         int ret;
 789
 790         cpuhp_lock_acquire(false);
 791         cpuhp_lock_release(false);
 792
 793         cpuhp_lock_acquire(true);
 794         cpuhp_lock_release(true);
 795
 796         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 797         ret = cpuhp_kick_ap(st, st->target);
 798         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 799
 800         return ret;
 801 }
 802
 803 static struct smp_hotplug_thread cpuhp_threads = {
 804         .store                  = &cpuhp_state.thread,
 805         .create                 = &cpuhp_create,
 806         .thread_should_run      = cpuhp_should_run,
 807         .thread_fn              = cpuhp_thread_fun,
 808         .thread_comm            = "cpuhp/%u",
 809         .selfparking            = true,
 810 };
 811
 812 void __init cpuhp_threads_init(void)
 813 {
 814         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 815         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 816 }
 817
 818 /*
 819  *
 820  * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
 821  * protected region.
 822  *
 823  * The operation is still serialized against concurrent CPU hotplug via
 824  * cpu_add_remove_lock, i.e. CPU map protection.  But it is _not_
 825  * serialized against other hotplug related activity like adding or
 826  * removing of state callbacks and state instances, which invoke either the
 827  * startup or the teardown callback of the affected state.
 828  *
 829  * This is required for subsystems which are unfixable vs. CPU hotplug and
 830  * evade lock inversion problems by scheduling work which has to be
 831  * completed _before_ cpu_up()/_cpu_down() returns.
 832  *
 833  * Don't even think about adding anything to this for any new code or even
 834  * drivers. It's only purpose is to keep existing lock order trainwrecks
 835  * working.
 836  *
 837  * For cpu_down() there might be valid reasons to finish cleanups which are
 838  * not required to be done under cpu_hotplug_lock, but that's a different
 839  * story and would be not invoked via this.
 840  */
 841 static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
 842 {
 843         /*
 844          * cpusets delegate hotplug operations to a worker to "solve" the
 845          * lock order problems. Wait for the worker, but only if tasks are
 846          * _not_ frozen (suspend, hibernate) as that would wait forever.
 847          *
 848          * The wait is required because otherwise the hotplug operation
 849          * returns with inconsistent state, which could even be observed in
 850          * user space when a new CPU is brought up. The CPU plug uevent
 851          * would be delivered and user space reacting on it would fail to
 852          * move tasks to the newly plugged CPU up to the point where the
 853          * work has finished because up to that point the newly plugged CPU
 854          * is not assignable in cpusets/cgroups. On unplug that's not
 855          * necessarily a visible issue, but it is still inconsistent state,
 856          * which is the real problem which needs to be "fixed". This can't
 857          * prevent the transient state between scheduling the work and
 858          * returning from waiting for it.
 859          */
 860         if (!tasks_frozen)
 861                 cpuset_wait_for_hotplug();
 862 }
 863
 864 #ifdef CONFIG_HOTPLUG_CPU
 865 #ifndef arch_clear_mm_cpumask_cpu
 866 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
 867 #endif
 868
 869 /**
 870  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 871  * @cpu: a CPU id
 872  *
 873  * This function walks all processes, finds a valid mm struct for each one and
 874  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 875  * trivial, there are various non-obvious corner cases, which this function
 876  * tries to solve in a safe manner.
 877  *
 878  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 879  * be called only for an already offlined CPU.
 880  */
 881 void clear_tasks_mm_cpumask(int cpu)
 882 {
 883         struct task_struct *p;
 884
 885         /*
 886          * This function is called after the cpu is taken down and marked
 887          * offline, so its not like new tasks will ever get this cpu set in
 888          * their mm mask. -- Peter Zijlstra
 889          * Thus, we may use rcu_read_lock() here, instead of grabbing
 890          * full-fledged tasklist_lock.
 891          */
 892         WARN_ON(cpu_online(cpu));
 893         rcu_read_lock();
 894         for_each_process(p) {
 895                 struct task_struct *t;
 896
 897                 /*
 898                  * Main thread might exit, but other threads may still have
 899                  * a valid mm. Find one.
 900                  */
 901                 t = find_lock_task_mm(p);
 902                 if (!t)
 903                         continue;
 904                 arch_clear_mm_cpumask_cpu(cpu, t->mm);
 905                 task_unlock(t);
 906         }
 907         rcu_read_unlock();
 908 }
 909
 910 /* Take this CPU down. */
 911 static int take_cpu_down(void *_param)
 912 {
 913         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 914         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 915         int err, cpu = smp_processor_id();
 916         int ret;
 917
 918         /* Ensure this CPU doesn't handle any more interrupts. */
 919         err = __cpu_disable();
 920         if (err < 0)
 921                 return err;
 922
 923         /*
 924          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 925          * do this step again.
 926          */
 927         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 928         st->state--;
 929         /* Invoke the former CPU_DYING callbacks */
 930         for (; st->state > target; st->state--) {
 931                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 932                 /*
 933                  * DYING must not fail!
 934                  */
 935                 WARN_ON_ONCE(ret);
 936         }
 937
 938         /* Give up timekeeping duties */
 939         tick_handover_do_timer();
 940         /* Remove CPU from timer broadcasting */
 941         tick_offline_cpu(cpu);
 942         /* Park the stopper thread */
 943         stop_machine_park(cpu);
 944         return 0;
 945 }
 946
 947 static int takedown_cpu(unsigned int cpu)
 948 {
 949         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 950         int err;
 951
 952         /* Park the smpboot threads */
 953         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 954
 955         /*
 956          * Prevent irq alloc/free while the dying cpu reorganizes the
 957          * interrupt affinities.
 958          */
 959         irq_lock_sparse();
 960
 961         /*
 962          * So now all preempt/rcu users must observe !cpu_active().
 963          */
 964         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 965         if (err) {
 966                 /* CPU refused to die */
 967                 irq_unlock_sparse();
 968                 /* Unpark the hotplug thread so we can rollback there */
 969                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 970                 return err;
 971         }
 972         BUG_ON(cpu_online(cpu));
 973
 974         /*
 975          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 976          * all runnable tasks from the CPU, there's only the idle task left now
 977          * that the migration thread is done doing the stop_machine thing.
 978          *
 979          * Wait for the stop thread to go away.
 980          */
 981         wait_for_ap_thread(st, false);
 982         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 983
 984         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 985         irq_unlock_sparse();
 986
 987         hotplug_cpu__broadcast_tick_pull(cpu);
 988         /* This actually kills the CPU. */
 989         __cpu_die(cpu);
 990
 991         tick_cleanup_dead_cpu(cpu);
 992         rcutree_migrate_callbacks(cpu);
 993         return 0;
 994 }
 995
 996 static void cpuhp_complete_idle_dead(void *arg)
 997 {
 998         struct cpuhp_cpu_state *st = arg;
 999
1000         complete_ap_thread(st, false);
1001 }
1002
1003 void cpuhp_report_idle_dead(void)
1004 {
1005         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1006
1007         BUG_ON(st->state != CPUHP_AP_OFFLINE);
1008         rcu_report_dead(smp_processor_id());
1009         st->state = CPUHP_AP_IDLE_DEAD;
1010         /*
1011          * We cannot call complete after rcu_report_dead() so we delegate it
1012          * to an online cpu.
1013          */
1014         smp_call_function_single(cpumask_first(cpu_online_mask),
1015                                  cpuhp_complete_idle_dead, st, 0);
1016 }
1017
1018 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
1019 {
1020         for (st->state++; st->state < st->target; st->state++)
1021                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1022 }
1023
1024 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1025                                 enum cpuhp_state target)
1026 {
1027         enum cpuhp_state prev_state = st->state;
1028         int ret = 0;
1029
1030         for (; st->state > target; st->state--) {
1031                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
1032                 if (ret) {
1033                         st->target = prev_state;
1034                         if (st->state < prev_state)
1035                                 undo_cpu_down(cpu, st);
1036                         break;
1037                 }
1038         }
1039         return ret;
1040 }
1041
1042 /* Requires cpu_add_remove_lock to be held */
1043 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1044                            enum cpuhp_state target)
1045 {
1046         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1047         int prev_state, ret = 0;
1048
1049         if (num_online_cpus() == 1)
1050                 return -EBUSY;
1051
1052         if (!cpu_present(cpu))
1053                 return -EINVAL;
1054
1055         cpus_write_lock();
1056
1057         cpuhp_tasks_frozen = tasks_frozen;
1058
1059         prev_state = cpuhp_set_state(st, target);
1060         /*
1061          * If the current CPU state is in the range of the AP hotplug thread,
1062          * then we need to kick the thread.
1063          */
1064         if (st->state > CPUHP_TEARDOWN_CPU) {
1065                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1066                 ret = cpuhp_kick_ap_work(cpu);
1067                 /*
1068                  * The AP side has done the error rollback already. Just
1069                  * return the error code..
1070                  */
1071                 if (ret)
1072                         goto out;
1073
1074                 /*
1075                  * We might have stopped still in the range of the AP hotplug
1076                  * thread. Nothing to do anymore.
1077                  */
1078                 if (st->state > CPUHP_TEARDOWN_CPU)
1079                         goto out;
1080
1081                 st->target = target;
1082         }
1083         /*
1084          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1085          * to do the further cleanups.
1086          */
1087         ret = cpuhp_down_callbacks(cpu, st, target);
1088         if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1089                 cpuhp_reset_state(st, prev_state);
1090                 __cpuhp_kick_ap(st);
1091         }
1092
1093 out:
1094         cpus_write_unlock();
1095         /*
1096          * Do post unplug cleanup. This is still protected against
1097          * concurrent CPU hotplug via cpu_add_remove_lock.
1098          */
1099         lockup_detector_cleanup();
1100         arch_smt_update();
1101         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1102         return ret;
1103 }
1104
1105 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1106 {
1107         if (cpu_hotplug_disabled)
1108                 return -EBUSY;
1109         return _cpu_down(cpu, 0, target);
1110 }
1111
1112 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1113 {
1114         int err;
1115
1116         cpu_maps_update_begin();
1117         err = cpu_down_maps_locked(cpu, target);
1118         cpu_maps_update_done();
1119         return err;
1120 }
1121
1122 /**
1123  * cpu_device_down - Bring down a cpu device
1124  * @dev: Pointer to the cpu device to offline
1125  *
1126  * This function is meant to be used by device core cpu subsystem only.
1127  *
1128  * Other subsystems should use remove_cpu() instead.
1129  */
1130 int cpu_device_down(struct device *dev)
1131 {
1132         return cpu_down(dev->id, CPUHP_OFFLINE);
1133 }
1134
1135 int remove_cpu(unsigned int cpu)
1136 {
1137         int ret;
1138
1139         lock_device_hotplug();
1140         ret = device_offline(get_cpu_device(cpu));
1141         unlock_device_hotplug();
1142
1143         return ret;
1144 }
1145 EXPORT_SYMBOL_GPL(remove_cpu);
1146
1147 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1148 {
1149         unsigned int cpu;
1150         int error;
1151
1152         cpu_maps_update_begin();
1153
1154         /*
1155          * Make certain the cpu I'm about to reboot on is online.
1156          *
1157          * This is inline to what migrate_to_reboot_cpu() already do.
1158          */
1159         if (!cpu_online(primary_cpu))
1160                 primary_cpu = cpumask_first(cpu_online_mask);
1161
1162         for_each_online_cpu(cpu) {
1163                 if (cpu == primary_cpu)
1164                         continue;
1165
1166                 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1167                 if (error) {
1168                         pr_err("Failed to offline CPU%d - error=%d",
1169                                 cpu, error);
1170                         break;
1171                 }
1172         }
1173
1174         /*
1175          * Ensure all but the reboot CPU are offline.
1176          */
1177         BUG_ON(num_online_cpus() > 1);
1178
1179         /*
1180          * Make sure the CPUs won't be enabled by someone else after this
1181          * point. Kexec will reboot to a new kernel shortly resetting
1182          * everything along the way.
1183          */
1184         cpu_hotplug_disabled++;
1185
1186         cpu_maps_update_done();
1187 }
1188
1189 #else
1190 #define takedown_cpu            NULL
1191 #endif /*CONFIG_HOTPLUG_CPU*/
1192
1193 /**
1194  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1195  * @cpu: cpu that just started
1196  *
1197  * It must be called by the arch code on the new cpu, before the new cpu
1198  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1199  */
1200 void notify_cpu_starting(unsigned int cpu)
1201 {
1202         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1203         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1204         int ret;
1205
1206         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1207         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1208         while (st->state < target) {
1209                 st->state++;
1210                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1211                 /*
1212                  * STARTING must not fail!
1213                  */
1214                 WARN_ON_ONCE(ret);
1215         }
1216 }
1217
1218 /*
1219  * Called from the idle task. Wake up the controlling task which brings the
1220  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1221  * online bringup to the hotplug thread.
1222  */
1223 void cpuhp_online_idle(enum cpuhp_state state)
1224 {
1225         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1226
1227         /* Happens for the boot cpu */
1228         if (state != CPUHP_AP_ONLINE_IDLE)
1229                 return;
1230
1231         /*
1232          * Unpart the stopper thread before we start the idle loop (and start
1233          * scheduling); this ensures the stopper task is always available.
1234          */
1235         stop_machine_unpark(smp_processor_id());
1236
1237         st->state = CPUHP_AP_ONLINE_IDLE;
1238         complete_ap_thread(st, true);
1239 }
1240
1241 /* Requires cpu_add_remove_lock to be held */
1242 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1243 {
1244         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1245         struct task_struct *idle;
1246         int ret = 0;
1247
1248         cpus_write_lock();
1249
1250         if (!cpu_present(cpu)) {
1251                 ret = -EINVAL;
1252                 goto out;
1253         }
1254
1255         /*
1256          * The caller of cpu_up() might have raced with another
1257          * caller. Nothing to do.
1258          */
1259         if (st->state >= target)
1260                 goto out;
1261
1262         if (st->state == CPUHP_OFFLINE) {
1263                 /* Let it fail before we try to bring the cpu up */
1264                 idle = idle_thread_get(cpu);
1265                 if (IS_ERR(idle)) {
1266                         ret = PTR_ERR(idle);
1267                         goto out;
1268                 }
1269         }
1270
1271         cpuhp_tasks_frozen = tasks_frozen;
1272
1273         cpuhp_set_state(st, target);
1274         /*
1275          * If the current CPU state is in the range of the AP hotplug thread,
1276          * then we need to kick the thread once more.
1277          */
1278         if (st->state > CPUHP_BRINGUP_CPU) {
1279                 ret = cpuhp_kick_ap_work(cpu);
1280                 /*
1281                  * The AP side has done the error rollback already. Just
1282                  * return the error code..
1283                  */
1284                 if (ret)
1285                         goto out;
1286         }
1287
1288         /*
1289          * Try to reach the target state. We max out on the BP at
1290          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1291          * responsible for bringing it up to the target state.
1292          */
1293         target = min((int)target, CPUHP_BRINGUP_CPU);
1294         ret = cpuhp_up_callbacks(cpu, st, target);
1295 out:
1296         cpus_write_unlock();
1297         arch_smt_update();
1298         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1299         return ret;
1300 }
1301
1302 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1303 {
1304         int err = 0;
1305
1306         if (!cpu_possible(cpu)) {
1307                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1308                        cpu);
1309 #if defined(CONFIG_IA64)
1310                 pr_err("please check additional_cpus= boot parameter\n");
1311 #endif
1312                 return -EINVAL;
1313         }
1314
1315         err = try_online_node(cpu_to_node(cpu));
1316         if (err)
1317                 return err;
1318
1319         cpu_maps_update_begin();
1320
1321         if (cpu_hotplug_disabled) {
1322                 err = -EBUSY;
1323                 goto out;
1324         }
1325         if (!cpu_smt_allowed(cpu)) {
1326                 err = -EPERM;
1327                 goto out;
1328         }
1329
1330         err = _cpu_up(cpu, 0, target);
1331 out:
1332         cpu_maps_update_done();
1333         return err;
1334 }
1335
1336 /**
1337  * cpu_device_up - Bring up a cpu device
1338  * @dev: Pointer to the cpu device to online
1339  *
1340  * This function is meant to be used by device core cpu subsystem only.
1341  *
1342  * Other subsystems should use add_cpu() instead.
1343  */
1344 int cpu_device_up(struct device *dev)
1345 {
1346         return cpu_up(dev->id, CPUHP_ONLINE);
1347 }
1348
1349 int add_cpu(unsigned int cpu)
1350 {
1351         int ret;
1352
1353         lock_device_hotplug();
1354         ret = device_online(get_cpu_device(cpu));
1355         unlock_device_hotplug();
1356
1357         return ret;
1358 }
1359 EXPORT_SYMBOL_GPL(add_cpu);
1360
1361 /**
1362  * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1363  * @sleep_cpu: The cpu we hibernated on and should be brought up.
1364  *
1365  * On some architectures like arm64, we can hibernate on any CPU, but on
1366  * wake up the CPU we hibernated on might be offline as a side effect of
1367  * using maxcpus= for example.
1368  */
1369 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1370 {
1371         int ret;
1372
1373         if (!cpu_online(sleep_cpu)) {
1374                 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1375                 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1376                 if (ret) {
1377                         pr_err("Failed to bring hibernate-CPU up!\n");
1378                         return ret;
1379                 }
1380         }
1381         return 0;
1382 }
1383
1384 void bringup_nonboot_cpus(unsigned int setup_max_cpus)
1385 {
1386         unsigned int cpu;
1387
1388         for_each_present_cpu(cpu) {
1389                 if (num_online_cpus() >= setup_max_cpus)
1390                         break;
1391                 if (!cpu_online(cpu))
1392                         cpu_up(cpu, CPUHP_ONLINE);
1393         }
1394 }
1395
1396 #ifdef CONFIG_PM_SLEEP_SMP
1397 static cpumask_var_t frozen_cpus;
1398
1399 int freeze_secondary_cpus(int primary)
1400 {
1401         int cpu, error = 0;
1402
1403         cpu_maps_update_begin();
1404         if (primary == -1) {
1405                 primary = cpumask_first(cpu_online_mask);
1406                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1407                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1408         } else {
1409                 if (!cpu_online(primary))
1410                         primary = cpumask_first(cpu_online_mask);
1411         }
1412
1413         /*
1414          * We take down all of the non-boot CPUs in one shot to avoid races
1415          * with the userspace trying to use the CPU hotplug at the same time
1416          */
1417         cpumask_clear(frozen_cpus);
1418
1419         pr_info("Disabling non-boot CPUs ...\n");
1420         for_each_online_cpu(cpu) {
1421                 if (cpu == primary)
1422                         continue;
1423
1424                 if (pm_wakeup_pending()) {
1425                         pr_info("Wakeup pending. Abort CPU freeze\n");
1426                         error = -EBUSY;
1427                         break;
1428                 }
1429
1430                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1431                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1432                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1433                 if (!error)
1434                         cpumask_set_cpu(cpu, frozen_cpus);
1435                 else {
1436                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1437                         break;
1438                 }
1439         }
1440
1441         if (!error)
1442                 BUG_ON(num_online_cpus() > 1);
1443         else
1444                 pr_err("Non-boot CPUs are not disabled\n");
1445
1446         /*
1447          * Make sure the CPUs won't be enabled by someone else. We need to do
1448          * this even in case of failure as all freeze_secondary_cpus() users are
1449          * supposed to do thaw_secondary_cpus() on the failure path.
1450          */
1451         cpu_hotplug_disabled++;
1452
1453         cpu_maps_update_done();
1454         return error;
1455 }
1456
1457 void __weak arch_thaw_secondary_cpus_begin(void)
1458 {
1459 }
1460
1461 void __weak arch_thaw_secondary_cpus_end(void)
1462 {
1463 }
1464
1465 void thaw_secondary_cpus(void)
1466 {
1467         int cpu, error;
1468
1469         /* Allow everyone to use the CPU hotplug again */
1470         cpu_maps_update_begin();
1471         __cpu_hotplug_enable();
1472         if (cpumask_empty(frozen_cpus))
1473                 goto out;
1474
1475         pr_info("Enabling non-boot CPUs ...\n");
1476
1477         arch_thaw_secondary_cpus_begin();
1478
1479         for_each_cpu(cpu, frozen_cpus) {
1480                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1481                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1482                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1483                 if (!error) {
1484                         pr_info("CPU%d is up\n", cpu);
1485                         continue;
1486                 }
1487                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1488         }
1489
1490         arch_thaw_secondary_cpus_end();
1491
1492         cpumask_clear(frozen_cpus);
1493 out:
1494         cpu_maps_update_done();
1495 }
1496
1497 static int __init alloc_frozen_cpus(void)
1498 {
1499         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1500                 return -ENOMEM;
1501         return 0;
1502 }
1503 core_initcall(alloc_frozen_cpus);
1504
1505 /*
1506  * When callbacks for CPU hotplug notifications are being executed, we must
1507  * ensure that the state of the system with respect to the tasks being frozen
1508  * or not, as reported by the notification, remains unchanged *throughout the
1509  * duration* of the execution of the callbacks.
1510  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1511  *
1512  * This synchronization is implemented by mutually excluding regular CPU
1513  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1514  * Hibernate notifications.
1515  */
1516 static int
1517 cpu_hotplug_pm_callback(struct notifier_block *nb,
1518                         unsigned long action, void *ptr)
1519 {
1520         switch (action) {
1521
1522         case PM_SUSPEND_PREPARE:
1523         case PM_HIBERNATION_PREPARE:
1524                 cpu_hotplug_disable();
1525                 break;
1526
1527         case PM_POST_SUSPEND:
1528         case PM_POST_HIBERNATION:
1529                 cpu_hotplug_enable();
1530                 break;
1531
1532         default:
1533                 return NOTIFY_DONE;
1534         }
1535
1536         return NOTIFY_OK;
1537 }
1538
1539
1540 static int __init cpu_hotplug_pm_sync_init(void)
1541 {
1542         /*
1543          * cpu_hotplug_pm_callback has higher priority than x86
1544          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1545          * to disable cpu hotplug to avoid cpu hotplug race.
1546          */
1547         pm_notifier(cpu_hotplug_pm_callback, 0);
1548         return 0;
1549 }
1550 core_initcall(cpu_hotplug_pm_sync_init);
1551
1552 #endif /* CONFIG_PM_SLEEP_SMP */
1553
1554 int __boot_cpu_id;
1555
1556 #endif /* CONFIG_SMP */
1557
1558 /* Boot processor state steps */
1559 static struct cpuhp_step cpuhp_hp_states[] = {
1560         [CPUHP_OFFLINE] = {
1561                 .name                   = "offline",
1562                 .startup.single         = NULL,
1563                 .teardown.single        = NULL,
1564         },
1565 #ifdef CONFIG_SMP
1566         [CPUHP_CREATE_THREADS]= {
1567                 .name                   = "threads:prepare",
1568                 .startup.single         = smpboot_create_threads,
1569                 .teardown.single        = NULL,
1570                 .cant_stop              = true,
1571         },
1572         [CPUHP_PERF_PREPARE] = {
1573                 .name                   = "perf:prepare",
1574                 .startup.single         = perf_event_init_cpu,
1575                 .teardown.single        = perf_event_exit_cpu,
1576         },
1577         [CPUHP_WORKQUEUE_PREP] = {
1578                 .name                   = "workqueue:prepare",
1579                 .startup.single         = workqueue_prepare_cpu,
1580                 .teardown.single        = NULL,
1581         },
1582         [CPUHP_HRTIMERS_PREPARE] = {
1583                 .name                   = "hrtimers:prepare",
1584                 .startup.single         = hrtimers_prepare_cpu,
1585                 .teardown.single        = hrtimers_dead_cpu,
1586         },
1587         [CPUHP_SMPCFD_PREPARE] = {
1588                 .name                   = "smpcfd:prepare",
1589                 .startup.single         = smpcfd_prepare_cpu,
1590                 .teardown.single        = smpcfd_dead_cpu,
1591         },
1592         [CPUHP_RELAY_PREPARE] = {
1593                 .name                   = "relay:prepare",
1594                 .startup.single         = relay_prepare_cpu,
1595                 .teardown.single        = NULL,
1596         },
1597         [CPUHP_SLAB_PREPARE] = {
1598                 .name                   = "slab:prepare",
1599                 .startup.single         = slab_prepare_cpu,
1600                 .teardown.single        = slab_dead_cpu,
1601         },
1602         [CPUHP_RCUTREE_PREP] = {
1603                 .name                   = "RCU/tree:prepare",
1604                 .startup.single         = rcutree_prepare_cpu,
1605                 .teardown.single        = rcutree_dead_cpu,
1606         },
1607         /*
1608          * On the tear-down path, timers_dead_cpu() must be invoked
1609          * before blk_mq_queue_reinit_notify() from notify_dead(),
1610          * otherwise a RCU stall occurs.
1611          */
1612         [CPUHP_TIMERS_PREPARE] = {
1613                 .name                   = "timers:prepare",
1614                 .startup.single         = timers_prepare_cpu,
1615                 .teardown.single        = timers_dead_cpu,
1616         },
1617         /* Kicks the plugged cpu into life */
1618         [CPUHP_BRINGUP_CPU] = {
1619                 .name                   = "cpu:bringup",
1620                 .startup.single         = bringup_cpu,
1621                 .teardown.single        = finish_cpu,
1622                 .cant_stop              = true,
1623         },
1624         /* Final state before CPU kills itself */
1625         [CPUHP_AP_IDLE_DEAD] = {
1626                 .name                   = "idle:dead",
1627         },
1628         /*
1629          * Last state before CPU enters the idle loop to die. Transient state
1630          * for synchronization.
1631          */
1632         [CPUHP_AP_OFFLINE] = {
1633                 .name                   = "ap:offline",
1634                 .cant_stop              = true,
1635         },
1636         /* First state is scheduler control. Interrupts are disabled */
1637         [CPUHP_AP_SCHED_STARTING] = {
1638                 .name                   = "sched:starting",
1639                 .startup.single         = sched_cpu_starting,
1640                 .teardown.single        = sched_cpu_dying,
1641         },
1642         [CPUHP_AP_RCUTREE_DYING] = {
1643                 .name                   = "RCU/tree:dying",
1644                 .startup.single         = NULL,
1645                 .teardown.single        = rcutree_dying_cpu,
1646         },
1647         [CPUHP_AP_SMPCFD_DYING] = {
1648                 .name                   = "smpcfd:dying",
1649                 .startup.single         = NULL,
1650                 .teardown.single        = smpcfd_dying_cpu,
1651         },
1652         /* Entry state on starting. Interrupts enabled from here on. Transient
1653          * state for synchronsization */
1654         [CPUHP_AP_ONLINE] = {
1655                 .name                   = "ap:online",
1656         },
1657         /*
1658          * Handled on controll processor until the plugged processor manages
1659          * this itself.
1660          */
1661         [CPUHP_TEARDOWN_CPU] = {
1662                 .name                   = "cpu:teardown",
1663                 .startup.single         = NULL,
1664                 .teardown.single        = takedown_cpu,
1665                 .cant_stop              = true,
1666         },
1667         /* Handle smpboot threads park/unpark */
1668         [CPUHP_AP_SMPBOOT_THREADS] = {
1669                 .name                   = "smpboot/threads:online",
1670                 .startup.single         = smpboot_unpark_threads,
1671                 .teardown.single        = smpboot_park_threads,
1672         },
1673         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1674                 .name                   = "irq/affinity:online",
1675                 .startup.single         = irq_affinity_online_cpu,
1676                 .teardown.single        = NULL,
1677         },
1678         [CPUHP_AP_PERF_ONLINE] = {
1679                 .name                   = "perf:online",
1680                 .startup.single         = perf_event_init_cpu,
1681                 .teardown.single        = perf_event_exit_cpu,
1682         },
1683         [CPUHP_AP_WATCHDOG_ONLINE] = {
1684                 .name                   = "lockup_detector:online",
1685                 .startup.single         = lockup_detector_online_cpu,
1686                 .teardown.single        = lockup_detector_offline_cpu,
1687         },
1688         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1689                 .name                   = "workqueue:online",
1690                 .startup.single         = workqueue_online_cpu,
1691                 .teardown.single        = workqueue_offline_cpu,
1692         },
1693         [CPUHP_AP_RCUTREE_ONLINE] = {
1694                 .name                   = "RCU/tree:online",
1695                 .startup.single         = rcutree_online_cpu,
1696                 .teardown.single        = rcutree_offline_cpu,
1697         },
1698 #endif
1699         /*
1700          * The dynamically registered state space is here
1701          */
1702
1703 #ifdef CONFIG_SMP
1704         /* Last state is scheduler control setting the cpu active */
1705         [CPUHP_AP_ACTIVE] = {
1706                 .name                   = "sched:active",
1707                 .startup.single         = sched_cpu_activate,
1708                 .teardown.single        = sched_cpu_deactivate,
1709         },
1710 #endif
1711
1712         /* CPU is fully up and running. */
1713         [CPUHP_ONLINE] = {
1714                 .name                   = "online",
1715                 .startup.single         = NULL,
1716                 .teardown.single        = NULL,
1717         },
1718 };
1719
1720 /* Sanity check for callbacks */
1721 static int cpuhp_cb_check(enum cpuhp_state state)
1722 {
1723         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1724                 return -EINVAL;
1725         return 0;
1726 }
1727
1728 /*
1729  * Returns a free for dynamic slot assignment of the Online state. The states
1730  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1731  * by having no name assigned.
1732  */
1733 static int cpuhp_reserve_state(enum cpuhp_state state)
1734 {
1735         enum cpuhp_state i, end;
1736         struct cpuhp_step *step;
1737
1738         switch (state) {
1739         case CPUHP_AP_ONLINE_DYN:
1740                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1741                 end = CPUHP_AP_ONLINE_DYN_END;
1742                 break;
1743         case CPUHP_BP_PREPARE_DYN:
1744                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1745                 end = CPUHP_BP_PREPARE_DYN_END;
1746                 break;
1747         default:
1748                 return -EINVAL;
1749         }
1750
1751         for (i = state; i <= end; i++, step++) {
1752                 if (!step->name)
1753                         return i;
1754         }
1755         WARN(1, "No more dynamic states available for CPU hotplug\n");
1756         return -ENOSPC;
1757 }
1758
1759 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1760                                  int (*startup)(unsigned int cpu),
1761                                  int (*teardown)(unsigned int cpu),
1762                                  bool multi_instance)
1763 {
1764         /* (Un)Install the callbacks for further cpu hotplug operations */
1765         struct cpuhp_step *sp;
1766         int ret = 0;
1767
1768         /*
1769          * If name is NULL, then the state gets removed.
1770          *
1771          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1772          * the first allocation from these dynamic ranges, so the removal
1773          * would trigger a new allocation and clear the wrong (already
1774          * empty) state, leaving the callbacks of the to be cleared state
1775          * dangling, which causes wreckage on the next hotplug operation.
1776          */
1777         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1778                      state == CPUHP_BP_PREPARE_DYN)) {
1779                 ret = cpuhp_reserve_state(state);
1780                 if (ret < 0)
1781                         return ret;
1782                 state = ret;
1783         }
1784         sp = cpuhp_get_step(state);
1785         if (name && sp->name)
1786                 return -EBUSY;
1787
1788         sp->startup.single = startup;
1789         sp->teardown.single = teardown;
1790         sp->name = name;
1791         sp->multi_instance = multi_instance;
1792         INIT_HLIST_HEAD(&sp->list);
1793         return ret;
1794 }
1795
1796 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1797 {
1798         return cpuhp_get_step(state)->teardown.single;
1799 }
1800
1801 /*
1802  * Call the startup/teardown function for a step either on the AP or
1803  * on the current CPU.
1804  */
1805 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1806                             struct hlist_node *node)
1807 {
1808         struct cpuhp_step *sp = cpuhp_get_step(state);
1809         int ret;
1810
1811         /*
1812          * If there's nothing to do, we done.
1813          * Relies on the union for multi_instance.
1814          */
1815         if ((bringup && !sp->startup.single) ||
1816             (!bringup && !sp->teardown.single))
1817                 return 0;
1818         /*
1819          * The non AP bound callbacks can fail on bringup. On teardown
1820          * e.g. module removal we crash for now.
1821          */
1822 #ifdef CONFIG_SMP
1823         if (cpuhp_is_ap_state(state))
1824                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1825         else
1826                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1827 #else
1828         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1829 #endif
1830         BUG_ON(ret && !bringup);
1831         return ret;
1832 }
1833
1834 /*
1835  * Called from __cpuhp_setup_state on a recoverable failure.
1836  *
1837  * Note: The teardown callbacks for rollback are not allowed to fail!
1838  */
1839 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1840                                    struct hlist_node *node)
1841 {
1842         int cpu;
1843
1844         /* Roll back the already executed steps on the other cpus */
1845         for_each_present_cpu(cpu) {
1846                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1847                 int cpustate = st->state;
1848
1849                 if (cpu >= failedcpu)
1850                         break;
1851
1852                 /* Did we invoke the startup call on that cpu ? */
1853                 if (cpustate >= state)
1854                         cpuhp_issue_call(cpu, state, false, node);
1855         }
1856 }
1857
1858 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1859                                           struct hlist_node *node,
1860                                           bool invoke)
1861 {
1862         struct cpuhp_step *sp;
1863         int cpu;
1864         int ret;
1865
1866         lockdep_assert_cpus_held();
1867
1868         sp = cpuhp_get_step(state);
1869         if (sp->multi_instance == false)
1870                 return -EINVAL;
1871
1872         mutex_lock(&cpuhp_state_mutex);
1873
1874         if (!invoke || !sp->startup.multi)
1875                 goto add_node;
1876
1877         /*
1878          * Try to call the startup callback for each present cpu
1879          * depending on the hotplug state of the cpu.
1880          */
1881         for_each_present_cpu(cpu) {
1882                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1883                 int cpustate = st->state;
1884
1885                 if (cpustate < state)
1886                         continue;
1887
1888                 ret = cpuhp_issue_call(cpu, state, true, node);
1889                 if (ret) {
1890                         if (sp->teardown.multi)
1891                                 cpuhp_rollback_install(cpu, state, node);
1892                         goto unlock;
1893                 }
1894         }
1895 add_node:
1896         ret = 0;
1897         hlist_add_head(node, &sp->list);
1898 unlock:
1899         mutex_unlock(&cpuhp_state_mutex);
1900         return ret;
1901 }
1902
1903 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1904                                bool invoke)
1905 {
1906         int ret;
1907
1908         cpus_read_lock();
1909         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1910         cpus_read_unlock();
1911         return ret;
1912 }
1913 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1914
1915 /**
1916  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1917  * @state:              The state to setup
1918  * @invoke:             If true, the startup function is invoked for cpus where
1919  *                      cpu state >= @state
1920  * @startup:            startup callback function
1921  * @teardown:           teardown callback function
1922  * @multi_instance:     State is set up for multiple instances which get
1923  *                      added afterwards.
1924  *
1925  * The caller needs to hold cpus read locked while calling this function.
1926  * Returns:
1927  *   On success:
1928  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1929  *      0 for all other states
1930  *   On failure: proper (negative) error code
1931  */
1932 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1933                                    const char *name, bool invoke,
1934                                    int (*startup)(unsigned int cpu),
1935                                    int (*teardown)(unsigned int cpu),
1936                                    bool multi_instance)
1937 {
1938         int cpu, ret = 0;
1939         bool dynstate;
1940
1941         lockdep_assert_cpus_held();
1942
1943         if (cpuhp_cb_check(state) || !name)
1944                 return -EINVAL;
1945
1946         mutex_lock(&cpuhp_state_mutex);
1947
1948         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1949                                     multi_instance);
1950
1951         dynstate = state == CPUHP_AP_ONLINE_DYN;
1952         if (ret > 0 && dynstate) {
1953                 state = ret;
1954                 ret = 0;
1955         }
1956
1957         if (ret || !invoke || !startup)
1958                 goto out;
1959
1960         /*
1961          * Try to call the startup callback for each present cpu
1962          * depending on the hotplug state of the cpu.
1963          */
1964         for_each_present_cpu(cpu) {
1965                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1966                 int cpustate = st->state;
1967
1968                 if (cpustate < state)
1969                         continue;
1970
1971                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1972                 if (ret) {
1973                         if (teardown)
1974                                 cpuhp_rollback_install(cpu, state, NULL);
1975                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1976                         goto out;
1977                 }
1978         }
1979 out:
1980         mutex_unlock(&cpuhp_state_mutex);
1981         /*
1982          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1983          * dynamically allocated state in case of success.
1984          */
1985         if (!ret && dynstate)
1986                 return state;
1987         return ret;
1988 }
1989 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1990
1991 int __cpuhp_setup_state(enum cpuhp_state state,
1992                         const char *name, bool invoke,
1993                         int (*startup)(unsigned int cpu),
1994                         int (*teardown)(unsigned int cpu),
1995                         bool multi_instance)
1996 {
1997         int ret;
1998
1999         cpus_read_lock();
2000         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2001                                              teardown, multi_instance);
2002         cpus_read_unlock();
2003         return ret;
2004 }
2005 EXPORT_SYMBOL(__cpuhp_setup_state);
2006
2007 int __cpuhp_state_remove_instance(enum cpuhp_state state,
2008                                   struct hlist_node *node, bool invoke)
2009 {
2010         struct cpuhp_step *sp = cpuhp_get_step(state);
2011         int cpu;
2012
2013         BUG_ON(cpuhp_cb_check(state));
2014
2015         if (!sp->multi_instance)
2016                 return -EINVAL;
2017
2018         cpus_read_lock();
2019         mutex_lock(&cpuhp_state_mutex);
2020
2021         if (!invoke || !cpuhp_get_teardown_cb(state))
2022                 goto remove;
2023         /*
2024          * Call the teardown callback for each present cpu depending
2025          * on the hotplug state of the cpu. This function is not
2026          * allowed to fail currently!
2027          */
2028         for_each_present_cpu(cpu) {
2029                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2030                 int cpustate = st->state;
2031
2032                 if (cpustate >= state)
2033                         cpuhp_issue_call(cpu, state, false, node);
2034         }
2035
2036 remove:
2037         hlist_del(node);
2038         mutex_unlock(&cpuhp_state_mutex);
2039         cpus_read_unlock();
2040
2041         return 0;
2042 }
2043 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2044
2045 /**
2046  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2047  * @state:      The state to remove
2048  * @invoke:     If true, the teardown function is invoked for cpus where
2049  *              cpu state >= @state
2050  *
2051  * The caller needs to hold cpus read locked while calling this function.
2052  * The teardown callback is currently not allowed to fail. Think
2053  * about module removal!
2054  */
2055 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2056 {
2057         struct cpuhp_step *sp = cpuhp_get_step(state);
2058         int cpu;
2059
2060         BUG_ON(cpuhp_cb_check(state));
2061
2062         lockdep_assert_cpus_held();
2063
2064         mutex_lock(&cpuhp_state_mutex);
2065         if (sp->multi_instance) {
2066                 WARN(!hlist_empty(&sp->list),
2067                      "Error: Removing state %d which has instances left.\n",
2068                      state);
2069                 goto remove;
2070         }
2071
2072         if (!invoke || !cpuhp_get_teardown_cb(state))
2073                 goto remove;
2074
2075         /*
2076          * Call the teardown callback for each present cpu depending
2077          * on the hotplug state of the cpu. This function is not
2078          * allowed to fail currently!
2079          */
2080         for_each_present_cpu(cpu) {
2081                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2082                 int cpustate = st->state;
2083
2084                 if (cpustate >= state)
2085                         cpuhp_issue_call(cpu, state, false, NULL);
2086         }
2087 remove:
2088         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2089         mutex_unlock(&cpuhp_state_mutex);
2090 }
2091 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2092
2093 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2094 {
2095         cpus_read_lock();
2096         __cpuhp_remove_state_cpuslocked(state, invoke);
2097         cpus_read_unlock();
2098 }
2099 EXPORT_SYMBOL(__cpuhp_remove_state);
2100
2101 #ifdef CONFIG_HOTPLUG_SMT
2102 static void cpuhp_offline_cpu_device(unsigned int cpu)
2103 {
2104         struct device *dev = get_cpu_device(cpu);
2105
2106         dev->offline = true;
2107         /* Tell user space about the state change */
2108         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2109 }
2110
2111 static void cpuhp_online_cpu_device(unsigned int cpu)
2112 {
2113         struct device *dev = get_cpu_device(cpu);
2114
2115         dev->offline = false;
2116         /* Tell user space about the state change */
2117         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2118 }
2119
2120 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2121 {
2122         int cpu, ret = 0;
2123
2124         cpu_maps_update_begin();
2125         for_each_online_cpu(cpu) {
2126                 if (topology_is_primary_thread(cpu))
2127                         continue;
2128                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2129                 if (ret)
2130                         break;
2131                 /*
2132                  * As this needs to hold the cpu maps lock it's impossible
2133                  * to call device_offline() because that ends up calling
2134                  * cpu_down() which takes cpu maps lock. cpu maps lock
2135                  * needs to be held as this might race against in kernel
2136                  * abusers of the hotplug machinery (thermal management).
2137                  *
2138                  * So nothing would update device:offline state. That would
2139                  * leave the sysfs entry stale and prevent onlining after
2140                  * smt control has been changed to 'off' again. This is
2141                  * called under the sysfs hotplug lock, so it is properly
2142                  * serialized against the regular offline usage.
2143                  */
2144                 cpuhp_offline_cpu_device(cpu);
2145         }
2146         if (!ret)
2147                 cpu_smt_control = ctrlval;
2148         cpu_maps_update_done();
2149         return ret;
2150 }
2151
2152 int cpuhp_smt_enable(void)
2153 {
2154         int cpu, ret = 0;
2155
2156         cpu_maps_update_begin();
2157         cpu_smt_control = CPU_SMT_ENABLED;
2158         for_each_present_cpu(cpu) {
2159                 /* Skip online CPUs and CPUs on offline nodes */
2160                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2161                         continue;
2162                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2163                 if (ret)
2164                         break;
2165                 /* See comment in cpuhp_smt_disable() */
2166                 cpuhp_online_cpu_device(cpu);
2167         }
2168         cpu_maps_update_done();
2169         return ret;
2170 }
2171 #endif
2172
2173 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2174 static ssize_t show_cpuhp_state(struct device *dev,
2175                                 struct device_attribute *attr, char *buf)
2176 {
2177         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2178
2179         return sprintf(buf, "%d\n", st->state);
2180 }
2181 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
2182
2183 static ssize_t write_cpuhp_target(struct device *dev,
2184                                   struct device_attribute *attr,
2185                                   const char *buf, size_t count)
2186 {
2187         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2188         struct cpuhp_step *sp;
2189         int target, ret;
2190
2191         ret = kstrtoint(buf, 10, &target);
2192         if (ret)
2193                 return ret;
2194
2195 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2196         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2197                 return -EINVAL;
2198 #else
2199         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2200                 return -EINVAL;
2201 #endif
2202
2203         ret = lock_device_hotplug_sysfs();
2204         if (ret)
2205                 return ret;
2206
2207         mutex_lock(&cpuhp_state_mutex);
2208         sp = cpuhp_get_step(target);
2209         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2210         mutex_unlock(&cpuhp_state_mutex);
2211         if (ret)
2212                 goto out;
2213
2214         if (st->state < target)
2215                 ret = cpu_up(dev->id, target);
2216         else
2217                 ret = cpu_down(dev->id, target);
2218 out:
2219         unlock_device_hotplug();
2220         return ret ? ret : count;
2221 }
2222
2223 static ssize_t show_cpuhp_target(struct device *dev,
2224                                  struct device_attribute *attr, char *buf)
2225 {
2226         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2227
2228         return sprintf(buf, "%d\n", st->target);
2229 }
2230 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
2231
2232
2233 static ssize_t write_cpuhp_fail(struct device *dev,
2234                                 struct device_attribute *attr,
2235                                 const char *buf, size_t count)
2236 {
2237         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2238         struct cpuhp_step *sp;
2239         int fail, ret;
2240
2241         ret = kstrtoint(buf, 10, &fail);
2242         if (ret)
2243                 return ret;
2244
2245         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2246                 return -EINVAL;
2247
2248         /*
2249          * Cannot fail STARTING/DYING callbacks.
2250          */
2251         if (cpuhp_is_atomic_state(fail))
2252                 return -EINVAL;
2253
2254         /*
2255          * Cannot fail anything that doesn't have callbacks.
2256          */
2257         mutex_lock(&cpuhp_state_mutex);
2258         sp = cpuhp_get_step(fail);
2259         if (!sp->startup.single && !sp->teardown.single)
2260                 ret = -EINVAL;
2261         mutex_unlock(&cpuhp_state_mutex);
2262         if (ret)
2263                 return ret;
2264
2265         st->fail = fail;
2266
2267         return count;
2268 }
2269
2270 static ssize_t show_cpuhp_fail(struct device *dev,
2271                                struct device_attribute *attr, char *buf)
2272 {
2273         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2274
2275         return sprintf(buf, "%d\n", st->fail);
2276 }
2277
2278 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2279
2280 static struct attribute *cpuhp_cpu_attrs[] = {
2281         &dev_attr_state.attr,
2282         &dev_attr_target.attr,
2283         &dev_attr_fail.attr,
2284         NULL
2285 };
2286
2287 static const struct attribute_group cpuhp_cpu_attr_group = {
2288         .attrs = cpuhp_cpu_attrs,
2289         .name = "hotplug",
2290         NULL
2291 };
2292
2293 static ssize_t show_cpuhp_states(struct device *dev,
2294                                  struct device_attribute *attr, char *buf)
2295 {
2296         ssize_t cur, res = 0;
2297         int i;
2298
2299         mutex_lock(&cpuhp_state_mutex);
2300         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2301                 struct cpuhp_step *sp = cpuhp_get_step(i);
2302
2303                 if (sp->name) {
2304                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2305                         buf += cur;
2306                         res += cur;
2307                 }
2308         }
2309         mutex_unlock(&cpuhp_state_mutex);
2310         return res;
2311 }
2312 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2313
2314 static struct attribute *cpuhp_cpu_root_attrs[] = {
2315         &dev_attr_states.attr,
2316         NULL
2317 };
2318
2319 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2320         .attrs = cpuhp_cpu_root_attrs,
2321         .name = "hotplug",
2322         NULL
2323 };
2324
2325 #ifdef CONFIG_HOTPLUG_SMT
2326
2327 static ssize_t
2328 __store_smt_control(struct device *dev, struct device_attribute *attr,
2329                     const char *buf, size_t count)
2330 {
2331         int ctrlval, ret;
2332
2333         if (sysfs_streq(buf, "on"))
2334                 ctrlval = CPU_SMT_ENABLED;
2335         else if (sysfs_streq(buf, "off"))
2336                 ctrlval = CPU_SMT_DISABLED;
2337         else if (sysfs_streq(buf, "forceoff"))
2338                 ctrlval = CPU_SMT_FORCE_DISABLED;
2339         else
2340                 return -EINVAL;
2341
2342         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2343                 return -EPERM;
2344
2345         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2346                 return -ENODEV;
2347
2348         ret = lock_device_hotplug_sysfs();
2349         if (ret)
2350                 return ret;
2351
2352         if (ctrlval != cpu_smt_control) {
2353                 switch (ctrlval) {
2354                 case CPU_SMT_ENABLED:
2355                         ret = cpuhp_smt_enable();
2356                         break;
2357                 case CPU_SMT_DISABLED:
2358                 case CPU_SMT_FORCE_DISABLED:
2359                         ret = cpuhp_smt_disable(ctrlval);
2360                         break;
2361                 }
2362         }
2363
2364         unlock_device_hotplug();
2365         return ret ? ret : count;
2366 }
2367
2368 #else /* !CONFIG_HOTPLUG_SMT */
2369 static ssize_t
2370 __store_smt_control(struct device *dev, struct device_attribute *attr,
2371                     const char *buf, size_t count)
2372 {
2373         return -ENODEV;
2374 }
2375 #endif /* CONFIG_HOTPLUG_SMT */
2376
2377 static const char *smt_states[] = {
2378         [CPU_SMT_ENABLED]               = "on",
2379         [CPU_SMT_DISABLED]              = "off",
2380         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2381         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2382         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2383 };
2384
2385 static ssize_t
2386 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2387 {
2388         const char *state = smt_states[cpu_smt_control];
2389
2390         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2391 }
2392
2393 static ssize_t
2394 store_smt_control(struct device *dev, struct device_attribute *attr,
2395                   const char *buf, size_t count)
2396 {
2397         return __store_smt_control(dev, attr, buf, count);
2398 }
2399 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2400
2401 static ssize_t
2402 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2403 {
2404         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2405 }
2406 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2407
2408 static struct attribute *cpuhp_smt_attrs[] = {
2409         &dev_attr_control.attr,
2410         &dev_attr_active.attr,
2411         NULL
2412 };
2413
2414 static const struct attribute_group cpuhp_smt_attr_group = {
2415         .attrs = cpuhp_smt_attrs,
2416         .name = "smt",
2417         NULL
2418 };
2419
2420 static int __init cpu_smt_sysfs_init(void)
2421 {
2422         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2423                                   &cpuhp_smt_attr_group);
2424 }
2425
2426 static int __init cpuhp_sysfs_init(void)
2427 {
2428         int cpu, ret;
2429
2430         ret = cpu_smt_sysfs_init();
2431         if (ret)
2432                 return ret;
2433
2434         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2435                                  &cpuhp_cpu_root_attr_group);
2436         if (ret)
2437                 return ret;
2438
2439         for_each_possible_cpu(cpu) {
2440                 struct device *dev = get_cpu_device(cpu);
2441
2442                 if (!dev)
2443                         continue;
2444                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2445                 if (ret)
2446                         return ret;
2447         }
2448         return 0;
2449 }
2450 device_initcall(cpuhp_sysfs_init);
2451 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2452
2453 /*
2454  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2455  * represents all NR_CPUS bits binary values of 1<<nr.
2456  *
2457  * It is used by cpumask_of() to get a constant address to a CPU
2458  * mask value that has a single bit set only.
2459  */
2460
2461 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2462 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2463 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2464 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2465 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2466
2467 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2468
2469         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2470         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2471 #if BITS_PER_LONG > 32
2472         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2473         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2474 #endif
2475 };
2476 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2477
2478 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2479 EXPORT_SYMBOL(cpu_all_bits);
2480
2481 #ifdef CONFIG_INIT_ALL_POSSIBLE
2482 struct cpumask __cpu_possible_mask __read_mostly
2483         = {CPU_BITS_ALL};
2484 #else
2485 struct cpumask __cpu_possible_mask __read_mostly;
2486 #endif
2487 EXPORT_SYMBOL(__cpu_possible_mask);
2488
2489 struct cpumask __cpu_online_mask __read_mostly;
2490 EXPORT_SYMBOL(__cpu_online_mask);
2491
2492 struct cpumask __cpu_present_mask __read_mostly;
2493 EXPORT_SYMBOL(__cpu_present_mask);
2494
2495 struct cpumask __cpu_active_mask __read_mostly;
2496 EXPORT_SYMBOL(__cpu_active_mask);
2497
2498 atomic_t __num_online_cpus __read_mostly;
2499 EXPORT_SYMBOL(__num_online_cpus);
2500
2501 void init_cpu_present(const struct cpumask *src)
2502 {
2503         cpumask_copy(&__cpu_present_mask, src);
2504 }
2505
2506 void init_cpu_possible(const struct cpumask *src)
2507 {
2508         cpumask_copy(&__cpu_possible_mask, src);
2509 }
2510
2511 void init_cpu_online(const struct cpumask *src)
2512 {
2513         cpumask_copy(&__cpu_online_mask, src);
2514 }
2515
2516 void set_cpu_online(unsigned int cpu, bool online)
2517 {
2518         /*
2519          * atomic_inc/dec() is required to handle the horrid abuse of this
2520          * function by the reboot and kexec code which invoke it from
2521          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2522          * regular CPU hotplug is properly serialized.
2523          *
2524          * Note, that the fact that __num_online_cpus is of type atomic_t
2525          * does not protect readers which are not serialized against
2526          * concurrent hotplug operations.
2527          */
2528         if (online) {
2529                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2530                         atomic_inc(&__num_online_cpus);
2531         } else {
2532                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2533                         atomic_dec(&__num_online_cpus);
2534         }
2535 }
2536
2537 /*
2538  * Activate the first processor.
2539  */
2540 void __init boot_cpu_init(void)
2541 {
2542         int cpu = smp_processor_id();
2543
2544         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2545         set_cpu_online(cpu, true);
2546         set_cpu_active(cpu, true);
2547         set_cpu_present(cpu, true);
2548         set_cpu_possible(cpu, true);
2549
2550 #ifdef CONFIG_SMP
2551         __boot_cpu_id = cpu;
2552 #endif
2553 }
2554
2555 /*
2556  * Must be called _AFTER_ setting up the per_cpu areas
2557  */
2558 void __init boot_cpu_hotplug_init(void)
2559 {
2560 #ifdef CONFIG_SMP
2561         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2562 #endif
2563         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2564 }
2565
2566 /*
2567  * These are used for a global "mitigations=" cmdline option for toggling
2568  * optional CPU mitigations.
2569  */
2570 enum cpu_mitigations {
2571         CPU_MITIGATIONS_OFF,
2572         CPU_MITIGATIONS_AUTO,
2573         CPU_MITIGATIONS_AUTO_NOSMT,
2574 };
2575
2576 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2577         CPU_MITIGATIONS_AUTO;
2578
2579 static int __init mitigations_parse_cmdline(char *arg)
2580 {
2581         if (!strcmp(arg, "off"))
2582                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2583         else if (!strcmp(arg, "auto"))
2584                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2585         else if (!strcmp(arg, "auto,nosmt"))
2586                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2587         else
2588                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2589                         arg);
2590
2591         return 0;
2592 }
2593 early_param("mitigations", mitigations_parse_cmdline);
2594
2595 /* mitigations=off */
2596 bool cpu_mitigations_off(void)
2597 {
2598         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2599 }
2600 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2601
2602 /* mitigations=auto,nosmt */
2603 bool cpu_mitigations_auto_nosmt(void)
2604 {
2605         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2606 }
2607 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);