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/unistd.h>
  14 #include <linux/cpu.h>
  15 #include <linux/oom.h>
  16 #include <linux/rcupdate.h>
  17 #include <linux/export.h>
  18 #include <linux/bug.h>
  19 #include <linux/kthread.h>
  20 #include <linux/stop_machine.h>
  21 #include <linux/mutex.h>
  22 #include <linux/gfp.h>
  23 #include <linux/suspend.h>
  24 #include <linux/lockdep.h>
  25 #include <linux/tick.h>
  26 #include <linux/irq.h>
  27 #include <linux/nmi.h>
  28 #include <linux/smpboot.h>
  29 #include <linux/relay.h>
  30 #include <linux/slab.h>
  31 #include <linux/percpu-rwsem.h>
  32
  33 #include <trace/events/power.h>
  34 #define CREATE_TRACE_POINTS
  35 #include <trace/events/cpuhp.h>
  36
  37 #include "smpboot.h"
  38
  39 /**
  40  * cpuhp_cpu_state - Per cpu hotplug state storage
  41  * @state:      The current cpu state
  42  * @target:     The target state
  43  * @thread:     Pointer to the hotplug thread
  44  * @should_run: Thread should execute
  45  * @rollback:   Perform a rollback
  46  * @single:     Single callback invocation
  47  * @bringup:    Single callback bringup or teardown selector
  48  * @cb_state:   The state for a single callback (install/uninstall)
  49  * @result:     Result of the operation
  50  * @done_up:    Signal completion to the issuer of the task for cpu-up
  51  * @done_down:  Signal completion to the issuer of the task for cpu-down
  52  */
  53 struct cpuhp_cpu_state {
  54         enum cpuhp_state        state;
  55         enum cpuhp_state        target;
  56         enum cpuhp_state        fail;
  57 #ifdef CONFIG_SMP
  58         struct task_struct      *thread;
  59         bool                    should_run;
  60         bool                    rollback;
  61         bool                    single;
  62         bool                    bringup;
  63         bool                    booted_once;
  64         struct hlist_node       *node;
  65         struct hlist_node       *last;
  66         enum cpuhp_state        cb_state;
  67         int                     result;
  68         struct completion       done_up;
  69         struct completion       done_down;
  70 #endif
  71 };
  72
  73 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  74         .fail = CPUHP_INVALID,
  75 };
  76
  77 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  78 static struct lockdep_map cpuhp_state_up_map =
  79         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  80 static struct lockdep_map cpuhp_state_down_map =
  81         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  82
  83
  84 static inline void cpuhp_lock_acquire(bool bringup)
  85 {
  86         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  87 }
  88
  89 static inline void cpuhp_lock_release(bool bringup)
  90 {
  91         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  92 }
  93 #else
  94
  95 static inline void cpuhp_lock_acquire(bool bringup) { }
  96 static inline void cpuhp_lock_release(bool bringup) { }
  97
  98 #endif
  99
 100 /**
 101  * cpuhp_step - Hotplug state machine step
 102  * @name:       Name of the step
 103  * @startup:    Startup function of the step
 104  * @teardown:   Teardown function of the step
 105  * @skip_onerr: Do not invoke the functions on error rollback
 106  *              Will go away once the notifiers are gone
 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                    skip_onerr;
 123         bool                    cant_stop;
 124         bool                    multi_instance;
 125 };
 126
 127 static DEFINE_MUTEX(cpuhp_state_mutex);
 128 static struct cpuhp_step cpuhp_hp_states[];
 129
 130 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 131 {
 132         return cpuhp_hp_states + state;
 133 }
 134
 135 /**
 136  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 137  * @cpu:        The cpu for which the callback should be invoked
 138  * @state:      The state to do callbacks for
 139  * @bringup:    True if the bringup callback should be invoked
 140  * @node:       For multi-instance, do a single entry callback for install/remove
 141  * @lastp:      For multi-instance rollback, remember how far we got
 142  *
 143  * Called from cpu hotplug and from the state register machinery.
 144  */
 145 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 146                                  bool bringup, struct hlist_node *node,
 147                                  struct hlist_node **lastp)
 148 {
 149         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 150         struct cpuhp_step *step = cpuhp_get_step(state);
 151         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 152         int (*cb)(unsigned int cpu);
 153         int ret, cnt;
 154
 155         if (st->fail == state) {
 156                 st->fail = CPUHP_INVALID;
 157
 158                 if (!(bringup ? step->startup.single : step->teardown.single))
 159                         return 0;
 160
 161                 return -EAGAIN;
 162         }
 163
 164         if (!step->multi_instance) {
 165                 WARN_ON_ONCE(lastp && *lastp);
 166                 cb = bringup ? step->startup.single : step->teardown.single;
 167                 if (!cb)
 168                         return 0;
 169                 trace_cpuhp_enter(cpu, st->target, state, cb);
 170                 ret = cb(cpu);
 171                 trace_cpuhp_exit(cpu, st->state, state, ret);
 172                 return ret;
 173         }
 174         cbm = bringup ? step->startup.multi : step->teardown.multi;
 175         if (!cbm)
 176                 return 0;
 177
 178         /* Single invocation for instance add/remove */
 179         if (node) {
 180                 WARN_ON_ONCE(lastp && *lastp);
 181                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 182                 ret = cbm(cpu, node);
 183                 trace_cpuhp_exit(cpu, st->state, state, ret);
 184                 return ret;
 185         }
 186
 187         /* State transition. Invoke on all instances */
 188         cnt = 0;
 189         hlist_for_each(node, &step->list) {
 190                 if (lastp && node == *lastp)
 191                         break;
 192
 193                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 194                 ret = cbm(cpu, node);
 195                 trace_cpuhp_exit(cpu, st->state, state, ret);
 196                 if (ret) {
 197                         if (!lastp)
 198                                 goto err;
 199
 200                         *lastp = node;
 201                         return ret;
 202                 }
 203                 cnt++;
 204         }
 205         if (lastp)
 206                 *lastp = NULL;
 207         return 0;
 208 err:
 209         /* Rollback the instances if one failed */
 210         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 211         if (!cbm)
 212                 return ret;
 213
 214         hlist_for_each(node, &step->list) {
 215                 if (!cnt--)
 216                         break;
 217
 218                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 219                 ret = cbm(cpu, node);
 220                 trace_cpuhp_exit(cpu, st->state, state, ret);
 221                 /*
 222                  * Rollback must not fail,
 223                  */
 224                 WARN_ON_ONCE(ret);
 225         }
 226         return ret;
 227 }
 228
 229 #ifdef CONFIG_SMP
 230 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 231 {
 232         /*
 233          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 234          * purposes as that state is handled explicitly in cpu_down.
 235          */
 236         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 237 }
 238
 239 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 240 {
 241         struct completion *done = bringup ? &st->done_up : &st->done_down;
 242         wait_for_completion(done);
 243 }
 244
 245 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 246 {
 247         struct completion *done = bringup ? &st->done_up : &st->done_down;
 248         complete(done);
 249 }
 250
 251 /*
 252  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 253  */
 254 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 255 {
 256         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 257 }
 258
 259 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 260 static DEFINE_MUTEX(cpu_add_remove_lock);
 261 bool cpuhp_tasks_frozen;
 262 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 263
 264 /*
 265  * The following two APIs (cpu_maps_update_begin/done) must be used when
 266  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 267  */
 268 void cpu_maps_update_begin(void)
 269 {
 270         mutex_lock(&cpu_add_remove_lock);
 271 }
 272
 273 void cpu_maps_update_done(void)
 274 {
 275         mutex_unlock(&cpu_add_remove_lock);
 276 }
 277
 278 /*
 279  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 280  * Should always be manipulated under cpu_add_remove_lock
 281  */
 282 static int cpu_hotplug_disabled;
 283
 284 #ifdef CONFIG_HOTPLUG_CPU
 285
 286 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 287
 288 void cpus_read_lock(void)
 289 {
 290         percpu_down_read(&cpu_hotplug_lock);
 291 }
 292 EXPORT_SYMBOL_GPL(cpus_read_lock);
 293
 294 void cpus_read_unlock(void)
 295 {
 296         percpu_up_read(&cpu_hotplug_lock);
 297 }
 298 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 299
 300 void cpus_write_lock(void)
 301 {
 302         percpu_down_write(&cpu_hotplug_lock);
 303 }
 304
 305 void cpus_write_unlock(void)
 306 {
 307         percpu_up_write(&cpu_hotplug_lock);
 308 }
 309
 310 void lockdep_assert_cpus_held(void)
 311 {
 312         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 313 }
 314
 315 /*
 316  * Wait for currently running CPU hotplug operations to complete (if any) and
 317  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 318  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 319  * hotplug path before performing hotplug operations. So acquiring that lock
 320  * guarantees mutual exclusion from any currently running hotplug operations.
 321  */
 322 void cpu_hotplug_disable(void)
 323 {
 324         cpu_maps_update_begin();
 325         cpu_hotplug_disabled++;
 326         cpu_maps_update_done();
 327 }
 328 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 329
 330 static void __cpu_hotplug_enable(void)
 331 {
 332         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 333                 return;
 334         cpu_hotplug_disabled--;
 335 }
 336
 337 void cpu_hotplug_enable(void)
 338 {
 339         cpu_maps_update_begin();
 340         __cpu_hotplug_enable();
 341         cpu_maps_update_done();
 342 }
 343 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 344 #endif  /* CONFIG_HOTPLUG_CPU */
 345
 346 #ifdef CONFIG_HOTPLUG_SMT
 347 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 348 EXPORT_SYMBOL_GPL(cpu_smt_control);
 349
 350 static bool cpu_smt_available __read_mostly;
 351
 352 void __init cpu_smt_disable(bool force)
 353 {
 354         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 355                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 356                 return;
 357
 358         if (force) {
 359                 pr_info("SMT: Force disabled\n");
 360                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 361         } else {
 362                 cpu_smt_control = CPU_SMT_DISABLED;
 363         }
 364 }
 365
 366 /*
 367  * The decision whether SMT is supported can only be done after the full
 368  * CPU identification. Called from architecture code before non boot CPUs
 369  * are brought up.
 370  */
 371 void __init cpu_smt_check_topology_early(void)
 372 {
 373         if (!topology_smt_supported())
 374                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 375 }
 376
 377 /*
 378  * If SMT was disabled by BIOS, detect it here, after the CPUs have been
 379  * brought online. This ensures the smt/l1tf sysfs entries are consistent
 380  * with reality. cpu_smt_available is set to true during the bringup of non
 381  * boot CPUs when a SMT sibling is detected. Note, this may overwrite
 382  * cpu_smt_control's previous setting.
 383  */
 384 void __init cpu_smt_check_topology(void)
 385 {
 386         if (!cpu_smt_available)
 387                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 388 }
 389
 390 static int __init smt_cmdline_disable(char *str)
 391 {
 392         cpu_smt_disable(str && !strcmp(str, "force"));
 393         return 0;
 394 }
 395 early_param("nosmt", smt_cmdline_disable);
 396
 397 static inline bool cpu_smt_allowed(unsigned int cpu)
 398 {
 399         if (topology_is_primary_thread(cpu))
 400                 return true;
 401
 402         /*
 403          * If the CPU is not a 'primary' thread and the booted_once bit is
 404          * set then the processor has SMT support. Store this information
 405          * for the late check of SMT support in cpu_smt_check_topology().
 406          */
 407         if (per_cpu(cpuhp_state, cpu).booted_once)
 408                 cpu_smt_available = true;
 409
 410         if (cpu_smt_control == CPU_SMT_ENABLED)
 411                 return true;
 412
 413         /*
 414          * On x86 it's required to boot all logical CPUs at least once so
 415          * that the init code can get a chance to set CR4.MCE on each
 416          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 417          * core will shutdown the machine.
 418          */
 419         return !per_cpu(cpuhp_state, cpu).booted_once;
 420 }
 421 #else
 422 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 423 #endif
 424
 425 static inline enum cpuhp_state
 426 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 427 {
 428         enum cpuhp_state prev_state = st->state;
 429
 430         st->rollback = false;
 431         st->last = NULL;
 432
 433         st->target = target;
 434         st->single = false;
 435         st->bringup = st->state < target;
 436
 437         return prev_state;
 438 }
 439
 440 static inline void
 441 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 442 {
 443         st->rollback = true;
 444
 445         /*
 446          * If we have st->last we need to undo partial multi_instance of this
 447          * state first. Otherwise start undo at the previous state.
 448          */
 449         if (!st->last) {
 450                 if (st->bringup)
 451                         st->state--;
 452                 else
 453                         st->state++;
 454         }
 455
 456         st->target = prev_state;
 457         st->bringup = !st->bringup;
 458 }
 459
 460 /* Regular hotplug invocation of the AP hotplug thread */
 461 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 462 {
 463         if (!st->single && st->state == st->target)
 464                 return;
 465
 466         st->result = 0;
 467         /*
 468          * Make sure the above stores are visible before should_run becomes
 469          * true. Paired with the mb() above in cpuhp_thread_fun()
 470          */
 471         smp_mb();
 472         st->should_run = true;
 473         wake_up_process(st->thread);
 474         wait_for_ap_thread(st, st->bringup);
 475 }
 476
 477 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 478 {
 479         enum cpuhp_state prev_state;
 480         int ret;
 481
 482         prev_state = cpuhp_set_state(st, target);
 483         __cpuhp_kick_ap(st);
 484         if ((ret = st->result)) {
 485                 cpuhp_reset_state(st, prev_state);
 486                 __cpuhp_kick_ap(st);
 487         }
 488
 489         return ret;
 490 }
 491
 492 static int bringup_wait_for_ap(unsigned int cpu)
 493 {
 494         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 495
 496         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 497         wait_for_ap_thread(st, true);
 498         if (WARN_ON_ONCE((!cpu_online(cpu))))
 499                 return -ECANCELED;
 500
 501         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 502         stop_machine_unpark(cpu);
 503         kthread_unpark(st->thread);
 504
 505         /*
 506          * SMT soft disabling on X86 requires to bring the CPU out of the
 507          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 508          * CPU marked itself as booted_once in cpu_notify_starting() so the
 509          * cpu_smt_allowed() check will now return false if this is not the
 510          * primary sibling.
 511          */
 512         if (!cpu_smt_allowed(cpu))
 513                 return -ECANCELED;
 514
 515         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 516                 return 0;
 517
 518         return cpuhp_kick_ap(st, st->target);
 519 }
 520
 521 static int bringup_cpu(unsigned int cpu)
 522 {
 523         struct task_struct *idle = idle_thread_get(cpu);
 524         int ret;
 525
 526         /*
 527          * Some architectures have to walk the irq descriptors to
 528          * setup the vector space for the cpu which comes online.
 529          * Prevent irq alloc/free across the bringup.
 530          */
 531         irq_lock_sparse();
 532
 533         /* Arch-specific enabling code. */
 534         ret = __cpu_up(cpu, idle);
 535         irq_unlock_sparse();
 536         if (ret)
 537                 return ret;
 538         return bringup_wait_for_ap(cpu);
 539 }
 540
 541 /*
 542  * Hotplug state machine related functions
 543  */
 544
 545 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 546 {
 547         for (st->state--; st->state > st->target; st->state--) {
 548                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 549
 550                 if (!step->skip_onerr)
 551                         cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 552         }
 553 }
 554
 555 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 556                               enum cpuhp_state target)
 557 {
 558         enum cpuhp_state prev_state = st->state;
 559         int ret = 0;
 560
 561         while (st->state < target) {
 562                 st->state++;
 563                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 564                 if (ret) {
 565                         st->target = prev_state;
 566                         undo_cpu_up(cpu, st);
 567                         break;
 568                 }
 569         }
 570         return ret;
 571 }
 572
 573 /*
 574  * The cpu hotplug threads manage the bringup and teardown of the cpus
 575  */
 576 static void cpuhp_create(unsigned int cpu)
 577 {
 578         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 579
 580         init_completion(&st->done_up);
 581         init_completion(&st->done_down);
 582 }
 583
 584 static int cpuhp_should_run(unsigned int cpu)
 585 {
 586         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 587
 588         return st->should_run;
 589 }
 590
 591 /*
 592  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 593  * callbacks when a state gets [un]installed at runtime.
 594  *
 595  * Each invocation of this function by the smpboot thread does a single AP
 596  * state callback.
 597  *
 598  * It has 3 modes of operation:
 599  *  - single: runs st->cb_state
 600  *  - up:     runs ++st->state, while st->state < st->target
 601  *  - down:   runs st->state--, while st->state > st->target
 602  *
 603  * When complete or on error, should_run is cleared and the completion is fired.
 604  */
 605 static void cpuhp_thread_fun(unsigned int cpu)
 606 {
 607         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 608         bool bringup = st->bringup;
 609         enum cpuhp_state state;
 610
 611         /*
 612          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 613          * that if we see ->should_run we also see the rest of the state.
 614          */
 615         smp_mb();
 616
 617         if (WARN_ON_ONCE(!st->should_run))
 618                 return;
 619
 620         cpuhp_lock_acquire(bringup);
 621
 622         if (st->single) {
 623                 state = st->cb_state;
 624                 st->should_run = false;
 625         } else {
 626                 if (bringup) {
 627                         st->state++;
 628                         state = st->state;
 629                         st->should_run = (st->state < st->target);
 630                         WARN_ON_ONCE(st->state > st->target);
 631                 } else {
 632                         state = st->state;
 633                         st->state--;
 634                         st->should_run = (st->state > st->target);
 635                         WARN_ON_ONCE(st->state < st->target);
 636                 }
 637         }
 638
 639         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 640
 641         if (st->rollback) {
 642                 struct cpuhp_step *step = cpuhp_get_step(state);
 643                 if (step->skip_onerr)
 644                         goto next;
 645         }
 646
 647         if (cpuhp_is_atomic_state(state)) {
 648                 local_irq_disable();
 649                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 650                 local_irq_enable();
 651
 652                 /*
 653                  * STARTING/DYING must not fail!
 654                  */
 655                 WARN_ON_ONCE(st->result);
 656         } else {
 657                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 658         }
 659
 660         if (st->result) {
 661                 /*
 662                  * If we fail on a rollback, we're up a creek without no
 663                  * paddle, no way forward, no way back. We loose, thanks for
 664                  * playing.
 665                  */
 666                 WARN_ON_ONCE(st->rollback);
 667                 st->should_run = false;
 668         }
 669
 670 next:
 671         cpuhp_lock_release(bringup);
 672
 673         if (!st->should_run)
 674                 complete_ap_thread(st, bringup);
 675 }
 676
 677 /* Invoke a single callback on a remote cpu */
 678 static int
 679 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 680                          struct hlist_node *node)
 681 {
 682         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 683         int ret;
 684
 685         if (!cpu_online(cpu))
 686                 return 0;
 687
 688         cpuhp_lock_acquire(false);
 689         cpuhp_lock_release(false);
 690
 691         cpuhp_lock_acquire(true);
 692         cpuhp_lock_release(true);
 693
 694         /*
 695          * If we are up and running, use the hotplug thread. For early calls
 696          * we invoke the thread function directly.
 697          */
 698         if (!st->thread)
 699                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 700
 701         st->rollback = false;
 702         st->last = NULL;
 703
 704         st->node = node;
 705         st->bringup = bringup;
 706         st->cb_state = state;
 707         st->single = true;
 708
 709         __cpuhp_kick_ap(st);
 710
 711         /*
 712          * If we failed and did a partial, do a rollback.
 713          */
 714         if ((ret = st->result) && st->last) {
 715                 st->rollback = true;
 716                 st->bringup = !bringup;
 717
 718                 __cpuhp_kick_ap(st);
 719         }
 720
 721         /*
 722          * Clean up the leftovers so the next hotplug operation wont use stale
 723          * data.
 724          */
 725         st->node = st->last = NULL;
 726         return ret;
 727 }
 728
 729 static int cpuhp_kick_ap_work(unsigned int cpu)
 730 {
 731         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 732         enum cpuhp_state prev_state = st->state;
 733         int ret;
 734
 735         cpuhp_lock_acquire(false);
 736         cpuhp_lock_release(false);
 737
 738         cpuhp_lock_acquire(true);
 739         cpuhp_lock_release(true);
 740
 741         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 742         ret = cpuhp_kick_ap(st, st->target);
 743         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 744
 745         return ret;
 746 }
 747
 748 static struct smp_hotplug_thread cpuhp_threads = {
 749         .store                  = &cpuhp_state.thread,
 750         .create                 = &cpuhp_create,
 751         .thread_should_run      = cpuhp_should_run,
 752         .thread_fn              = cpuhp_thread_fun,
 753         .thread_comm            = "cpuhp/%u",
 754         .selfparking            = true,
 755 };
 756
 757 void __init cpuhp_threads_init(void)
 758 {
 759         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 760         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 761 }
 762
 763 #ifdef CONFIG_HOTPLUG_CPU
 764 /**
 765  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 766  * @cpu: a CPU id
 767  *
 768  * This function walks all processes, finds a valid mm struct for each one and
 769  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 770  * trivial, there are various non-obvious corner cases, which this function
 771  * tries to solve in a safe manner.
 772  *
 773  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 774  * be called only for an already offlined CPU.
 775  */
 776 void clear_tasks_mm_cpumask(int cpu)
 777 {
 778         struct task_struct *p;
 779
 780         /*
 781          * This function is called after the cpu is taken down and marked
 782          * offline, so its not like new tasks will ever get this cpu set in
 783          * their mm mask. -- Peter Zijlstra
 784          * Thus, we may use rcu_read_lock() here, instead of grabbing
 785          * full-fledged tasklist_lock.
 786          */
 787         WARN_ON(cpu_online(cpu));
 788         rcu_read_lock();
 789         for_each_process(p) {
 790                 struct task_struct *t;
 791
 792                 /*
 793                  * Main thread might exit, but other threads may still have
 794                  * a valid mm. Find one.
 795                  */
 796                 t = find_lock_task_mm(p);
 797                 if (!t)
 798                         continue;
 799                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 800                 task_unlock(t);
 801         }
 802         rcu_read_unlock();
 803 }
 804
 805 /* Take this CPU down. */
 806 static int take_cpu_down(void *_param)
 807 {
 808         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 809         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 810         int err, cpu = smp_processor_id();
 811         int ret;
 812
 813         /* Ensure this CPU doesn't handle any more interrupts. */
 814         err = __cpu_disable();
 815         if (err < 0)
 816                 return err;
 817
 818         /*
 819          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 820          * do this step again.
 821          */
 822         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 823         st->state--;
 824         /* Invoke the former CPU_DYING callbacks */
 825         for (; st->state > target; st->state--) {
 826                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 827                 /*
 828                  * DYING must not fail!
 829                  */
 830                 WARN_ON_ONCE(ret);
 831         }
 832
 833         /* Give up timekeeping duties */
 834         tick_handover_do_timer();
 835         /* Park the stopper thread */
 836         stop_machine_park(cpu);
 837         return 0;
 838 }
 839
 840 static int takedown_cpu(unsigned int cpu)
 841 {
 842         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 843         int err;
 844
 845         /* Park the smpboot threads */
 846         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 847
 848         /*
 849          * Prevent irq alloc/free while the dying cpu reorganizes the
 850          * interrupt affinities.
 851          */
 852         irq_lock_sparse();
 853
 854         /*
 855          * So now all preempt/rcu users must observe !cpu_active().
 856          */
 857         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 858         if (err) {
 859                 /* CPU refused to die */
 860                 irq_unlock_sparse();
 861                 /* Unpark the hotplug thread so we can rollback there */
 862                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 863                 return err;
 864         }
 865         BUG_ON(cpu_online(cpu));
 866
 867         /*
 868          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 869          * all runnable tasks from the CPU, there's only the idle task left now
 870          * that the migration thread is done doing the stop_machine thing.
 871          *
 872          * Wait for the stop thread to go away.
 873          */
 874         wait_for_ap_thread(st, false);
 875         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 876
 877         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 878         irq_unlock_sparse();
 879
 880         hotplug_cpu__broadcast_tick_pull(cpu);
 881         /* This actually kills the CPU. */
 882         __cpu_die(cpu);
 883
 884         tick_cleanup_dead_cpu(cpu);
 885         rcutree_migrate_callbacks(cpu);
 886         return 0;
 887 }
 888
 889 static void cpuhp_complete_idle_dead(void *arg)
 890 {
 891         struct cpuhp_cpu_state *st = arg;
 892
 893         complete_ap_thread(st, false);
 894 }
 895
 896 void cpuhp_report_idle_dead(void)
 897 {
 898         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 899
 900         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 901         rcu_report_dead(smp_processor_id());
 902         st->state = CPUHP_AP_IDLE_DEAD;
 903         /*
 904          * We cannot call complete after rcu_report_dead() so we delegate it
 905          * to an online cpu.
 906          */
 907         smp_call_function_single(cpumask_first(cpu_online_mask),
 908                                  cpuhp_complete_idle_dead, st, 0);
 909 }
 910
 911 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 912 {
 913         for (st->state++; st->state < st->target; st->state++) {
 914                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 915
 916                 if (!step->skip_onerr)
 917                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 918         }
 919 }
 920
 921 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 922                                 enum cpuhp_state target)
 923 {
 924         enum cpuhp_state prev_state = st->state;
 925         int ret = 0;
 926
 927         for (; st->state > target; st->state--) {
 928                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 929                 if (ret) {
 930                         st->target = 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         return ret;
 997 }
 998
 999 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1000 {
1001         if (cpu_hotplug_disabled)
1002                 return -EBUSY;
1003         return _cpu_down(cpu, 0, target);
1004 }
1005
1006 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1007 {
1008         int err;
1009
1010         cpu_maps_update_begin();
1011         err = cpu_down_maps_locked(cpu, target);
1012         cpu_maps_update_done();
1013         return err;
1014 }
1015
1016 int cpu_down(unsigned int cpu)
1017 {
1018         return do_cpu_down(cpu, CPUHP_OFFLINE);
1019 }
1020 EXPORT_SYMBOL(cpu_down);
1021
1022 #else
1023 #define takedown_cpu            NULL
1024 #endif /*CONFIG_HOTPLUG_CPU*/
1025
1026 /**
1027  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1028  * @cpu: cpu that just started
1029  *
1030  * It must be called by the arch code on the new cpu, before the new cpu
1031  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1032  */
1033 void notify_cpu_starting(unsigned int cpu)
1034 {
1035         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1036         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1037         int ret;
1038
1039         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1040         st->booted_once = true;
1041         while (st->state < target) {
1042                 st->state++;
1043                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1044                 /*
1045                  * STARTING must not fail!
1046                  */
1047                 WARN_ON_ONCE(ret);
1048         }
1049 }
1050
1051 /*
1052  * Called from the idle task. Wake up the controlling task which brings the
1053  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1054  * the rest of the online bringup to the hotplug thread.
1055  */
1056 void cpuhp_online_idle(enum cpuhp_state state)
1057 {
1058         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1059
1060         /* Happens for the boot cpu */
1061         if (state != CPUHP_AP_ONLINE_IDLE)
1062                 return;
1063
1064         st->state = CPUHP_AP_ONLINE_IDLE;
1065         complete_ap_thread(st, true);
1066 }
1067
1068 /* Requires cpu_add_remove_lock to be held */
1069 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1070 {
1071         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1072         struct task_struct *idle;
1073         int ret = 0;
1074
1075         cpus_write_lock();
1076
1077         if (!cpu_present(cpu)) {
1078                 ret = -EINVAL;
1079                 goto out;
1080         }
1081
1082         /*
1083          * The caller of do_cpu_up might have raced with another
1084          * caller. Ignore it for now.
1085          */
1086         if (st->state >= target)
1087                 goto out;
1088
1089         if (st->state == CPUHP_OFFLINE) {
1090                 /* Let it fail before we try to bring the cpu up */
1091                 idle = idle_thread_get(cpu);
1092                 if (IS_ERR(idle)) {
1093                         ret = PTR_ERR(idle);
1094                         goto out;
1095                 }
1096         }
1097
1098         cpuhp_tasks_frozen = tasks_frozen;
1099
1100         cpuhp_set_state(st, target);
1101         /*
1102          * If the current CPU state is in the range of the AP hotplug thread,
1103          * then we need to kick the thread once more.
1104          */
1105         if (st->state > CPUHP_BRINGUP_CPU) {
1106                 ret = cpuhp_kick_ap_work(cpu);
1107                 /*
1108                  * The AP side has done the error rollback already. Just
1109                  * return the error code..
1110                  */
1111                 if (ret)
1112                         goto out;
1113         }
1114
1115         /*
1116          * Try to reach the target state. We max out on the BP at
1117          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1118          * responsible for bringing it up to the target state.
1119          */
1120         target = min((int)target, CPUHP_BRINGUP_CPU);
1121         ret = cpuhp_up_callbacks(cpu, st, target);
1122 out:
1123         cpus_write_unlock();
1124         return ret;
1125 }
1126
1127 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1128 {
1129         int err = 0;
1130
1131         if (!cpu_possible(cpu)) {
1132                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1133                        cpu);
1134 #if defined(CONFIG_IA64)
1135                 pr_err("please check additional_cpus= boot parameter\n");
1136 #endif
1137                 return -EINVAL;
1138         }
1139
1140         err = try_online_node(cpu_to_node(cpu));
1141         if (err)
1142                 return err;
1143
1144         cpu_maps_update_begin();
1145
1146         if (cpu_hotplug_disabled) {
1147                 err = -EBUSY;
1148                 goto out;
1149         }
1150         if (!cpu_smt_allowed(cpu)) {
1151                 err = -EPERM;
1152                 goto out;
1153         }
1154
1155         err = _cpu_up(cpu, 0, target);
1156 out:
1157         cpu_maps_update_done();
1158         return err;
1159 }
1160
1161 int cpu_up(unsigned int cpu)
1162 {
1163         return do_cpu_up(cpu, CPUHP_ONLINE);
1164 }
1165 EXPORT_SYMBOL_GPL(cpu_up);
1166
1167 #ifdef CONFIG_PM_SLEEP_SMP
1168 static cpumask_var_t frozen_cpus;
1169
1170 int freeze_secondary_cpus(int primary)
1171 {
1172         int cpu, error = 0;
1173
1174         cpu_maps_update_begin();
1175         if (!cpu_online(primary))
1176                 primary = cpumask_first(cpu_online_mask);
1177         /*
1178          * We take down all of the non-boot CPUs in one shot to avoid races
1179          * with the userspace trying to use the CPU hotplug at the same time
1180          */
1181         cpumask_clear(frozen_cpus);
1182
1183         pr_info("Disabling non-boot CPUs ...\n");
1184         for_each_online_cpu(cpu) {
1185                 if (cpu == primary)
1186                         continue;
1187                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1188                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1189                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1190                 if (!error)
1191                         cpumask_set_cpu(cpu, frozen_cpus);
1192                 else {
1193                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1194                         break;
1195                 }
1196         }
1197
1198         if (!error)
1199                 BUG_ON(num_online_cpus() > 1);
1200         else
1201                 pr_err("Non-boot CPUs are not disabled\n");
1202
1203         /*
1204          * Make sure the CPUs won't be enabled by someone else. We need to do
1205          * this even in case of failure as all disable_nonboot_cpus() users are
1206          * supposed to do enable_nonboot_cpus() on the failure path.
1207          */
1208         cpu_hotplug_disabled++;
1209
1210         cpu_maps_update_done();
1211         return error;
1212 }
1213
1214 void __weak arch_enable_nonboot_cpus_begin(void)
1215 {
1216 }
1217
1218 void __weak arch_enable_nonboot_cpus_end(void)
1219 {
1220 }
1221
1222 void enable_nonboot_cpus(void)
1223 {
1224         int cpu, error;
1225
1226         /* Allow everyone to use the CPU hotplug again */
1227         cpu_maps_update_begin();
1228         __cpu_hotplug_enable();
1229         if (cpumask_empty(frozen_cpus))
1230                 goto out;
1231
1232         pr_info("Enabling non-boot CPUs ...\n");
1233
1234         arch_enable_nonboot_cpus_begin();
1235
1236         for_each_cpu(cpu, frozen_cpus) {
1237                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1238                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1239                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1240                 if (!error) {
1241                         pr_info("CPU%d is up\n", cpu);
1242                         continue;
1243                 }
1244                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1245         }
1246
1247         arch_enable_nonboot_cpus_end();
1248
1249         cpumask_clear(frozen_cpus);
1250 out:
1251         cpu_maps_update_done();
1252 }
1253
1254 static int __init alloc_frozen_cpus(void)
1255 {
1256         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1257                 return -ENOMEM;
1258         return 0;
1259 }
1260 core_initcall(alloc_frozen_cpus);
1261
1262 /*
1263  * When callbacks for CPU hotplug notifications are being executed, we must
1264  * ensure that the state of the system with respect to the tasks being frozen
1265  * or not, as reported by the notification, remains unchanged *throughout the
1266  * duration* of the execution of the callbacks.
1267  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1268  *
1269  * This synchronization is implemented by mutually excluding regular CPU
1270  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1271  * Hibernate notifications.
1272  */
1273 static int
1274 cpu_hotplug_pm_callback(struct notifier_block *nb,
1275                         unsigned long action, void *ptr)
1276 {
1277         switch (action) {
1278
1279         case PM_SUSPEND_PREPARE:
1280         case PM_HIBERNATION_PREPARE:
1281                 cpu_hotplug_disable();
1282                 break;
1283
1284         case PM_POST_SUSPEND:
1285         case PM_POST_HIBERNATION:
1286                 cpu_hotplug_enable();
1287                 break;
1288
1289         default:
1290                 return NOTIFY_DONE;
1291         }
1292
1293         return NOTIFY_OK;
1294 }
1295
1296
1297 static int __init cpu_hotplug_pm_sync_init(void)
1298 {
1299         /*
1300          * cpu_hotplug_pm_callback has higher priority than x86
1301          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1302          * to disable cpu hotplug to avoid cpu hotplug race.
1303          */
1304         pm_notifier(cpu_hotplug_pm_callback, 0);
1305         return 0;
1306 }
1307 core_initcall(cpu_hotplug_pm_sync_init);
1308
1309 #endif /* CONFIG_PM_SLEEP_SMP */
1310
1311 int __boot_cpu_id;
1312
1313 #endif /* CONFIG_SMP */
1314
1315 /* Boot processor state steps */
1316 static struct cpuhp_step cpuhp_hp_states[] = {
1317         [CPUHP_OFFLINE] = {
1318                 .name                   = "offline",
1319                 .startup.single         = NULL,
1320                 .teardown.single        = NULL,
1321         },
1322 #ifdef CONFIG_SMP
1323         [CPUHP_CREATE_THREADS]= {
1324                 .name                   = "threads:prepare",
1325                 .startup.single         = smpboot_create_threads,
1326                 .teardown.single        = NULL,
1327                 .cant_stop              = true,
1328         },
1329         [CPUHP_PERF_PREPARE] = {
1330                 .name                   = "perf:prepare",
1331                 .startup.single         = perf_event_init_cpu,
1332                 .teardown.single        = perf_event_exit_cpu,
1333         },
1334         [CPUHP_WORKQUEUE_PREP] = {
1335                 .name                   = "workqueue:prepare",
1336                 .startup.single         = workqueue_prepare_cpu,
1337                 .teardown.single        = NULL,
1338         },
1339         [CPUHP_HRTIMERS_PREPARE] = {
1340                 .name                   = "hrtimers:prepare",
1341                 .startup.single         = hrtimers_prepare_cpu,
1342                 .teardown.single        = hrtimers_dead_cpu,
1343         },
1344         [CPUHP_SMPCFD_PREPARE] = {
1345                 .name                   = "smpcfd:prepare",
1346                 .startup.single         = smpcfd_prepare_cpu,
1347                 .teardown.single        = smpcfd_dead_cpu,
1348         },
1349         [CPUHP_RELAY_PREPARE] = {
1350                 .name                   = "relay:prepare",
1351                 .startup.single         = relay_prepare_cpu,
1352                 .teardown.single        = NULL,
1353         },
1354         [CPUHP_SLAB_PREPARE] = {
1355                 .name                   = "slab:prepare",
1356                 .startup.single         = slab_prepare_cpu,
1357                 .teardown.single        = slab_dead_cpu,
1358         },
1359         [CPUHP_RCUTREE_PREP] = {
1360                 .name                   = "RCU/tree:prepare",
1361                 .startup.single         = rcutree_prepare_cpu,
1362                 .teardown.single        = rcutree_dead_cpu,
1363         },
1364         /*
1365          * On the tear-down path, timers_dead_cpu() must be invoked
1366          * before blk_mq_queue_reinit_notify() from notify_dead(),
1367          * otherwise a RCU stall occurs.
1368          */
1369         [CPUHP_TIMERS_PREPARE] = {
1370                 .name                   = "timers:dead",
1371                 .startup.single         = timers_prepare_cpu,
1372                 .teardown.single        = timers_dead_cpu,
1373         },
1374         /* Kicks the plugged cpu into life */
1375         [CPUHP_BRINGUP_CPU] = {
1376                 .name                   = "cpu:bringup",
1377                 .startup.single         = bringup_cpu,
1378                 .teardown.single        = NULL,
1379                 .cant_stop              = true,
1380         },
1381         /* Final state before CPU kills itself */
1382         [CPUHP_AP_IDLE_DEAD] = {
1383                 .name                   = "idle:dead",
1384         },
1385         /*
1386          * Last state before CPU enters the idle loop to die. Transient state
1387          * for synchronization.
1388          */
1389         [CPUHP_AP_OFFLINE] = {
1390                 .name                   = "ap:offline",
1391                 .cant_stop              = true,
1392         },
1393         /* First state is scheduler control. Interrupts are disabled */
1394         [CPUHP_AP_SCHED_STARTING] = {
1395                 .name                   = "sched:starting",
1396                 .startup.single         = sched_cpu_starting,
1397                 .teardown.single        = sched_cpu_dying,
1398         },
1399         [CPUHP_AP_RCUTREE_DYING] = {
1400                 .name                   = "RCU/tree:dying",
1401                 .startup.single         = NULL,
1402                 .teardown.single        = rcutree_dying_cpu,
1403         },
1404         [CPUHP_AP_SMPCFD_DYING] = {
1405                 .name                   = "smpcfd:dying",
1406                 .startup.single         = NULL,
1407                 .teardown.single        = smpcfd_dying_cpu,
1408         },
1409         /* Entry state on starting. Interrupts enabled from here on. Transient
1410          * state for synchronsization */
1411         [CPUHP_AP_ONLINE] = {
1412                 .name                   = "ap:online",
1413         },
1414         /*
1415          * Handled on controll processor until the plugged processor manages
1416          * this itself.
1417          */
1418         [CPUHP_TEARDOWN_CPU] = {
1419                 .name                   = "cpu:teardown",
1420                 .startup.single         = NULL,
1421                 .teardown.single        = takedown_cpu,
1422                 .cant_stop              = true,
1423         },
1424         /* Handle smpboot threads park/unpark */
1425         [CPUHP_AP_SMPBOOT_THREADS] = {
1426                 .name                   = "smpboot/threads:online",
1427                 .startup.single         = smpboot_unpark_threads,
1428                 .teardown.single        = smpboot_park_threads,
1429         },
1430         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1431                 .name                   = "irq/affinity:online",
1432                 .startup.single         = irq_affinity_online_cpu,
1433                 .teardown.single        = NULL,
1434         },
1435         [CPUHP_AP_PERF_ONLINE] = {
1436                 .name                   = "perf:online",
1437                 .startup.single         = perf_event_init_cpu,
1438                 .teardown.single        = perf_event_exit_cpu,
1439         },
1440         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1441                 .name                   = "workqueue:online",
1442                 .startup.single         = workqueue_online_cpu,
1443                 .teardown.single        = workqueue_offline_cpu,
1444         },
1445         [CPUHP_AP_RCUTREE_ONLINE] = {
1446                 .name                   = "RCU/tree:online",
1447                 .startup.single         = rcutree_online_cpu,
1448                 .teardown.single        = rcutree_offline_cpu,
1449         },
1450 #endif
1451         /*
1452          * The dynamically registered state space is here
1453          */
1454
1455 #ifdef CONFIG_SMP
1456         /* Last state is scheduler control setting the cpu active */
1457         [CPUHP_AP_ACTIVE] = {
1458                 .name                   = "sched:active",
1459                 .startup.single         = sched_cpu_activate,
1460                 .teardown.single        = sched_cpu_deactivate,
1461         },
1462 #endif
1463
1464         /* CPU is fully up and running. */
1465         [CPUHP_ONLINE] = {
1466                 .name                   = "online",
1467                 .startup.single         = NULL,
1468                 .teardown.single        = NULL,
1469         },
1470 };
1471
1472 /* Sanity check for callbacks */
1473 static int cpuhp_cb_check(enum cpuhp_state state)
1474 {
1475         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1476                 return -EINVAL;
1477         return 0;
1478 }
1479
1480 /*
1481  * Returns a free for dynamic slot assignment of the Online state. The states
1482  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1483  * by having no name assigned.
1484  */
1485 static int cpuhp_reserve_state(enum cpuhp_state state)
1486 {
1487         enum cpuhp_state i, end;
1488         struct cpuhp_step *step;
1489
1490         switch (state) {
1491         case CPUHP_AP_ONLINE_DYN:
1492                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1493                 end = CPUHP_AP_ONLINE_DYN_END;
1494                 break;
1495         case CPUHP_BP_PREPARE_DYN:
1496                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1497                 end = CPUHP_BP_PREPARE_DYN_END;
1498                 break;
1499         default:
1500                 return -EINVAL;
1501         }
1502
1503         for (i = state; i <= end; i++, step++) {
1504                 if (!step->name)
1505                         return i;
1506         }
1507         WARN(1, "No more dynamic states available for CPU hotplug\n");
1508         return -ENOSPC;
1509 }
1510
1511 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1512                                  int (*startup)(unsigned int cpu),
1513                                  int (*teardown)(unsigned int cpu),
1514                                  bool multi_instance)
1515 {
1516         /* (Un)Install the callbacks for further cpu hotplug operations */
1517         struct cpuhp_step *sp;
1518         int ret = 0;
1519
1520         /*
1521          * If name is NULL, then the state gets removed.
1522          *
1523          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1524          * the first allocation from these dynamic ranges, so the removal
1525          * would trigger a new allocation and clear the wrong (already
1526          * empty) state, leaving the callbacks of the to be cleared state
1527          * dangling, which causes wreckage on the next hotplug operation.
1528          */
1529         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1530                      state == CPUHP_BP_PREPARE_DYN)) {
1531                 ret = cpuhp_reserve_state(state);
1532                 if (ret < 0)
1533                         return ret;
1534                 state = ret;
1535         }
1536         sp = cpuhp_get_step(state);
1537         if (name && sp->name)
1538                 return -EBUSY;
1539
1540         sp->startup.single = startup;
1541         sp->teardown.single = teardown;
1542         sp->name = name;
1543         sp->multi_instance = multi_instance;
1544         INIT_HLIST_HEAD(&sp->list);
1545         return ret;
1546 }
1547
1548 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1549 {
1550         return cpuhp_get_step(state)->teardown.single;
1551 }
1552
1553 /*
1554  * Call the startup/teardown function for a step either on the AP or
1555  * on the current CPU.
1556  */
1557 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1558                             struct hlist_node *node)
1559 {
1560         struct cpuhp_step *sp = cpuhp_get_step(state);
1561         int ret;
1562
1563         /*
1564          * If there's nothing to do, we done.
1565          * Relies on the union for multi_instance.
1566          */
1567         if ((bringup && !sp->startup.single) ||
1568             (!bringup && !sp->teardown.single))
1569                 return 0;
1570         /*
1571          * The non AP bound callbacks can fail on bringup. On teardown
1572          * e.g. module removal we crash for now.
1573          */
1574 #ifdef CONFIG_SMP
1575         if (cpuhp_is_ap_state(state))
1576                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1577         else
1578                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1579 #else
1580         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1581 #endif
1582         BUG_ON(ret && !bringup);
1583         return ret;
1584 }
1585
1586 /*
1587  * Called from __cpuhp_setup_state on a recoverable failure.
1588  *
1589  * Note: The teardown callbacks for rollback are not allowed to fail!
1590  */
1591 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1592                                    struct hlist_node *node)
1593 {
1594         int cpu;
1595
1596         /* Roll back the already executed steps on the other cpus */
1597         for_each_present_cpu(cpu) {
1598                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1599                 int cpustate = st->state;
1600
1601                 if (cpu >= failedcpu)
1602                         break;
1603
1604                 /* Did we invoke the startup call on that cpu ? */
1605                 if (cpustate >= state)
1606                         cpuhp_issue_call(cpu, state, false, node);
1607         }
1608 }
1609
1610 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1611                                           struct hlist_node *node,
1612                                           bool invoke)
1613 {
1614         struct cpuhp_step *sp;
1615         int cpu;
1616         int ret;
1617
1618         lockdep_assert_cpus_held();
1619
1620         sp = cpuhp_get_step(state);
1621         if (sp->multi_instance == false)
1622                 return -EINVAL;
1623
1624         mutex_lock(&cpuhp_state_mutex);
1625
1626         if (!invoke || !sp->startup.multi)
1627                 goto add_node;
1628
1629         /*
1630          * Try to call the startup callback for each present cpu
1631          * depending on the hotplug state of the cpu.
1632          */
1633         for_each_present_cpu(cpu) {
1634                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1635                 int cpustate = st->state;
1636
1637                 if (cpustate < state)
1638                         continue;
1639
1640                 ret = cpuhp_issue_call(cpu, state, true, node);
1641                 if (ret) {
1642                         if (sp->teardown.multi)
1643                                 cpuhp_rollback_install(cpu, state, node);
1644                         goto unlock;
1645                 }
1646         }
1647 add_node:
1648         ret = 0;
1649         hlist_add_head(node, &sp->list);
1650 unlock:
1651         mutex_unlock(&cpuhp_state_mutex);
1652         return ret;
1653 }
1654
1655 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1656                                bool invoke)
1657 {
1658         int ret;
1659
1660         cpus_read_lock();
1661         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1662         cpus_read_unlock();
1663         return ret;
1664 }
1665 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1666
1667 /**
1668  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1669  * @state:              The state to setup
1670  * @invoke:             If true, the startup function is invoked for cpus where
1671  *                      cpu state >= @state
1672  * @startup:            startup callback function
1673  * @teardown:           teardown callback function
1674  * @multi_instance:     State is set up for multiple instances which get
1675  *                      added afterwards.
1676  *
1677  * The caller needs to hold cpus read locked while calling this function.
1678  * Returns:
1679  *   On success:
1680  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1681  *      0 for all other states
1682  *   On failure: proper (negative) error code
1683  */
1684 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1685                                    const char *name, bool invoke,
1686                                    int (*startup)(unsigned int cpu),
1687                                    int (*teardown)(unsigned int cpu),
1688                                    bool multi_instance)
1689 {
1690         int cpu, ret = 0;
1691         bool dynstate;
1692
1693         lockdep_assert_cpus_held();
1694
1695         if (cpuhp_cb_check(state) || !name)
1696                 return -EINVAL;
1697
1698         mutex_lock(&cpuhp_state_mutex);
1699
1700         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1701                                     multi_instance);
1702
1703         dynstate = state == CPUHP_AP_ONLINE_DYN;
1704         if (ret > 0 && dynstate) {
1705                 state = ret;
1706                 ret = 0;
1707         }
1708
1709         if (ret || !invoke || !startup)
1710                 goto out;
1711
1712         /*
1713          * Try to call the startup callback for each present cpu
1714          * depending on the hotplug state of the cpu.
1715          */
1716         for_each_present_cpu(cpu) {
1717                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1718                 int cpustate = st->state;
1719
1720                 if (cpustate < state)
1721                         continue;
1722
1723                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1724                 if (ret) {
1725                         if (teardown)
1726                                 cpuhp_rollback_install(cpu, state, NULL);
1727                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1728                         goto out;
1729                 }
1730         }
1731 out:
1732         mutex_unlock(&cpuhp_state_mutex);
1733         /*
1734          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1735          * dynamically allocated state in case of success.
1736          */
1737         if (!ret && dynstate)
1738                 return state;
1739         return ret;
1740 }
1741 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1742
1743 int __cpuhp_setup_state(enum cpuhp_state state,
1744                         const char *name, bool invoke,
1745                         int (*startup)(unsigned int cpu),
1746                         int (*teardown)(unsigned int cpu),
1747                         bool multi_instance)
1748 {
1749         int ret;
1750
1751         cpus_read_lock();
1752         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1753                                              teardown, multi_instance);
1754         cpus_read_unlock();
1755         return ret;
1756 }
1757 EXPORT_SYMBOL(__cpuhp_setup_state);
1758
1759 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1760                                   struct hlist_node *node, bool invoke)
1761 {
1762         struct cpuhp_step *sp = cpuhp_get_step(state);
1763         int cpu;
1764
1765         BUG_ON(cpuhp_cb_check(state));
1766
1767         if (!sp->multi_instance)
1768                 return -EINVAL;
1769
1770         cpus_read_lock();
1771         mutex_lock(&cpuhp_state_mutex);
1772
1773         if (!invoke || !cpuhp_get_teardown_cb(state))
1774                 goto remove;
1775         /*
1776          * Call the teardown callback for each present cpu depending
1777          * on the hotplug state of the cpu. This function is not
1778          * allowed to fail currently!
1779          */
1780         for_each_present_cpu(cpu) {
1781                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1782                 int cpustate = st->state;
1783
1784                 if (cpustate >= state)
1785                         cpuhp_issue_call(cpu, state, false, node);
1786         }
1787
1788 remove:
1789         hlist_del(node);
1790         mutex_unlock(&cpuhp_state_mutex);
1791         cpus_read_unlock();
1792
1793         return 0;
1794 }
1795 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1796
1797 /**
1798  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1799  * @state:      The state to remove
1800  * @invoke:     If true, the teardown function is invoked for cpus where
1801  *              cpu state >= @state
1802  *
1803  * The caller needs to hold cpus read locked while calling this function.
1804  * The teardown callback is currently not allowed to fail. Think
1805  * about module removal!
1806  */
1807 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1808 {
1809         struct cpuhp_step *sp = cpuhp_get_step(state);
1810         int cpu;
1811
1812         BUG_ON(cpuhp_cb_check(state));
1813
1814         lockdep_assert_cpus_held();
1815
1816         mutex_lock(&cpuhp_state_mutex);
1817         if (sp->multi_instance) {
1818                 WARN(!hlist_empty(&sp->list),
1819                      "Error: Removing state %d which has instances left.\n",
1820                      state);
1821                 goto remove;
1822         }
1823
1824         if (!invoke || !cpuhp_get_teardown_cb(state))
1825                 goto remove;
1826
1827         /*
1828          * Call the teardown callback for each present cpu depending
1829          * on the hotplug state of the cpu. This function is not
1830          * allowed to fail currently!
1831          */
1832         for_each_present_cpu(cpu) {
1833                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1834                 int cpustate = st->state;
1835
1836                 if (cpustate >= state)
1837                         cpuhp_issue_call(cpu, state, false, NULL);
1838         }
1839 remove:
1840         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1841         mutex_unlock(&cpuhp_state_mutex);
1842 }
1843 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1844
1845 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1846 {
1847         cpus_read_lock();
1848         __cpuhp_remove_state_cpuslocked(state, invoke);
1849         cpus_read_unlock();
1850 }
1851 EXPORT_SYMBOL(__cpuhp_remove_state);
1852
1853 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1854 static ssize_t show_cpuhp_state(struct device *dev,
1855                                 struct device_attribute *attr, char *buf)
1856 {
1857         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1858
1859         return sprintf(buf, "%d\n", st->state);
1860 }
1861 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1862
1863 static ssize_t write_cpuhp_target(struct device *dev,
1864                                   struct device_attribute *attr,
1865                                   const char *buf, size_t count)
1866 {
1867         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1868         struct cpuhp_step *sp;
1869         int target, ret;
1870
1871         ret = kstrtoint(buf, 10, &target);
1872         if (ret)
1873                 return ret;
1874
1875 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1876         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1877                 return -EINVAL;
1878 #else
1879         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1880                 return -EINVAL;
1881 #endif
1882
1883         ret = lock_device_hotplug_sysfs();
1884         if (ret)
1885                 return ret;
1886
1887         mutex_lock(&cpuhp_state_mutex);
1888         sp = cpuhp_get_step(target);
1889         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1890         mutex_unlock(&cpuhp_state_mutex);
1891         if (ret)
1892                 goto out;
1893
1894         if (st->state < target)
1895                 ret = do_cpu_up(dev->id, target);
1896         else
1897                 ret = do_cpu_down(dev->id, target);
1898 out:
1899         unlock_device_hotplug();
1900         return ret ? ret : count;
1901 }
1902
1903 static ssize_t show_cpuhp_target(struct device *dev,
1904                                  struct device_attribute *attr, char *buf)
1905 {
1906         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1907
1908         return sprintf(buf, "%d\n", st->target);
1909 }
1910 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1911
1912
1913 static ssize_t write_cpuhp_fail(struct device *dev,
1914                                 struct device_attribute *attr,
1915                                 const char *buf, size_t count)
1916 {
1917         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1918         struct cpuhp_step *sp;
1919         int fail, ret;
1920
1921         ret = kstrtoint(buf, 10, &fail);
1922         if (ret)
1923                 return ret;
1924
1925         /*
1926          * Cannot fail STARTING/DYING callbacks.
1927          */
1928         if (cpuhp_is_atomic_state(fail))
1929                 return -EINVAL;
1930
1931         /*
1932          * Cannot fail anything that doesn't have callbacks.
1933          */
1934         mutex_lock(&cpuhp_state_mutex);
1935         sp = cpuhp_get_step(fail);
1936         if (!sp->startup.single && !sp->teardown.single)
1937                 ret = -EINVAL;
1938         mutex_unlock(&cpuhp_state_mutex);
1939         if (ret)
1940                 return ret;
1941
1942         st->fail = fail;
1943
1944         return count;
1945 }
1946
1947 static ssize_t show_cpuhp_fail(struct device *dev,
1948                                struct device_attribute *attr, char *buf)
1949 {
1950         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1951
1952         return sprintf(buf, "%d\n", st->fail);
1953 }
1954
1955 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1956
1957 static struct attribute *cpuhp_cpu_attrs[] = {
1958         &dev_attr_state.attr,
1959         &dev_attr_target.attr,
1960         &dev_attr_fail.attr,
1961         NULL
1962 };
1963
1964 static const struct attribute_group cpuhp_cpu_attr_group = {
1965         .attrs = cpuhp_cpu_attrs,
1966         .name = "hotplug",
1967         NULL
1968 };
1969
1970 static ssize_t show_cpuhp_states(struct device *dev,
1971                                  struct device_attribute *attr, char *buf)
1972 {
1973         ssize_t cur, res = 0;
1974         int i;
1975
1976         mutex_lock(&cpuhp_state_mutex);
1977         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1978                 struct cpuhp_step *sp = cpuhp_get_step(i);
1979
1980                 if (sp->name) {
1981                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1982                         buf += cur;
1983                         res += cur;
1984                 }
1985         }
1986         mutex_unlock(&cpuhp_state_mutex);
1987         return res;
1988 }
1989 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1990
1991 static struct attribute *cpuhp_cpu_root_attrs[] = {
1992         &dev_attr_states.attr,
1993         NULL
1994 };
1995
1996 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1997         .attrs = cpuhp_cpu_root_attrs,
1998         .name = "hotplug",
1999         NULL
2000 };
2001
2002 #ifdef CONFIG_HOTPLUG_SMT
2003
2004 static const char *smt_states[] = {
2005         [CPU_SMT_ENABLED]               = "on",
2006         [CPU_SMT_DISABLED]              = "off",
2007         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2008         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2009 };
2010
2011 static ssize_t
2012 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2013 {
2014         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2015 }
2016
2017 static void cpuhp_offline_cpu_device(unsigned int cpu)
2018 {
2019         struct device *dev = get_cpu_device(cpu);
2020
2021         dev->offline = true;
2022         /* Tell user space about the state change */
2023         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2024 }
2025
2026 static void cpuhp_online_cpu_device(unsigned int cpu)
2027 {
2028         struct device *dev = get_cpu_device(cpu);
2029
2030         dev->offline = false;
2031         /* Tell user space about the state change */
2032         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2033 }
2034
2035 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2036 {
2037         int cpu, ret = 0;
2038
2039         cpu_maps_update_begin();
2040         for_each_online_cpu(cpu) {
2041                 if (topology_is_primary_thread(cpu))
2042                         continue;
2043                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2044                 if (ret)
2045                         break;
2046                 /*
2047                  * As this needs to hold the cpu maps lock it's impossible
2048                  * to call device_offline() because that ends up calling
2049                  * cpu_down() which takes cpu maps lock. cpu maps lock
2050                  * needs to be held as this might race against in kernel
2051                  * abusers of the hotplug machinery (thermal management).
2052                  *
2053                  * So nothing would update device:offline state. That would
2054                  * leave the sysfs entry stale and prevent onlining after
2055                  * smt control has been changed to 'off' again. This is
2056                  * called under the sysfs hotplug lock, so it is properly
2057                  * serialized against the regular offline usage.
2058                  */
2059                 cpuhp_offline_cpu_device(cpu);
2060         }
2061         if (!ret)
2062                 cpu_smt_control = ctrlval;
2063         cpu_maps_update_done();
2064         return ret;
2065 }
2066
2067 static int cpuhp_smt_enable(void)
2068 {
2069         int cpu, ret = 0;
2070
2071         cpu_maps_update_begin();
2072         cpu_smt_control = CPU_SMT_ENABLED;
2073         for_each_present_cpu(cpu) {
2074                 /* Skip online CPUs and CPUs on offline nodes */
2075                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2076                         continue;
2077                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2078                 if (ret)
2079                         break;
2080                 /* See comment in cpuhp_smt_disable() */
2081                 cpuhp_online_cpu_device(cpu);
2082         }
2083         cpu_maps_update_done();
2084         return ret;
2085 }
2086
2087 static ssize_t
2088 store_smt_control(struct device *dev, struct device_attribute *attr,
2089                   const char *buf, size_t count)
2090 {
2091         int ctrlval, ret;
2092
2093         if (sysfs_streq(buf, "on"))
2094                 ctrlval = CPU_SMT_ENABLED;
2095         else if (sysfs_streq(buf, "off"))
2096                 ctrlval = CPU_SMT_DISABLED;
2097         else if (sysfs_streq(buf, "forceoff"))
2098                 ctrlval = CPU_SMT_FORCE_DISABLED;
2099         else
2100                 return -EINVAL;
2101
2102         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2103                 return -EPERM;
2104
2105         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2106                 return -ENODEV;
2107
2108         ret = lock_device_hotplug_sysfs();
2109         if (ret)
2110                 return ret;
2111
2112         if (ctrlval != cpu_smt_control) {
2113                 switch (ctrlval) {
2114                 case CPU_SMT_ENABLED:
2115                         ret = cpuhp_smt_enable();
2116                         break;
2117                 case CPU_SMT_DISABLED:
2118                 case CPU_SMT_FORCE_DISABLED:
2119                         ret = cpuhp_smt_disable(ctrlval);
2120                         break;
2121                 }
2122         }
2123
2124         unlock_device_hotplug();
2125         return ret ? ret : count;
2126 }
2127 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2128
2129 static ssize_t
2130 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2131 {
2132         bool active = topology_max_smt_threads() > 1;
2133
2134         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2135 }
2136 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2137
2138 static struct attribute *cpuhp_smt_attrs[] = {
2139         &dev_attr_control.attr,
2140         &dev_attr_active.attr,
2141         NULL
2142 };
2143
2144 static const struct attribute_group cpuhp_smt_attr_group = {
2145         .attrs = cpuhp_smt_attrs,
2146         .name = "smt",
2147         NULL
2148 };
2149
2150 static int __init cpu_smt_state_init(void)
2151 {
2152         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2153                                   &cpuhp_smt_attr_group);
2154 }
2155
2156 #else
2157 static inline int cpu_smt_state_init(void) { return 0; }
2158 #endif
2159
2160 static int __init cpuhp_sysfs_init(void)
2161 {
2162         int cpu, ret;
2163
2164         ret = cpu_smt_state_init();
2165         if (ret)
2166                 return ret;
2167
2168         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2169                                  &cpuhp_cpu_root_attr_group);
2170         if (ret)
2171                 return ret;
2172
2173         for_each_possible_cpu(cpu) {
2174                 struct device *dev = get_cpu_device(cpu);
2175
2176                 if (!dev)
2177                         continue;
2178                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2179                 if (ret)
2180                         return ret;
2181         }
2182         return 0;
2183 }
2184 device_initcall(cpuhp_sysfs_init);
2185 #endif
2186
2187 /*
2188  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2189  * represents all NR_CPUS bits binary values of 1<<nr.
2190  *
2191  * It is used by cpumask_of() to get a constant address to a CPU
2192  * mask value that has a single bit set only.
2193  */
2194
2195 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2196 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2197 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2198 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2199 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2200
2201 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2202
2203         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2204         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2205 #if BITS_PER_LONG > 32
2206         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2207         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2208 #endif
2209 };
2210 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2211
2212 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2213 EXPORT_SYMBOL(cpu_all_bits);
2214
2215 #ifdef CONFIG_INIT_ALL_POSSIBLE
2216 struct cpumask __cpu_possible_mask __read_mostly
2217         = {CPU_BITS_ALL};
2218 #else
2219 struct cpumask __cpu_possible_mask __read_mostly;
2220 #endif
2221 EXPORT_SYMBOL(__cpu_possible_mask);
2222
2223 struct cpumask __cpu_online_mask __read_mostly;
2224 EXPORT_SYMBOL(__cpu_online_mask);
2225
2226 struct cpumask __cpu_present_mask __read_mostly;
2227 EXPORT_SYMBOL(__cpu_present_mask);
2228
2229 struct cpumask __cpu_active_mask __read_mostly;
2230 EXPORT_SYMBOL(__cpu_active_mask);
2231
2232 void init_cpu_present(const struct cpumask *src)
2233 {
2234         cpumask_copy(&__cpu_present_mask, src);
2235 }
2236
2237 void init_cpu_possible(const struct cpumask *src)
2238 {
2239         cpumask_copy(&__cpu_possible_mask, src);
2240 }
2241
2242 void init_cpu_online(const struct cpumask *src)
2243 {
2244         cpumask_copy(&__cpu_online_mask, src);
2245 }
2246
2247 /*
2248  * Activate the first processor.
2249  */
2250 void __init boot_cpu_init(void)
2251 {
2252         int cpu = smp_processor_id();
2253
2254         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2255         set_cpu_online(cpu, true);
2256         set_cpu_active(cpu, true);
2257         set_cpu_present(cpu, true);
2258         set_cpu_possible(cpu, true);
2259
2260 #ifdef CONFIG_SMP
2261         __boot_cpu_id = cpu;
2262 #endif
2263 }
2264
2265 /*
2266  * Must be called _AFTER_ setting up the per_cpu areas
2267  */
2268 void __init boot_cpu_state_init(void)
2269 {
2270         this_cpu_write(cpuhp_state.booted_once, true);
2271         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2272 }