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