kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/sched/mm.h>
   7 #include <linux/proc_fs.h>
   8 #include <linux/smp.h>
   9 #include <linux/init.h>
  10 #include <linux/notifier.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/sched/hotplug.h>
  13 #include <linux/sched/isolation.h>
  14 #include <linux/sched/task.h>
  15 #include <linux/sched/smt.h>
  16 #include <linux/unistd.h>
  17 #include <linux/cpu.h>
  18 #include <linux/oom.h>
  19 #include <linux/rcupdate.h>
  20 #include <linux/export.h>
  21 #include <linux/bug.h>
  22 #include <linux/kthread.h>
  23 #include <linux/stop_machine.h>
  24 #include <linux/mutex.h>
  25 #include <linux/gfp.h>
  26 #include <linux/suspend.h>
  27 #include <linux/lockdep.h>
  28 #include <linux/tick.h>
  29 #include <linux/irq.h>
  30 #include <linux/nmi.h>
  31 #include <linux/smpboot.h>
  32 #include <linux/relay.h>
  33 #include <linux/slab.h>
  34 #include <linux/percpu-rwsem.h>
  35 #include <linux/cpuset.h>
  36
  37 #include <trace/events/power.h>
  38 #define CREATE_TRACE_POINTS
  39 #include <trace/events/cpuhp.h>
  40
  41 #include "smpboot.h"
  42
  43 /**
  44  * cpuhp_cpu_state - Per cpu hotplug state storage
  45  * @state:      The current cpu state
  46  * @target:     The target state
  47  * @thread:     Pointer to the hotplug thread
  48  * @should_run: Thread should execute
  49  * @rollback:   Perform a rollback
  50  * @single:     Single callback invocation
  51  * @bringup:    Single callback bringup or teardown selector
  52  * @cb_state:   The state for a single callback (install/uninstall)
  53  * @result:     Result of the operation
  54  * @done_up:    Signal completion to the issuer of the task for cpu-up
  55  * @done_down:  Signal completion to the issuer of the task for cpu-down
  56  */
  57 struct cpuhp_cpu_state {
  58         enum cpuhp_state        state;
  59         enum cpuhp_state        target;
  60         enum cpuhp_state        fail;
  61 #ifdef CONFIG_SMP
  62         struct task_struct      *thread;
  63         bool                    should_run;
  64         bool                    rollback;
  65         bool                    single;
  66         bool                    bringup;
  67         int                     cpu;
  68         struct hlist_node       *node;
  69         struct hlist_node       *last;
  70         enum cpuhp_state        cb_state;
  71         int                     result;
  72         struct completion       done_up;
  73         struct completion       done_down;
  74 #endif
  75 };
  76
  77 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  78         .fail = CPUHP_INVALID,
  79 };
  80
  81 #ifdef CONFIG_SMP
  82 cpumask_t cpus_booted_once_mask;
  83 #endif
  84
  85 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  86 static struct lockdep_map cpuhp_state_up_map =
  87         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  88 static struct lockdep_map cpuhp_state_down_map =
  89         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  90
  91
  92 static inline void cpuhp_lock_acquire(bool bringup)
  93 {
  94         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  95 }
  96
  97 static inline void cpuhp_lock_release(bool bringup)
  98 {
  99         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
 100 }
 101 #else
 102
 103 static inline void cpuhp_lock_acquire(bool bringup) { }
 104 static inline void cpuhp_lock_release(bool bringup) { }
 105
 106 #endif
 107
 108 /**
 109  * cpuhp_step - Hotplug state machine step
 110  * @name:       Name of the step
 111  * @startup:    Startup function of the step
 112  * @teardown:   Teardown function of the step
 113  * @cant_stop:  Bringup/teardown can't be stopped at this step
 114  */
 115 struct cpuhp_step {
 116         const char              *name;
 117         union {
 118                 int             (*single)(unsigned int cpu);
 119                 int             (*multi)(unsigned int cpu,
 120                                          struct hlist_node *node);
 121         } startup;
 122         union {
 123                 int             (*single)(unsigned int cpu);
 124                 int             (*multi)(unsigned int cpu,
 125                                          struct hlist_node *node);
 126         } teardown;
 127         struct hlist_head       list;
 128         bool                    cant_stop;
 129         bool                    multi_instance;
 130 };
 131
 132 static DEFINE_MUTEX(cpuhp_state_mutex);
 133 static struct cpuhp_step cpuhp_hp_states[];
 134
 135 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 136 {
 137         return cpuhp_hp_states + state;
 138 }
 139
 140 static bool cpuhp_step_empty(bool bringup, struct cpuhp_step *step)
 141 {
 142         return bringup ? !step->startup.single : !step->teardown.single;
 143 }
 144
 145 /**
 146  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 147  * @cpu:        The cpu for which the callback should be invoked
 148  * @state:      The state to do callbacks for
 149  * @bringup:    True if the bringup callback should be invoked
 150  * @node:       For multi-instance, do a single entry callback for install/remove
 151  * @lastp:      For multi-instance rollback, remember how far we got
 152  *
 153  * Called from cpu hotplug and from the state register machinery.
 154  */
 155 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 156                                  bool bringup, struct hlist_node *node,
 157                                  struct hlist_node **lastp)
 158 {
 159         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 160         struct cpuhp_step *step = cpuhp_get_step(state);
 161         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 162         int (*cb)(unsigned int cpu);
 163         int ret, cnt;
 164
 165         if (st->fail == state) {
 166                 st->fail = CPUHP_INVALID;
 167                 return -EAGAIN;
 168         }
 169
 170         if (cpuhp_step_empty(bringup, step)) {
 171                 WARN_ON_ONCE(1);
 172                 return 0;
 173         }
 174
 175         if (!step->multi_instance) {
 176                 WARN_ON_ONCE(lastp && *lastp);
 177                 cb = bringup ? step->startup.single : step->teardown.single;
 178
 179                 trace_cpuhp_enter(cpu, st->target, state, cb);
 180                 ret = cb(cpu);
 181                 trace_cpuhp_exit(cpu, st->state, state, ret);
 182                 return ret;
 183         }
 184         cbm = bringup ? step->startup.multi : step->teardown.multi;
 185
 186         /* Single invocation for instance add/remove */
 187         if (node) {
 188                 WARN_ON_ONCE(lastp && *lastp);
 189                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 190                 ret = cbm(cpu, node);
 191                 trace_cpuhp_exit(cpu, st->state, state, ret);
 192                 return ret;
 193         }
 194
 195         /* State transition. Invoke on all instances */
 196         cnt = 0;
 197         hlist_for_each(node, &step->list) {
 198                 if (lastp && node == *lastp)
 199                         break;
 200
 201                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 202                 ret = cbm(cpu, node);
 203                 trace_cpuhp_exit(cpu, st->state, state, ret);
 204                 if (ret) {
 205                         if (!lastp)
 206                                 goto err;
 207
 208                         *lastp = node;
 209                         return ret;
 210                 }
 211                 cnt++;
 212         }
 213         if (lastp)
 214                 *lastp = NULL;
 215         return 0;
 216 err:
 217         /* Rollback the instances if one failed */
 218         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 219         if (!cbm)
 220                 return ret;
 221
 222         hlist_for_each(node, &step->list) {
 223                 if (!cnt--)
 224                         break;
 225
 226                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 227                 ret = cbm(cpu, node);
 228                 trace_cpuhp_exit(cpu, st->state, state, ret);
 229                 /*
 230                  * Rollback must not fail,
 231                  */
 232                 WARN_ON_ONCE(ret);
 233         }
 234         return ret;
 235 }
 236
 237 #ifdef CONFIG_SMP
 238 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 239 {
 240         /*
 241          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 242          * purposes as that state is handled explicitly in cpu_down.
 243          */
 244         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 245 }
 246
 247 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 248 {
 249         struct completion *done = bringup ? &st->done_up : &st->done_down;
 250         wait_for_completion(done);
 251 }
 252
 253 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 254 {
 255         struct completion *done = bringup ? &st->done_up : &st->done_down;
 256         complete(done);
 257 }
 258
 259 /*
 260  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 261  */
 262 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 263 {
 264         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 265 }
 266
 267 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 268 static DEFINE_MUTEX(cpu_add_remove_lock);
 269 bool cpuhp_tasks_frozen;
 270 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 271
 272 /*
 273  * The following two APIs (cpu_maps_update_begin/done) must be used when
 274  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 275  */
 276 void cpu_maps_update_begin(void)
 277 {
 278         mutex_lock(&cpu_add_remove_lock);
 279 }
 280
 281 void cpu_maps_update_done(void)
 282 {
 283         mutex_unlock(&cpu_add_remove_lock);
 284 }
 285
 286 /*
 287  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 288  * Should always be manipulated under cpu_add_remove_lock
 289  */
 290 static int cpu_hotplug_disabled;
 291
 292 #ifdef CONFIG_HOTPLUG_CPU
 293
 294 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 295
 296 void cpus_read_lock(void)
 297 {
 298         percpu_down_read(&cpu_hotplug_lock);
 299 }
 300 EXPORT_SYMBOL_GPL(cpus_read_lock);
 301
 302 int cpus_read_trylock(void)
 303 {
 304         return percpu_down_read_trylock(&cpu_hotplug_lock);
 305 }
 306 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 307
 308 void cpus_read_unlock(void)
 309 {
 310         percpu_up_read(&cpu_hotplug_lock);
 311 }
 312 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 313
 314 void cpus_write_lock(void)
 315 {
 316         percpu_down_write(&cpu_hotplug_lock);
 317 }
 318
 319 void cpus_write_unlock(void)
 320 {
 321         percpu_up_write(&cpu_hotplug_lock);
 322 }
 323
 324 void lockdep_assert_cpus_held(void)
 325 {
 326         /*
 327          * We can't have hotplug operations before userspace starts running,
 328          * and some init codepaths will knowingly not take the hotplug lock.
 329          * This is all valid, so mute lockdep until it makes sense to report
 330          * unheld locks.
 331          */
 332         if (system_state < SYSTEM_RUNNING)
 333                 return;
 334
 335         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 336 }
 337
 338 #ifdef CONFIG_LOCKDEP
 339 int lockdep_is_cpus_held(void)
 340 {
 341         return percpu_rwsem_is_held(&cpu_hotplug_lock);
 342 }
 343 #endif
 344
 345 static void lockdep_acquire_cpus_lock(void)
 346 {
 347         rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
 348 }
 349
 350 static void lockdep_release_cpus_lock(void)
 351 {
 352         rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
 353 }
 354
 355 /*
 356  * Wait for currently running CPU hotplug operations to complete (if any) and
 357  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 358  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 359  * hotplug path before performing hotplug operations. So acquiring that lock
 360  * guarantees mutual exclusion from any currently running hotplug operations.
 361  */
 362 void cpu_hotplug_disable(void)
 363 {
 364         cpu_maps_update_begin();
 365         cpu_hotplug_disabled++;
 366         cpu_maps_update_done();
 367 }
 368 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 369
 370 static void __cpu_hotplug_enable(void)
 371 {
 372         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 373                 return;
 374         cpu_hotplug_disabled--;
 375 }
 376
 377 void cpu_hotplug_enable(void)
 378 {
 379         cpu_maps_update_begin();
 380         __cpu_hotplug_enable();
 381         cpu_maps_update_done();
 382 }
 383 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 384
 385 #else
 386
 387 static void lockdep_acquire_cpus_lock(void)
 388 {
 389 }
 390
 391 static void lockdep_release_cpus_lock(void)
 392 {
 393 }
 394
 395 #endif  /* CONFIG_HOTPLUG_CPU */
 396
 397 /*
 398  * Architectures that need SMT-specific errata handling during SMT hotplug
 399  * should override this.
 400  */
 401 void __weak arch_smt_update(void) { }
 402
 403 #ifdef CONFIG_HOTPLUG_SMT
 404 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 405
 406 void __init cpu_smt_disable(bool force)
 407 {
 408         if (!cpu_smt_possible())
 409                 return;
 410
 411         if (force) {
 412                 pr_info("SMT: Force disabled\n");
 413                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 414         } else {
 415                 pr_info("SMT: disabled\n");
 416                 cpu_smt_control = CPU_SMT_DISABLED;
 417         }
 418 }
 419
 420 /*
 421  * The decision whether SMT is supported can only be done after the full
 422  * CPU identification. Called from architecture code.
 423  */
 424 void __init cpu_smt_check_topology(void)
 425 {
 426         if (!topology_smt_supported())
 427                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 428 }
 429
 430 static int __init smt_cmdline_disable(char *str)
 431 {
 432         cpu_smt_disable(str && !strcmp(str, "force"));
 433         return 0;
 434 }
 435 early_param("nosmt", smt_cmdline_disable);
 436
 437 static inline bool cpu_smt_allowed(unsigned int cpu)
 438 {
 439         if (cpu_smt_control == CPU_SMT_ENABLED)
 440                 return true;
 441
 442         if (topology_is_primary_thread(cpu))
 443                 return true;
 444
 445         /*
 446          * On x86 it's required to boot all logical CPUs at least once so
 447          * that the init code can get a chance to set CR4.MCE on each
 448          * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
 449          * core will shutdown the machine.
 450          */
 451         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
 452 }
 453
 454 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
 455 bool cpu_smt_possible(void)
 456 {
 457         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
 458                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
 459 }
 460 EXPORT_SYMBOL_GPL(cpu_smt_possible);
 461 #else
 462 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 463 #endif
 464
 465 static inline enum cpuhp_state
 466 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 467 {
 468         enum cpuhp_state prev_state = st->state;
 469         bool bringup = st->state < target;
 470
 471         st->rollback = false;
 472         st->last = NULL;
 473
 474         st->target = target;
 475         st->single = false;
 476         st->bringup = bringup;
 477         if (cpu_dying(st->cpu) != !bringup)
 478                 set_cpu_dying(st->cpu, !bringup);
 479
 480         return prev_state;
 481 }
 482
 483 static inline void
 484 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 485 {
 486         bool bringup = !st->bringup;
 487
 488         st->target = prev_state;
 489
 490         /*
 491          * Already rolling back. No need invert the bringup value or to change
 492          * the current state.
 493          */
 494         if (st->rollback)
 495                 return;
 496
 497         st->rollback = true;
 498
 499         /*
 500          * If we have st->last we need to undo partial multi_instance of this
 501          * state first. Otherwise start undo at the previous state.
 502          */
 503         if (!st->last) {
 504                 if (st->bringup)
 505                         st->state--;
 506                 else
 507                         st->state++;
 508         }
 509
 510         st->bringup = bringup;
 511         if (cpu_dying(st->cpu) != !bringup)
 512                 set_cpu_dying(st->cpu, !bringup);
 513 }
 514
 515 /* Regular hotplug invocation of the AP hotplug thread */
 516 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 517 {
 518         if (!st->single && st->state == st->target)
 519                 return;
 520
 521         st->result = 0;
 522         /*
 523          * Make sure the above stores are visible before should_run becomes
 524          * true. Paired with the mb() above in cpuhp_thread_fun()
 525          */
 526         smp_mb();
 527         st->should_run = true;
 528         wake_up_process(st->thread);
 529         wait_for_ap_thread(st, st->bringup);
 530 }
 531
 532 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 533 {
 534         enum cpuhp_state prev_state;
 535         int ret;
 536
 537         prev_state = cpuhp_set_state(st, target);
 538         __cpuhp_kick_ap(st);
 539         if ((ret = st->result)) {
 540                 cpuhp_reset_state(st, prev_state);
 541                 __cpuhp_kick_ap(st);
 542         }
 543
 544         return ret;
 545 }
 546
 547 static int bringup_wait_for_ap(unsigned int cpu)
 548 {
 549         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 550
 551         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 552         wait_for_ap_thread(st, true);
 553         if (WARN_ON_ONCE((!cpu_online(cpu))))
 554                 return -ECANCELED;
 555
 556         /* Unpark the hotplug thread of the target cpu */
 557         kthread_unpark(st->thread);
 558
 559         /*
 560          * SMT soft disabling on X86 requires to bring the CPU out of the
 561          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 562          * CPU marked itself as booted_once in notify_cpu_starting() so the
 563          * cpu_smt_allowed() check will now return false if this is not the
 564          * primary sibling.
 565          */
 566         if (!cpu_smt_allowed(cpu))
 567                 return -ECANCELED;
 568
 569         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 570                 return 0;
 571
 572         return cpuhp_kick_ap(st, st->target);
 573 }
 574
 575 static int bringup_cpu(unsigned int cpu)
 576 {
 577         struct task_struct *idle = idle_thread_get(cpu);
 578         int ret;
 579
 580         /*
 581          * Some architectures have to walk the irq descriptors to
 582          * setup the vector space for the cpu which comes online.
 583          * Prevent irq alloc/free across the bringup.
 584          */
 585         irq_lock_sparse();
 586
 587         /* Arch-specific enabling code. */
 588         ret = __cpu_up(cpu, idle);
 589         irq_unlock_sparse();
 590         if (ret)
 591                 return ret;
 592         return bringup_wait_for_ap(cpu);
 593 }
 594
 595 static int finish_cpu(unsigned int cpu)
 596 {
 597         struct task_struct *idle = idle_thread_get(cpu);
 598         struct mm_struct *mm = idle->active_mm;
 599
 600         /*
 601          * idle_task_exit() will have switched to &init_mm, now
 602          * clean up any remaining active_mm state.
 603          */
 604         if (mm != &init_mm)
 605                 idle->active_mm = &init_mm;
 606         mmdrop(mm);
 607         return 0;
 608 }
 609
 610 /*
 611  * Hotplug state machine related functions
 612  */
 613
 614 /*
 615  * Get the next state to run. Empty ones will be skipped. Returns true if a
 616  * state must be run.
 617  *
 618  * st->state will be modified ahead of time, to match state_to_run, as if it
 619  * has already ran.
 620  */
 621 static bool cpuhp_next_state(bool bringup,
 622                              enum cpuhp_state *state_to_run,
 623                              struct cpuhp_cpu_state *st,
 624                              enum cpuhp_state target)
 625 {
 626         do {
 627                 if (bringup) {
 628                         if (st->state >= target)
 629                                 return false;
 630
 631                         *state_to_run = ++st->state;
 632                 } else {
 633                         if (st->state <= target)
 634                                 return false;
 635
 636                         *state_to_run = st->state--;
 637                 }
 638
 639                 if (!cpuhp_step_empty(bringup, cpuhp_get_step(*state_to_run)))
 640                         break;
 641         } while (true);
 642
 643         return true;
 644 }
 645
 646 static int cpuhp_invoke_callback_range(bool bringup,
 647                                        unsigned int cpu,
 648                                        struct cpuhp_cpu_state *st,
 649                                        enum cpuhp_state target)
 650 {
 651         enum cpuhp_state state;
 652         int err = 0;
 653
 654         while (cpuhp_next_state(bringup, &state, st, target)) {
 655                 err = cpuhp_invoke_callback(cpu, state, bringup, NULL, NULL);
 656                 if (err)
 657                         break;
 658         }
 659
 660         return err;
 661 }
 662
 663 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 664 {
 665         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 666                 return true;
 667         /*
 668          * When CPU hotplug is disabled, then taking the CPU down is not
 669          * possible because takedown_cpu() and the architecture and
 670          * subsystem specific mechanisms are not available. So the CPU
 671          * which would be completely unplugged again needs to stay around
 672          * in the current state.
 673          */
 674         return st->state <= CPUHP_BRINGUP_CPU;
 675 }
 676
 677 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 678                               enum cpuhp_state target)
 679 {
 680         enum cpuhp_state prev_state = st->state;
 681         int ret = 0;
 682
 683         ret = cpuhp_invoke_callback_range(true, cpu, st, target);
 684         if (ret) {
 685                 cpuhp_reset_state(st, prev_state);
 686                 if (can_rollback_cpu(st))
 687                         WARN_ON(cpuhp_invoke_callback_range(false, cpu, st,
 688                                                             prev_state));
 689         }
 690         return ret;
 691 }
 692
 693 /*
 694  * The cpu hotplug threads manage the bringup and teardown of the cpus
 695  */
 696 static void cpuhp_create(unsigned int cpu)
 697 {
 698         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 699
 700         init_completion(&st->done_up);
 701         init_completion(&st->done_down);
 702         st->cpu = cpu;
 703 }
 704
 705 static int cpuhp_should_run(unsigned int cpu)
 706 {
 707         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 708
 709         return st->should_run;
 710 }
 711
 712 /*
 713  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 714  * callbacks when a state gets [un]installed at runtime.
 715  *
 716  * Each invocation of this function by the smpboot thread does a single AP
 717  * state callback.
 718  *
 719  * It has 3 modes of operation:
 720  *  - single: runs st->cb_state
 721  *  - up:     runs ++st->state, while st->state < st->target
 722  *  - down:   runs st->state--, while st->state > st->target
 723  *
 724  * When complete or on error, should_run is cleared and the completion is fired.
 725  */
 726 static void cpuhp_thread_fun(unsigned int cpu)
 727 {
 728         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 729         bool bringup = st->bringup;
 730         enum cpuhp_state state;
 731
 732         if (WARN_ON_ONCE(!st->should_run))
 733                 return;
 734
 735         /*
 736          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 737          * that if we see ->should_run we also see the rest of the state.
 738          */
 739         smp_mb();
 740
 741         /*
 742          * The BP holds the hotplug lock, but we're now running on the AP,
 743          * ensure that anybody asserting the lock is held, will actually find
 744          * it so.
 745          */
 746         lockdep_acquire_cpus_lock();
 747         cpuhp_lock_acquire(bringup);
 748
 749         if (st->single) {
 750                 state = st->cb_state;
 751                 st->should_run = false;
 752         } else {
 753                 st->should_run = cpuhp_next_state(bringup, &state, st, st->target);
 754                 if (!st->should_run)
 755                         goto end;
 756         }
 757
 758         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 759
 760         if (cpuhp_is_atomic_state(state)) {
 761                 local_irq_disable();
 762                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 763                 local_irq_enable();
 764
 765                 /*
 766                  * STARTING/DYING must not fail!
 767                  */
 768                 WARN_ON_ONCE(st->result);
 769         } else {
 770                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 771         }
 772
 773         if (st->result) {
 774                 /*
 775                  * If we fail on a rollback, we're up a creek without no
 776                  * paddle, no way forward, no way back. We loose, thanks for
 777                  * playing.
 778                  */
 779                 WARN_ON_ONCE(st->rollback);
 780                 st->should_run = false;
 781         }
 782
 783 end:
 784         cpuhp_lock_release(bringup);
 785         lockdep_release_cpus_lock();
 786
 787         if (!st->should_run)
 788                 complete_ap_thread(st, bringup);
 789 }
 790
 791 /* Invoke a single callback on a remote cpu */
 792 static int
 793 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 794                          struct hlist_node *node)
 795 {
 796         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 797         int ret;
 798
 799         if (!cpu_online(cpu))
 800                 return 0;
 801
 802         cpuhp_lock_acquire(false);
 803         cpuhp_lock_release(false);
 804
 805         cpuhp_lock_acquire(true);
 806         cpuhp_lock_release(true);
 807
 808         /*
 809          * If we are up and running, use the hotplug thread. For early calls
 810          * we invoke the thread function directly.
 811          */
 812         if (!st->thread)
 813                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 814
 815         st->rollback = false;
 816         st->last = NULL;
 817
 818         st->node = node;
 819         st->bringup = bringup;
 820         st->cb_state = state;
 821         st->single = true;
 822
 823         __cpuhp_kick_ap(st);
 824
 825         /*
 826          * If we failed and did a partial, do a rollback.
 827          */
 828         if ((ret = st->result) && st->last) {
 829                 st->rollback = true;
 830                 st->bringup = !bringup;
 831
 832                 __cpuhp_kick_ap(st);
 833         }
 834
 835         /*
 836          * Clean up the leftovers so the next hotplug operation wont use stale
 837          * data.
 838          */
 839         st->node = st->last = NULL;
 840         return ret;
 841 }
 842
 843 static int cpuhp_kick_ap_work(unsigned int cpu)
 844 {
 845         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 846         enum cpuhp_state prev_state = st->state;
 847         int ret;
 848
 849         cpuhp_lock_acquire(false);
 850         cpuhp_lock_release(false);
 851
 852         cpuhp_lock_acquire(true);
 853         cpuhp_lock_release(true);
 854
 855         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 856         ret = cpuhp_kick_ap(st, st->target);
 857         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 858
 859         return ret;
 860 }
 861
 862 static struct smp_hotplug_thread cpuhp_threads = {
 863         .store                  = &cpuhp_state.thread,
 864         .create                 = &cpuhp_create,
 865         .thread_should_run      = cpuhp_should_run,
 866         .thread_fn              = cpuhp_thread_fun,
 867         .thread_comm            = "cpuhp/%u",
 868         .selfparking            = true,
 869 };
 870
 871 void __init cpuhp_threads_init(void)
 872 {
 873         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 874         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 875 }
 876
 877 /*
 878  *
 879  * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
 880  * protected region.
 881  *
 882  * The operation is still serialized against concurrent CPU hotplug via
 883  * cpu_add_remove_lock, i.e. CPU map protection.  But it is _not_
 884  * serialized against other hotplug related activity like adding or
 885  * removing of state callbacks and state instances, which invoke either the
 886  * startup or the teardown callback of the affected state.
 887  *
 888  * This is required for subsystems which are unfixable vs. CPU hotplug and
 889  * evade lock inversion problems by scheduling work which has to be
 890  * completed _before_ cpu_up()/_cpu_down() returns.
 891  *
 892  * Don't even think about adding anything to this for any new code or even
 893  * drivers. It's only purpose is to keep existing lock order trainwrecks
 894  * working.
 895  *
 896  * For cpu_down() there might be valid reasons to finish cleanups which are
 897  * not required to be done under cpu_hotplug_lock, but that's a different
 898  * story and would be not invoked via this.
 899  */
 900 static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
 901 {
 902         /*
 903          * cpusets delegate hotplug operations to a worker to "solve" the
 904          * lock order problems. Wait for the worker, but only if tasks are
 905          * _not_ frozen (suspend, hibernate) as that would wait forever.
 906          *
 907          * The wait is required because otherwise the hotplug operation
 908          * returns with inconsistent state, which could even be observed in
 909          * user space when a new CPU is brought up. The CPU plug uevent
 910          * would be delivered and user space reacting on it would fail to
 911          * move tasks to the newly plugged CPU up to the point where the
 912          * work has finished because up to that point the newly plugged CPU
 913          * is not assignable in cpusets/cgroups. On unplug that's not
 914          * necessarily a visible issue, but it is still inconsistent state,
 915          * which is the real problem which needs to be "fixed". This can't
 916          * prevent the transient state between scheduling the work and
 917          * returning from waiting for it.
 918          */
 919         if (!tasks_frozen)
 920                 cpuset_wait_for_hotplug();
 921 }
 922
 923 #ifdef CONFIG_HOTPLUG_CPU
 924 #ifndef arch_clear_mm_cpumask_cpu
 925 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
 926 #endif
 927
 928 /**
 929  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 930  * @cpu: a CPU id
 931  *
 932  * This function walks all processes, finds a valid mm struct for each one and
 933  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 934  * trivial, there are various non-obvious corner cases, which this function
 935  * tries to solve in a safe manner.
 936  *
 937  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 938  * be called only for an already offlined CPU.
 939  */
 940 void clear_tasks_mm_cpumask(int cpu)
 941 {
 942         struct task_struct *p;
 943
 944         /*
 945          * This function is called after the cpu is taken down and marked
 946          * offline, so its not like new tasks will ever get this cpu set in
 947          * their mm mask. -- Peter Zijlstra
 948          * Thus, we may use rcu_read_lock() here, instead of grabbing
 949          * full-fledged tasklist_lock.
 950          */
 951         WARN_ON(cpu_online(cpu));
 952         rcu_read_lock();
 953         for_each_process(p) {
 954                 struct task_struct *t;
 955
 956                 /*
 957                  * Main thread might exit, but other threads may still have
 958                  * a valid mm. Find one.
 959                  */
 960                 t = find_lock_task_mm(p);
 961                 if (!t)
 962                         continue;
 963                 arch_clear_mm_cpumask_cpu(cpu, t->mm);
 964                 task_unlock(t);
 965         }
 966         rcu_read_unlock();
 967 }
 968
 969 /* Take this CPU down. */
 970 static int take_cpu_down(void *_param)
 971 {
 972         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 973         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 974         int err, cpu = smp_processor_id();
 975         int ret;
 976
 977         /* Ensure this CPU doesn't handle any more interrupts. */
 978         err = __cpu_disable();
 979         if (err < 0)
 980                 return err;
 981
 982         /*
 983          * Must be called from CPUHP_TEARDOWN_CPU, which means, as we are going
 984          * down, that the current state is CPUHP_TEARDOWN_CPU - 1.
 985          */
 986         WARN_ON(st->state != (CPUHP_TEARDOWN_CPU - 1));
 987
 988         /* Invoke the former CPU_DYING callbacks */
 989         ret = cpuhp_invoke_callback_range(false, cpu, st, target);
 990
 991         /*
 992          * DYING must not fail!
 993          */
 994         WARN_ON_ONCE(ret);
 995
 996         /* Give up timekeeping duties */
 997         tick_handover_do_timer();
 998         /* Remove CPU from timer broadcasting */
 999         tick_offline_cpu(cpu);
1000         /* Park the stopper thread */
1001         stop_machine_park(cpu);
1002         return 0;
1003 }
1004
1005 static int takedown_cpu(unsigned int cpu)
1006 {
1007         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1008         int err;
1009
1010         /* Park the smpboot threads */
1011         kthread_park(st->thread);
1012
1013         /*
1014          * Prevent irq alloc/free while the dying cpu reorganizes the
1015          * interrupt affinities.
1016          */
1017         irq_lock_sparse();
1018
1019         /*
1020          * So now all preempt/rcu users must observe !cpu_active().
1021          */
1022         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
1023         if (err) {
1024                 /* CPU refused to die */
1025                 irq_unlock_sparse();
1026                 /* Unpark the hotplug thread so we can rollback there */
1027                 kthread_unpark(st->thread);
1028                 return err;
1029         }
1030         BUG_ON(cpu_online(cpu));
1031
1032         /*
1033          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
1034          * all runnable tasks from the CPU, there's only the idle task left now
1035          * that the migration thread is done doing the stop_machine thing.
1036          *
1037          * Wait for the stop thread to go away.
1038          */
1039         wait_for_ap_thread(st, false);
1040         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
1041
1042         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
1043         irq_unlock_sparse();
1044
1045         hotplug_cpu__broadcast_tick_pull(cpu);
1046         /* This actually kills the CPU. */
1047         __cpu_die(cpu);
1048
1049         tick_cleanup_dead_cpu(cpu);
1050         rcutree_migrate_callbacks(cpu);
1051         return 0;
1052 }
1053
1054 static void cpuhp_complete_idle_dead(void *arg)
1055 {
1056         struct cpuhp_cpu_state *st = arg;
1057
1058         complete_ap_thread(st, false);
1059 }
1060
1061 void cpuhp_report_idle_dead(void)
1062 {
1063         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1064
1065         BUG_ON(st->state != CPUHP_AP_OFFLINE);
1066         rcu_report_dead(smp_processor_id());
1067         st->state = CPUHP_AP_IDLE_DEAD;
1068         /*
1069          * We cannot call complete after rcu_report_dead() so we delegate it
1070          * to an online cpu.
1071          */
1072         smp_call_function_single(cpumask_first(cpu_online_mask),
1073                                  cpuhp_complete_idle_dead, st, 0);
1074 }
1075
1076 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1077                                 enum cpuhp_state target)
1078 {
1079         enum cpuhp_state prev_state = st->state;
1080         int ret = 0;
1081
1082         ret = cpuhp_invoke_callback_range(false, cpu, st, target);
1083         if (ret) {
1084
1085                 cpuhp_reset_state(st, prev_state);
1086
1087                 if (st->state < prev_state)
1088                         WARN_ON(cpuhp_invoke_callback_range(true, cpu, st,
1089                                                             prev_state));
1090         }
1091
1092         return ret;
1093 }
1094
1095 /* Requires cpu_add_remove_lock to be held */
1096 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1097                            enum cpuhp_state target)
1098 {
1099         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1100         int prev_state, ret = 0;
1101
1102         if (num_online_cpus() == 1)
1103                 return -EBUSY;
1104
1105         if (!cpu_present(cpu))
1106                 return -EINVAL;
1107
1108         cpus_write_lock();
1109
1110         cpuhp_tasks_frozen = tasks_frozen;
1111
1112         prev_state = cpuhp_set_state(st, target);
1113         /*
1114          * If the current CPU state is in the range of the AP hotplug thread,
1115          * then we need to kick the thread.
1116          */
1117         if (st->state > CPUHP_TEARDOWN_CPU) {
1118                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1119                 ret = cpuhp_kick_ap_work(cpu);
1120                 /*
1121                  * The AP side has done the error rollback already. Just
1122                  * return the error code..
1123                  */
1124                 if (ret)
1125                         goto out;
1126
1127                 /*
1128                  * We might have stopped still in the range of the AP hotplug
1129                  * thread. Nothing to do anymore.
1130                  */
1131                 if (st->state > CPUHP_TEARDOWN_CPU)
1132                         goto out;
1133
1134                 st->target = target;
1135         }
1136         /*
1137          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1138          * to do the further cleanups.
1139          */
1140         ret = cpuhp_down_callbacks(cpu, st, target);
1141         if (ret && st->state < prev_state) {
1142                 if (st->state == CPUHP_TEARDOWN_CPU) {
1143                         cpuhp_reset_state(st, prev_state);
1144                         __cpuhp_kick_ap(st);
1145                 } else {
1146                         WARN(1, "DEAD callback error for CPU%d", cpu);
1147                 }
1148         }
1149
1150 out:
1151         cpus_write_unlock();
1152         /*
1153          * Do post unplug cleanup. This is still protected against
1154          * concurrent CPU hotplug via cpu_add_remove_lock.
1155          */
1156         lockup_detector_cleanup();
1157         arch_smt_update();
1158         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1159         return ret;
1160 }
1161
1162 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1163 {
1164         if (cpu_hotplug_disabled)
1165                 return -EBUSY;
1166         return _cpu_down(cpu, 0, target);
1167 }
1168
1169 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1170 {
1171         int err;
1172
1173         cpu_maps_update_begin();
1174         err = cpu_down_maps_locked(cpu, target);
1175         cpu_maps_update_done();
1176         return err;
1177 }
1178
1179 /**
1180  * cpu_device_down - Bring down a cpu device
1181  * @dev: Pointer to the cpu device to offline
1182  *
1183  * This function is meant to be used by device core cpu subsystem only.
1184  *
1185  * Other subsystems should use remove_cpu() instead.
1186  */
1187 int cpu_device_down(struct device *dev)
1188 {
1189         return cpu_down(dev->id, CPUHP_OFFLINE);
1190 }
1191
1192 int remove_cpu(unsigned int cpu)
1193 {
1194         int ret;
1195
1196         lock_device_hotplug();
1197         ret = device_offline(get_cpu_device(cpu));
1198         unlock_device_hotplug();
1199
1200         return ret;
1201 }
1202 EXPORT_SYMBOL_GPL(remove_cpu);
1203
1204 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1205 {
1206         unsigned int cpu;
1207         int error;
1208
1209         cpu_maps_update_begin();
1210
1211         /*
1212          * Make certain the cpu I'm about to reboot on is online.
1213          *
1214          * This is inline to what migrate_to_reboot_cpu() already do.
1215          */
1216         if (!cpu_online(primary_cpu))
1217                 primary_cpu = cpumask_first(cpu_online_mask);
1218
1219         for_each_online_cpu(cpu) {
1220                 if (cpu == primary_cpu)
1221                         continue;
1222
1223                 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1224                 if (error) {
1225                         pr_err("Failed to offline CPU%d - error=%d",
1226                                 cpu, error);
1227                         break;
1228                 }
1229         }
1230
1231         /*
1232          * Ensure all but the reboot CPU are offline.
1233          */
1234         BUG_ON(num_online_cpus() > 1);
1235
1236         /*
1237          * Make sure the CPUs won't be enabled by someone else after this
1238          * point. Kexec will reboot to a new kernel shortly resetting
1239          * everything along the way.
1240          */
1241         cpu_hotplug_disabled++;
1242
1243         cpu_maps_update_done();
1244 }
1245
1246 #else
1247 #define takedown_cpu            NULL
1248 #endif /*CONFIG_HOTPLUG_CPU*/
1249
1250 /**
1251  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1252  * @cpu: cpu that just started
1253  *
1254  * It must be called by the arch code on the new cpu, before the new cpu
1255  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1256  */
1257 void notify_cpu_starting(unsigned int cpu)
1258 {
1259         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1260         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1261         int ret;
1262
1263         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1264         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1265         ret = cpuhp_invoke_callback_range(true, cpu, st, target);
1266
1267         /*
1268          * STARTING must not fail!
1269          */
1270         WARN_ON_ONCE(ret);
1271 }
1272
1273 /*
1274  * Called from the idle task. Wake up the controlling task which brings the
1275  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1276  * online bringup to the hotplug thread.
1277  */
1278 void cpuhp_online_idle(enum cpuhp_state state)
1279 {
1280         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1281
1282         /* Happens for the boot cpu */
1283         if (state != CPUHP_AP_ONLINE_IDLE)
1284                 return;
1285
1286         /*
1287          * Unpart the stopper thread before we start the idle loop (and start
1288          * scheduling); this ensures the stopper task is always available.
1289          */
1290         stop_machine_unpark(smp_processor_id());
1291
1292         st->state = CPUHP_AP_ONLINE_IDLE;
1293         complete_ap_thread(st, true);
1294 }
1295
1296 /* Requires cpu_add_remove_lock to be held */
1297 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1298 {
1299         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1300         struct task_struct *idle;
1301         int ret = 0;
1302
1303         cpus_write_lock();
1304
1305         if (!cpu_present(cpu)) {
1306                 ret = -EINVAL;
1307                 goto out;
1308         }
1309
1310         /*
1311          * The caller of cpu_up() might have raced with another
1312          * caller. Nothing to do.
1313          */
1314         if (st->state >= target)
1315                 goto out;
1316
1317         if (st->state == CPUHP_OFFLINE) {
1318                 /* Let it fail before we try to bring the cpu up */
1319                 idle = idle_thread_get(cpu);
1320                 if (IS_ERR(idle)) {
1321                         ret = PTR_ERR(idle);
1322                         goto out;
1323                 }
1324         }
1325
1326         cpuhp_tasks_frozen = tasks_frozen;
1327
1328         cpuhp_set_state(st, target);
1329         /*
1330          * If the current CPU state is in the range of the AP hotplug thread,
1331          * then we need to kick the thread once more.
1332          */
1333         if (st->state > CPUHP_BRINGUP_CPU) {
1334                 ret = cpuhp_kick_ap_work(cpu);
1335                 /*
1336                  * The AP side has done the error rollback already. Just
1337                  * return the error code..
1338                  */
1339                 if (ret)
1340                         goto out;
1341         }
1342
1343         /*
1344          * Try to reach the target state. We max out on the BP at
1345          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1346          * responsible for bringing it up to the target state.
1347          */
1348         target = min((int)target, CPUHP_BRINGUP_CPU);
1349         ret = cpuhp_up_callbacks(cpu, st, target);
1350 out:
1351         cpus_write_unlock();
1352         arch_smt_update();
1353         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1354         return ret;
1355 }
1356
1357 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1358 {
1359         int err = 0;
1360
1361         if (!cpu_possible(cpu)) {
1362                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1363                        cpu);
1364 #if defined(CONFIG_IA64)
1365                 pr_err("please check additional_cpus= boot parameter\n");
1366 #endif
1367                 return -EINVAL;
1368         }
1369
1370         err = try_online_node(cpu_to_node(cpu));
1371         if (err)
1372                 return err;
1373
1374         cpu_maps_update_begin();
1375
1376         if (cpu_hotplug_disabled) {
1377                 err = -EBUSY;
1378                 goto out;
1379         }
1380         if (!cpu_smt_allowed(cpu)) {
1381                 err = -EPERM;
1382                 goto out;
1383         }
1384
1385         err = _cpu_up(cpu, 0, target);
1386 out:
1387         cpu_maps_update_done();
1388         return err;
1389 }
1390
1391 /**
1392  * cpu_device_up - Bring up a cpu device
1393  * @dev: Pointer to the cpu device to online
1394  *
1395  * This function is meant to be used by device core cpu subsystem only.
1396  *
1397  * Other subsystems should use add_cpu() instead.
1398  */
1399 int cpu_device_up(struct device *dev)
1400 {
1401         return cpu_up(dev->id, CPUHP_ONLINE);
1402 }
1403
1404 int add_cpu(unsigned int cpu)
1405 {
1406         int ret;
1407
1408         lock_device_hotplug();
1409         ret = device_online(get_cpu_device(cpu));
1410         unlock_device_hotplug();
1411
1412         return ret;
1413 }
1414 EXPORT_SYMBOL_GPL(add_cpu);
1415
1416 /**
1417  * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1418  * @sleep_cpu: The cpu we hibernated on and should be brought up.
1419  *
1420  * On some architectures like arm64, we can hibernate on any CPU, but on
1421  * wake up the CPU we hibernated on might be offline as a side effect of
1422  * using maxcpus= for example.
1423  */
1424 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1425 {
1426         int ret;
1427
1428         if (!cpu_online(sleep_cpu)) {
1429                 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1430                 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1431                 if (ret) {
1432                         pr_err("Failed to bring hibernate-CPU up!\n");
1433                         return ret;
1434                 }
1435         }
1436         return 0;
1437 }
1438
1439 void bringup_nonboot_cpus(unsigned int setup_max_cpus)
1440 {
1441         unsigned int cpu;
1442
1443         for_each_present_cpu(cpu) {
1444                 if (num_online_cpus() >= setup_max_cpus)
1445                         break;
1446                 if (!cpu_online(cpu))
1447                         cpu_up(cpu, CPUHP_ONLINE);
1448         }
1449 }
1450
1451 #ifdef CONFIG_PM_SLEEP_SMP
1452 static cpumask_var_t frozen_cpus;
1453
1454 int freeze_secondary_cpus(int primary)
1455 {
1456         int cpu, error = 0;
1457
1458         cpu_maps_update_begin();
1459         if (primary == -1) {
1460                 primary = cpumask_first(cpu_online_mask);
1461                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1462                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1463         } else {
1464                 if (!cpu_online(primary))
1465                         primary = cpumask_first(cpu_online_mask);
1466         }
1467
1468         /*
1469          * We take down all of the non-boot CPUs in one shot to avoid races
1470          * with the userspace trying to use the CPU hotplug at the same time
1471          */
1472         cpumask_clear(frozen_cpus);
1473
1474         pr_info("Disabling non-boot CPUs ...\n");
1475         for_each_online_cpu(cpu) {
1476                 if (cpu == primary)
1477                         continue;
1478
1479                 if (pm_wakeup_pending()) {
1480                         pr_info("Wakeup pending. Abort CPU freeze\n");
1481                         error = -EBUSY;
1482                         break;
1483                 }
1484
1485                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1486                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1487                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1488                 if (!error)
1489                         cpumask_set_cpu(cpu, frozen_cpus);
1490                 else {
1491                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1492                         break;
1493                 }
1494         }
1495
1496         if (!error)
1497                 BUG_ON(num_online_cpus() > 1);
1498         else
1499                 pr_err("Non-boot CPUs are not disabled\n");
1500
1501         /*
1502          * Make sure the CPUs won't be enabled by someone else. We need to do
1503          * this even in case of failure as all freeze_secondary_cpus() users are
1504          * supposed to do thaw_secondary_cpus() on the failure path.
1505          */
1506         cpu_hotplug_disabled++;
1507
1508         cpu_maps_update_done();
1509         return error;
1510 }
1511
1512 void __weak arch_thaw_secondary_cpus_begin(void)
1513 {
1514 }
1515
1516 void __weak arch_thaw_secondary_cpus_end(void)
1517 {
1518 }
1519
1520 void thaw_secondary_cpus(void)
1521 {
1522         int cpu, error;
1523
1524         /* Allow everyone to use the CPU hotplug again */
1525         cpu_maps_update_begin();
1526         __cpu_hotplug_enable();
1527         if (cpumask_empty(frozen_cpus))
1528                 goto out;
1529
1530         pr_info("Enabling non-boot CPUs ...\n");
1531
1532         arch_thaw_secondary_cpus_begin();
1533
1534         for_each_cpu(cpu, frozen_cpus) {
1535                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1536                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1537                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1538                 if (!error) {
1539                         pr_info("CPU%d is up\n", cpu);
1540                         continue;
1541                 }
1542                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1543         }
1544
1545         arch_thaw_secondary_cpus_end();
1546
1547         cpumask_clear(frozen_cpus);
1548 out:
1549         cpu_maps_update_done();
1550 }
1551
1552 static int __init alloc_frozen_cpus(void)
1553 {
1554         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1555                 return -ENOMEM;
1556         return 0;
1557 }
1558 core_initcall(alloc_frozen_cpus);
1559
1560 /*
1561  * When callbacks for CPU hotplug notifications are being executed, we must
1562  * ensure that the state of the system with respect to the tasks being frozen
1563  * or not, as reported by the notification, remains unchanged *throughout the
1564  * duration* of the execution of the callbacks.
1565  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1566  *
1567  * This synchronization is implemented by mutually excluding regular CPU
1568  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1569  * Hibernate notifications.
1570  */
1571 static int
1572 cpu_hotplug_pm_callback(struct notifier_block *nb,
1573                         unsigned long action, void *ptr)
1574 {
1575         switch (action) {
1576
1577         case PM_SUSPEND_PREPARE:
1578         case PM_HIBERNATION_PREPARE:
1579                 cpu_hotplug_disable();
1580                 break;
1581
1582         case PM_POST_SUSPEND:
1583         case PM_POST_HIBERNATION:
1584                 cpu_hotplug_enable();
1585                 break;
1586
1587         default:
1588                 return NOTIFY_DONE;
1589         }
1590
1591         return NOTIFY_OK;
1592 }
1593
1594
1595 static int __init cpu_hotplug_pm_sync_init(void)
1596 {
1597         /*
1598          * cpu_hotplug_pm_callback has higher priority than x86
1599          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1600          * to disable cpu hotplug to avoid cpu hotplug race.
1601          */
1602         pm_notifier(cpu_hotplug_pm_callback, 0);
1603         return 0;
1604 }
1605 core_initcall(cpu_hotplug_pm_sync_init);
1606
1607 #endif /* CONFIG_PM_SLEEP_SMP */
1608
1609 int __boot_cpu_id;
1610
1611 #endif /* CONFIG_SMP */
1612
1613 /* Boot processor state steps */
1614 static struct cpuhp_step cpuhp_hp_states[] = {
1615         [CPUHP_OFFLINE] = {
1616                 .name                   = "offline",
1617                 .startup.single         = NULL,
1618                 .teardown.single        = NULL,
1619         },
1620 #ifdef CONFIG_SMP
1621         [CPUHP_CREATE_THREADS]= {
1622                 .name                   = "threads:prepare",
1623                 .startup.single         = smpboot_create_threads,
1624                 .teardown.single        = NULL,
1625                 .cant_stop              = true,
1626         },
1627         [CPUHP_PERF_PREPARE] = {
1628                 .name                   = "perf:prepare",
1629                 .startup.single         = perf_event_init_cpu,
1630                 .teardown.single        = perf_event_exit_cpu,
1631         },
1632         [CPUHP_WORKQUEUE_PREP] = {
1633                 .name                   = "workqueue:prepare",
1634                 .startup.single         = workqueue_prepare_cpu,
1635                 .teardown.single        = NULL,
1636         },
1637         [CPUHP_HRTIMERS_PREPARE] = {
1638                 .name                   = "hrtimers:prepare",
1639                 .startup.single         = hrtimers_prepare_cpu,
1640                 .teardown.single        = hrtimers_dead_cpu,
1641         },
1642         [CPUHP_SMPCFD_PREPARE] = {
1643                 .name                   = "smpcfd:prepare",
1644                 .startup.single         = smpcfd_prepare_cpu,
1645                 .teardown.single        = smpcfd_dead_cpu,
1646         },
1647         [CPUHP_RELAY_PREPARE] = {
1648                 .name                   = "relay:prepare",
1649                 .startup.single         = relay_prepare_cpu,
1650                 .teardown.single        = NULL,
1651         },
1652         [CPUHP_SLAB_PREPARE] = {
1653                 .name                   = "slab:prepare",
1654                 .startup.single         = slab_prepare_cpu,
1655                 .teardown.single        = slab_dead_cpu,
1656         },
1657         [CPUHP_RCUTREE_PREP] = {
1658                 .name                   = "RCU/tree:prepare",
1659                 .startup.single         = rcutree_prepare_cpu,
1660                 .teardown.single        = rcutree_dead_cpu,
1661         },
1662         /*
1663          * On the tear-down path, timers_dead_cpu() must be invoked
1664          * before blk_mq_queue_reinit_notify() from notify_dead(),
1665          * otherwise a RCU stall occurs.
1666          */
1667         [CPUHP_TIMERS_PREPARE] = {
1668                 .name                   = "timers:prepare",
1669                 .startup.single         = timers_prepare_cpu,
1670                 .teardown.single        = timers_dead_cpu,
1671         },
1672         /* Kicks the plugged cpu into life */
1673         [CPUHP_BRINGUP_CPU] = {
1674                 .name                   = "cpu:bringup",
1675                 .startup.single         = bringup_cpu,
1676                 .teardown.single        = finish_cpu,
1677                 .cant_stop              = true,
1678         },
1679         /* Final state before CPU kills itself */
1680         [CPUHP_AP_IDLE_DEAD] = {
1681                 .name                   = "idle:dead",
1682         },
1683         /*
1684          * Last state before CPU enters the idle loop to die. Transient state
1685          * for synchronization.
1686          */
1687         [CPUHP_AP_OFFLINE] = {
1688                 .name                   = "ap:offline",
1689                 .cant_stop              = true,
1690         },
1691         /* First state is scheduler control. Interrupts are disabled */
1692         [CPUHP_AP_SCHED_STARTING] = {
1693                 .name                   = "sched:starting",
1694                 .startup.single         = sched_cpu_starting,
1695                 .teardown.single        = sched_cpu_dying,
1696         },
1697         [CPUHP_AP_RCUTREE_DYING] = {
1698                 .name                   = "RCU/tree:dying",
1699                 .startup.single         = NULL,
1700                 .teardown.single        = rcutree_dying_cpu,
1701         },
1702         [CPUHP_AP_SMPCFD_DYING] = {
1703                 .name                   = "smpcfd:dying",
1704                 .startup.single         = NULL,
1705                 .teardown.single        = smpcfd_dying_cpu,
1706         },
1707         /* Entry state on starting. Interrupts enabled from here on. Transient
1708          * state for synchronsization */
1709         [CPUHP_AP_ONLINE] = {
1710                 .name                   = "ap:online",
1711         },
1712         /*
1713          * Handled on control processor until the plugged processor manages
1714          * this itself.
1715          */
1716         [CPUHP_TEARDOWN_CPU] = {
1717                 .name                   = "cpu:teardown",
1718                 .startup.single         = NULL,
1719                 .teardown.single        = takedown_cpu,
1720                 .cant_stop              = true,
1721         },
1722
1723         [CPUHP_AP_SCHED_WAIT_EMPTY] = {
1724                 .name                   = "sched:waitempty",
1725                 .startup.single         = NULL,
1726                 .teardown.single        = sched_cpu_wait_empty,
1727         },
1728
1729         /* Handle smpboot threads park/unpark */
1730         [CPUHP_AP_SMPBOOT_THREADS] = {
1731                 .name                   = "smpboot/threads:online",
1732                 .startup.single         = smpboot_unpark_threads,
1733                 .teardown.single        = smpboot_park_threads,
1734         },
1735         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1736                 .name                   = "irq/affinity:online",
1737                 .startup.single         = irq_affinity_online_cpu,
1738                 .teardown.single        = NULL,
1739         },
1740         [CPUHP_AP_PERF_ONLINE] = {
1741                 .name                   = "perf:online",
1742                 .startup.single         = perf_event_init_cpu,
1743                 .teardown.single        = perf_event_exit_cpu,
1744         },
1745         [CPUHP_AP_WATCHDOG_ONLINE] = {
1746                 .name                   = "lockup_detector:online",
1747                 .startup.single         = lockup_detector_online_cpu,
1748                 .teardown.single        = lockup_detector_offline_cpu,
1749         },
1750         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1751                 .name                   = "workqueue:online",
1752                 .startup.single         = workqueue_online_cpu,
1753                 .teardown.single        = workqueue_offline_cpu,
1754         },
1755         [CPUHP_AP_RCUTREE_ONLINE] = {
1756                 .name                   = "RCU/tree:online",
1757                 .startup.single         = rcutree_online_cpu,
1758                 .teardown.single        = rcutree_offline_cpu,
1759         },
1760 #endif
1761         /*
1762          * The dynamically registered state space is here
1763          */
1764
1765 #ifdef CONFIG_SMP
1766         /* Last state is scheduler control setting the cpu active */
1767         [CPUHP_AP_ACTIVE] = {
1768                 .name                   = "sched:active",
1769                 .startup.single         = sched_cpu_activate,
1770                 .teardown.single        = sched_cpu_deactivate,
1771         },
1772 #endif
1773
1774         /* CPU is fully up and running. */
1775         [CPUHP_ONLINE] = {
1776                 .name                   = "online",
1777                 .startup.single         = NULL,
1778                 .teardown.single        = NULL,
1779         },
1780 };
1781
1782 /* Sanity check for callbacks */
1783 static int cpuhp_cb_check(enum cpuhp_state state)
1784 {
1785         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1786                 return -EINVAL;
1787         return 0;
1788 }
1789
1790 /*
1791  * Returns a free for dynamic slot assignment of the Online state. The states
1792  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1793  * by having no name assigned.
1794  */
1795 static int cpuhp_reserve_state(enum cpuhp_state state)
1796 {
1797         enum cpuhp_state i, end;
1798         struct cpuhp_step *step;
1799
1800         switch (state) {
1801         case CPUHP_AP_ONLINE_DYN:
1802                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1803                 end = CPUHP_AP_ONLINE_DYN_END;
1804                 break;
1805         case CPUHP_BP_PREPARE_DYN:
1806                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1807                 end = CPUHP_BP_PREPARE_DYN_END;
1808                 break;
1809         default:
1810                 return -EINVAL;
1811         }
1812
1813         for (i = state; i <= end; i++, step++) {
1814                 if (!step->name)
1815                         return i;
1816         }
1817         WARN(1, "No more dynamic states available for CPU hotplug\n");
1818         return -ENOSPC;
1819 }
1820
1821 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1822                                  int (*startup)(unsigned int cpu),
1823                                  int (*teardown)(unsigned int cpu),
1824                                  bool multi_instance)
1825 {
1826         /* (Un)Install the callbacks for further cpu hotplug operations */
1827         struct cpuhp_step *sp;
1828         int ret = 0;
1829
1830         /*
1831          * If name is NULL, then the state gets removed.
1832          *
1833          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1834          * the first allocation from these dynamic ranges, so the removal
1835          * would trigger a new allocation and clear the wrong (already
1836          * empty) state, leaving the callbacks of the to be cleared state
1837          * dangling, which causes wreckage on the next hotplug operation.
1838          */
1839         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1840                      state == CPUHP_BP_PREPARE_DYN)) {
1841                 ret = cpuhp_reserve_state(state);
1842                 if (ret < 0)
1843                         return ret;
1844                 state = ret;
1845         }
1846         sp = cpuhp_get_step(state);
1847         if (name && sp->name)
1848                 return -EBUSY;
1849
1850         sp->startup.single = startup;
1851         sp->teardown.single = teardown;
1852         sp->name = name;
1853         sp->multi_instance = multi_instance;
1854         INIT_HLIST_HEAD(&sp->list);
1855         return ret;
1856 }
1857
1858 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1859 {
1860         return cpuhp_get_step(state)->teardown.single;
1861 }
1862
1863 /*
1864  * Call the startup/teardown function for a step either on the AP or
1865  * on the current CPU.
1866  */
1867 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1868                             struct hlist_node *node)
1869 {
1870         struct cpuhp_step *sp = cpuhp_get_step(state);
1871         int ret;
1872
1873         /*
1874          * If there's nothing to do, we done.
1875          * Relies on the union for multi_instance.
1876          */
1877         if (cpuhp_step_empty(bringup, sp))
1878                 return 0;
1879         /*
1880          * The non AP bound callbacks can fail on bringup. On teardown
1881          * e.g. module removal we crash for now.
1882          */
1883 #ifdef CONFIG_SMP
1884         if (cpuhp_is_ap_state(state))
1885                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1886         else
1887                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1888 #else
1889         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1890 #endif
1891         BUG_ON(ret && !bringup);
1892         return ret;
1893 }
1894
1895 /*
1896  * Called from __cpuhp_setup_state on a recoverable failure.
1897  *
1898  * Note: The teardown callbacks for rollback are not allowed to fail!
1899  */
1900 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1901                                    struct hlist_node *node)
1902 {
1903         int cpu;
1904
1905         /* Roll back the already executed steps on the other cpus */
1906         for_each_present_cpu(cpu) {
1907                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1908                 int cpustate = st->state;
1909
1910                 if (cpu >= failedcpu)
1911                         break;
1912
1913                 /* Did we invoke the startup call on that cpu ? */
1914                 if (cpustate >= state)
1915                         cpuhp_issue_call(cpu, state, false, node);
1916         }
1917 }
1918
1919 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1920                                           struct hlist_node *node,
1921                                           bool invoke)
1922 {
1923         struct cpuhp_step *sp;
1924         int cpu;
1925         int ret;
1926
1927         lockdep_assert_cpus_held();
1928
1929         sp = cpuhp_get_step(state);
1930         if (sp->multi_instance == false)
1931                 return -EINVAL;
1932
1933         mutex_lock(&cpuhp_state_mutex);
1934
1935         if (!invoke || !sp->startup.multi)
1936                 goto add_node;
1937
1938         /*
1939          * Try to call the startup callback for each present cpu
1940          * depending on the hotplug state of the cpu.
1941          */
1942         for_each_present_cpu(cpu) {
1943                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1944                 int cpustate = st->state;
1945
1946                 if (cpustate < state)
1947                         continue;
1948
1949                 ret = cpuhp_issue_call(cpu, state, true, node);
1950                 if (ret) {
1951                         if (sp->teardown.multi)
1952                                 cpuhp_rollback_install(cpu, state, node);
1953                         goto unlock;
1954                 }
1955         }
1956 add_node:
1957         ret = 0;
1958         hlist_add_head(node, &sp->list);
1959 unlock:
1960         mutex_unlock(&cpuhp_state_mutex);
1961         return ret;
1962 }
1963
1964 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1965                                bool invoke)
1966 {
1967         int ret;
1968
1969         cpus_read_lock();
1970         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1971         cpus_read_unlock();
1972         return ret;
1973 }
1974 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1975
1976 /**
1977  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1978  * @state:              The state to setup
1979  * @invoke:             If true, the startup function is invoked for cpus where
1980  *                      cpu state >= @state
1981  * @startup:            startup callback function
1982  * @teardown:           teardown callback function
1983  * @multi_instance:     State is set up for multiple instances which get
1984  *                      added afterwards.
1985  *
1986  * The caller needs to hold cpus read locked while calling this function.
1987  * Returns:
1988  *   On success:
1989  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1990  *      0 for all other states
1991  *   On failure: proper (negative) error code
1992  */
1993 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1994                                    const char *name, bool invoke,
1995                                    int (*startup)(unsigned int cpu),
1996                                    int (*teardown)(unsigned int cpu),
1997                                    bool multi_instance)
1998 {
1999         int cpu, ret = 0;
2000         bool dynstate;
2001
2002         lockdep_assert_cpus_held();
2003
2004         if (cpuhp_cb_check(state) || !name)
2005                 return -EINVAL;
2006
2007         mutex_lock(&cpuhp_state_mutex);
2008
2009         ret = cpuhp_store_callbacks(state, name, startup, teardown,
2010                                     multi_instance);
2011
2012         dynstate = state == CPUHP_AP_ONLINE_DYN;
2013         if (ret > 0 && dynstate) {
2014                 state = ret;
2015                 ret = 0;
2016         }
2017
2018         if (ret || !invoke || !startup)
2019                 goto out;
2020
2021         /*
2022          * Try to call the startup callback for each present cpu
2023          * depending on the hotplug state of the cpu.
2024          */
2025         for_each_present_cpu(cpu) {
2026                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2027                 int cpustate = st->state;
2028
2029                 if (cpustate < state)
2030                         continue;
2031
2032                 ret = cpuhp_issue_call(cpu, state, true, NULL);
2033                 if (ret) {
2034                         if (teardown)
2035                                 cpuhp_rollback_install(cpu, state, NULL);
2036                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2037                         goto out;
2038                 }
2039         }
2040 out:
2041         mutex_unlock(&cpuhp_state_mutex);
2042         /*
2043          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
2044          * dynamically allocated state in case of success.
2045          */
2046         if (!ret && dynstate)
2047                 return state;
2048         return ret;
2049 }
2050 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2051
2052 int __cpuhp_setup_state(enum cpuhp_state state,
2053                         const char *name, bool invoke,
2054                         int (*startup)(unsigned int cpu),
2055                         int (*teardown)(unsigned int cpu),
2056                         bool multi_instance)
2057 {
2058         int ret;
2059
2060         cpus_read_lock();
2061         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2062                                              teardown, multi_instance);
2063         cpus_read_unlock();
2064         return ret;
2065 }
2066 EXPORT_SYMBOL(__cpuhp_setup_state);
2067
2068 int __cpuhp_state_remove_instance(enum cpuhp_state state,
2069                                   struct hlist_node *node, bool invoke)
2070 {
2071         struct cpuhp_step *sp = cpuhp_get_step(state);
2072         int cpu;
2073
2074         BUG_ON(cpuhp_cb_check(state));
2075
2076         if (!sp->multi_instance)
2077                 return -EINVAL;
2078
2079         cpus_read_lock();
2080         mutex_lock(&cpuhp_state_mutex);
2081
2082         if (!invoke || !cpuhp_get_teardown_cb(state))
2083                 goto remove;
2084         /*
2085          * Call the teardown callback for each present cpu depending
2086          * on the hotplug state of the cpu. This function is not
2087          * allowed to fail currently!
2088          */
2089         for_each_present_cpu(cpu) {
2090                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2091                 int cpustate = st->state;
2092
2093                 if (cpustate >= state)
2094                         cpuhp_issue_call(cpu, state, false, node);
2095         }
2096
2097 remove:
2098         hlist_del(node);
2099         mutex_unlock(&cpuhp_state_mutex);
2100         cpus_read_unlock();
2101
2102         return 0;
2103 }
2104 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2105
2106 /**
2107  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2108  * @state:      The state to remove
2109  * @invoke:     If true, the teardown function is invoked for cpus where
2110  *              cpu state >= @state
2111  *
2112  * The caller needs to hold cpus read locked while calling this function.
2113  * The teardown callback is currently not allowed to fail. Think
2114  * about module removal!
2115  */
2116 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2117 {
2118         struct cpuhp_step *sp = cpuhp_get_step(state);
2119         int cpu;
2120
2121         BUG_ON(cpuhp_cb_check(state));
2122
2123         lockdep_assert_cpus_held();
2124
2125         mutex_lock(&cpuhp_state_mutex);
2126         if (sp->multi_instance) {
2127                 WARN(!hlist_empty(&sp->list),
2128                      "Error: Removing state %d which has instances left.\n",
2129                      state);
2130                 goto remove;
2131         }
2132
2133         if (!invoke || !cpuhp_get_teardown_cb(state))
2134                 goto remove;
2135
2136         /*
2137          * Call the teardown callback for each present cpu depending
2138          * on the hotplug state of the cpu. This function is not
2139          * allowed to fail currently!
2140          */
2141         for_each_present_cpu(cpu) {
2142                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2143                 int cpustate = st->state;
2144
2145                 if (cpustate >= state)
2146                         cpuhp_issue_call(cpu, state, false, NULL);
2147         }
2148 remove:
2149         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2150         mutex_unlock(&cpuhp_state_mutex);
2151 }
2152 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2153
2154 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2155 {
2156         cpus_read_lock();
2157         __cpuhp_remove_state_cpuslocked(state, invoke);
2158         cpus_read_unlock();
2159 }
2160 EXPORT_SYMBOL(__cpuhp_remove_state);
2161
2162 #ifdef CONFIG_HOTPLUG_SMT
2163 static void cpuhp_offline_cpu_device(unsigned int cpu)
2164 {
2165         struct device *dev = get_cpu_device(cpu);
2166
2167         dev->offline = true;
2168         /* Tell user space about the state change */
2169         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2170 }
2171
2172 static void cpuhp_online_cpu_device(unsigned int cpu)
2173 {
2174         struct device *dev = get_cpu_device(cpu);
2175
2176         dev->offline = false;
2177         /* Tell user space about the state change */
2178         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2179 }
2180
2181 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2182 {
2183         int cpu, ret = 0;
2184
2185         cpu_maps_update_begin();
2186         for_each_online_cpu(cpu) {
2187                 if (topology_is_primary_thread(cpu))
2188                         continue;
2189                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2190                 if (ret)
2191                         break;
2192                 /*
2193                  * As this needs to hold the cpu maps lock it's impossible
2194                  * to call device_offline() because that ends up calling
2195                  * cpu_down() which takes cpu maps lock. cpu maps lock
2196                  * needs to be held as this might race against in kernel
2197                  * abusers of the hotplug machinery (thermal management).
2198                  *
2199                  * So nothing would update device:offline state. That would
2200                  * leave the sysfs entry stale and prevent onlining after
2201                  * smt control has been changed to 'off' again. This is
2202                  * called under the sysfs hotplug lock, so it is properly
2203                  * serialized against the regular offline usage.
2204                  */
2205                 cpuhp_offline_cpu_device(cpu);
2206         }
2207         if (!ret)
2208                 cpu_smt_control = ctrlval;
2209         cpu_maps_update_done();
2210         return ret;
2211 }
2212
2213 int cpuhp_smt_enable(void)
2214 {
2215         int cpu, ret = 0;
2216
2217         cpu_maps_update_begin();
2218         cpu_smt_control = CPU_SMT_ENABLED;
2219         for_each_present_cpu(cpu) {
2220                 /* Skip online CPUs and CPUs on offline nodes */
2221                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2222                         continue;
2223                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2224                 if (ret)
2225                         break;
2226                 /* See comment in cpuhp_smt_disable() */
2227                 cpuhp_online_cpu_device(cpu);
2228         }
2229         cpu_maps_update_done();
2230         return ret;
2231 }
2232 #endif
2233
2234 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2235 static ssize_t show_cpuhp_state(struct device *dev,
2236                                 struct device_attribute *attr, char *buf)
2237 {
2238         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2239
2240         return sprintf(buf, "%d\n", st->state);
2241 }
2242 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
2243
2244 static ssize_t write_cpuhp_target(struct device *dev,
2245                                   struct device_attribute *attr,
2246                                   const char *buf, size_t count)
2247 {
2248         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2249         struct cpuhp_step *sp;
2250         int target, ret;
2251
2252         ret = kstrtoint(buf, 10, &target);
2253         if (ret)
2254                 return ret;
2255
2256 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2257         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2258                 return -EINVAL;
2259 #else
2260         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2261                 return -EINVAL;
2262 #endif
2263
2264         ret = lock_device_hotplug_sysfs();
2265         if (ret)
2266                 return ret;
2267
2268         mutex_lock(&cpuhp_state_mutex);
2269         sp = cpuhp_get_step(target);
2270         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2271         mutex_unlock(&cpuhp_state_mutex);
2272         if (ret)
2273                 goto out;
2274
2275         if (st->state < target)
2276                 ret = cpu_up(dev->id, target);
2277         else
2278                 ret = cpu_down(dev->id, target);
2279 out:
2280         unlock_device_hotplug();
2281         return ret ? ret : count;
2282 }
2283
2284 static ssize_t show_cpuhp_target(struct device *dev,
2285                                  struct device_attribute *attr, char *buf)
2286 {
2287         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2288
2289         return sprintf(buf, "%d\n", st->target);
2290 }
2291 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
2292
2293
2294 static ssize_t write_cpuhp_fail(struct device *dev,
2295                                 struct device_attribute *attr,
2296                                 const char *buf, size_t count)
2297 {
2298         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2299         struct cpuhp_step *sp;
2300         int fail, ret;
2301
2302         ret = kstrtoint(buf, 10, &fail);
2303         if (ret)
2304                 return ret;
2305
2306         if (fail == CPUHP_INVALID) {
2307                 st->fail = fail;
2308                 return count;
2309         }
2310
2311         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2312                 return -EINVAL;
2313
2314         /*
2315          * Cannot fail STARTING/DYING callbacks.
2316          */
2317         if (cpuhp_is_atomic_state(fail))
2318                 return -EINVAL;
2319
2320         /*
2321          * DEAD callbacks cannot fail...
2322          * ... neither can CPUHP_BRINGUP_CPU during hotunplug. The latter
2323          * triggering STARTING callbacks, a failure in this state would
2324          * hinder rollback.
2325          */
2326         if (fail <= CPUHP_BRINGUP_CPU && st->state > CPUHP_BRINGUP_CPU)
2327                 return -EINVAL;
2328
2329         /*
2330          * Cannot fail anything that doesn't have callbacks.
2331          */
2332         mutex_lock(&cpuhp_state_mutex);
2333         sp = cpuhp_get_step(fail);
2334         if (!sp->startup.single && !sp->teardown.single)
2335                 ret = -EINVAL;
2336         mutex_unlock(&cpuhp_state_mutex);
2337         if (ret)
2338                 return ret;
2339
2340         st->fail = fail;
2341
2342         return count;
2343 }
2344
2345 static ssize_t show_cpuhp_fail(struct device *dev,
2346                                struct device_attribute *attr, char *buf)
2347 {
2348         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2349
2350         return sprintf(buf, "%d\n", st->fail);
2351 }
2352
2353 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2354
2355 static struct attribute *cpuhp_cpu_attrs[] = {
2356         &dev_attr_state.attr,
2357         &dev_attr_target.attr,
2358         &dev_attr_fail.attr,
2359         NULL
2360 };
2361
2362 static const struct attribute_group cpuhp_cpu_attr_group = {
2363         .attrs = cpuhp_cpu_attrs,
2364         .name = "hotplug",
2365         NULL
2366 };
2367
2368 static ssize_t show_cpuhp_states(struct device *dev,
2369                                  struct device_attribute *attr, char *buf)
2370 {
2371         ssize_t cur, res = 0;
2372         int i;
2373
2374         mutex_lock(&cpuhp_state_mutex);
2375         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2376                 struct cpuhp_step *sp = cpuhp_get_step(i);
2377
2378                 if (sp->name) {
2379                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2380                         buf += cur;
2381                         res += cur;
2382                 }
2383         }
2384         mutex_unlock(&cpuhp_state_mutex);
2385         return res;
2386 }
2387 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2388
2389 static struct attribute *cpuhp_cpu_root_attrs[] = {
2390         &dev_attr_states.attr,
2391         NULL
2392 };
2393
2394 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2395         .attrs = cpuhp_cpu_root_attrs,
2396         .name = "hotplug",
2397         NULL
2398 };
2399
2400 #ifdef CONFIG_HOTPLUG_SMT
2401
2402 static ssize_t
2403 __store_smt_control(struct device *dev, struct device_attribute *attr,
2404                     const char *buf, size_t count)
2405 {
2406         int ctrlval, ret;
2407
2408         if (sysfs_streq(buf, "on"))
2409                 ctrlval = CPU_SMT_ENABLED;
2410         else if (sysfs_streq(buf, "off"))
2411                 ctrlval = CPU_SMT_DISABLED;
2412         else if (sysfs_streq(buf, "forceoff"))
2413                 ctrlval = CPU_SMT_FORCE_DISABLED;
2414         else
2415                 return -EINVAL;
2416
2417         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2418                 return -EPERM;
2419
2420         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2421                 return -ENODEV;
2422
2423         ret = lock_device_hotplug_sysfs();
2424         if (ret)
2425                 return ret;
2426
2427         if (ctrlval != cpu_smt_control) {
2428                 switch (ctrlval) {
2429                 case CPU_SMT_ENABLED:
2430                         ret = cpuhp_smt_enable();
2431                         break;
2432                 case CPU_SMT_DISABLED:
2433                 case CPU_SMT_FORCE_DISABLED:
2434                         ret = cpuhp_smt_disable(ctrlval);
2435                         break;
2436                 }
2437         }
2438
2439         unlock_device_hotplug();
2440         return ret ? ret : count;
2441 }
2442
2443 #else /* !CONFIG_HOTPLUG_SMT */
2444 static ssize_t
2445 __store_smt_control(struct device *dev, struct device_attribute *attr,
2446                     const char *buf, size_t count)
2447 {
2448         return -ENODEV;
2449 }
2450 #endif /* CONFIG_HOTPLUG_SMT */
2451
2452 static const char *smt_states[] = {
2453         [CPU_SMT_ENABLED]               = "on",
2454         [CPU_SMT_DISABLED]              = "off",
2455         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2456         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2457         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2458 };
2459
2460 static ssize_t
2461 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2462 {
2463         const char *state = smt_states[cpu_smt_control];
2464
2465         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2466 }
2467
2468 static ssize_t
2469 store_smt_control(struct device *dev, struct device_attribute *attr,
2470                   const char *buf, size_t count)
2471 {
2472         return __store_smt_control(dev, attr, buf, count);
2473 }
2474 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2475
2476 static ssize_t
2477 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2478 {
2479         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2480 }
2481 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2482
2483 static struct attribute *cpuhp_smt_attrs[] = {
2484         &dev_attr_control.attr,
2485         &dev_attr_active.attr,
2486         NULL
2487 };
2488
2489 static const struct attribute_group cpuhp_smt_attr_group = {
2490         .attrs = cpuhp_smt_attrs,
2491         .name = "smt",
2492         NULL
2493 };
2494
2495 static int __init cpu_smt_sysfs_init(void)
2496 {
2497         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2498                                   &cpuhp_smt_attr_group);
2499 }
2500
2501 static int __init cpuhp_sysfs_init(void)
2502 {
2503         int cpu, ret;
2504
2505         ret = cpu_smt_sysfs_init();
2506         if (ret)
2507                 return ret;
2508
2509         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2510                                  &cpuhp_cpu_root_attr_group);
2511         if (ret)
2512                 return ret;
2513
2514         for_each_possible_cpu(cpu) {
2515                 struct device *dev = get_cpu_device(cpu);
2516
2517                 if (!dev)
2518                         continue;
2519                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2520                 if (ret)
2521                         return ret;
2522         }
2523         return 0;
2524 }
2525 device_initcall(cpuhp_sysfs_init);
2526 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2527
2528 /*
2529  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2530  * represents all NR_CPUS bits binary values of 1<<nr.
2531  *
2532  * It is used by cpumask_of() to get a constant address to a CPU
2533  * mask value that has a single bit set only.
2534  */
2535
2536 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2537 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2538 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2539 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2540 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2541
2542 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2543
2544         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2545         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2546 #if BITS_PER_LONG > 32
2547         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2548         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2549 #endif
2550 };
2551 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2552
2553 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2554 EXPORT_SYMBOL(cpu_all_bits);
2555
2556 #ifdef CONFIG_INIT_ALL_POSSIBLE
2557 struct cpumask __cpu_possible_mask __read_mostly
2558         = {CPU_BITS_ALL};
2559 #else
2560 struct cpumask __cpu_possible_mask __read_mostly;
2561 #endif
2562 EXPORT_SYMBOL(__cpu_possible_mask);
2563
2564 struct cpumask __cpu_online_mask __read_mostly;
2565 EXPORT_SYMBOL(__cpu_online_mask);
2566
2567 struct cpumask __cpu_present_mask __read_mostly;
2568 EXPORT_SYMBOL(__cpu_present_mask);
2569
2570 struct cpumask __cpu_active_mask __read_mostly;
2571 EXPORT_SYMBOL(__cpu_active_mask);
2572
2573 struct cpumask __cpu_dying_mask __read_mostly;
2574 EXPORT_SYMBOL(__cpu_dying_mask);
2575
2576 atomic_t __num_online_cpus __read_mostly;
2577 EXPORT_SYMBOL(__num_online_cpus);
2578
2579 void init_cpu_present(const struct cpumask *src)
2580 {
2581         cpumask_copy(&__cpu_present_mask, src);
2582 }
2583
2584 void init_cpu_possible(const struct cpumask *src)
2585 {
2586         cpumask_copy(&__cpu_possible_mask, src);
2587 }
2588
2589 void init_cpu_online(const struct cpumask *src)
2590 {
2591         cpumask_copy(&__cpu_online_mask, src);
2592 }
2593
2594 void set_cpu_online(unsigned int cpu, bool online)
2595 {
2596         /*
2597          * atomic_inc/dec() is required to handle the horrid abuse of this
2598          * function by the reboot and kexec code which invoke it from
2599          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2600          * regular CPU hotplug is properly serialized.
2601          *
2602          * Note, that the fact that __num_online_cpus is of type atomic_t
2603          * does not protect readers which are not serialized against
2604          * concurrent hotplug operations.
2605          */
2606         if (online) {
2607                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2608                         atomic_inc(&__num_online_cpus);
2609         } else {
2610                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2611                         atomic_dec(&__num_online_cpus);
2612         }
2613 }
2614
2615 /*
2616  * Activate the first processor.
2617  */
2618 void __init boot_cpu_init(void)
2619 {
2620         int cpu = smp_processor_id();
2621
2622         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2623         set_cpu_online(cpu, true);
2624         set_cpu_active(cpu, true);
2625         set_cpu_present(cpu, true);
2626         set_cpu_possible(cpu, true);
2627
2628 #ifdef CONFIG_SMP
2629         __boot_cpu_id = cpu;
2630 #endif
2631 }
2632
2633 /*
2634  * Must be called _AFTER_ setting up the per_cpu areas
2635  */
2636 void __init boot_cpu_hotplug_init(void)
2637 {
2638 #ifdef CONFIG_SMP
2639         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2640 #endif
2641         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2642 }
2643
2644 /*
2645  * These are used for a global "mitigations=" cmdline option for toggling
2646  * optional CPU mitigations.
2647  */
2648 enum cpu_mitigations {
2649         CPU_MITIGATIONS_OFF,
2650         CPU_MITIGATIONS_AUTO,
2651         CPU_MITIGATIONS_AUTO_NOSMT,
2652 };
2653
2654 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2655         CPU_MITIGATIONS_AUTO;
2656
2657 static int __init mitigations_parse_cmdline(char *arg)
2658 {
2659         if (!strcmp(arg, "off"))
2660                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2661         else if (!strcmp(arg, "auto"))
2662                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2663         else if (!strcmp(arg, "auto,nosmt"))
2664                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2665         else
2666                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2667                         arg);
2668
2669         return 0;
2670 }
2671 early_param("mitigations", mitigations_parse_cmdline);
2672
2673 /* mitigations=off */
2674 bool cpu_mitigations_off(void)
2675 {
2676         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2677 }
2678 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2679
2680 /* mitigations=auto,nosmt */
2681 bool cpu_mitigations_auto_nosmt(void)
2682 {
2683         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2684 }
2685 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);