kernel/cpu.c

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