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  * struct cpuhp_cpu_state - Per cpu hotplug state storage
  45  * @state:      The current cpu state
  46  * @target:     The target state
  47  * @fail:       Current CPU hotplug callback state
  48  * @thread:     Pointer to the hotplug thread
  49  * @should_run: Thread should execute
  50  * @rollback:   Perform a rollback
  51  * @single:     Single callback invocation
  52  * @bringup:    Single callback bringup or teardown selector
  53  * @cpu:        CPU number
  54  * @node:       Remote CPU node; for multi-instance, do a
  55  *              single entry callback for install/remove
  56  * @last:       For multi-instance rollback, remember how far we got
  57  * @cb_state:   The state for a single callback (install/uninstall)
  58  * @result:     Result of the operation
  59  * @done_up:    Signal completion to the issuer of the task for cpu-up
  60  * @done_down:  Signal completion to the issuer of the task for cpu-down
  61  */
  62 struct cpuhp_cpu_state {
  63         enum cpuhp_state        state;
  64         enum cpuhp_state        target;
  65         enum cpuhp_state        fail;
  66 #ifdef CONFIG_SMP
  67         struct task_struct      *thread;
  68         bool                    should_run;
  69         bool                    rollback;
  70         bool                    single;
  71         bool                    bringup;
  72         int                     cpu;
  73         struct hlist_node       *node;
  74         struct hlist_node       *last;
  75         enum cpuhp_state        cb_state;
  76         int                     result;
  77         struct completion       done_up;
  78         struct completion       done_down;
  79 #endif
  80 };
  81
  82 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  83         .fail = CPUHP_INVALID,
  84 };
  85
  86 #ifdef CONFIG_SMP
  87 cpumask_t cpus_booted_once_mask;
  88 #endif
  89
  90 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  91 static struct lockdep_map cpuhp_state_up_map =
  92         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  93 static struct lockdep_map cpuhp_state_down_map =
  94         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  95
  96
  97 static inline void cpuhp_lock_acquire(bool bringup)
  98 {
  99         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
 100 }
 101
 102 static inline void cpuhp_lock_release(bool bringup)
 103 {
 104         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
 105 }
 106 #else
 107
 108 static inline void cpuhp_lock_acquire(bool bringup) { }
 109 static inline void cpuhp_lock_release(bool bringup) { }
 110
 111 #endif
 112
 113 /**
 114  * struct cpuhp_step - Hotplug state machine step
 115  * @name:       Name of the step
 116  * @startup:    Startup function of the step
 117  * @teardown:   Teardown function of the step
 118  * @cant_stop:  Bringup/teardown can't be stopped at this step
 119  * @multi_instance:     State has multiple instances which get added afterwards
 120  */
 121 struct cpuhp_step {
 122         const char              *name;
 123         union {
 124                 int             (*single)(unsigned int cpu);
 125                 int             (*multi)(unsigned int cpu,
 126                                          struct hlist_node *node);
 127         } startup;
 128         union {
 129                 int             (*single)(unsigned int cpu);
 130                 int             (*multi)(unsigned int cpu,
 131                                          struct hlist_node *node);
 132         } teardown;
 133         /* private: */
 134         struct hlist_head       list;
 135         /* public: */
 136         bool                    cant_stop;
 137         bool                    multi_instance;
 138 };
 139
 140 static DEFINE_MUTEX(cpuhp_state_mutex);
 141 static struct cpuhp_step cpuhp_hp_states[];
 142
 143 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 144 {
 145         return cpuhp_hp_states + state;
 146 }
 147
 148 static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
 149 {
 150         return bringup ? !step->startup.single : !step->teardown.single;
 151 }
 152
 153 /**
 154  * cpuhp_invoke_callback - Invoke the callbacks for a given state
 155  * @cpu:        The cpu for which the callback should be invoked
 156  * @state:      The state to do callbacks for
 157  * @bringup:    True if the bringup callback should be invoked
 158  * @node:       For multi-instance, do a single entry callback for install/remove
 159  * @lastp:      For multi-instance rollback, remember how far we got
 160  *
 161  * Called from cpu hotplug and from the state register machinery.
 162  *
 163  * Return: %0 on success or a negative errno code
 164  */
 165 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 166                                  bool bringup, struct hlist_node *node,
 167                                  struct hlist_node **lastp)
 168 {
 169         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 170         struct cpuhp_step *step = cpuhp_get_step(state);
 171         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 172         int (*cb)(unsigned int cpu);
 173         int ret, cnt;
 174
 175         if (st->fail == state) {
 176                 st->fail = CPUHP_INVALID;
 177                 return -EAGAIN;
 178         }
 179
 180         if (cpuhp_step_empty(bringup, step)) {
 181                 WARN_ON_ONCE(1);
 182                 return 0;
 183         }
 184
 185         if (!step->multi_instance) {
 186                 WARN_ON_ONCE(lastp && *lastp);
 187                 cb = bringup ? step->startup.single : step->teardown.single;
 188
 189                 trace_cpuhp_enter(cpu, st->target, state, cb);
 190                 ret = cb(cpu);
 191                 trace_cpuhp_exit(cpu, st->state, state, ret);
 192                 return ret;
 193         }
 194         cbm = bringup ? step->startup.multi : step->teardown.multi;
 195
 196         /* Single invocation for instance add/remove */
 197         if (node) {
 198                 WARN_ON_ONCE(lastp && *lastp);
 199                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 200                 ret = cbm(cpu, node);
 201                 trace_cpuhp_exit(cpu, st->state, state, ret);
 202                 return ret;
 203         }
 204
 205         /* State transition. Invoke on all instances */
 206         cnt = 0;
 207         hlist_for_each(node, &step->list) {
 208                 if (lastp && node == *lastp)
 209                         break;
 210
 211                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 212                 ret = cbm(cpu, node);
 213                 trace_cpuhp_exit(cpu, st->state, state, ret);
 214                 if (ret) {
 215                         if (!lastp)
 216                                 goto err;
 217
 218                         *lastp = node;
 219                         return ret;
 220                 }
 221                 cnt++;
 222         }
 223         if (lastp)
 224                 *lastp = NULL;
 225         return 0;
 226 err:
 227         /* Rollback the instances if one failed */
 228         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 229         if (!cbm)
 230                 return ret;
 231
 232         hlist_for_each(node, &step->list) {
 233                 if (!cnt--)
 234                         break;
 235
 236                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 237                 ret = cbm(cpu, node);
 238                 trace_cpuhp_exit(cpu, st->state, state, ret);
 239                 /*
 240                  * Rollback must not fail,
 241                  */
 242                 WARN_ON_ONCE(ret);
 243         }
 244         return ret;
 245 }
 246
 247 #ifdef CONFIG_SMP
 248 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 249 {
 250         /*
 251          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 252          * purposes as that state is handled explicitly in cpu_down.
 253          */
 254         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 255 }
 256
 257 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 258 {
 259         struct completion *done = bringup ? &st->done_up : &st->done_down;
 260         wait_for_completion(done);
 261 }
 262
 263 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 264 {
 265         struct completion *done = bringup ? &st->done_up : &st->done_down;
 266         complete(done);
 267 }
 268
 269 /*
 270  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 271  */
 272 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 273 {
 274         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 275 }
 276
 277 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 278 static DEFINE_MUTEX(cpu_add_remove_lock);
 279 bool cpuhp_tasks_frozen;
 280 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 281
 282 /*
 283  * The following two APIs (cpu_maps_update_begin/done) must be used when
 284  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 285  */
 286 void cpu_maps_update_begin(void)
 287 {
 288         mutex_lock(&cpu_add_remove_lock);
 289 }
 290
 291 void cpu_maps_update_done(void)
 292 {
 293         mutex_unlock(&cpu_add_remove_lock);
 294 }
 295
 296 /*
 297  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 298  * Should always be manipulated under cpu_add_remove_lock
 299  */
 300 static int cpu_hotplug_disabled;
 301
 302 #ifdef CONFIG_HOTPLUG_CPU
 303
 304 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 305
 306 void cpus_read_lock(void)
 307 {
 308         percpu_down_read(&cpu_hotplug_lock);
 309 }
 310 EXPORT_SYMBOL_GPL(cpus_read_lock);
 311
 312 int cpus_read_trylock(void)
 313 {
 314         return percpu_down_read_trylock(&cpu_hotplug_lock);
 315 }
 316 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 317
 318 void cpus_read_unlock(void)
 319 {
 320         percpu_up_read(&cpu_hotplug_lock);
 321 }
 322 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 323
 324 void cpus_write_lock(void)
 325 {
 326         percpu_down_write(&cpu_hotplug_lock);
 327 }
 328
 329 void cpus_write_unlock(void)
 330 {
 331         percpu_up_write(&cpu_hotplug_lock);
 332 }
 333
 334 void lockdep_assert_cpus_held(void)
 335 {
 336         /*
 337          * We can't have hotplug operations before userspace starts running,
 338          * and some init codepaths will knowingly not take the hotplug lock.
 339          * This is all valid, so mute lockdep until it makes sense to report
 340          * unheld locks.
 341          */
 342         if (system_state < SYSTEM_RUNNING)
 343                 return;
 344
 345         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 346 }
 347
 348 #ifdef CONFIG_LOCKDEP
 349 int lockdep_is_cpus_held(void)
 350 {
 351         return percpu_rwsem_is_held(&cpu_hotplug_lock);
 352 }
 353 #endif
 354
 355 static void lockdep_acquire_cpus_lock(void)
 356 {
 357         rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
 358 }
 359
 360 static void lockdep_release_cpus_lock(void)
 361 {
 362         rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
 363 }
 364
 365 /*
 366  * Wait for currently running CPU hotplug operations to complete (if any) and
 367  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 368  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 369  * hotplug path before performing hotplug operations. So acquiring that lock
 370  * guarantees mutual exclusion from any currently running hotplug operations.
 371  */
 372 void cpu_hotplug_disable(void)
 373 {
 374         cpu_maps_update_begin();
 375         cpu_hotplug_disabled++;
 376         cpu_maps_update_done();
 377 }
 378 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 379
 380 static void __cpu_hotplug_enable(void)
 381 {
 382         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 383                 return;
 384         cpu_hotplug_disabled--;
 385 }
 386
 387 void cpu_hotplug_enable(void)
 388 {
 389         cpu_maps_update_begin();
 390         __cpu_hotplug_enable();
 391         cpu_maps_update_done();
 392 }
 393 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 394
 395 #else
 396
 397 static void lockdep_acquire_cpus_lock(void)
 398 {
 399 }
 400
 401 static void lockdep_release_cpus_lock(void)
 402 {
 403 }
 404
 405 #endif  /* CONFIG_HOTPLUG_CPU */
 406
 407 /*
 408  * Architectures that need SMT-specific errata handling during SMT hotplug
 409  * should override this.
 410  */
 411 void __weak arch_smt_update(void) { }
 412
 413 #ifdef CONFIG_HOTPLUG_SMT
 414 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 415
 416 void __init cpu_smt_disable(bool force)
 417 {
 418         if (!cpu_smt_possible())
 419                 return;
 420
 421         if (force) {
 422                 pr_info("SMT: Force disabled\n");
 423                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 424         } else {
 425                 pr_info("SMT: disabled\n");
 426                 cpu_smt_control = CPU_SMT_DISABLED;
 427         }
 428 }
 429
 430 /*
 431  * The decision whether SMT is supported can only be done after the full
 432  * CPU identification. Called from architecture code.
 433  */
 434 void __init cpu_smt_check_topology(void)
 435 {
 436         if (!topology_smt_supported())
 437                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 438 }
 439
 440 static int __init smt_cmdline_disable(char *str)
 441 {
 442         cpu_smt_disable(str && !strcmp(str, "force"));
 443         return 0;
 444 }
 445 early_param("nosmt", smt_cmdline_disable);
 446
 447 static inline bool cpu_smt_allowed(unsigned int cpu)
 448 {
 449         if (cpu_smt_control == CPU_SMT_ENABLED)
 450                 return true;
 451
 452         if (topology_is_primary_thread(cpu))
 453                 return true;
 454
 455         /*
 456          * On x86 it's required to boot all logical CPUs at least once so
 457          * that the init code can get a chance to set CR4.MCE on each
 458          * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
 459          * core will shutdown the machine.
 460          */
 461         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
 462 }
 463
 464 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
 465 bool cpu_smt_possible(void)
 466 {
 467         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
 468                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
 469 }
 470 EXPORT_SYMBOL_GPL(cpu_smt_possible);
 471 #else
 472 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 473 #endif
 474
 475 static inline enum cpuhp_state
 476 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 477 {
 478         enum cpuhp_state prev_state = st->state;
 479         bool bringup = st->state < target;
 480
 481         st->rollback = false;
 482         st->last = NULL;
 483
 484         st->target = target;
 485         st->single = false;
 486         st->bringup = bringup;
 487         if (cpu_dying(st->cpu) != !bringup)
 488                 set_cpu_dying(st->cpu, !bringup);
 489
 490         return prev_state;
 491 }
 492
 493 static inline void
 494 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 495 {
 496         bool bringup = !st->bringup;
 497
 498         st->target = prev_state;
 499
 500         /*
 501          * Already rolling back. No need invert the bringup value or to change
 502          * the current state.
 503          */
 504         if (st->rollback)
 505                 return;
 506
 507         st->rollback = true;
 508
 509         /*
 510          * If we have st->last we need to undo partial multi_instance of this
 511          * state first. Otherwise start undo at the previous state.
 512          */
 513         if (!st->last) {
 514                 if (st->bringup)
 515                         st->state--;
 516                 else
 517                         st->state++;
 518         }
 519
 520         st->bringup = bringup;
 521         if (cpu_dying(st->cpu) != !bringup)
 522                 set_cpu_dying(st->cpu, !bringup);
 523 }
 524
 525 /* Regular hotplug invocation of the AP hotplug thread */
 526 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 527 {
 528         if (!st->single && st->state == st->target)
 529                 return;
 530
 531         st->result = 0;
 532         /*
 533          * Make sure the above stores are visible before should_run becomes
 534          * true. Paired with the mb() above in cpuhp_thread_fun()
 535          */
 536         smp_mb();
 537         st->should_run = true;
 538         wake_up_process(st->thread);
 539         wait_for_ap_thread(st, st->bringup);
 540 }
 541
 542 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 543 {
 544         enum cpuhp_state prev_state;
 545         int ret;
 546
 547         prev_state = cpuhp_set_state(st, target);
 548         __cpuhp_kick_ap(st);
 549         if ((ret = st->result)) {
 550                 cpuhp_reset_state(st, prev_state);
 551                 __cpuhp_kick_ap(st);
 552         }
 553
 554         return ret;
 555 }
 556
 557 static int bringup_wait_for_ap(unsigned int cpu)
 558 {
 559         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 560
 561         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 562         wait_for_ap_thread(st, true);
 563         if (WARN_ON_ONCE((!cpu_online(cpu))))
 564                 return -ECANCELED;
 565
 566         /* Unpark the hotplug thread of the target cpu */
 567         kthread_unpark(st->thread);
 568
 569         /*
 570          * SMT soft disabling on X86 requires to bring the CPU out of the
 571          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 572          * CPU marked itself as booted_once in notify_cpu_starting() so the
 573          * cpu_smt_allowed() check will now return false if this is not the
 574          * primary sibling.
 575          */
 576         if (!cpu_smt_allowed(cpu))
 577                 return -ECANCELED;
 578
 579         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 580                 return 0;
 581
 582         return cpuhp_kick_ap(st, st->target);
 583 }
 584
 585 static int bringup_cpu(unsigned int cpu)
 586 {
 587         struct task_struct *idle = idle_thread_get(cpu);
 588         int ret;
 589
 590         /*
 591          * Some architectures have to walk the irq descriptors to
 592          * setup the vector space for the cpu which comes online.
 593          * Prevent irq alloc/free across the bringup.
 594          */
 595         irq_lock_sparse();
 596
 597         /* Arch-specific enabling code. */
 598         ret = __cpu_up(cpu, idle);
 599         irq_unlock_sparse();
 600         if (ret)
 601                 return ret;
 602         return bringup_wait_for_ap(cpu);
 603 }
 604
 605 static int finish_cpu(unsigned int cpu)
 606 {
 607         struct task_struct *idle = idle_thread_get(cpu);
 608         struct mm_struct *mm = idle->active_mm;
 609
 610         /*
 611          * idle_task_exit() will have switched to &init_mm, now
 612          * clean up any remaining active_mm state.
 613          */
 614         if (mm != &init_mm)
 615                 idle->active_mm = &init_mm;
 616         mmdrop(mm);
 617         return 0;
 618 }
 619
 620 /*
 621  * Hotplug state machine related functions
 622  */
 623
 624 /*
 625  * Get the next state to run. Empty ones will be skipped. Returns true if a
 626  * state must be run.
 627  *
 628  * st->state will be modified ahead of time, to match state_to_run, as if it
 629  * has already ran.
 630  */
 631 static bool cpuhp_next_state(bool bringup,
 632                              enum cpuhp_state *state_to_run,
 633                              struct cpuhp_cpu_state *st,
 634                              enum cpuhp_state target)
 635 {
 636         do {
 637                 if (bringup) {
 638                         if (st->state >= target)
 639                                 return false;
 640
 641                         *state_to_run = ++st->state;
 642                 } else {
 643                         if (st->state <= target)
 644                                 return false;
 645
 646                         *state_to_run = st->state--;
 647                 }
 648
 649                 if (!cpuhp_step_empty(bringup, cpuhp_get_step(*state_to_run)))
 650                         break;
 651         } while (true);
 652
 653         return true;
 654 }
 655
 656 static int cpuhp_invoke_callback_range(bool bringup,
 657                                        unsigned int cpu,
 658                                        struct cpuhp_cpu_state *st,
 659                                        enum cpuhp_state target)
 660 {
 661         enum cpuhp_state state;
 662         int err = 0;
 663
 664         while (cpuhp_next_state(bringup, &state, st, target)) {
 665                 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
 666                 if (err)
 667                         break;
 668         }
 669
 670         return err;
 671 }
 672
 673 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 674 {
 675         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 676                 return true;
 677         /*
 678          * When CPU hotplug is disabled, then taking the CPU down is not
 679          * possible because takedown_cpu() and the architecture and
 680          * subsystem specific mechanisms are not available. So the CPU
 681          * which would be completely unplugged again needs to stay around
 682          * in the current state.
 683          */
 684         return st->state <= CPUHP_BRINGUP_CPU;
 685 }
 686
 687 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 688                               enum cpuhp_state target)
 689 {
 690         enum cpuhp_state prev_state = st->state;
 691         int ret = 0;
 692
 693         ret = cpuhp_invoke_callback_range(true, cpu, st, target);
 694         if (ret) {
 695                 pr_debug("CPU UP failed (%d) CPU %u state %s (%d)\n",
 696                          ret, cpu, cpuhp_get_step(st->state)->name,
 697                          st->state);
 698
 699                 cpuhp_reset_state(st, prev_state);
 700                 if (can_rollback_cpu(st))
 701                         WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
 702                                                             prev_state));
 703         }
 704         return ret;
 705 }
 706
 707 /*
 708  * The cpu hotplug threads manage the bringup and teardown of the cpus
 709  */
 710 static void cpuhp_create(unsigned int cpu)
 711 {
 712         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 713
 714         init_completion(&st->done_up);
 715         init_completion(&st->done_down);
 716         st->cpu = cpu;
 717 }
 718
 719 static int cpuhp_should_run(unsigned int cpu)
 720 {
 721         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 722
 723         return st->should_run;
 724 }
 725
 726 /*
 727  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 728  * callbacks when a state gets [un]installed at runtime.
 729  *
 730  * Each invocation of this function by the smpboot thread does a single AP
 731  * state callback.
 732  *
 733  * It has 3 modes of operation:
 734  *  - single: runs st->cb_state
 735  *  - up:     runs ++st->state, while st->state < st->target
 736  *  - down:   runs st->state--, while st->state > st->target
 737  *
 738  * When complete or on error, should_run is cleared and the completion is fired.
 739  */
 740 static void cpuhp_thread_fun(unsigned int cpu)
 741 {
 742         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 743         bool bringup = st->bringup;
 744         enum cpuhp_state state;
 745
 746         if (WARN_ON_ONCE(!st->should_run))
 747                 return;
 748
 749         /*
 750          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 751          * that if we see ->should_run we also see the rest of the state.
 752          */
 753         smp_mb();
 754
 755         /*
 756          * The BP holds the hotplug lock, but we're now running on the AP,
 757          * ensure that anybody asserting the lock is held, will actually find
 758          * it so.
 759          */
 760         lockdep_acquire_cpus_lock();
 761         cpuhp_lock_acquire(bringup);
 762
 763         if (st->single) {
 764                 state = st->cb_state;
 765                 st->should_run = false;
 766         } else {
 767                 st->should_run = cpuhp_next_state(bringup, &state, st, st->target);
 768                 if (!st->should_run)
 769                         goto end;
 770         }
 771
 772         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 773
 774         if (cpuhp_is_atomic_state(state)) {
 775                 local_irq_disable();
 776                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 777                 local_irq_enable();
 778
 779                 /*
 780                  * STARTING/DYING must not fail!
 781                  */
 782                 WARN_ON_ONCE(st->result);
 783         } else {
 784                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 785         }
 786
 787         if (st->result) {
 788                 /*
 789                  * If we fail on a rollback, we're up a creek without no
 790                  * paddle, no way forward, no way back. We loose, thanks for
 791                  * playing.
 792                  */
 793                 WARN_ON_ONCE(st->rollback);
 794                 st->should_run = false;
 795         }
 796
 797 end:
 798         cpuhp_lock_release(bringup);
 799         lockdep_release_cpus_lock();
 800
 801         if (!st->should_run)
 802                 complete_ap_thread(st, bringup);
 803 }
 804
 805 /* Invoke a single callback on a remote cpu */
 806 static int
 807 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 808                          struct hlist_node *node)
 809 {
 810         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 811         int ret;
 812
 813         if (!cpu_online(cpu))
 814                 return 0;
 815
 816         cpuhp_lock_acquire(false);
 817         cpuhp_lock_release(false);
 818
 819         cpuhp_lock_acquire(true);
 820         cpuhp_lock_release(true);
 821
 822         /*
 823          * If we are up and running, use the hotplug thread. For early calls
 824          * we invoke the thread function directly.
 825          */
 826         if (!st->thread)
 827                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 828
 829         st->rollback = false;
 830         st->last = NULL;
 831
 832         st->node = node;
 833         st->bringup = bringup;
 834         st->cb_state = state;
 835         st->single = true;
 836
 837         __cpuhp_kick_ap(st);
 838
 839         /*
 840          * If we failed and did a partial, do a rollback.
 841          */
 842         if ((ret = st->result) && st->last) {
 843                 st->rollback = true;
 844                 st->bringup = !bringup;
 845
 846                 __cpuhp_kick_ap(st);
 847         }
 848
 849         /*
 850          * Clean up the leftovers so the next hotplug operation wont use stale
 851          * data.
 852          */
 853         st->node = st->last = NULL;
 854         return ret;
 855 }
 856
 857 static int cpuhp_kick_ap_work(unsigned int cpu)
 858 {
 859         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 860         enum cpuhp_state prev_state = st->state;
 861         int ret;
 862
 863         cpuhp_lock_acquire(false);
 864         cpuhp_lock_release(false);
 865
 866         cpuhp_lock_acquire(true);
 867         cpuhp_lock_release(true);
 868
 869         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 870         ret = cpuhp_kick_ap(st, st->target);
 871         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 872
 873         return ret;
 874 }
 875
 876 static struct smp_hotplug_thread cpuhp_threads = {
 877         .store                  = &cpuhp_state.thread,
 878         .create                 = &cpuhp_create,
 879         .thread_should_run      = cpuhp_should_run,
 880         .thread_fn              = cpuhp_thread_fun,
 881         .thread_comm            = "cpuhp/%u",
 882         .selfparking            = true,
 883 };
 884
 885 void __init cpuhp_threads_init(void)
 886 {
 887         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 888         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 889 }
 890
 891 /*
 892  *
 893  * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
 894  * protected region.
 895  *
 896  * The operation is still serialized against concurrent CPU hotplug via
 897  * cpu_add_remove_lock, i.e. CPU map protection.  But it is _not_
 898  * serialized against other hotplug related activity like adding or
 899  * removing of state callbacks and state instances, which invoke either the
 900  * startup or the teardown callback of the affected state.
 901  *
 902  * This is required for subsystems which are unfixable vs. CPU hotplug and
 903  * evade lock inversion problems by scheduling work which has to be
 904  * completed _before_ cpu_up()/_cpu_down() returns.
 905  *
 906  * Don't even think about adding anything to this for any new code or even
 907  * drivers. It's only purpose is to keep existing lock order trainwrecks
 908  * working.
 909  *
 910  * For cpu_down() there might be valid reasons to finish cleanups which are
 911  * not required to be done under cpu_hotplug_lock, but that's a different
 912  * story and would be not invoked via this.
 913  */
 914 static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
 915 {
 916         /*
 917          * cpusets delegate hotplug operations to a worker to "solve" the
 918          * lock order problems. Wait for the worker, but only if tasks are
 919          * _not_ frozen (suspend, hibernate) as that would wait forever.
 920          *
 921          * The wait is required because otherwise the hotplug operation
 922          * returns with inconsistent state, which could even be observed in
 923          * user space when a new CPU is brought up. The CPU plug uevent
 924          * would be delivered and user space reacting on it would fail to
 925          * move tasks to the newly plugged CPU up to the point where the
 926          * work has finished because up to that point the newly plugged CPU
 927          * is not assignable in cpusets/cgroups. On unplug that's not
 928          * necessarily a visible issue, but it is still inconsistent state,
 929          * which is the real problem which needs to be "fixed". This can't
 930          * prevent the transient state between scheduling the work and
 931          * returning from waiting for it.
 932          */
 933         if (!tasks_frozen)
 934                 cpuset_wait_for_hotplug();
 935 }
 936
 937 #ifdef CONFIG_HOTPLUG_CPU
 938 #ifndef arch_clear_mm_cpumask_cpu
 939 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
 940 #endif
 941
 942 /**
 943  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 944  * @cpu: a CPU id
 945  *
 946  * This function walks all processes, finds a valid mm struct for each one and
 947  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 948  * trivial, there are various non-obvious corner cases, which this function
 949  * tries to solve in a safe manner.
 950  *
 951  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 952  * be called only for an already offlined CPU.
 953  */
 954 void clear_tasks_mm_cpumask(int cpu)
 955 {
 956         struct task_struct *p;
 957
 958         /*
 959          * This function is called after the cpu is taken down and marked
 960          * offline, so its not like new tasks will ever get this cpu set in
 961          * their mm mask. -- Peter Zijlstra
 962          * Thus, we may use rcu_read_lock() here, instead of grabbing
 963          * full-fledged tasklist_lock.
 964          */
 965         WARN_ON(cpu_online(cpu));
 966         rcu_read_lock();
 967         for_each_process(p) {
 968                 struct task_struct *t;
 969
 970                 /*
 971                  * Main thread might exit, but other threads may still have
 972                  * a valid mm. Find one.
 973                  */
 974                 t = find_lock_task_mm(p);
 975                 if (!t)
 976                         continue;
 977                 arch_clear_mm_cpumask_cpu(cpu, t->mm);
 978                 task_unlock(t);
 979         }
 980         rcu_read_unlock();
 981 }
 982
 983 /* Take this CPU down. */
 984 static int take_cpu_down(void *_param)
 985 {
 986         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 987         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 988         int err, cpu = smp_processor_id();
 989         int ret;
 990
 991         /* Ensure this CPU doesn't handle any more interrupts. */
 992         err = __cpu_disable();
 993         if (err < 0)
 994                 return err;
 995
 996         /*
 997          * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
 998          * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
 999          */
1000         WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
1001
1002         /* Invoke the former CPU_DYING callbacks */
1003         ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1004
1005         /*
1006          * DYING must not fail!
1007          */
1008         WARN_ON_ONCE(ret);
1009
1010         /* Give up timekeeping duties */
1011         tick_handover_do_timer();
1012         /* Remove CPU from timer broadcasting */
1013         tick_offline_cpu(cpu);
1014         /* Park the stopper thread */
1015         stop_machine_park(cpu);
1016         return 0;
1017 }
1018
1019 static int takedown_cpu(unsigned int cpu)
1020 {
1021         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1022         int err;
1023
1024         /* Park the smpboot threads */
1025         kthread_park(st->thread);
1026
1027         /*
1028          * Prevent irq alloc/free while the dying cpu reorganizes the
1029          * interrupt affinities.
1030          */
1031         irq_lock_sparse();
1032
1033         /*
1034          * So now all preempt/rcu users must observe !cpu_active().
1035          */
1036         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
1037         if (err) {
1038                 /* CPU refused to die */
1039                 irq_unlock_sparse();
1040                 /* Unpark the hotplug thread so we can rollback there */
1041                 kthread_unpark(st->thread);
1042                 return err;
1043         }
1044         BUG_ON(cpu_online(cpu));
1045
1046         /*
1047          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1048          * all runnable tasks from the CPU, there's only the idle task left now
1049          * that the migration thread is done doing the stop_machine thing.
1050          *
1051          * Wait for the stop thread to go away.
1052          */
1053         wait_for_ap_thread(st, false);
1054         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1055
1056         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1057         irq_unlock_sparse();
1058
1059         hotplug_cpu__broadcast_tick_pull(cpu);
1060         /* This actually kills the CPU. */
1061         __cpu_die(cpu);
1062
1063         tick_cleanup_dead_cpu(cpu);
1064         rcutree_migrate_callbacks(cpu);
1065         return 0;
1066 }
1067
1068 static void cpuhp_complete_idle_dead(void *arg)
1069 {
1070         struct cpuhp_cpu_state *st = arg;
1071
1072         complete_ap_thread(st, false);
1073 }
1074
1075 void cpuhp_report_idle_dead(void)
1076 {
1077         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1078
1079         BUG_ON(st->state != CPUHP_AP_OFFLINE);
1080         rcu_report_dead(smp_processor_id());
1081         st->state = CPUHP_AP_IDLE_DEAD;
1082         /*
1083          * We cannot call complete after rcu_report_dead() so we delegate it
1084          * to an online cpu.
1085          */
1086         smp_call_function_single(cpumask_first(cpu_online_mask),
1087                                  cpuhp_complete_idle_dead, st, 0);
1088 }
1089
1090 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1091                                 enum cpuhp_state target)
1092 {
1093         enum cpuhp_state prev_state = st->state;
1094         int ret = 0;
1095
1096         ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1097         if (ret) {
1098                 pr_debug("CPU DOWN failed (%d) CPU %u state %s (%d)\n",
1099                          ret, cpu, cpuhp_get_step(st->state)->name,
1100                          st->state);
1101
1102                 cpuhp_reset_state(st, prev_state);
1103
1104                 if (st->state < prev_state)
1105                         WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1106                                                             prev_state));
1107         }
1108
1109         return ret;
1110 }
1111
1112 /* Requires cpu_add_remove_lock to be held */
1113 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1114                            enum cpuhp_state target)
1115 {
1116         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1117         int prev_state, ret = 0;
1118
1119         if (num_online_cpus() == 1)
1120                 return -EBUSY;
1121
1122         if (!cpu_present(cpu))
1123                 return -EINVAL;
1124
1125         cpus_write_lock();
1126
1127         cpuhp_tasks_frozen = tasks_frozen;
1128
1129         prev_state = cpuhp_set_state(st, target);
1130         /*
1131          * If the current CPU state is in the range of the AP hotplug thread,
1132          * then we need to kick the thread.
1133          */
1134         if (st->state > CPUHP_TEARDOWN_CPU) {
1135                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1136                 ret = cpuhp_kick_ap_work(cpu);
1137                 /*
1138                  * The AP side has done the error rollback already. Just
1139                  * return the error code..
1140                  */
1141                 if (ret)
1142                         goto out;
1143
1144                 /*
1145                  * We might have stopped still in the range of the AP hotplug
1146                  * thread. Nothing to do anymore.
1147                  */
1148                 if (st->state > CPUHP_TEARDOWN_CPU)
1149                         goto out;
1150
1151                 st->target = target;
1152         }
1153         /*
1154          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1155          * to do the further cleanups.
1156          */
1157         ret = cpuhp_down_callbacks(cpu, st, target);
1158         if (ret && st->state < prev_state) {
1159                 if (st->state == CPUHP_TEARDOWN_CPU) {
1160                         cpuhp_reset_state(st, prev_state);
1161                         __cpuhp_kick_ap(st);
1162                 } else {
1163                         WARN(1, "DEAD callback error for CPU%d", cpu);
1164                 }
1165         }
1166
1167 out:
1168         cpus_write_unlock();
1169         /*
1170          * Do post unplug cleanup. This is still protected against
1171          * concurrent CPU hotplug via cpu_add_remove_lock.
1172          */
1173         lockup_detector_cleanup();
1174         arch_smt_update();
1175         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1176         return ret;
1177 }
1178
1179 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1180 {
1181         if (cpu_hotplug_disabled)
1182                 return -EBUSY;
1183         return _cpu_down(cpu, 0, target);
1184 }
1185
1186 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1187 {
1188         int err;
1189
1190         cpu_maps_update_begin();
1191         err = cpu_down_maps_locked(cpu, target);
1192         cpu_maps_update_done();
1193         return err;
1194 }
1195
1196 /**
1197  * cpu_device_down - Bring down a cpu device
1198  * @dev: Pointer to the cpu device to offline
1199  *
1200  * This function is meant to be used by device core cpu subsystem only.
1201  *
1202  * Other subsystems should use remove_cpu() instead.
1203  *
1204  * Return: %0 on success or a negative errno code
1205  */
1206 int cpu_device_down(struct device *dev)
1207 {
1208         return cpu_down(dev->id, CPUHP_OFFLINE);
1209 }
1210
1211 int remove_cpu(unsigned int cpu)
1212 {
1213         int ret;
1214
1215         lock_device_hotplug();
1216         ret = device_offline(get_cpu_device(cpu));
1217         unlock_device_hotplug();
1218
1219         return ret;
1220 }
1221 EXPORT_SYMBOL_GPL(remove_cpu);
1222
1223 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1224 {
1225         unsigned int cpu;
1226         int error;
1227
1228         cpu_maps_update_begin();
1229
1230         /*
1231          * Make certain the cpu I'm about to reboot on is online.
1232          *
1233          * This is inline to what migrate_to_reboot_cpu() already do.
1234          */
1235         if (!cpu_online(primary_cpu))
1236                 primary_cpu = cpumask_first(cpu_online_mask);
1237
1238         for_each_online_cpu(cpu) {
1239                 if (cpu == primary_cpu)
1240                         continue;
1241
1242                 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1243                 if (error) {
1244                         pr_err("Failed to offline CPU%d - error=%d",
1245                                 cpu, error);
1246                         break;
1247                 }
1248         }
1249
1250         /*
1251          * Ensure all but the reboot CPU are offline.
1252          */
1253         BUG_ON(num_online_cpus() > 1);
1254
1255         /*
1256          * Make sure the CPUs won't be enabled by someone else after this
1257          * point. Kexec will reboot to a new kernel shortly resetting
1258          * everything along the way.
1259          */
1260         cpu_hotplug_disabled++;
1261
1262         cpu_maps_update_done();
1263 }
1264
1265 #else
1266 #define takedown_cpu            NULL
1267 #endif /*CONFIG_HOTPLUG_CPU*/
1268
1269 /**
1270  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1271  * @cpu: cpu that just started
1272  *
1273  * It must be called by the arch code on the new cpu, before the new cpu
1274  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1275  */
1276 void notify_cpu_starting(unsigned int cpu)
1277 {
1278         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1279         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1280         int ret;
1281
1282         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1283         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1284         ret = cpuhp_invoke_callback_range(true, cpu, st, target);
1285
1286         /*
1287          * STARTING must not fail!
1288          */
1289         WARN_ON_ONCE(ret);
1290 }
1291
1292 /*
1293  * Called from the idle task. Wake up the controlling task which brings the
1294  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1295  * online bringup to the hotplug thread.
1296  */
1297 void cpuhp_online_idle(enum cpuhp_state state)
1298 {
1299         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1300
1301         /* Happens for the boot cpu */
1302         if (state != CPUHP_AP_ONLINE_IDLE)
1303                 return;
1304
1305         /*
1306          * Unpart the stopper thread before we start the idle loop (and start
1307          * scheduling); this ensures the stopper task is always available.
1308          */
1309         stop_machine_unpark(smp_processor_id());
1310
1311         st->state = CPUHP_AP_ONLINE_IDLE;
1312         complete_ap_thread(st, true);
1313 }
1314
1315 /* Requires cpu_add_remove_lock to be held */
1316 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1317 {
1318         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1319         struct task_struct *idle;
1320         int ret = 0;
1321
1322         cpus_write_lock();
1323
1324         if (!cpu_present(cpu)) {
1325                 ret = -EINVAL;
1326                 goto out;
1327         }
1328
1329         /*
1330          * The caller of cpu_up() might have raced with another
1331          * caller. Nothing to do.
1332          */
1333         if (st->state >= target)
1334                 goto out;
1335
1336         if (st->state == CPUHP_OFFLINE) {
1337                 /* Let it fail before we try to bring the cpu up */
1338                 idle = idle_thread_get(cpu);
1339                 if (IS_ERR(idle)) {
1340                         ret = PTR_ERR(idle);
1341                         goto out;
1342                 }
1343         }
1344
1345         cpuhp_tasks_frozen = tasks_frozen;
1346
1347         cpuhp_set_state(st, target);
1348         /*
1349          * If the current CPU state is in the range of the AP hotplug thread,
1350          * then we need to kick the thread once more.
1351          */
1352         if (st->state > CPUHP_BRINGUP_CPU) {
1353                 ret = cpuhp_kick_ap_work(cpu);
1354                 /*
1355                  * The AP side has done the error rollback already. Just
1356                  * return the error code..
1357                  */
1358                 if (ret)
1359                         goto out;
1360         }
1361
1362         /*
1363          * Try to reach the target state. We max out on the BP at
1364          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1365          * responsible for bringing it up to the target state.
1366          */
1367         target = min((int)target, CPUHP_BRINGUP_CPU);
1368         ret = cpuhp_up_callbacks(cpu, st, target);
1369 out:
1370         cpus_write_unlock();
1371         arch_smt_update();
1372         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1373         return ret;
1374 }
1375
1376 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1377 {
1378         int err = 0;
1379
1380         if (!cpu_possible(cpu)) {
1381                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1382                        cpu);
1383 #if defined(CONFIG_IA64)
1384                 pr_err("please check additional_cpus= boot parameter\n");
1385 #endif
1386                 return -EINVAL;
1387         }
1388
1389         err = try_online_node(cpu_to_node(cpu));
1390         if (err)
1391                 return err;
1392
1393         cpu_maps_update_begin();
1394
1395         if (cpu_hotplug_disabled) {
1396                 err = -EBUSY;
1397                 goto out;
1398         }
1399         if (!cpu_smt_allowed(cpu)) {
1400                 err = -EPERM;
1401                 goto out;
1402         }
1403
1404         err = _cpu_up(cpu, 0, target);
1405 out:
1406         cpu_maps_update_done();
1407         return err;
1408 }
1409
1410 /**
1411  * cpu_device_up - Bring up a cpu device
1412  * @dev: Pointer to the cpu device to online
1413  *
1414  * This function is meant to be used by device core cpu subsystem only.
1415  *
1416  * Other subsystems should use add_cpu() instead.
1417  *
1418  * Return: %0 on success or a negative errno code
1419  */
1420 int cpu_device_up(struct device *dev)
1421 {
1422         return cpu_up(dev->id, CPUHP_ONLINE);
1423 }
1424
1425 int add_cpu(unsigned int cpu)
1426 {
1427         int ret;
1428
1429         lock_device_hotplug();
1430         ret = device_online(get_cpu_device(cpu));
1431         unlock_device_hotplug();
1432
1433         return ret;
1434 }
1435 EXPORT_SYMBOL_GPL(add_cpu);
1436
1437 /**
1438  * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1439  * @sleep_cpu: The cpu we hibernated on and should be brought up.
1440  *
1441  * On some architectures like arm64, we can hibernate on any CPU, but on
1442  * wake up the CPU we hibernated on might be offline as a side effect of
1443  * using maxcpus= for example.
1444  *
1445  * Return: %0 on success or a negative errno code
1446  */
1447 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1448 {
1449         int ret;
1450
1451         if (!cpu_online(sleep_cpu)) {
1452                 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1453                 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1454                 if (ret) {
1455                         pr_err("Failed to bring hibernate-CPU up!\n");
1456                         return ret;
1457                 }
1458         }
1459         return 0;
1460 }
1461
1462 void bringup_nonboot_cpus(unsigned int setup_max_cpus)
1463 {
1464         unsigned int cpu;
1465
1466         for_each_present_cpu(cpu) {
1467                 if (num_online_cpus() >= setup_max_cpus)
1468                         break;
1469                 if (!cpu_online(cpu))
1470                         cpu_up(cpu, CPUHP_ONLINE);
1471         }
1472 }
1473
1474 #ifdef CONFIG_PM_SLEEP_SMP
1475 static cpumask_var_t frozen_cpus;
1476
1477 int freeze_secondary_cpus(int primary)
1478 {
1479         int cpu, error = 0;
1480
1481         cpu_maps_update_begin();
1482         if (primary == -1) {
1483                 primary = cpumask_first(cpu_online_mask);
1484                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1485                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1486         } else {
1487                 if (!cpu_online(primary))
1488                         primary = cpumask_first(cpu_online_mask);
1489         }
1490
1491         /*
1492          * We take down all of the non-boot CPUs in one shot to avoid races
1493          * with the userspace trying to use the CPU hotplug at the same time
1494          */
1495         cpumask_clear(frozen_cpus);
1496
1497         pr_info("Disabling non-boot CPUs ...\n");
1498         for_each_online_cpu(cpu) {
1499                 if (cpu == primary)
1500                         continue;
1501
1502                 if (pm_wakeup_pending()) {
1503                         pr_info("Wakeup pending. Abort CPU freeze\n");
1504                         error = -EBUSY;
1505                         break;
1506                 }
1507
1508                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1509                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1510                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1511                 if (!error)
1512                         cpumask_set_cpu(cpu, frozen_cpus);
1513                 else {
1514                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1515                         break;
1516                 }
1517         }
1518
1519         if (!error)
1520                 BUG_ON(num_online_cpus() > 1);
1521         else
1522                 pr_err("Non-boot CPUs are not disabled\n");
1523
1524         /*
1525          * Make sure the CPUs won't be enabled by someone else. We need to do
1526          * this even in case of failure as all freeze_secondary_cpus() users are
1527          * supposed to do thaw_secondary_cpus() on the failure path.
1528          */
1529         cpu_hotplug_disabled++;
1530
1531         cpu_maps_update_done();
1532         return error;
1533 }
1534
1535 void __weak arch_thaw_secondary_cpus_begin(void)
1536 {
1537 }
1538
1539 void __weak arch_thaw_secondary_cpus_end(void)
1540 {
1541 }
1542
1543 void thaw_secondary_cpus(void)
1544 {
1545         int cpu, error;
1546
1547         /* Allow everyone to use the CPU hotplug again */
1548         cpu_maps_update_begin();
1549         __cpu_hotplug_enable();
1550         if (cpumask_empty(frozen_cpus))
1551                 goto out;
1552
1553         pr_info("Enabling non-boot CPUs ...\n");
1554
1555         arch_thaw_secondary_cpus_begin();
1556
1557         for_each_cpu(cpu, frozen_cpus) {
1558                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1559                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1560                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1561                 if (!error) {
1562                         pr_info("CPU%d is up\n", cpu);
1563                         continue;
1564                 }
1565                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1566         }
1567
1568         arch_thaw_secondary_cpus_end();
1569
1570         cpumask_clear(frozen_cpus);
1571 out:
1572         cpu_maps_update_done();
1573 }
1574
1575 static int __init alloc_frozen_cpus(void)
1576 {
1577         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1578                 return -ENOMEM;
1579         return 0;
1580 }
1581 core_initcall(alloc_frozen_cpus);
1582
1583 /*
1584  * When callbacks for CPU hotplug notifications are being executed, we must
1585  * ensure that the state of the system with respect to the tasks being frozen
1586  * or not, as reported by the notification, remains unchanged *throughout the
1587  * duration* of the execution of the callbacks.
1588  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1589  *
1590  * This synchronization is implemented by mutually excluding regular CPU
1591  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1592  * Hibernate notifications.
1593  */
1594 static int
1595 cpu_hotplug_pm_callback(struct notifier_block *nb,
1596                         unsigned long action, void *ptr)
1597 {
1598         switch (action) {
1599
1600         case PM_SUSPEND_PREPARE:
1601         case PM_HIBERNATION_PREPARE:
1602                 cpu_hotplug_disable();
1603                 break;
1604
1605         case PM_POST_SUSPEND:
1606         case PM_POST_HIBERNATION:
1607                 cpu_hotplug_enable();
1608                 break;
1609
1610         default:
1611                 return NOTIFY_DONE;
1612         }
1613
1614         return NOTIFY_OK;
1615 }
1616
1617
1618 static int __init cpu_hotplug_pm_sync_init(void)
1619 {
1620         /*
1621          * cpu_hotplug_pm_callback has higher priority than x86
1622          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1623          * to disable cpu hotplug to avoid cpu hotplug race.
1624          */
1625         pm_notifier(cpu_hotplug_pm_callback, 0);
1626         return 0;
1627 }
1628 core_initcall(cpu_hotplug_pm_sync_init);
1629
1630 #endif /* CONFIG_PM_SLEEP_SMP */
1631
1632 int __boot_cpu_id;
1633
1634 #endif /* CONFIG_SMP */
1635
1636 /* Boot processor state steps */
1637 static struct cpuhp_step cpuhp_hp_states[] = {
1638         [CPUHP_OFFLINE] = {
1639                 .name                   = "offline",
1640                 .startup.single         = NULL,
1641                 .teardown.single        = NULL,
1642         },
1643 #ifdef CONFIG_SMP
1644         [CPUHP_CREATE_THREADS]= {
1645                 .name                   = "threads:prepare",
1646                 .startup.single         = smpboot_create_threads,
1647                 .teardown.single        = NULL,
1648                 .cant_stop              = true,
1649         },
1650         [CPUHP_PERF_PREPARE] = {
1651                 .name                   = "perf:prepare",
1652                 .startup.single         = perf_event_init_cpu,
1653                 .teardown.single        = perf_event_exit_cpu,
1654         },
1655         [CPUHP_WORKQUEUE_PREP] = {
1656                 .name                   = "workqueue:prepare",
1657                 .startup.single         = workqueue_prepare_cpu,
1658                 .teardown.single        = NULL,
1659         },
1660         [CPUHP_HRTIMERS_PREPARE] = {
1661                 .name                   = "hrtimers:prepare",
1662                 .startup.single         = hrtimers_prepare_cpu,
1663                 .teardown.single        = hrtimers_dead_cpu,
1664         },
1665         [CPUHP_SMPCFD_PREPARE] = {
1666                 .name                   = "smpcfd:prepare",
1667                 .startup.single         = smpcfd_prepare_cpu,
1668                 .teardown.single        = smpcfd_dead_cpu,
1669         },
1670         [CPUHP_RELAY_PREPARE] = {
1671                 .name                   = "relay:prepare",
1672                 .startup.single         = relay_prepare_cpu,
1673                 .teardown.single        = NULL,
1674         },
1675         [CPUHP_SLAB_PREPARE] = {
1676                 .name                   = "slab:prepare",
1677                 .startup.single         = slab_prepare_cpu,
1678                 .teardown.single        = slab_dead_cpu,
1679         },
1680         [CPUHP_RCUTREE_PREP] = {
1681                 .name                   = "RCU/tree:prepare",
1682                 .startup.single         = rcutree_prepare_cpu,
1683                 .teardown.single        = rcutree_dead_cpu,
1684         },
1685         /*
1686          * On the tear-down path, timers_dead_cpu() must be invoked
1687          * before blk_mq_queue_reinit_notify() from notify_dead(),
1688          * otherwise a RCU stall occurs.
1689          */
1690         [CPUHP_TIMERS_PREPARE] = {
1691                 .name                   = "timers:prepare",
1692                 .startup.single         = timers_prepare_cpu,
1693                 .teardown.single        = timers_dead_cpu,
1694         },
1695         /* Kicks the plugged cpu into life */
1696         [CPUHP_BRINGUP_CPU] = {
1697                 .name                   = "cpu:bringup",
1698                 .startup.single         = bringup_cpu,
1699                 .teardown.single        = finish_cpu,
1700                 .cant_stop              = true,
1701         },
1702         /* Final state before CPU kills itself */
1703         [CPUHP_AP_IDLE_DEAD] = {
1704                 .name                   = "idle:dead",
1705         },
1706         /*
1707          * Last state before CPU enters the idle loop to die. Transient state
1708          * for synchronization.
1709          */
1710         [CPUHP_AP_OFFLINE] = {
1711                 .name                   = "ap:offline",
1712                 .cant_stop              = true,
1713         },
1714         /* First state is scheduler control. Interrupts are disabled */
1715         [CPUHP_AP_SCHED_STARTING] = {
1716                 .name                   = "sched:starting",
1717                 .startup.single         = sched_cpu_starting,
1718                 .teardown.single        = sched_cpu_dying,
1719         },
1720         [CPUHP_AP_RCUTREE_DYING] = {
1721                 .name                   = "RCU/tree:dying",
1722                 .startup.single         = NULL,
1723                 .teardown.single        = rcutree_dying_cpu,
1724         },
1725         [CPUHP_AP_SMPCFD_DYING] = {
1726                 .name                   = "smpcfd:dying",
1727                 .startup.single         = NULL,
1728                 .teardown.single        = smpcfd_dying_cpu,
1729         },
1730         /* Entry state on starting. Interrupts enabled from here on. Transient
1731          * state for synchronsization */
1732         [CPUHP_AP_ONLINE] = {
1733                 .name                   = "ap:online",
1734         },
1735         /*
1736          * Handled on control processor until the plugged processor manages
1737          * this itself.
1738          */
1739         [CPUHP_TEARDOWN_CPU] = {
1740                 .name                   = "cpu:teardown",
1741                 .startup.single         = NULL,
1742                 .teardown.single        = takedown_cpu,
1743                 .cant_stop              = true,
1744         },
1745
1746         [CPUHP_AP_SCHED_WAIT_EMPTY] = {
1747                 .name                   = "sched:waitempty",
1748                 .startup.single         = NULL,
1749                 .teardown.single        = sched_cpu_wait_empty,
1750         },
1751
1752         /* Handle smpboot threads park/unpark */
1753         [CPUHP_AP_SMPBOOT_THREADS] = {
1754                 .name                   = "smpboot/threads:online",
1755                 .startup.single         = smpboot_unpark_threads,
1756                 .teardown.single        = smpboot_park_threads,
1757         },
1758         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1759                 .name                   = "irq/affinity:online",
1760                 .startup.single         = irq_affinity_online_cpu,
1761                 .teardown.single        = NULL,
1762         },
1763         [CPUHP_AP_PERF_ONLINE] = {
1764                 .name                   = "perf:online",
1765                 .startup.single         = perf_event_init_cpu,
1766                 .teardown.single        = perf_event_exit_cpu,
1767         },
1768         [CPUHP_AP_WATCHDOG_ONLINE] = {
1769                 .name                   = "lockup_detector:online",
1770                 .startup.single         = lockup_detector_online_cpu,
1771                 .teardown.single        = lockup_detector_offline_cpu,
1772         },
1773         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1774                 .name                   = "workqueue:online",
1775                 .startup.single         = workqueue_online_cpu,
1776                 .teardown.single        = workqueue_offline_cpu,
1777         },
1778         [CPUHP_AP_RCUTREE_ONLINE] = {
1779                 .name                   = "RCU/tree:online",
1780                 .startup.single         = rcutree_online_cpu,
1781                 .teardown.single        = rcutree_offline_cpu,
1782         },
1783 #endif
1784         /*
1785          * The dynamically registered state space is here
1786          */
1787
1788 #ifdef CONFIG_SMP
1789         /* Last state is scheduler control setting the cpu active */
1790         [CPUHP_AP_ACTIVE] = {
1791                 .name                   = "sched:active",
1792                 .startup.single         = sched_cpu_activate,
1793                 .teardown.single        = sched_cpu_deactivate,
1794         },
1795 #endif
1796
1797         /* CPU is fully up and running. */
1798         [CPUHP_ONLINE] = {
1799                 .name                   = "online",
1800                 .startup.single         = NULL,
1801                 .teardown.single        = NULL,
1802         },
1803 };
1804
1805 /* Sanity check for callbacks */
1806 static int cpuhp_cb_check(enum cpuhp_state state)
1807 {
1808         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1809                 return -EINVAL;
1810         return 0;
1811 }
1812
1813 /*
1814  * Returns a free for dynamic slot assignment of the Online state. The states
1815  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1816  * by having no name assigned.
1817  */
1818 static int cpuhp_reserve_state(enum cpuhp_state state)
1819 {
1820         enum cpuhp_state i, end;
1821         struct cpuhp_step *step;
1822
1823         switch (state) {
1824         case CPUHP_AP_ONLINE_DYN:
1825                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1826                 end = CPUHP_AP_ONLINE_DYN_END;
1827                 break;
1828         case CPUHP_BP_PREPARE_DYN:
1829                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1830                 end = CPUHP_BP_PREPARE_DYN_END;
1831                 break;
1832         default:
1833                 return -EINVAL;
1834         }
1835
1836         for (i = state; i <= end; i++, step++) {
1837                 if (!step->name)
1838                         return i;
1839         }
1840         WARN(1, "No more dynamic states available for CPU hotplug\n");
1841         return -ENOSPC;
1842 }
1843
1844 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1845                                  int (*startup)(unsigned int cpu),
1846                                  int (*teardown)(unsigned int cpu),
1847                                  bool multi_instance)
1848 {
1849         /* (Un)Install the callbacks for further cpu hotplug operations */
1850         struct cpuhp_step *sp;
1851         int ret = 0;
1852
1853         /*
1854          * If name is NULL, then the state gets removed.
1855          *
1856          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1857          * the first allocation from these dynamic ranges, so the removal
1858          * would trigger a new allocation and clear the wrong (already
1859          * empty) state, leaving the callbacks of the to be cleared state
1860          * dangling, which causes wreckage on the next hotplug operation.
1861          */
1862         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1863                      state == CPUHP_BP_PREPARE_DYN)) {
1864                 ret = cpuhp_reserve_state(state);
1865                 if (ret < 0)
1866                         return ret;
1867                 state = ret;
1868         }
1869         sp = cpuhp_get_step(state);
1870         if (name && sp->name)
1871                 return -EBUSY;
1872
1873         sp->startup.single = startup;
1874         sp->teardown.single = teardown;
1875         sp->name = name;
1876         sp->multi_instance = multi_instance;
1877         INIT_HLIST_HEAD(&sp->list);
1878         return ret;
1879 }
1880
1881 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1882 {
1883         return cpuhp_get_step(state)->teardown.single;
1884 }
1885
1886 /*
1887  * Call the startup/teardown function for a step either on the AP or
1888  * on the current CPU.
1889  */
1890 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1891                             struct hlist_node *node)
1892 {
1893         struct cpuhp_step *sp = cpuhp_get_step(state);
1894         int ret;
1895
1896         /*
1897          * If there's nothing to do, we done.
1898          * Relies on the union for multi_instance.
1899          */
1900         if (cpuhp_step_empty(bringup, sp))
1901                 return 0;
1902         /*
1903          * The non AP bound callbacks can fail on bringup. On teardown
1904          * e.g. module removal we crash for now.
1905          */
1906 #ifdef CONFIG_SMP
1907         if (cpuhp_is_ap_state(state))
1908                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1909         else
1910                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1911 #else
1912         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1913 #endif
1914         BUG_ON(ret && !bringup);
1915         return ret;
1916 }
1917
1918 /*
1919  * Called from __cpuhp_setup_state on a recoverable failure.
1920  *
1921  * Note: The teardown callbacks for rollback are not allowed to fail!
1922  */
1923 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1924                                    struct hlist_node *node)
1925 {
1926         int cpu;
1927
1928         /* Roll back the already executed steps on the other cpus */
1929         for_each_present_cpu(cpu) {
1930                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1931                 int cpustate = st->state;
1932
1933                 if (cpu >= failedcpu)
1934                         break;
1935
1936                 /* Did we invoke the startup call on that cpu ? */
1937                 if (cpustate >= state)
1938                         cpuhp_issue_call(cpu, state, false, node);
1939         }
1940 }
1941
1942 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1943                                           struct hlist_node *node,
1944                                           bool invoke)
1945 {
1946         struct cpuhp_step *sp;
1947         int cpu;
1948         int ret;
1949
1950         lockdep_assert_cpus_held();
1951
1952         sp = cpuhp_get_step(state);
1953         if (sp->multi_instance == false)
1954                 return -EINVAL;
1955
1956         mutex_lock(&cpuhp_state_mutex);
1957
1958         if (!invoke || !sp->startup.multi)
1959                 goto add_node;
1960
1961         /*
1962          * Try to call the startup callback for each present cpu
1963          * depending on the hotplug state of the cpu.
1964          */
1965         for_each_present_cpu(cpu) {
1966                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1967                 int cpustate = st->state;
1968
1969                 if (cpustate < state)
1970                         continue;
1971
1972                 ret = cpuhp_issue_call(cpu, state, true, node);
1973                 if (ret) {
1974                         if (sp->teardown.multi)
1975                                 cpuhp_rollback_install(cpu, state, node);
1976                         goto unlock;
1977                 }
1978         }
1979 add_node:
1980         ret = 0;
1981         hlist_add_head(node, &sp->list);
1982 unlock:
1983         mutex_unlock(&cpuhp_state_mutex);
1984         return ret;
1985 }
1986
1987 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1988                                bool invoke)
1989 {
1990         int ret;
1991
1992         cpus_read_lock();
1993         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1994         cpus_read_unlock();
1995         return ret;
1996 }
1997 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1998
1999 /**
2000  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
2001  * @state:              The state to setup
2002  * @name:               Name of the step
2003  * @invoke:             If true, the startup function is invoked for cpus where
2004  *                      cpu state >= @state
2005  * @startup:            startup callback function
2006  * @teardown:           teardown callback function
2007  * @multi_instance:     State is set up for multiple instances which get
2008  *                      added afterwards.
2009  *
2010  * The caller needs to hold cpus read locked while calling this function.
2011  * Return:
2012  *   On success:
2013  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN;
2014  *      0 for all other states
2015  *   On failure: proper (negative) error code
2016  */
2017 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
2018                                    const char *name, bool invoke,
2019                                    int (*startup)(unsigned int cpu),
2020                                    int (*teardown)(unsigned int cpu),
2021                                    bool multi_instance)
2022 {
2023         int cpu, ret = 0;
2024         bool dynstate;
2025
2026         lockdep_assert_cpus_held();
2027
2028         if (cpuhp_cb_check(state) || !name)
2029                 return -EINVAL;
2030
2031         mutex_lock(&cpuhp_state_mutex);
2032
2033         ret = cpuhp_store_callbacks(state, name, startup, teardown,
2034                                     multi_instance);
2035
2036         dynstate = state == CPUHP_AP_ONLINE_DYN;
2037         if (ret > 0 && dynstate) {
2038                 state = ret;
2039                 ret = 0;
2040         }
2041
2042         if (ret || !invoke || !startup)
2043                 goto out;
2044
2045         /*
2046          * Try to call the startup callback for each present cpu
2047          * depending on the hotplug state of the cpu.
2048          */
2049         for_each_present_cpu(cpu) {
2050                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2051                 int cpustate = st->state;
2052
2053                 if (cpustate < state)
2054                         continue;
2055
2056                 ret = cpuhp_issue_call(cpu, state, true, NULL);
2057                 if (ret) {
2058                         if (teardown)
2059                                 cpuhp_rollback_install(cpu, state, NULL);
2060                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2061                         goto out;
2062                 }
2063         }
2064 out:
2065         mutex_unlock(&cpuhp_state_mutex);
2066         /*
2067          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
2068          * dynamically allocated state in case of success.
2069          */
2070         if (!ret && dynstate)
2071                 return state;
2072         return ret;
2073 }
2074 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2075
2076 int __cpuhp_setup_state(enum cpuhp_state state,
2077                         const char *name, bool invoke,
2078                         int (*startup)(unsigned int cpu),
2079                         int (*teardown)(unsigned int cpu),
2080                         bool multi_instance)
2081 {
2082         int ret;
2083
2084         cpus_read_lock();
2085         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2086                                              teardown, multi_instance);
2087         cpus_read_unlock();
2088         return ret;
2089 }
2090 EXPORT_SYMBOL(__cpuhp_setup_state);
2091
2092 int __cpuhp_state_remove_instance(enum cpuhp_state state,
2093                                   struct hlist_node *node, bool invoke)
2094 {
2095         struct cpuhp_step *sp = cpuhp_get_step(state);
2096         int cpu;
2097
2098         BUG_ON(cpuhp_cb_check(state));
2099
2100         if (!sp->multi_instance)
2101                 return -EINVAL;
2102
2103         cpus_read_lock();
2104         mutex_lock(&cpuhp_state_mutex);
2105
2106         if (!invoke || !cpuhp_get_teardown_cb(state))
2107                 goto remove;
2108         /*
2109          * Call the teardown callback for each present cpu depending
2110          * on the hotplug state of the cpu. This function is not
2111          * allowed to fail currently!
2112          */
2113         for_each_present_cpu(cpu) {
2114                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2115                 int cpustate = st->state;
2116
2117                 if (cpustate >= state)
2118                         cpuhp_issue_call(cpu, state, false, node);
2119         }
2120
2121 remove:
2122         hlist_del(node);
2123         mutex_unlock(&cpuhp_state_mutex);
2124         cpus_read_unlock();
2125
2126         return 0;
2127 }
2128 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2129
2130 /**
2131  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2132  * @state:      The state to remove
2133  * @invoke:     If true, the teardown function is invoked for cpus where
2134  *              cpu state >= @state
2135  *
2136  * The caller needs to hold cpus read locked while calling this function.
2137  * The teardown callback is currently not allowed to fail. Think
2138  * about module removal!
2139  */
2140 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2141 {
2142         struct cpuhp_step *sp = cpuhp_get_step(state);
2143         int cpu;
2144
2145         BUG_ON(cpuhp_cb_check(state));
2146
2147         lockdep_assert_cpus_held();
2148
2149         mutex_lock(&cpuhp_state_mutex);
2150         if (sp->multi_instance) {
2151                 WARN(!hlist_empty(&sp->list),
2152                      "Error: Removing state %d which has instances left.\n",
2153                      state);
2154                 goto remove;
2155         }
2156
2157         if (!invoke || !cpuhp_get_teardown_cb(state))
2158                 goto remove;
2159
2160         /*
2161          * Call the teardown callback for each present cpu depending
2162          * on the hotplug state of the cpu. This function is not
2163          * allowed to fail currently!
2164          */
2165         for_each_present_cpu(cpu) {
2166                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2167                 int cpustate = st->state;
2168
2169                 if (cpustate >= state)
2170                         cpuhp_issue_call(cpu, state, false, NULL);
2171         }
2172 remove:
2173         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2174         mutex_unlock(&cpuhp_state_mutex);
2175 }
2176 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2177
2178 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2179 {
2180         cpus_read_lock();
2181         __cpuhp_remove_state_cpuslocked(state, invoke);
2182         cpus_read_unlock();
2183 }
2184 EXPORT_SYMBOL(__cpuhp_remove_state);
2185
2186 #ifdef CONFIG_HOTPLUG_SMT
2187 static void cpuhp_offline_cpu_device(unsigned int cpu)
2188 {
2189         struct device *dev = get_cpu_device(cpu);
2190
2191         dev->offline = true;
2192         /* Tell user space about the state change */
2193         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2194 }
2195
2196 static void cpuhp_online_cpu_device(unsigned int cpu)
2197 {
2198         struct device *dev = get_cpu_device(cpu);
2199
2200         dev->offline = false;
2201         /* Tell user space about the state change */
2202         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2203 }
2204
2205 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2206 {
2207         int cpu, ret = 0;
2208
2209         cpu_maps_update_begin();
2210         for_each_online_cpu(cpu) {
2211                 if (topology_is_primary_thread(cpu))
2212                         continue;
2213                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2214                 if (ret)
2215                         break;
2216                 /*
2217                  * As this needs to hold the cpu maps lock it's impossible
2218                  * to call device_offline() because that ends up calling
2219                  * cpu_down() which takes cpu maps lock. cpu maps lock
2220                  * needs to be held as this might race against in kernel
2221                  * abusers of the hotplug machinery (thermal management).
2222                  *
2223                  * So nothing would update device:offline state. That would
2224                  * leave the sysfs entry stale and prevent onlining after
2225                  * smt control has been changed to 'off' again. This is
2226                  * called under the sysfs hotplug lock, so it is properly
2227                  * serialized against the regular offline usage.
2228                  */
2229                 cpuhp_offline_cpu_device(cpu);
2230         }
2231         if (!ret)
2232                 cpu_smt_control = ctrlval;
2233         cpu_maps_update_done();
2234         return ret;
2235 }
2236
2237 int cpuhp_smt_enable(void)
2238 {
2239         int cpu, ret = 0;
2240
2241         cpu_maps_update_begin();
2242         cpu_smt_control = CPU_SMT_ENABLED;
2243         for_each_present_cpu(cpu) {
2244                 /* Skip online CPUs and CPUs on offline nodes */
2245                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2246                         continue;
2247                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2248                 if (ret)
2249                         break;
2250                 /* See comment in cpuhp_smt_disable() */
2251                 cpuhp_online_cpu_device(cpu);
2252         }
2253         cpu_maps_update_done();
2254         return ret;
2255 }
2256 #endif
2257
2258 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2259 static ssize_t state_show(struct device *dev,
2260                           struct device_attribute *attr, char *buf)
2261 {
2262         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2263
2264         return sprintf(buf, "%d\n", st->state);
2265 }
2266 static DEVICE_ATTR_RO(state);
2267
2268 static ssize_t target_store(struct device *dev, struct device_attribute *attr,
2269                             const char *buf, size_t count)
2270 {
2271         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2272         struct cpuhp_step *sp;
2273         int target, ret;
2274
2275         ret = kstrtoint(buf, 10, &target);
2276         if (ret)
2277                 return ret;
2278
2279 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2280         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2281                 return -EINVAL;
2282 #else
2283         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2284                 return -EINVAL;
2285 #endif
2286
2287         ret = lock_device_hotplug_sysfs();
2288         if (ret)
2289                 return ret;
2290
2291         mutex_lock(&cpuhp_state_mutex);
2292         sp = cpuhp_get_step(target);
2293         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2294         mutex_unlock(&cpuhp_state_mutex);
2295         if (ret)
2296                 goto out;
2297
2298         if (st->state < target)
2299                 ret = cpu_up(dev->id, target);
2300         else
2301                 ret = cpu_down(dev->id, target);
2302 out:
2303         unlock_device_hotplug();
2304         return ret ? ret : count;
2305 }
2306
2307 static ssize_t target_show(struct device *dev,
2308                            struct device_attribute *attr, char *buf)
2309 {
2310         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2311
2312         return sprintf(buf, "%d\n", st->target);
2313 }
2314 static DEVICE_ATTR_RW(target);
2315
2316 static ssize_t fail_store(struct device *dev, struct device_attribute *attr,
2317                           const char *buf, size_t count)
2318 {
2319         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2320         struct cpuhp_step *sp;
2321         int fail, ret;
2322
2323         ret = kstrtoint(buf, 10, &fail);
2324         if (ret)
2325                 return ret;
2326
2327         if (fail == CPUHP_INVALID) {
2328                 st->fail = fail;
2329                 return count;
2330         }
2331
2332         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2333                 return -EINVAL;
2334
2335         /*
2336          * Cannot fail STARTING/DYING callbacks.
2337          */
2338         if (cpuhp_is_atomic_state(fail))
2339                 return -EINVAL;
2340
2341         /*
2342          * DEAD callbacks cannot fail...
2343          * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2344          * triggering STARTING callbacks, a failure in this state would
2345          * hinder rollback.
2346          */
2347         if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2348                 return -EINVAL;
2349
2350         /*
2351          * Cannot fail anything that doesn't have callbacks.
2352          */
2353         mutex_lock(&cpuhp_state_mutex);
2354         sp = cpuhp_get_step(fail);
2355         if (!sp->startup.single && !sp->teardown.single)
2356                 ret = -EINVAL;
2357         mutex_unlock(&cpuhp_state_mutex);
2358         if (ret)
2359                 return ret;
2360
2361         st->fail = fail;
2362
2363         return count;
2364 }
2365
2366 static ssize_t fail_show(struct device *dev,
2367                          struct device_attribute *attr, char *buf)
2368 {
2369         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2370
2371         return sprintf(buf, "%d\n", st->fail);
2372 }
2373
2374 static DEVICE_ATTR_RW(fail);
2375
2376 static struct attribute *cpuhp_cpu_attrs[] = {
2377         &dev_attr_state.attr,
2378         &dev_attr_target.attr,
2379         &dev_attr_fail.attr,
2380         NULL
2381 };
2382
2383 static const struct attribute_group cpuhp_cpu_attr_group = {
2384         .attrs = cpuhp_cpu_attrs,
2385         .name = "hotplug",
2386         NULL
2387 };
2388
2389 static ssize_t states_show(struct device *dev,
2390                                  struct device_attribute *attr, char *buf)
2391 {
2392         ssize_t cur, res = 0;
2393         int i;
2394
2395         mutex_lock(&cpuhp_state_mutex);
2396         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2397                 struct cpuhp_step *sp = cpuhp_get_step(i);
2398
2399                 if (sp->name) {
2400                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2401                         buf += cur;
2402                         res += cur;
2403                 }
2404         }
2405         mutex_unlock(&cpuhp_state_mutex);
2406         return res;
2407 }
2408 static DEVICE_ATTR_RO(states);
2409
2410 static struct attribute *cpuhp_cpu_root_attrs[] = {
2411         &dev_attr_states.attr,
2412         NULL
2413 };
2414
2415 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2416         .attrs = cpuhp_cpu_root_attrs,
2417         .name = "hotplug",
2418         NULL
2419 };
2420
2421 #ifdef CONFIG_HOTPLUG_SMT
2422
2423 static ssize_t
2424 __store_smt_control(struct device *dev, struct device_attribute *attr,
2425                     const char *buf, size_t count)
2426 {
2427         int ctrlval, ret;
2428
2429         if (sysfs_streq(buf, "on"))
2430                 ctrlval = CPU_SMT_ENABLED;
2431         else if (sysfs_streq(buf, "off"))
2432                 ctrlval = CPU_SMT_DISABLED;
2433         else if (sysfs_streq(buf, "forceoff"))
2434                 ctrlval = CPU_SMT_FORCE_DISABLED;
2435         else
2436                 return -EINVAL;
2437
2438         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2439                 return -EPERM;
2440
2441         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2442                 return -ENODEV;
2443
2444         ret = lock_device_hotplug_sysfs();
2445         if (ret)
2446                 return ret;
2447
2448         if (ctrlval != cpu_smt_control) {
2449                 switch (ctrlval) {
2450                 case CPU_SMT_ENABLED:
2451                         ret = cpuhp_smt_enable();
2452                         break;
2453                 case CPU_SMT_DISABLED:
2454                 case CPU_SMT_FORCE_DISABLED:
2455                         ret = cpuhp_smt_disable(ctrlval);
2456                         break;
2457                 }
2458         }
2459
2460         unlock_device_hotplug();
2461         return ret ? ret : count;
2462 }
2463
2464 #else /* !CONFIG_HOTPLUG_SMT */
2465 static ssize_t
2466 __store_smt_control(struct device *dev, struct device_attribute *attr,
2467                     const char *buf, size_t count)
2468 {
2469         return -ENODEV;
2470 }
2471 #endif /* CONFIG_HOTPLUG_SMT */
2472
2473 static const char *smt_states[] = {
2474         [CPU_SMT_ENABLED]               = "on",
2475         [CPU_SMT_DISABLED]              = "off",
2476         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2477         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2478         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2479 };
2480
2481 static ssize_t control_show(struct device *dev,
2482                             struct device_attribute *attr, char *buf)
2483 {
2484         const char *state = smt_states[cpu_smt_control];
2485
2486         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2487 }
2488
2489 static ssize_t control_store(struct device *dev, struct device_attribute *attr,
2490                              const char *buf, size_t count)
2491 {
2492         return __store_smt_control(dev, attr, buf, count);
2493 }
2494 static DEVICE_ATTR_RW(control);
2495
2496 static ssize_t active_show(struct device *dev,
2497                            struct device_attribute *attr, char *buf)
2498 {
2499         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2500 }
2501 static DEVICE_ATTR_RO(active);
2502
2503 static struct attribute *cpuhp_smt_attrs[] = {
2504         &dev_attr_control.attr,
2505         &dev_attr_active.attr,
2506         NULL
2507 };
2508
2509 static const struct attribute_group cpuhp_smt_attr_group = {
2510         .attrs = cpuhp_smt_attrs,
2511         .name = "smt",
2512         NULL
2513 };
2514
2515 static int __init cpu_smt_sysfs_init(void)
2516 {
2517         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2518                                   &cpuhp_smt_attr_group);
2519 }
2520
2521 static int __init cpuhp_sysfs_init(void)
2522 {
2523         int cpu, ret;
2524
2525         ret = cpu_smt_sysfs_init();
2526         if (ret)
2527                 return ret;
2528
2529         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2530                                  &cpuhp_cpu_root_attr_group);
2531         if (ret)
2532                 return ret;
2533
2534         for_each_possible_cpu(cpu) {
2535                 struct device *dev = get_cpu_device(cpu);
2536
2537                 if (!dev)
2538                         continue;
2539                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2540                 if (ret)
2541                         return ret;
2542         }
2543         return 0;
2544 }
2545 device_initcall(cpuhp_sysfs_init);
2546 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2547
2548 /*
2549  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2550  * represents all NR_CPUS bits binary values of 1<<nr.
2551  *
2552  * It is used by cpumask_of() to get a constant address to a CPU
2553  * mask value that has a single bit set only.
2554  */
2555
2556 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2557 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2558 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2559 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2560 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2561
2562 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2563
2564         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2565         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2566 #if BITS_PER_LONG > 32
2567         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2568         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2569 #endif
2570 };
2571 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2572
2573 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2574 EXPORT_SYMBOL(cpu_all_bits);
2575
2576 #ifdef CONFIG_INIT_ALL_POSSIBLE
2577 struct cpumask __cpu_possible_mask __read_mostly
2578         = {CPU_BITS_ALL};
2579 #else
2580 struct cpumask __cpu_possible_mask __read_mostly;
2581 #endif
2582 EXPORT_SYMBOL(__cpu_possible_mask);
2583
2584 struct cpumask __cpu_online_mask __read_mostly;
2585 EXPORT_SYMBOL(__cpu_online_mask);
2586
2587 struct cpumask __cpu_present_mask __read_mostly;
2588 EXPORT_SYMBOL(__cpu_present_mask);
2589
2590 struct cpumask __cpu_active_mask __read_mostly;
2591 EXPORT_SYMBOL(__cpu_active_mask);
2592
2593 struct cpumask __cpu_dying_mask __read_mostly;
2594 EXPORT_SYMBOL(__cpu_dying_mask);
2595
2596 atomic_t __num_online_cpus __read_mostly;
2597 EXPORT_SYMBOL(__num_online_cpus);
2598
2599 void init_cpu_present(const struct cpumask *src)
2600 {
2601         cpumask_copy(&__cpu_present_mask, src);
2602 }
2603
2604 void init_cpu_possible(const struct cpumask *src)
2605 {
2606         cpumask_copy(&__cpu_possible_mask, src);
2607 }
2608
2609 void init_cpu_online(const struct cpumask *src)
2610 {
2611         cpumask_copy(&__cpu_online_mask, src);
2612 }
2613
2614 void set_cpu_online(unsigned int cpu, bool online)
2615 {
2616         /*
2617          * atomic_inc/dec() is required to handle the horrid abuse of this
2618          * function by the reboot and kexec code which invoke it from
2619          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2620          * regular CPU hotplug is properly serialized.
2621          *
2622          * Note, that the fact that __num_online_cpus is of type atomic_t
2623          * does not protect readers which are not serialized against
2624          * concurrent hotplug operations.
2625          */
2626         if (online) {
2627                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2628                         atomic_inc(&__num_online_cpus);
2629         } else {
2630                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2631                         atomic_dec(&__num_online_cpus);
2632         }
2633 }
2634
2635 /*
2636  * Activate the first processor.
2637  */
2638 void __init boot_cpu_init(void)
2639 {
2640         int cpu = smp_processor_id();
2641
2642         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2643         set_cpu_online(cpu, true);
2644         set_cpu_active(cpu, true);
2645         set_cpu_present(cpu, true);
2646         set_cpu_possible(cpu, true);
2647
2648 #ifdef CONFIG_SMP
2649         __boot_cpu_id = cpu;
2650 #endif
2651 }
2652
2653 /*
2654  * Must be called _AFTER_ setting up the per_cpu areas
2655  */
2656 void __init boot_cpu_hotplug_init(void)
2657 {
2658 #ifdef CONFIG_SMP
2659         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2660 #endif
2661         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2662 }
2663
2664 /*
2665  * These are used for a global "mitigations=" cmdline option for toggling
2666  * optional CPU mitigations.
2667  */
2668 enum cpu_mitigations {
2669         CPU_MITIGATIONS_OFF,
2670         CPU_MITIGATIONS_AUTO,
2671         CPU_MITIGATIONS_AUTO_NOSMT,
2672 };
2673
2674 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2675         CPU_MITIGATIONS_AUTO;
2676
2677 static int __init mitigations_parse_cmdline(char *arg)
2678 {
2679         if (!strcmp(arg, "off"))
2680                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2681         else if (!strcmp(arg, "auto"))
2682                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2683         else if (!strcmp(arg, "auto,nosmt"))
2684                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2685         else
2686                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2687                         arg);
2688
2689         return 0;
2690 }
2691 early_param("mitigations", mitigations_parse_cmdline);
2692
2693 /* mitigations=off */
2694 bool cpu_mitigations_off(void)
2695 {
2696         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2697 }
2698 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2699
2700 /* mitigations=auto,nosmt */
2701 bool cpu_mitigations_auto_nosmt(void)
2702 {
2703         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2704 }
2705 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);