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
  36 #include <trace/events/power.h>
  37 #define CREATE_TRACE_POINTS
  38 #include <trace/events/cpuhp.h>
  39
  40 #include "smpboot.h"
  41
  42 /**
  43  * cpuhp_cpu_state - Per cpu hotplug state storage
  44  * @state:      The current cpu state
  45  * @target:     The target state
  46  * @thread:     Pointer to the hotplug thread
  47  * @should_run: Thread should execute
  48  * @rollback:   Perform a rollback
  49  * @single:     Single callback invocation
  50  * @bringup:    Single callback bringup or teardown selector
  51  * @cb_state:   The state for a single callback (install/uninstall)
  52  * @result:     Result of the operation
  53  * @done_up:    Signal completion to the issuer of the task for cpu-up
  54  * @done_down:  Signal completion to the issuer of the task for cpu-down
  55  */
  56 struct cpuhp_cpu_state {
  57         enum cpuhp_state        state;
  58         enum cpuhp_state        target;
  59         enum cpuhp_state        fail;
  60 #ifdef CONFIG_SMP
  61         struct task_struct      *thread;
  62         bool                    should_run;
  63         bool                    rollback;
  64         bool                    single;
  65         bool                    bringup;
  66         struct hlist_node       *node;
  67         struct hlist_node       *last;
  68         enum cpuhp_state        cb_state;
  69         int                     result;
  70         struct completion       done_up;
  71         struct completion       done_down;
  72 #endif
  73 };
  74
  75 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  76         .fail = CPUHP_INVALID,
  77 };
  78
  79 #ifdef CONFIG_SMP
  80 cpumask_t cpus_booted_once_mask;
  81 #endif
  82
  83 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  84 static struct lockdep_map cpuhp_state_up_map =
  85         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  86 static struct lockdep_map cpuhp_state_down_map =
  87         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  88
  89
  90 static inline void cpuhp_lock_acquire(bool bringup)
  91 {
  92         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  93 }
  94
  95 static inline void cpuhp_lock_release(bool bringup)
  96 {
  97         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  98 }
  99 #else
 100
 101 static inline void cpuhp_lock_acquire(bool bringup) { }
 102 static inline void cpuhp_lock_release(bool bringup) { }
 103
 104 #endif
 105
 106 /**
 107  * cpuhp_step - Hotplug state machine step
 108  * @name:       Name of the step
 109  * @startup:    Startup function of the step
 110  * @teardown:   Teardown function of the step
 111  * @cant_stop:  Bringup/teardown can't be stopped at this step
 112  */
 113 struct cpuhp_step {
 114         const char              *name;
 115         union {
 116                 int             (*single)(unsigned int cpu);
 117                 int             (*multi)(unsigned int cpu,
 118                                          struct hlist_node *node);
 119         } startup;
 120         union {
 121                 int             (*single)(unsigned int cpu);
 122                 int             (*multi)(unsigned int cpu,
 123                                          struct hlist_node *node);
 124         } teardown;
 125         struct hlist_head       list;
 126         bool                    cant_stop;
 127         bool                    multi_instance;
 128 };
 129
 130 static DEFINE_MUTEX(cpuhp_state_mutex);
 131 static struct cpuhp_step cpuhp_hp_states[];
 132
 133 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 134 {
 135         return cpuhp_hp_states + state;
 136 }
 137
 138 /**
 139  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 140  * @cpu:        The cpu for which the callback should be invoked
 141  * @state:      The state to do callbacks for
 142  * @bringup:    True if the bringup callback should be invoked
 143  * @node:       For multi-instance, do a single entry callback for install/remove
 144  * @lastp:      For multi-instance rollback, remember how far we got
 145  *
 146  * Called from cpu hotplug and from the state register machinery.
 147  */
 148 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 149                                  bool bringup, struct hlist_node *node,
 150                                  struct hlist_node **lastp)
 151 {
 152         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 153         struct cpuhp_step *step = cpuhp_get_step(state);
 154         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 155         int (*cb)(unsigned int cpu);
 156         int ret, cnt;
 157
 158         if (st->fail == state) {
 159                 st->fail = CPUHP_INVALID;
 160
 161                 if (!(bringup ? step->startup.single : step->teardown.single))
 162                         return 0;
 163
 164                 return -EAGAIN;
 165         }
 166
 167         if (!step->multi_instance) {
 168                 WARN_ON_ONCE(lastp && *lastp);
 169                 cb = bringup ? step->startup.single : step->teardown.single;
 170                 if (!cb)
 171                         return 0;
 172                 trace_cpuhp_enter(cpu, st->target, state, cb);
 173                 ret = cb(cpu);
 174                 trace_cpuhp_exit(cpu, st->state, state, ret);
 175                 return ret;
 176         }
 177         cbm = bringup ? step->startup.multi : step->teardown.multi;
 178         if (!cbm)
 179                 return 0;
 180
 181         /* Single invocation for instance add/remove */
 182         if (node) {
 183                 WARN_ON_ONCE(lastp && *lastp);
 184                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 185                 ret = cbm(cpu, node);
 186                 trace_cpuhp_exit(cpu, st->state, state, ret);
 187                 return ret;
 188         }
 189
 190         /* State transition. Invoke on all instances */
 191         cnt = 0;
 192         hlist_for_each(node, &step->list) {
 193                 if (lastp && node == *lastp)
 194                         break;
 195
 196                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 197                 ret = cbm(cpu, node);
 198                 trace_cpuhp_exit(cpu, st->state, state, ret);
 199                 if (ret) {
 200                         if (!lastp)
 201                                 goto err;
 202
 203                         *lastp = node;
 204                         return ret;
 205                 }
 206                 cnt++;
 207         }
 208         if (lastp)
 209                 *lastp = NULL;
 210         return 0;
 211 err:
 212         /* Rollback the instances if one failed */
 213         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 214         if (!cbm)
 215                 return ret;
 216
 217         hlist_for_each(node, &step->list) {
 218                 if (!cnt--)
 219                         break;
 220
 221                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 222                 ret = cbm(cpu, node);
 223                 trace_cpuhp_exit(cpu, st->state, state, ret);
 224                 /*
 225                  * Rollback must not fail,
 226                  */
 227                 WARN_ON_ONCE(ret);
 228         }
 229         return ret;
 230 }
 231
 232 #ifdef CONFIG_SMP
 233 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 234 {
 235         /*
 236          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 237          * purposes as that state is handled explicitly in cpu_down.
 238          */
 239         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 240 }
 241
 242 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 243 {
 244         struct completion *done = bringup ? &st->done_up : &st->done_down;
 245         wait_for_completion(done);
 246 }
 247
 248 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 249 {
 250         struct completion *done = bringup ? &st->done_up : &st->done_down;
 251         complete(done);
 252 }
 253
 254 /*
 255  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 256  */
 257 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 258 {
 259         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 260 }
 261
 262 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 263 static DEFINE_MUTEX(cpu_add_remove_lock);
 264 bool cpuhp_tasks_frozen;
 265 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 266
 267 /*
 268  * The following two APIs (cpu_maps_update_begin/done) must be used when
 269  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 270  */
 271 void cpu_maps_update_begin(void)
 272 {
 273         mutex_lock(&cpu_add_remove_lock);
 274 }
 275
 276 void cpu_maps_update_done(void)
 277 {
 278         mutex_unlock(&cpu_add_remove_lock);
 279 }
 280
 281 /*
 282  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 283  * Should always be manipulated under cpu_add_remove_lock
 284  */
 285 static int cpu_hotplug_disabled;
 286
 287 #ifdef CONFIG_HOTPLUG_CPU
 288
 289 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 290
 291 void cpus_read_lock(void)
 292 {
 293         percpu_down_read(&cpu_hotplug_lock);
 294 }
 295 EXPORT_SYMBOL_GPL(cpus_read_lock);
 296
 297 int cpus_read_trylock(void)
 298 {
 299         return percpu_down_read_trylock(&cpu_hotplug_lock);
 300 }
 301 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 302
 303 void cpus_read_unlock(void)
 304 {
 305         percpu_up_read(&cpu_hotplug_lock);
 306 }
 307 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 308
 309 void cpus_write_lock(void)
 310 {
 311         percpu_down_write(&cpu_hotplug_lock);
 312 }
 313
 314 void cpus_write_unlock(void)
 315 {
 316         percpu_up_write(&cpu_hotplug_lock);
 317 }
 318
 319 void lockdep_assert_cpus_held(void)
 320 {
 321         /*
 322          * We can't have hotplug operations before userspace starts running,
 323          * and some init codepaths will knowingly not take the hotplug lock.
 324          * This is all valid, so mute lockdep until it makes sense to report
 325          * unheld locks.
 326          */
 327         if (system_state < SYSTEM_RUNNING)
 328                 return;
 329
 330         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 331 }
 332
 333 static void lockdep_acquire_cpus_lock(void)
 334 {
 335         rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
 336 }
 337
 338 static void lockdep_release_cpus_lock(void)
 339 {
 340         rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
 341 }
 342
 343 /*
 344  * Wait for currently running CPU hotplug operations to complete (if any) and
 345  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 346  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 347  * hotplug path before performing hotplug operations. So acquiring that lock
 348  * guarantees mutual exclusion from any currently running hotplug operations.
 349  */
 350 void cpu_hotplug_disable(void)
 351 {
 352         cpu_maps_update_begin();
 353         cpu_hotplug_disabled++;
 354         cpu_maps_update_done();
 355 }
 356 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 357
 358 static void __cpu_hotplug_enable(void)
 359 {
 360         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 361                 return;
 362         cpu_hotplug_disabled--;
 363 }
 364
 365 void cpu_hotplug_enable(void)
 366 {
 367         cpu_maps_update_begin();
 368         __cpu_hotplug_enable();
 369         cpu_maps_update_done();
 370 }
 371 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 372
 373 #else
 374
 375 static void lockdep_acquire_cpus_lock(void)
 376 {
 377 }
 378
 379 static void lockdep_release_cpus_lock(void)
 380 {
 381 }
 382
 383 #endif  /* CONFIG_HOTPLUG_CPU */
 384
 385 /*
 386  * Architectures that need SMT-specific errata handling during SMT hotplug
 387  * should override this.
 388  */
 389 void __weak arch_smt_update(void) { }
 390
 391 #ifdef CONFIG_HOTPLUG_SMT
 392 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 393
 394 void __init cpu_smt_disable(bool force)
 395 {
 396         if (!cpu_smt_possible())
 397                 return;
 398
 399         if (force) {
 400                 pr_info("SMT: Force disabled\n");
 401                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 402         } else {
 403                 pr_info("SMT: disabled\n");
 404                 cpu_smt_control = CPU_SMT_DISABLED;
 405         }
 406 }
 407
 408 /*
 409  * The decision whether SMT is supported can only be done after the full
 410  * CPU identification. Called from architecture code.
 411  */
 412 void __init cpu_smt_check_topology(void)
 413 {
 414         if (!topology_smt_supported())
 415                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 416 }
 417
 418 static int __init smt_cmdline_disable(char *str)
 419 {
 420         cpu_smt_disable(str && !strcmp(str, "force"));
 421         return 0;
 422 }
 423 early_param("nosmt", smt_cmdline_disable);
 424
 425 static inline bool cpu_smt_allowed(unsigned int cpu)
 426 {
 427         if (cpu_smt_control == CPU_SMT_ENABLED)
 428                 return true;
 429
 430         if (topology_is_primary_thread(cpu))
 431                 return true;
 432
 433         /*
 434          * On x86 it's required to boot all logical CPUs at least once so
 435          * that the init code can get a chance to set CR4.MCE on each
 436          * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
 437          * core will shutdown the machine.
 438          */
 439         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
 440 }
 441
 442 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
 443 bool cpu_smt_possible(void)
 444 {
 445         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
 446                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
 447 }
 448 EXPORT_SYMBOL_GPL(cpu_smt_possible);
 449 #else
 450 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 451 #endif
 452
 453 static inline enum cpuhp_state
 454 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 455 {
 456         enum cpuhp_state prev_state = st->state;
 457
 458         st->rollback = false;
 459         st->last = NULL;
 460
 461         st->target = target;
 462         st->single = false;
 463         st->bringup = st->state < target;
 464
 465         return prev_state;
 466 }
 467
 468 static inline void
 469 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 470 {
 471         st->rollback = true;
 472
 473         /*
 474          * If we have st->last we need to undo partial multi_instance of this
 475          * state first. Otherwise start undo at the previous state.
 476          */
 477         if (!st->last) {
 478                 if (st->bringup)
 479                         st->state--;
 480                 else
 481                         st->state++;
 482         }
 483
 484         st->target = prev_state;
 485         st->bringup = !st->bringup;
 486 }
 487
 488 /* Regular hotplug invocation of the AP hotplug thread */
 489 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 490 {
 491         if (!st->single && st->state == st->target)
 492                 return;
 493
 494         st->result = 0;
 495         /*
 496          * Make sure the above stores are visible before should_run becomes
 497          * true. Paired with the mb() above in cpuhp_thread_fun()
 498          */
 499         smp_mb();
 500         st->should_run = true;
 501         wake_up_process(st->thread);
 502         wait_for_ap_thread(st, st->bringup);
 503 }
 504
 505 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 506 {
 507         enum cpuhp_state prev_state;
 508         int ret;
 509
 510         prev_state = cpuhp_set_state(st, target);
 511         __cpuhp_kick_ap(st);
 512         if ((ret = st->result)) {
 513                 cpuhp_reset_state(st, prev_state);
 514                 __cpuhp_kick_ap(st);
 515         }
 516
 517         return ret;
 518 }
 519
 520 static int bringup_wait_for_ap(unsigned int cpu)
 521 {
 522         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 523
 524         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 525         wait_for_ap_thread(st, true);
 526         if (WARN_ON_ONCE((!cpu_online(cpu))))
 527                 return -ECANCELED;
 528
 529         /* Unpark the hotplug thread of the target cpu */
 530         kthread_unpark(st->thread);
 531
 532         /*
 533          * SMT soft disabling on X86 requires to bring the CPU out of the
 534          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 535          * CPU marked itself as booted_once in notify_cpu_starting() so the
 536          * cpu_smt_allowed() check will now return false if this is not the
 537          * primary sibling.
 538          */
 539         if (!cpu_smt_allowed(cpu))
 540                 return -ECANCELED;
 541
 542         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 543                 return 0;
 544
 545         return cpuhp_kick_ap(st, st->target);
 546 }
 547
 548 static int bringup_cpu(unsigned int cpu)
 549 {
 550         struct task_struct *idle = idle_thread_get(cpu);
 551         int ret;
 552
 553         /*
 554          * Some architectures have to walk the irq descriptors to
 555          * setup the vector space for the cpu which comes online.
 556          * Prevent irq alloc/free across the bringup.
 557          */
 558         irq_lock_sparse();
 559
 560         /* Arch-specific enabling code. */
 561         ret = __cpu_up(cpu, idle);
 562         irq_unlock_sparse();
 563         if (ret)
 564                 return ret;
 565         return bringup_wait_for_ap(cpu);
 566 }
 567
 568 static int finish_cpu(unsigned int cpu)
 569 {
 570         struct task_struct *idle = idle_thread_get(cpu);
 571         struct mm_struct *mm = idle->active_mm;
 572
 573         /*
 574          * idle_task_exit() will have switched to &init_mm, now
 575          * clean up any remaining active_mm state.
 576          */
 577         if (mm != &init_mm)
 578                 idle->active_mm = &init_mm;
 579         mmdrop(mm);
 580         return 0;
 581 }
 582
 583 /*
 584  * Hotplug state machine related functions
 585  */
 586
 587 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 588 {
 589         for (st->state--; st->state > st->target; st->state--)
 590                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 591 }
 592
 593 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 594 {
 595         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 596                 return true;
 597         /*
 598          * When CPU hotplug is disabled, then taking the CPU down is not
 599          * possible because takedown_cpu() and the architecture and
 600          * subsystem specific mechanisms are not available. So the CPU
 601          * which would be completely unplugged again needs to stay around
 602          * in the current state.
 603          */
 604         return st->state <= CPUHP_BRINGUP_CPU;
 605 }
 606
 607 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 608                               enum cpuhp_state target)
 609 {
 610         enum cpuhp_state prev_state = st->state;
 611         int ret = 0;
 612
 613         while (st->state < target) {
 614                 st->state++;
 615                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 616                 if (ret) {
 617                         if (can_rollback_cpu(st)) {
 618                                 st->target = prev_state;
 619                                 undo_cpu_up(cpu, st);
 620                         }
 621                         break;
 622                 }
 623         }
 624         return ret;
 625 }
 626
 627 /*
 628  * The cpu hotplug threads manage the bringup and teardown of the cpus
 629  */
 630 static void cpuhp_create(unsigned int cpu)
 631 {
 632         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 633
 634         init_completion(&st->done_up);
 635         init_completion(&st->done_down);
 636 }
 637
 638 static int cpuhp_should_run(unsigned int cpu)
 639 {
 640         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 641
 642         return st->should_run;
 643 }
 644
 645 /*
 646  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 647  * callbacks when a state gets [un]installed at runtime.
 648  *
 649  * Each invocation of this function by the smpboot thread does a single AP
 650  * state callback.
 651  *
 652  * It has 3 modes of operation:
 653  *  - single: runs st->cb_state
 654  *  - up:     runs ++st->state, while st->state < st->target
 655  *  - down:   runs st->state--, while st->state > st->target
 656  *
 657  * When complete or on error, should_run is cleared and the completion is fired.
 658  */
 659 static void cpuhp_thread_fun(unsigned int cpu)
 660 {
 661         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 662         bool bringup = st->bringup;
 663         enum cpuhp_state state;
 664
 665         if (WARN_ON_ONCE(!st->should_run))
 666                 return;
 667
 668         /*
 669          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 670          * that if we see ->should_run we also see the rest of the state.
 671          */
 672         smp_mb();
 673
 674         /*
 675          * The BP holds the hotplug lock, but we're now running on the AP,
 676          * ensure that anybody asserting the lock is held, will actually find
 677          * it so.
 678          */
 679         lockdep_acquire_cpus_lock();
 680         cpuhp_lock_acquire(bringup);
 681
 682         if (st->single) {
 683                 state = st->cb_state;
 684                 st->should_run = false;
 685         } else {
 686                 if (bringup) {
 687                         st->state++;
 688                         state = st->state;
 689                         st->should_run = (st->state < st->target);
 690                         WARN_ON_ONCE(st->state > st->target);
 691                 } else {
 692                         state = st->state;
 693                         st->state--;
 694                         st->should_run = (st->state > st->target);
 695                         WARN_ON_ONCE(st->state < st->target);
 696                 }
 697         }
 698
 699         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 700
 701         if (cpuhp_is_atomic_state(state)) {
 702                 local_irq_disable();
 703                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 704                 local_irq_enable();
 705
 706                 /*
 707                  * STARTING/DYING must not fail!
 708                  */
 709                 WARN_ON_ONCE(st->result);
 710         } else {
 711                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 712         }
 713
 714         if (st->result) {
 715                 /*
 716                  * If we fail on a rollback, we're up a creek without no
 717                  * paddle, no way forward, no way back. We loose, thanks for
 718                  * playing.
 719                  */
 720                 WARN_ON_ONCE(st->rollback);
 721                 st->should_run = false;
 722         }
 723
 724         cpuhp_lock_release(bringup);
 725         lockdep_release_cpus_lock();
 726
 727         if (!st->should_run)
 728                 complete_ap_thread(st, bringup);
 729 }
 730
 731 /* Invoke a single callback on a remote cpu */
 732 static int
 733 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 734                          struct hlist_node *node)
 735 {
 736         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 737         int ret;
 738
 739         if (!cpu_online(cpu))
 740                 return 0;
 741
 742         cpuhp_lock_acquire(false);
 743         cpuhp_lock_release(false);
 744
 745         cpuhp_lock_acquire(true);
 746         cpuhp_lock_release(true);
 747
 748         /*
 749          * If we are up and running, use the hotplug thread. For early calls
 750          * we invoke the thread function directly.
 751          */
 752         if (!st->thread)
 753                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 754
 755         st->rollback = false;
 756         st->last = NULL;
 757
 758         st->node = node;
 759         st->bringup = bringup;
 760         st->cb_state = state;
 761         st->single = true;
 762
 763         __cpuhp_kick_ap(st);
 764
 765         /*
 766          * If we failed and did a partial, do a rollback.
 767          */
 768         if ((ret = st->result) && st->last) {
 769                 st->rollback = true;
 770                 st->bringup = !bringup;
 771
 772                 __cpuhp_kick_ap(st);
 773         }
 774
 775         /*
 776          * Clean up the leftovers so the next hotplug operation wont use stale
 777          * data.
 778          */
 779         st->node = st->last = NULL;
 780         return ret;
 781 }
 782
 783 static int cpuhp_kick_ap_work(unsigned int cpu)
 784 {
 785         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 786         enum cpuhp_state prev_state = st->state;
 787         int ret;
 788
 789         cpuhp_lock_acquire(false);
 790         cpuhp_lock_release(false);
 791
 792         cpuhp_lock_acquire(true);
 793         cpuhp_lock_release(true);
 794
 795         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 796         ret = cpuhp_kick_ap(st, st->target);
 797         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 798
 799         return ret;
 800 }
 801
 802 static struct smp_hotplug_thread cpuhp_threads = {
 803         .store                  = &cpuhp_state.thread,
 804         .create                 = &cpuhp_create,
 805         .thread_should_run      = cpuhp_should_run,
 806         .thread_fn              = cpuhp_thread_fun,
 807         .thread_comm            = "cpuhp/%u",
 808         .selfparking            = true,
 809 };
 810
 811 void __init cpuhp_threads_init(void)
 812 {
 813         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 814         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 815 }
 816
 817 #ifdef CONFIG_HOTPLUG_CPU
 818 /**
 819  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 820  * @cpu: a CPU id
 821  *
 822  * This function walks all processes, finds a valid mm struct for each one and
 823  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 824  * trivial, there are various non-obvious corner cases, which this function
 825  * tries to solve in a safe manner.
 826  *
 827  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 828  * be called only for an already offlined CPU.
 829  */
 830 void clear_tasks_mm_cpumask(int cpu)
 831 {
 832         struct task_struct *p;
 833
 834         /*
 835          * This function is called after the cpu is taken down and marked
 836          * offline, so its not like new tasks will ever get this cpu set in
 837          * their mm mask. -- Peter Zijlstra
 838          * Thus, we may use rcu_read_lock() here, instead of grabbing
 839          * full-fledged tasklist_lock.
 840          */
 841         WARN_ON(cpu_online(cpu));
 842         rcu_read_lock();
 843         for_each_process(p) {
 844                 struct task_struct *t;
 845
 846                 /*
 847                  * Main thread might exit, but other threads may still have
 848                  * a valid mm. Find one.
 849                  */
 850                 t = find_lock_task_mm(p);
 851                 if (!t)
 852                         continue;
 853                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 854                 task_unlock(t);
 855         }
 856         rcu_read_unlock();
 857 }
 858
 859 /* Take this CPU down. */
 860 static int take_cpu_down(void *_param)
 861 {
 862         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 863         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 864         int err, cpu = smp_processor_id();
 865         int ret;
 866
 867         /* Ensure this CPU doesn't handle any more interrupts. */
 868         err = __cpu_disable();
 869         if (err < 0)
 870                 return err;
 871
 872         /*
 873          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 874          * do this step again.
 875          */
 876         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 877         st->state--;
 878         /* Invoke the former CPU_DYING callbacks */
 879         for (; st->state > target; st->state--) {
 880                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 881                 /*
 882                  * DYING must not fail!
 883                  */
 884                 WARN_ON_ONCE(ret);
 885         }
 886
 887         /* Give up timekeeping duties */
 888         tick_handover_do_timer();
 889         /* Remove CPU from timer broadcasting */
 890         tick_offline_cpu(cpu);
 891         /* Park the stopper thread */
 892         stop_machine_park(cpu);
 893         return 0;
 894 }
 895
 896 static int takedown_cpu(unsigned int cpu)
 897 {
 898         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 899         int err;
 900
 901         /* Park the smpboot threads */
 902         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 903
 904         /*
 905          * Prevent irq alloc/free while the dying cpu reorganizes the
 906          * interrupt affinities.
 907          */
 908         irq_lock_sparse();
 909
 910         /*
 911          * So now all preempt/rcu users must observe !cpu_active().
 912          */
 913         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 914         if (err) {
 915                 /* CPU refused to die */
 916                 irq_unlock_sparse();
 917                 /* Unpark the hotplug thread so we can rollback there */
 918                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 919                 return err;
 920         }
 921         BUG_ON(cpu_online(cpu));
 922
 923         /*
 924          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 925          * all runnable tasks from the CPU, there's only the idle task left now
 926          * that the migration thread is done doing the stop_machine thing.
 927          *
 928          * Wait for the stop thread to go away.
 929          */
 930         wait_for_ap_thread(st, false);
 931         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 932
 933         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 934         irq_unlock_sparse();
 935
 936         hotplug_cpu__broadcast_tick_pull(cpu);
 937         /* This actually kills the CPU. */
 938         __cpu_die(cpu);
 939
 940         tick_cleanup_dead_cpu(cpu);
 941         rcutree_migrate_callbacks(cpu);
 942         return 0;
 943 }
 944
 945 static void cpuhp_complete_idle_dead(void *arg)
 946 {
 947         struct cpuhp_cpu_state *st = arg;
 948
 949         complete_ap_thread(st, false);
 950 }
 951
 952 void cpuhp_report_idle_dead(void)
 953 {
 954         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 955
 956         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 957         rcu_report_dead(smp_processor_id());
 958         st->state = CPUHP_AP_IDLE_DEAD;
 959         /*
 960          * We cannot call complete after rcu_report_dead() so we delegate it
 961          * to an online cpu.
 962          */
 963         smp_call_function_single(cpumask_first(cpu_online_mask),
 964                                  cpuhp_complete_idle_dead, st, 0);
 965 }
 966
 967 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 968 {
 969         for (st->state++; st->state < st->target; st->state++)
 970                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 971 }
 972
 973 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 974                                 enum cpuhp_state target)
 975 {
 976         enum cpuhp_state prev_state = st->state;
 977         int ret = 0;
 978
 979         for (; st->state > target; st->state--) {
 980                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 981                 if (ret) {
 982                         st->target = prev_state;
 983                         if (st->state < prev_state)
 984                                 undo_cpu_down(cpu, st);
 985                         break;
 986                 }
 987         }
 988         return ret;
 989 }
 990
 991 /* Requires cpu_add_remove_lock to be held */
 992 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 993                            enum cpuhp_state target)
 994 {
 995         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 996         int prev_state, ret = 0;
 997
 998         if (num_online_cpus() == 1)
 999                 return -EBUSY;
1000
1001         if (!cpu_present(cpu))
1002                 return -EINVAL;
1003
1004         cpus_write_lock();
1005
1006         cpuhp_tasks_frozen = tasks_frozen;
1007
1008         prev_state = cpuhp_set_state(st, target);
1009         /*
1010          * If the current CPU state is in the range of the AP hotplug thread,
1011          * then we need to kick the thread.
1012          */
1013         if (st->state > CPUHP_TEARDOWN_CPU) {
1014                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1015                 ret = cpuhp_kick_ap_work(cpu);
1016                 /*
1017                  * The AP side has done the error rollback already. Just
1018                  * return the error code..
1019                  */
1020                 if (ret)
1021                         goto out;
1022
1023                 /*
1024                  * We might have stopped still in the range of the AP hotplug
1025                  * thread. Nothing to do anymore.
1026                  */
1027                 if (st->state > CPUHP_TEARDOWN_CPU)
1028                         goto out;
1029
1030                 st->target = target;
1031         }
1032         /*
1033          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1034          * to do the further cleanups.
1035          */
1036         ret = cpuhp_down_callbacks(cpu, st, target);
1037         if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1038                 cpuhp_reset_state(st, prev_state);
1039                 __cpuhp_kick_ap(st);
1040         }
1041
1042 out:
1043         cpus_write_unlock();
1044         /*
1045          * Do post unplug cleanup. This is still protected against
1046          * concurrent CPU hotplug via cpu_add_remove_lock.
1047          */
1048         lockup_detector_cleanup();
1049         arch_smt_update();
1050         return ret;
1051 }
1052
1053 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1054 {
1055         if (cpu_hotplug_disabled)
1056                 return -EBUSY;
1057         return _cpu_down(cpu, 0, target);
1058 }
1059
1060 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1061 {
1062         int err;
1063
1064         cpu_maps_update_begin();
1065         err = cpu_down_maps_locked(cpu, target);
1066         cpu_maps_update_done();
1067         return err;
1068 }
1069
1070 /**
1071  * cpu_device_down - Bring down a cpu device
1072  * @dev: Pointer to the cpu device to offline
1073  *
1074  * This function is meant to be used by device core cpu subsystem only.
1075  *
1076  * Other subsystems should use remove_cpu() instead.
1077  */
1078 int cpu_device_down(struct device *dev)
1079 {
1080         return cpu_down(dev->id, CPUHP_OFFLINE);
1081 }
1082
1083 int remove_cpu(unsigned int cpu)
1084 {
1085         int ret;
1086
1087         lock_device_hotplug();
1088         ret = device_offline(get_cpu_device(cpu));
1089         unlock_device_hotplug();
1090
1091         return ret;
1092 }
1093 EXPORT_SYMBOL_GPL(remove_cpu);
1094
1095 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1096 {
1097         unsigned int cpu;
1098         int error;
1099
1100         cpu_maps_update_begin();
1101
1102         /*
1103          * Make certain the cpu I'm about to reboot on is online.
1104          *
1105          * This is inline to what migrate_to_reboot_cpu() already do.
1106          */
1107         if (!cpu_online(primary_cpu))
1108                 primary_cpu = cpumask_first(cpu_online_mask);
1109
1110         for_each_online_cpu(cpu) {
1111                 if (cpu == primary_cpu)
1112                         continue;
1113
1114                 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1115                 if (error) {
1116                         pr_err("Failed to offline CPU%d - error=%d",
1117                                 cpu, error);
1118                         break;
1119                 }
1120         }
1121
1122         /*
1123          * Ensure all but the reboot CPU are offline.
1124          */
1125         BUG_ON(num_online_cpus() > 1);
1126
1127         /*
1128          * Make sure the CPUs won't be enabled by someone else after this
1129          * point. Kexec will reboot to a new kernel shortly resetting
1130          * everything along the way.
1131          */
1132         cpu_hotplug_disabled++;
1133
1134         cpu_maps_update_done();
1135 }
1136
1137 #else
1138 #define takedown_cpu            NULL
1139 #endif /*CONFIG_HOTPLUG_CPU*/
1140
1141 /**
1142  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1143  * @cpu: cpu that just started
1144  *
1145  * It must be called by the arch code on the new cpu, before the new cpu
1146  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1147  */
1148 void notify_cpu_starting(unsigned int cpu)
1149 {
1150         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1151         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1152         int ret;
1153
1154         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1155         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1156         while (st->state < target) {
1157                 st->state++;
1158                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1159                 /*
1160                  * STARTING must not fail!
1161                  */
1162                 WARN_ON_ONCE(ret);
1163         }
1164 }
1165
1166 /*
1167  * Called from the idle task. Wake up the controlling task which brings the
1168  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1169  * online bringup to the hotplug thread.
1170  */
1171 void cpuhp_online_idle(enum cpuhp_state state)
1172 {
1173         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1174
1175         /* Happens for the boot cpu */
1176         if (state != CPUHP_AP_ONLINE_IDLE)
1177                 return;
1178
1179         /*
1180          * Unpart the stopper thread before we start the idle loop (and start
1181          * scheduling); this ensures the stopper task is always available.
1182          */
1183         stop_machine_unpark(smp_processor_id());
1184
1185         st->state = CPUHP_AP_ONLINE_IDLE;
1186         complete_ap_thread(st, true);
1187 }
1188
1189 /* Requires cpu_add_remove_lock to be held */
1190 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1191 {
1192         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1193         struct task_struct *idle;
1194         int ret = 0;
1195
1196         cpus_write_lock();
1197
1198         if (!cpu_present(cpu)) {
1199                 ret = -EINVAL;
1200                 goto out;
1201         }
1202
1203         /*
1204          * The caller of cpu_up() might have raced with another
1205          * caller. Nothing to do.
1206          */
1207         if (st->state >= target)
1208                 goto out;
1209
1210         if (st->state == CPUHP_OFFLINE) {
1211                 /* Let it fail before we try to bring the cpu up */
1212                 idle = idle_thread_get(cpu);
1213                 if (IS_ERR(idle)) {
1214                         ret = PTR_ERR(idle);
1215                         goto out;
1216                 }
1217         }
1218
1219         cpuhp_tasks_frozen = tasks_frozen;
1220
1221         cpuhp_set_state(st, target);
1222         /*
1223          * If the current CPU state is in the range of the AP hotplug thread,
1224          * then we need to kick the thread once more.
1225          */
1226         if (st->state > CPUHP_BRINGUP_CPU) {
1227                 ret = cpuhp_kick_ap_work(cpu);
1228                 /*
1229                  * The AP side has done the error rollback already. Just
1230                  * return the error code..
1231                  */
1232                 if (ret)
1233                         goto out;
1234         }
1235
1236         /*
1237          * Try to reach the target state. We max out on the BP at
1238          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1239          * responsible for bringing it up to the target state.
1240          */
1241         target = min((int)target, CPUHP_BRINGUP_CPU);
1242         ret = cpuhp_up_callbacks(cpu, st, target);
1243 out:
1244         cpus_write_unlock();
1245         arch_smt_update();
1246         return ret;
1247 }
1248
1249 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1250 {
1251         int err = 0;
1252
1253         if (!cpu_possible(cpu)) {
1254                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1255                        cpu);
1256 #if defined(CONFIG_IA64)
1257                 pr_err("please check additional_cpus= boot parameter\n");
1258 #endif
1259                 return -EINVAL;
1260         }
1261
1262         err = try_online_node(cpu_to_node(cpu));
1263         if (err)
1264                 return err;
1265
1266         cpu_maps_update_begin();
1267
1268         if (cpu_hotplug_disabled) {
1269                 err = -EBUSY;
1270                 goto out;
1271         }
1272         if (!cpu_smt_allowed(cpu)) {
1273                 err = -EPERM;
1274                 goto out;
1275         }
1276
1277         err = _cpu_up(cpu, 0, target);
1278 out:
1279         cpu_maps_update_done();
1280         return err;
1281 }
1282
1283 /**
1284  * cpu_device_up - Bring up a cpu device
1285  * @dev: Pointer to the cpu device to online
1286  *
1287  * This function is meant to be used by device core cpu subsystem only.
1288  *
1289  * Other subsystems should use add_cpu() instead.
1290  */
1291 int cpu_device_up(struct device *dev)
1292 {
1293         return cpu_up(dev->id, CPUHP_ONLINE);
1294 }
1295
1296 int add_cpu(unsigned int cpu)
1297 {
1298         int ret;
1299
1300         lock_device_hotplug();
1301         ret = device_online(get_cpu_device(cpu));
1302         unlock_device_hotplug();
1303
1304         return ret;
1305 }
1306 EXPORT_SYMBOL_GPL(add_cpu);
1307
1308 /**
1309  * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1310  * @sleep_cpu: The cpu we hibernated on and should be brought up.
1311  *
1312  * On some architectures like arm64, we can hibernate on any CPU, but on
1313  * wake up the CPU we hibernated on might be offline as a side effect of
1314  * using maxcpus= for example.
1315  */
1316 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1317 {
1318         int ret;
1319
1320         if (!cpu_online(sleep_cpu)) {
1321                 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1322                 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1323                 if (ret) {
1324                         pr_err("Failed to bring hibernate-CPU up!\n");
1325                         return ret;
1326                 }
1327         }
1328         return 0;
1329 }
1330
1331 void bringup_nonboot_cpus(unsigned int setup_max_cpus)
1332 {
1333         unsigned int cpu;
1334
1335         for_each_present_cpu(cpu) {
1336                 if (num_online_cpus() >= setup_max_cpus)
1337                         break;
1338                 if (!cpu_online(cpu))
1339                         cpu_up(cpu, CPUHP_ONLINE);
1340         }
1341 }
1342
1343 #ifdef CONFIG_PM_SLEEP_SMP
1344 static cpumask_var_t frozen_cpus;
1345
1346 int freeze_secondary_cpus(int primary)
1347 {
1348         int cpu, error = 0;
1349
1350         cpu_maps_update_begin();
1351         if (primary == -1) {
1352                 primary = cpumask_first(cpu_online_mask);
1353                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1354                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1355         } else {
1356                 if (!cpu_online(primary))
1357                         primary = cpumask_first(cpu_online_mask);
1358         }
1359
1360         /*
1361          * We take down all of the non-boot CPUs in one shot to avoid races
1362          * with the userspace trying to use the CPU hotplug at the same time
1363          */
1364         cpumask_clear(frozen_cpus);
1365
1366         pr_info("Disabling non-boot CPUs ...\n");
1367         for_each_online_cpu(cpu) {
1368                 if (cpu == primary)
1369                         continue;
1370
1371                 if (pm_wakeup_pending()) {
1372                         pr_info("Wakeup pending. Abort CPU freeze\n");
1373                         error = -EBUSY;
1374                         break;
1375                 }
1376
1377                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1378                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1379                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1380                 if (!error)
1381                         cpumask_set_cpu(cpu, frozen_cpus);
1382                 else {
1383                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1384                         break;
1385                 }
1386         }
1387
1388         if (!error)
1389                 BUG_ON(num_online_cpus() > 1);
1390         else
1391                 pr_err("Non-boot CPUs are not disabled\n");
1392
1393         /*
1394          * Make sure the CPUs won't be enabled by someone else. We need to do
1395          * this even in case of failure as all freeze_secondary_cpus() users are
1396          * supposed to do thaw_secondary_cpus() on the failure path.
1397          */
1398         cpu_hotplug_disabled++;
1399
1400         cpu_maps_update_done();
1401         return error;
1402 }
1403
1404 void __weak arch_thaw_secondary_cpus_begin(void)
1405 {
1406 }
1407
1408 void __weak arch_thaw_secondary_cpus_end(void)
1409 {
1410 }
1411
1412 void thaw_secondary_cpus(void)
1413 {
1414         int cpu, error;
1415
1416         /* Allow everyone to use the CPU hotplug again */
1417         cpu_maps_update_begin();
1418         __cpu_hotplug_enable();
1419         if (cpumask_empty(frozen_cpus))
1420                 goto out;
1421
1422         pr_info("Enabling non-boot CPUs ...\n");
1423
1424         arch_thaw_secondary_cpus_begin();
1425
1426         for_each_cpu(cpu, frozen_cpus) {
1427                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1428                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1429                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1430                 if (!error) {
1431                         pr_info("CPU%d is up\n", cpu);
1432                         continue;
1433                 }
1434                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1435         }
1436
1437         arch_thaw_secondary_cpus_end();
1438
1439         cpumask_clear(frozen_cpus);
1440 out:
1441         cpu_maps_update_done();
1442 }
1443
1444 static int __init alloc_frozen_cpus(void)
1445 {
1446         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1447                 return -ENOMEM;
1448         return 0;
1449 }
1450 core_initcall(alloc_frozen_cpus);
1451
1452 /*
1453  * When callbacks for CPU hotplug notifications are being executed, we must
1454  * ensure that the state of the system with respect to the tasks being frozen
1455  * or not, as reported by the notification, remains unchanged *throughout the
1456  * duration* of the execution of the callbacks.
1457  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1458  *
1459  * This synchronization is implemented by mutually excluding regular CPU
1460  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1461  * Hibernate notifications.
1462  */
1463 static int
1464 cpu_hotplug_pm_callback(struct notifier_block *nb,
1465                         unsigned long action, void *ptr)
1466 {
1467         switch (action) {
1468
1469         case PM_SUSPEND_PREPARE:
1470         case PM_HIBERNATION_PREPARE:
1471                 cpu_hotplug_disable();
1472                 break;
1473
1474         case PM_POST_SUSPEND:
1475         case PM_POST_HIBERNATION:
1476                 cpu_hotplug_enable();
1477                 break;
1478
1479         default:
1480                 return NOTIFY_DONE;
1481         }
1482
1483         return NOTIFY_OK;
1484 }
1485
1486
1487 static int __init cpu_hotplug_pm_sync_init(void)
1488 {
1489         /*
1490          * cpu_hotplug_pm_callback has higher priority than x86
1491          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1492          * to disable cpu hotplug to avoid cpu hotplug race.
1493          */
1494         pm_notifier(cpu_hotplug_pm_callback, 0);
1495         return 0;
1496 }
1497 core_initcall(cpu_hotplug_pm_sync_init);
1498
1499 #endif /* CONFIG_PM_SLEEP_SMP */
1500
1501 int __boot_cpu_id;
1502
1503 #endif /* CONFIG_SMP */
1504
1505 /* Boot processor state steps */
1506 static struct cpuhp_step cpuhp_hp_states[] = {
1507         [CPUHP_OFFLINE] = {
1508                 .name                   = "offline",
1509                 .startup.single         = NULL,
1510                 .teardown.single        = NULL,
1511         },
1512 #ifdef CONFIG_SMP
1513         [CPUHP_CREATE_THREADS]= {
1514                 .name                   = "threads:prepare",
1515                 .startup.single         = smpboot_create_threads,
1516                 .teardown.single        = NULL,
1517                 .cant_stop              = true,
1518         },
1519         [CPUHP_PERF_PREPARE] = {
1520                 .name                   = "perf:prepare",
1521                 .startup.single         = perf_event_init_cpu,
1522                 .teardown.single        = perf_event_exit_cpu,
1523         },
1524         [CPUHP_WORKQUEUE_PREP] = {
1525                 .name                   = "workqueue:prepare",
1526                 .startup.single         = workqueue_prepare_cpu,
1527                 .teardown.single        = NULL,
1528         },
1529         [CPUHP_HRTIMERS_PREPARE] = {
1530                 .name                   = "hrtimers:prepare",
1531                 .startup.single         = hrtimers_prepare_cpu,
1532                 .teardown.single        = hrtimers_dead_cpu,
1533         },
1534         [CPUHP_SMPCFD_PREPARE] = {
1535                 .name                   = "smpcfd:prepare",
1536                 .startup.single         = smpcfd_prepare_cpu,
1537                 .teardown.single        = smpcfd_dead_cpu,
1538         },
1539         [CPUHP_RELAY_PREPARE] = {
1540                 .name                   = "relay:prepare",
1541                 .startup.single         = relay_prepare_cpu,
1542                 .teardown.single        = NULL,
1543         },
1544         [CPUHP_SLAB_PREPARE] = {
1545                 .name                   = "slab:prepare",
1546                 .startup.single         = slab_prepare_cpu,
1547                 .teardown.single        = slab_dead_cpu,
1548         },
1549         [CPUHP_RCUTREE_PREP] = {
1550                 .name                   = "RCU/tree:prepare",
1551                 .startup.single         = rcutree_prepare_cpu,
1552                 .teardown.single        = rcutree_dead_cpu,
1553         },
1554         /*
1555          * On the tear-down path, timers_dead_cpu() must be invoked
1556          * before blk_mq_queue_reinit_notify() from notify_dead(),
1557          * otherwise a RCU stall occurs.
1558          */
1559         [CPUHP_TIMERS_PREPARE] = {
1560                 .name                   = "timers:prepare",
1561                 .startup.single         = timers_prepare_cpu,
1562                 .teardown.single        = timers_dead_cpu,
1563         },
1564         /* Kicks the plugged cpu into life */
1565         [CPUHP_BRINGUP_CPU] = {
1566                 .name                   = "cpu:bringup",
1567                 .startup.single         = bringup_cpu,
1568                 .teardown.single        = finish_cpu,
1569                 .cant_stop              = true,
1570         },
1571         /* Final state before CPU kills itself */
1572         [CPUHP_AP_IDLE_DEAD] = {
1573                 .name                   = "idle:dead",
1574         },
1575         /*
1576          * Last state before CPU enters the idle loop to die. Transient state
1577          * for synchronization.
1578          */
1579         [CPUHP_AP_OFFLINE] = {
1580                 .name                   = "ap:offline",
1581                 .cant_stop              = true,
1582         },
1583         /* First state is scheduler control. Interrupts are disabled */
1584         [CPUHP_AP_SCHED_STARTING] = {
1585                 .name                   = "sched:starting",
1586                 .startup.single         = sched_cpu_starting,
1587                 .teardown.single        = sched_cpu_dying,
1588         },
1589         [CPUHP_AP_RCUTREE_DYING] = {
1590                 .name                   = "RCU/tree:dying",
1591                 .startup.single         = NULL,
1592                 .teardown.single        = rcutree_dying_cpu,
1593         },
1594         [CPUHP_AP_SMPCFD_DYING] = {
1595                 .name                   = "smpcfd:dying",
1596                 .startup.single         = NULL,
1597                 .teardown.single        = smpcfd_dying_cpu,
1598         },
1599         /* Entry state on starting. Interrupts enabled from here on. Transient
1600          * state for synchronsization */
1601         [CPUHP_AP_ONLINE] = {
1602                 .name                   = "ap:online",
1603         },
1604         /*
1605          * Handled on controll processor until the plugged processor manages
1606          * this itself.
1607          */
1608         [CPUHP_TEARDOWN_CPU] = {
1609                 .name                   = "cpu:teardown",
1610                 .startup.single         = NULL,
1611                 .teardown.single        = takedown_cpu,
1612                 .cant_stop              = true,
1613         },
1614         /* Handle smpboot threads park/unpark */
1615         [CPUHP_AP_SMPBOOT_THREADS] = {
1616                 .name                   = "smpboot/threads:online",
1617                 .startup.single         = smpboot_unpark_threads,
1618                 .teardown.single        = smpboot_park_threads,
1619         },
1620         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1621                 .name                   = "irq/affinity:online",
1622                 .startup.single         = irq_affinity_online_cpu,
1623                 .teardown.single        = NULL,
1624         },
1625         [CPUHP_AP_PERF_ONLINE] = {
1626                 .name                   = "perf:online",
1627                 .startup.single         = perf_event_init_cpu,
1628                 .teardown.single        = perf_event_exit_cpu,
1629         },
1630         [CPUHP_AP_WATCHDOG_ONLINE] = {
1631                 .name                   = "lockup_detector:online",
1632                 .startup.single         = lockup_detector_online_cpu,
1633                 .teardown.single        = lockup_detector_offline_cpu,
1634         },
1635         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1636                 .name                   = "workqueue:online",
1637                 .startup.single         = workqueue_online_cpu,
1638                 .teardown.single        = workqueue_offline_cpu,
1639         },
1640         [CPUHP_AP_RCUTREE_ONLINE] = {
1641                 .name                   = "RCU/tree:online",
1642                 .startup.single         = rcutree_online_cpu,
1643                 .teardown.single        = rcutree_offline_cpu,
1644         },
1645 #endif
1646         /*
1647          * The dynamically registered state space is here
1648          */
1649
1650 #ifdef CONFIG_SMP
1651         /* Last state is scheduler control setting the cpu active */
1652         [CPUHP_AP_ACTIVE] = {
1653                 .name                   = "sched:active",
1654                 .startup.single         = sched_cpu_activate,
1655                 .teardown.single        = sched_cpu_deactivate,
1656         },
1657 #endif
1658
1659         /* CPU is fully up and running. */
1660         [CPUHP_ONLINE] = {
1661                 .name                   = "online",
1662                 .startup.single         = NULL,
1663                 .teardown.single        = NULL,
1664         },
1665 };
1666
1667 /* Sanity check for callbacks */
1668 static int cpuhp_cb_check(enum cpuhp_state state)
1669 {
1670         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1671                 return -EINVAL;
1672         return 0;
1673 }
1674
1675 /*
1676  * Returns a free for dynamic slot assignment of the Online state. The states
1677  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1678  * by having no name assigned.
1679  */
1680 static int cpuhp_reserve_state(enum cpuhp_state state)
1681 {
1682         enum cpuhp_state i, end;
1683         struct cpuhp_step *step;
1684
1685         switch (state) {
1686         case CPUHP_AP_ONLINE_DYN:
1687                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1688                 end = CPUHP_AP_ONLINE_DYN_END;
1689                 break;
1690         case CPUHP_BP_PREPARE_DYN:
1691                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1692                 end = CPUHP_BP_PREPARE_DYN_END;
1693                 break;
1694         default:
1695                 return -EINVAL;
1696         }
1697
1698         for (i = state; i <= end; i++, step++) {
1699                 if (!step->name)
1700                         return i;
1701         }
1702         WARN(1, "No more dynamic states available for CPU hotplug\n");
1703         return -ENOSPC;
1704 }
1705
1706 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1707                                  int (*startup)(unsigned int cpu),
1708                                  int (*teardown)(unsigned int cpu),
1709                                  bool multi_instance)
1710 {
1711         /* (Un)Install the callbacks for further cpu hotplug operations */
1712         struct cpuhp_step *sp;
1713         int ret = 0;
1714
1715         /*
1716          * If name is NULL, then the state gets removed.
1717          *
1718          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1719          * the first allocation from these dynamic ranges, so the removal
1720          * would trigger a new allocation and clear the wrong (already
1721          * empty) state, leaving the callbacks of the to be cleared state
1722          * dangling, which causes wreckage on the next hotplug operation.
1723          */
1724         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1725                      state == CPUHP_BP_PREPARE_DYN)) {
1726                 ret = cpuhp_reserve_state(state);
1727                 if (ret < 0)
1728                         return ret;
1729                 state = ret;
1730         }
1731         sp = cpuhp_get_step(state);
1732         if (name && sp->name)
1733                 return -EBUSY;
1734
1735         sp->startup.single = startup;
1736         sp->teardown.single = teardown;
1737         sp->name = name;
1738         sp->multi_instance = multi_instance;
1739         INIT_HLIST_HEAD(&sp->list);
1740         return ret;
1741 }
1742
1743 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1744 {
1745         return cpuhp_get_step(state)->teardown.single;
1746 }
1747
1748 /*
1749  * Call the startup/teardown function for a step either on the AP or
1750  * on the current CPU.
1751  */
1752 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1753                             struct hlist_node *node)
1754 {
1755         struct cpuhp_step *sp = cpuhp_get_step(state);
1756         int ret;
1757
1758         /*
1759          * If there's nothing to do, we done.
1760          * Relies on the union for multi_instance.
1761          */
1762         if ((bringup && !sp->startup.single) ||
1763             (!bringup && !sp->teardown.single))
1764                 return 0;
1765         /*
1766          * The non AP bound callbacks can fail on bringup. On teardown
1767          * e.g. module removal we crash for now.
1768          */
1769 #ifdef CONFIG_SMP
1770         if (cpuhp_is_ap_state(state))
1771                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1772         else
1773                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1774 #else
1775         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1776 #endif
1777         BUG_ON(ret && !bringup);
1778         return ret;
1779 }
1780
1781 /*
1782  * Called from __cpuhp_setup_state on a recoverable failure.
1783  *
1784  * Note: The teardown callbacks for rollback are not allowed to fail!
1785  */
1786 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1787                                    struct hlist_node *node)
1788 {
1789         int cpu;
1790
1791         /* Roll back the already executed steps on the other cpus */
1792         for_each_present_cpu(cpu) {
1793                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1794                 int cpustate = st->state;
1795
1796                 if (cpu >= failedcpu)
1797                         break;
1798
1799                 /* Did we invoke the startup call on that cpu ? */
1800                 if (cpustate >= state)
1801                         cpuhp_issue_call(cpu, state, false, node);
1802         }
1803 }
1804
1805 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1806                                           struct hlist_node *node,
1807                                           bool invoke)
1808 {
1809         struct cpuhp_step *sp;
1810         int cpu;
1811         int ret;
1812
1813         lockdep_assert_cpus_held();
1814
1815         sp = cpuhp_get_step(state);
1816         if (sp->multi_instance == false)
1817                 return -EINVAL;
1818
1819         mutex_lock(&cpuhp_state_mutex);
1820
1821         if (!invoke || !sp->startup.multi)
1822                 goto add_node;
1823
1824         /*
1825          * Try to call the startup callback for each present cpu
1826          * depending on the hotplug state of the cpu.
1827          */
1828         for_each_present_cpu(cpu) {
1829                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1830                 int cpustate = st->state;
1831
1832                 if (cpustate < state)
1833                         continue;
1834
1835                 ret = cpuhp_issue_call(cpu, state, true, node);
1836                 if (ret) {
1837                         if (sp->teardown.multi)
1838                                 cpuhp_rollback_install(cpu, state, node);
1839                         goto unlock;
1840                 }
1841         }
1842 add_node:
1843         ret = 0;
1844         hlist_add_head(node, &sp->list);
1845 unlock:
1846         mutex_unlock(&cpuhp_state_mutex);
1847         return ret;
1848 }
1849
1850 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1851                                bool invoke)
1852 {
1853         int ret;
1854
1855         cpus_read_lock();
1856         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1857         cpus_read_unlock();
1858         return ret;
1859 }
1860 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1861
1862 /**
1863  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1864  * @state:              The state to setup
1865  * @invoke:             If true, the startup function is invoked for cpus where
1866  *                      cpu state >= @state
1867  * @startup:            startup callback function
1868  * @teardown:           teardown callback function
1869  * @multi_instance:     State is set up for multiple instances which get
1870  *                      added afterwards.
1871  *
1872  * The caller needs to hold cpus read locked while calling this function.
1873  * Returns:
1874  *   On success:
1875  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1876  *      0 for all other states
1877  *   On failure: proper (negative) error code
1878  */
1879 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1880                                    const char *name, bool invoke,
1881                                    int (*startup)(unsigned int cpu),
1882                                    int (*teardown)(unsigned int cpu),
1883                                    bool multi_instance)
1884 {
1885         int cpu, ret = 0;
1886         bool dynstate;
1887
1888         lockdep_assert_cpus_held();
1889
1890         if (cpuhp_cb_check(state) || !name)
1891                 return -EINVAL;
1892
1893         mutex_lock(&cpuhp_state_mutex);
1894
1895         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1896                                     multi_instance);
1897
1898         dynstate = state == CPUHP_AP_ONLINE_DYN;
1899         if (ret > 0 && dynstate) {
1900                 state = ret;
1901                 ret = 0;
1902         }
1903
1904         if (ret || !invoke || !startup)
1905                 goto out;
1906
1907         /*
1908          * Try to call the startup callback for each present cpu
1909          * depending on the hotplug state of the cpu.
1910          */
1911         for_each_present_cpu(cpu) {
1912                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1913                 int cpustate = st->state;
1914
1915                 if (cpustate < state)
1916                         continue;
1917
1918                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1919                 if (ret) {
1920                         if (teardown)
1921                                 cpuhp_rollback_install(cpu, state, NULL);
1922                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1923                         goto out;
1924                 }
1925         }
1926 out:
1927         mutex_unlock(&cpuhp_state_mutex);
1928         /*
1929          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1930          * dynamically allocated state in case of success.
1931          */
1932         if (!ret && dynstate)
1933                 return state;
1934         return ret;
1935 }
1936 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1937
1938 int __cpuhp_setup_state(enum cpuhp_state state,
1939                         const char *name, bool invoke,
1940                         int (*startup)(unsigned int cpu),
1941                         int (*teardown)(unsigned int cpu),
1942                         bool multi_instance)
1943 {
1944         int ret;
1945
1946         cpus_read_lock();
1947         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1948                                              teardown, multi_instance);
1949         cpus_read_unlock();
1950         return ret;
1951 }
1952 EXPORT_SYMBOL(__cpuhp_setup_state);
1953
1954 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1955                                   struct hlist_node *node, bool invoke)
1956 {
1957         struct cpuhp_step *sp = cpuhp_get_step(state);
1958         int cpu;
1959
1960         BUG_ON(cpuhp_cb_check(state));
1961
1962         if (!sp->multi_instance)
1963                 return -EINVAL;
1964
1965         cpus_read_lock();
1966         mutex_lock(&cpuhp_state_mutex);
1967
1968         if (!invoke || !cpuhp_get_teardown_cb(state))
1969                 goto remove;
1970         /*
1971          * Call the teardown callback for each present cpu depending
1972          * on the hotplug state of the cpu. This function is not
1973          * allowed to fail currently!
1974          */
1975         for_each_present_cpu(cpu) {
1976                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1977                 int cpustate = st->state;
1978
1979                 if (cpustate >= state)
1980                         cpuhp_issue_call(cpu, state, false, node);
1981         }
1982
1983 remove:
1984         hlist_del(node);
1985         mutex_unlock(&cpuhp_state_mutex);
1986         cpus_read_unlock();
1987
1988         return 0;
1989 }
1990 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1991
1992 /**
1993  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1994  * @state:      The state to remove
1995  * @invoke:     If true, the teardown function is invoked for cpus where
1996  *              cpu state >= @state
1997  *
1998  * The caller needs to hold cpus read locked while calling this function.
1999  * The teardown callback is currently not allowed to fail. Think
2000  * about module removal!
2001  */
2002 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2003 {
2004         struct cpuhp_step *sp = cpuhp_get_step(state);
2005         int cpu;
2006
2007         BUG_ON(cpuhp_cb_check(state));
2008
2009         lockdep_assert_cpus_held();
2010
2011         mutex_lock(&cpuhp_state_mutex);
2012         if (sp->multi_instance) {
2013                 WARN(!hlist_empty(&sp->list),
2014                      "Error: Removing state %d which has instances left.\n",
2015                      state);
2016                 goto remove;
2017         }
2018
2019         if (!invoke || !cpuhp_get_teardown_cb(state))
2020                 goto remove;
2021
2022         /*
2023          * Call the teardown callback for each present cpu depending
2024          * on the hotplug state of the cpu. This function is not
2025          * allowed to fail currently!
2026          */
2027         for_each_present_cpu(cpu) {
2028                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2029                 int cpustate = st->state;
2030
2031                 if (cpustate >= state)
2032                         cpuhp_issue_call(cpu, state, false, NULL);
2033         }
2034 remove:
2035         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2036         mutex_unlock(&cpuhp_state_mutex);
2037 }
2038 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2039
2040 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2041 {
2042         cpus_read_lock();
2043         __cpuhp_remove_state_cpuslocked(state, invoke);
2044         cpus_read_unlock();
2045 }
2046 EXPORT_SYMBOL(__cpuhp_remove_state);
2047
2048 #ifdef CONFIG_HOTPLUG_SMT
2049 static void cpuhp_offline_cpu_device(unsigned int cpu)
2050 {
2051         struct device *dev = get_cpu_device(cpu);
2052
2053         dev->offline = true;
2054         /* Tell user space about the state change */
2055         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2056 }
2057
2058 static void cpuhp_online_cpu_device(unsigned int cpu)
2059 {
2060         struct device *dev = get_cpu_device(cpu);
2061
2062         dev->offline = false;
2063         /* Tell user space about the state change */
2064         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2065 }
2066
2067 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2068 {
2069         int cpu, ret = 0;
2070
2071         cpu_maps_update_begin();
2072         for_each_online_cpu(cpu) {
2073                 if (topology_is_primary_thread(cpu))
2074                         continue;
2075                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2076                 if (ret)
2077                         break;
2078                 /*
2079                  * As this needs to hold the cpu maps lock it's impossible
2080                  * to call device_offline() because that ends up calling
2081                  * cpu_down() which takes cpu maps lock. cpu maps lock
2082                  * needs to be held as this might race against in kernel
2083                  * abusers of the hotplug machinery (thermal management).
2084                  *
2085                  * So nothing would update device:offline state. That would
2086                  * leave the sysfs entry stale and prevent onlining after
2087                  * smt control has been changed to 'off' again. This is
2088                  * called under the sysfs hotplug lock, so it is properly
2089                  * serialized against the regular offline usage.
2090                  */
2091                 cpuhp_offline_cpu_device(cpu);
2092         }
2093         if (!ret)
2094                 cpu_smt_control = ctrlval;
2095         cpu_maps_update_done();
2096         return ret;
2097 }
2098
2099 int cpuhp_smt_enable(void)
2100 {
2101         int cpu, ret = 0;
2102
2103         cpu_maps_update_begin();
2104         cpu_smt_control = CPU_SMT_ENABLED;
2105         for_each_present_cpu(cpu) {
2106                 /* Skip online CPUs and CPUs on offline nodes */
2107                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2108                         continue;
2109                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2110                 if (ret)
2111                         break;
2112                 /* See comment in cpuhp_smt_disable() */
2113                 cpuhp_online_cpu_device(cpu);
2114         }
2115         cpu_maps_update_done();
2116         return ret;
2117 }
2118 #endif
2119
2120 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2121 static ssize_t show_cpuhp_state(struct device *dev,
2122                                 struct device_attribute *attr, char *buf)
2123 {
2124         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2125
2126         return sprintf(buf, "%d\n", st->state);
2127 }
2128 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
2129
2130 static ssize_t write_cpuhp_target(struct device *dev,
2131                                   struct device_attribute *attr,
2132                                   const char *buf, size_t count)
2133 {
2134         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2135         struct cpuhp_step *sp;
2136         int target, ret;
2137
2138         ret = kstrtoint(buf, 10, &target);
2139         if (ret)
2140                 return ret;
2141
2142 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2143         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2144                 return -EINVAL;
2145 #else
2146         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2147                 return -EINVAL;
2148 #endif
2149
2150         ret = lock_device_hotplug_sysfs();
2151         if (ret)
2152                 return ret;
2153
2154         mutex_lock(&cpuhp_state_mutex);
2155         sp = cpuhp_get_step(target);
2156         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2157         mutex_unlock(&cpuhp_state_mutex);
2158         if (ret)
2159                 goto out;
2160
2161         if (st->state < target)
2162                 ret = cpu_up(dev->id, target);
2163         else
2164                 ret = cpu_down(dev->id, target);
2165 out:
2166         unlock_device_hotplug();
2167         return ret ? ret : count;
2168 }
2169
2170 static ssize_t show_cpuhp_target(struct device *dev,
2171                                  struct device_attribute *attr, char *buf)
2172 {
2173         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2174
2175         return sprintf(buf, "%d\n", st->target);
2176 }
2177 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
2178
2179
2180 static ssize_t write_cpuhp_fail(struct device *dev,
2181                                 struct device_attribute *attr,
2182                                 const char *buf, size_t count)
2183 {
2184         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2185         struct cpuhp_step *sp;
2186         int fail, ret;
2187
2188         ret = kstrtoint(buf, 10, &fail);
2189         if (ret)
2190                 return ret;
2191
2192         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2193                 return -EINVAL;
2194
2195         /*
2196          * Cannot fail STARTING/DYING callbacks.
2197          */
2198         if (cpuhp_is_atomic_state(fail))
2199                 return -EINVAL;
2200
2201         /*
2202          * Cannot fail anything that doesn't have callbacks.
2203          */
2204         mutex_lock(&cpuhp_state_mutex);
2205         sp = cpuhp_get_step(fail);
2206         if (!sp->startup.single && !sp->teardown.single)
2207                 ret = -EINVAL;
2208         mutex_unlock(&cpuhp_state_mutex);
2209         if (ret)
2210                 return ret;
2211
2212         st->fail = fail;
2213
2214         return count;
2215 }
2216
2217 static ssize_t show_cpuhp_fail(struct device *dev,
2218                                struct device_attribute *attr, char *buf)
2219 {
2220         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2221
2222         return sprintf(buf, "%d\n", st->fail);
2223 }
2224
2225 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2226
2227 static struct attribute *cpuhp_cpu_attrs[] = {
2228         &dev_attr_state.attr,
2229         &dev_attr_target.attr,
2230         &dev_attr_fail.attr,
2231         NULL
2232 };
2233
2234 static const struct attribute_group cpuhp_cpu_attr_group = {
2235         .attrs = cpuhp_cpu_attrs,
2236         .name = "hotplug",
2237         NULL
2238 };
2239
2240 static ssize_t show_cpuhp_states(struct device *dev,
2241                                  struct device_attribute *attr, char *buf)
2242 {
2243         ssize_t cur, res = 0;
2244         int i;
2245
2246         mutex_lock(&cpuhp_state_mutex);
2247         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2248                 struct cpuhp_step *sp = cpuhp_get_step(i);
2249
2250                 if (sp->name) {
2251                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2252                         buf += cur;
2253                         res += cur;
2254                 }
2255         }
2256         mutex_unlock(&cpuhp_state_mutex);
2257         return res;
2258 }
2259 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2260
2261 static struct attribute *cpuhp_cpu_root_attrs[] = {
2262         &dev_attr_states.attr,
2263         NULL
2264 };
2265
2266 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2267         .attrs = cpuhp_cpu_root_attrs,
2268         .name = "hotplug",
2269         NULL
2270 };
2271
2272 #ifdef CONFIG_HOTPLUG_SMT
2273
2274 static ssize_t
2275 __store_smt_control(struct device *dev, struct device_attribute *attr,
2276                     const char *buf, size_t count)
2277 {
2278         int ctrlval, ret;
2279
2280         if (sysfs_streq(buf, "on"))
2281                 ctrlval = CPU_SMT_ENABLED;
2282         else if (sysfs_streq(buf, "off"))
2283                 ctrlval = CPU_SMT_DISABLED;
2284         else if (sysfs_streq(buf, "forceoff"))
2285                 ctrlval = CPU_SMT_FORCE_DISABLED;
2286         else
2287                 return -EINVAL;
2288
2289         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2290                 return -EPERM;
2291
2292         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2293                 return -ENODEV;
2294
2295         ret = lock_device_hotplug_sysfs();
2296         if (ret)
2297                 return ret;
2298
2299         if (ctrlval != cpu_smt_control) {
2300                 switch (ctrlval) {
2301                 case CPU_SMT_ENABLED:
2302                         ret = cpuhp_smt_enable();
2303                         break;
2304                 case CPU_SMT_DISABLED:
2305                 case CPU_SMT_FORCE_DISABLED:
2306                         ret = cpuhp_smt_disable(ctrlval);
2307                         break;
2308                 }
2309         }
2310
2311         unlock_device_hotplug();
2312         return ret ? ret : count;
2313 }
2314
2315 #else /* !CONFIG_HOTPLUG_SMT */
2316 static ssize_t
2317 __store_smt_control(struct device *dev, struct device_attribute *attr,
2318                     const char *buf, size_t count)
2319 {
2320         return -ENODEV;
2321 }
2322 #endif /* CONFIG_HOTPLUG_SMT */
2323
2324 static const char *smt_states[] = {
2325         [CPU_SMT_ENABLED]               = "on",
2326         [CPU_SMT_DISABLED]              = "off",
2327         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2328         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2329         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2330 };
2331
2332 static ssize_t
2333 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2334 {
2335         const char *state = smt_states[cpu_smt_control];
2336
2337         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2338 }
2339
2340 static ssize_t
2341 store_smt_control(struct device *dev, struct device_attribute *attr,
2342                   const char *buf, size_t count)
2343 {
2344         return __store_smt_control(dev, attr, buf, count);
2345 }
2346 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2347
2348 static ssize_t
2349 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2350 {
2351         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2352 }
2353 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2354
2355 static struct attribute *cpuhp_smt_attrs[] = {
2356         &dev_attr_control.attr,
2357         &dev_attr_active.attr,
2358         NULL
2359 };
2360
2361 static const struct attribute_group cpuhp_smt_attr_group = {
2362         .attrs = cpuhp_smt_attrs,
2363         .name = "smt",
2364         NULL
2365 };
2366
2367 static int __init cpu_smt_sysfs_init(void)
2368 {
2369         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2370                                   &cpuhp_smt_attr_group);
2371 }
2372
2373 static int __init cpuhp_sysfs_init(void)
2374 {
2375         int cpu, ret;
2376
2377         ret = cpu_smt_sysfs_init();
2378         if (ret)
2379                 return ret;
2380
2381         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2382                                  &cpuhp_cpu_root_attr_group);
2383         if (ret)
2384                 return ret;
2385
2386         for_each_possible_cpu(cpu) {
2387                 struct device *dev = get_cpu_device(cpu);
2388
2389                 if (!dev)
2390                         continue;
2391                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2392                 if (ret)
2393                         return ret;
2394         }
2395         return 0;
2396 }
2397 device_initcall(cpuhp_sysfs_init);
2398 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2399
2400 /*
2401  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2402  * represents all NR_CPUS bits binary values of 1<<nr.
2403  *
2404  * It is used by cpumask_of() to get a constant address to a CPU
2405  * mask value that has a single bit set only.
2406  */
2407
2408 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2409 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2410 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2411 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2412 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2413
2414 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2415
2416         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2417         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2418 #if BITS_PER_LONG > 32
2419         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2420         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2421 #endif
2422 };
2423 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2424
2425 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2426 EXPORT_SYMBOL(cpu_all_bits);
2427
2428 #ifdef CONFIG_INIT_ALL_POSSIBLE
2429 struct cpumask __cpu_possible_mask __read_mostly
2430         = {CPU_BITS_ALL};
2431 #else
2432 struct cpumask __cpu_possible_mask __read_mostly;
2433 #endif
2434 EXPORT_SYMBOL(__cpu_possible_mask);
2435
2436 struct cpumask __cpu_online_mask __read_mostly;
2437 EXPORT_SYMBOL(__cpu_online_mask);
2438
2439 struct cpumask __cpu_present_mask __read_mostly;
2440 EXPORT_SYMBOL(__cpu_present_mask);
2441
2442 struct cpumask __cpu_active_mask __read_mostly;
2443 EXPORT_SYMBOL(__cpu_active_mask);
2444
2445 atomic_t __num_online_cpus __read_mostly;
2446 EXPORT_SYMBOL(__num_online_cpus);
2447
2448 void init_cpu_present(const struct cpumask *src)
2449 {
2450         cpumask_copy(&__cpu_present_mask, src);
2451 }
2452
2453 void init_cpu_possible(const struct cpumask *src)
2454 {
2455         cpumask_copy(&__cpu_possible_mask, src);
2456 }
2457
2458 void init_cpu_online(const struct cpumask *src)
2459 {
2460         cpumask_copy(&__cpu_online_mask, src);
2461 }
2462
2463 void set_cpu_online(unsigned int cpu, bool online)
2464 {
2465         /*
2466          * atomic_inc/dec() is required to handle the horrid abuse of this
2467          * function by the reboot and kexec code which invoke it from
2468          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2469          * regular CPU hotplug is properly serialized.
2470          *
2471          * Note, that the fact that __num_online_cpus is of type atomic_t
2472          * does not protect readers which are not serialized against
2473          * concurrent hotplug operations.
2474          */
2475         if (online) {
2476                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2477                         atomic_inc(&__num_online_cpus);
2478         } else {
2479                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2480                         atomic_dec(&__num_online_cpus);
2481         }
2482 }
2483
2484 /*
2485  * Activate the first processor.
2486  */
2487 void __init boot_cpu_init(void)
2488 {
2489         int cpu = smp_processor_id();
2490
2491         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2492         set_cpu_online(cpu, true);
2493         set_cpu_active(cpu, true);
2494         set_cpu_present(cpu, true);
2495         set_cpu_possible(cpu, true);
2496
2497 #ifdef CONFIG_SMP
2498         __boot_cpu_id = cpu;
2499 #endif
2500 }
2501
2502 /*
2503  * Must be called _AFTER_ setting up the per_cpu areas
2504  */
2505 void __init boot_cpu_hotplug_init(void)
2506 {
2507 #ifdef CONFIG_SMP
2508         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2509 #endif
2510         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2511 }
2512
2513 /*
2514  * These are used for a global "mitigations=" cmdline option for toggling
2515  * optional CPU mitigations.
2516  */
2517 enum cpu_mitigations {
2518         CPU_MITIGATIONS_OFF,
2519         CPU_MITIGATIONS_AUTO,
2520         CPU_MITIGATIONS_AUTO_NOSMT,
2521 };
2522
2523 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2524         CPU_MITIGATIONS_AUTO;
2525
2526 static int __init mitigations_parse_cmdline(char *arg)
2527 {
2528         if (!strcmp(arg, "off"))
2529                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2530         else if (!strcmp(arg, "auto"))
2531                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2532         else if (!strcmp(arg, "auto,nosmt"))
2533                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2534         else
2535                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2536                         arg);
2537
2538         return 0;
2539 }
2540 early_param("mitigations", mitigations_parse_cmdline);
2541
2542 /* mitigations=off */
2543 bool cpu_mitigations_off(void)
2544 {
2545         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2546 }
2547 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2548
2549 /* mitigations=auto,nosmt */
2550 bool cpu_mitigations_auto_nosmt(void)
2551 {
2552         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2553 }
2554 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);