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 void inline cpuhp_lock_acquire(bool bringup)
  85 {
  86         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  87 }
  88
  89 static void inline 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 void inline cpuhp_lock_acquire(bool bringup) { }
  96 static void inline 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_bp_states[];
 129 static struct cpuhp_step cpuhp_ap_states[];
 130
 131 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 132 {
 133         /*
 134          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 135          * purposes as that state is handled explicitly in cpu_down.
 136          */
 137         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 138 }
 139
 140 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 141 {
 142         struct cpuhp_step *sp;
 143
 144         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 145         return sp + state;
 146 }
 147
 148 /**
 149  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 150  * @cpu:        The cpu for which the callback should be invoked
 151  * @state:      The state to do callbacks for
 152  * @bringup:    True if the bringup callback should be invoked
 153  * @node:       For multi-instance, do a single entry callback for install/remove
 154  * @lastp:      For multi-instance rollback, remember how far we got
 155  *
 156  * Called from cpu hotplug and from the state register machinery.
 157  */
 158 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 159                                  bool bringup, struct hlist_node *node,
 160                                  struct hlist_node **lastp)
 161 {
 162         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 163         struct cpuhp_step *step = cpuhp_get_step(state);
 164         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 165         int (*cb)(unsigned int cpu);
 166         int ret, cnt;
 167
 168         if (st->fail == state) {
 169                 st->fail = CPUHP_INVALID;
 170
 171                 if (!(bringup ? step->startup.single : step->teardown.single))
 172                         return 0;
 173
 174                 return -EAGAIN;
 175         }
 176
 177         if (!step->multi_instance) {
 178                 WARN_ON_ONCE(lastp && *lastp);
 179                 cb = bringup ? step->startup.single : step->teardown.single;
 180                 if (!cb)
 181                         return 0;
 182                 trace_cpuhp_enter(cpu, st->target, state, cb);
 183                 ret = cb(cpu);
 184                 trace_cpuhp_exit(cpu, st->state, state, ret);
 185                 return ret;
 186         }
 187         cbm = bringup ? step->startup.multi : step->teardown.multi;
 188         if (!cbm)
 189                 return 0;
 190
 191         /* Single invocation for instance add/remove */
 192         if (node) {
 193                 WARN_ON_ONCE(lastp && *lastp);
 194                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 195                 ret = cbm(cpu, node);
 196                 trace_cpuhp_exit(cpu, st->state, state, ret);
 197                 return ret;
 198         }
 199
 200         /* State transition. Invoke on all instances */
 201         cnt = 0;
 202         hlist_for_each(node, &step->list) {
 203                 if (lastp && node == *lastp)
 204                         break;
 205
 206                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 207                 ret = cbm(cpu, node);
 208                 trace_cpuhp_exit(cpu, st->state, state, ret);
 209                 if (ret) {
 210                         if (!lastp)
 211                                 goto err;
 212
 213                         *lastp = node;
 214                         return ret;
 215                 }
 216                 cnt++;
 217         }
 218         if (lastp)
 219                 *lastp = NULL;
 220         return 0;
 221 err:
 222         /* Rollback the instances if one failed */
 223         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 224         if (!cbm)
 225                 return ret;
 226
 227         hlist_for_each(node, &step->list) {
 228                 if (!cnt--)
 229                         break;
 230
 231                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 232                 ret = cbm(cpu, node);
 233                 trace_cpuhp_exit(cpu, st->state, state, ret);
 234                 /*
 235                  * Rollback must not fail,
 236                  */
 237                 WARN_ON_ONCE(ret);
 238         }
 239         return ret;
 240 }
 241
 242 #ifdef CONFIG_SMP
 243 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 244 {
 245         struct completion *done = bringup ? &st->done_up : &st->done_down;
 246         wait_for_completion(done);
 247 }
 248
 249 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 250 {
 251         struct completion *done = bringup ? &st->done_up : &st->done_down;
 252         complete(done);
 253 }
 254
 255 /*
 256  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 257  */
 258 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 259 {
 260         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 261 }
 262
 263 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 264 static DEFINE_MUTEX(cpu_add_remove_lock);
 265 bool cpuhp_tasks_frozen;
 266 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 267
 268 /*
 269  * The following two APIs (cpu_maps_update_begin/done) must be used when
 270  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 271  */
 272 void cpu_maps_update_begin(void)
 273 {
 274         mutex_lock(&cpu_add_remove_lock);
 275 }
 276
 277 void cpu_maps_update_done(void)
 278 {
 279         mutex_unlock(&cpu_add_remove_lock);
 280 }
 281
 282 /*
 283  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 284  * Should always be manipulated under cpu_add_remove_lock
 285  */
 286 static int cpu_hotplug_disabled;
 287
 288 #ifdef CONFIG_HOTPLUG_CPU
 289
 290 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 291
 292 void cpus_read_lock(void)
 293 {
 294         percpu_down_read(&cpu_hotplug_lock);
 295 }
 296 EXPORT_SYMBOL_GPL(cpus_read_lock);
 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 /*
 320  * Wait for currently running CPU hotplug operations to complete (if any) and
 321  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 322  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 323  * hotplug path before performing hotplug operations. So acquiring that lock
 324  * guarantees mutual exclusion from any currently running hotplug operations.
 325  */
 326 void cpu_hotplug_disable(void)
 327 {
 328         cpu_maps_update_begin();
 329         cpu_hotplug_disabled++;
 330         cpu_maps_update_done();
 331 }
 332 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 333
 334 static void __cpu_hotplug_enable(void)
 335 {
 336         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 337                 return;
 338         cpu_hotplug_disabled--;
 339 }
 340
 341 void cpu_hotplug_enable(void)
 342 {
 343         cpu_maps_update_begin();
 344         __cpu_hotplug_enable();
 345         cpu_maps_update_done();
 346 }
 347 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 348 #endif  /* CONFIG_HOTPLUG_CPU */
 349
 350 #ifdef CONFIG_HOTPLUG_SMT
 351 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 352 EXPORT_SYMBOL_GPL(cpu_smt_control);
 353
 354 void __init cpu_smt_disable(bool force)
 355 {
 356         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 357                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 358                 return;
 359
 360         if (force) {
 361                 pr_info("SMT: Force disabled\n");
 362                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 363         } else {
 364                 cpu_smt_control = CPU_SMT_DISABLED;
 365         }
 366 }
 367
 368 /*
 369  * The decision whether SMT is supported can only be done after the full
 370  * CPU identification. Called from architecture code.
 371  */
 372 void __init cpu_smt_check_topology(void)
 373 {
 374         if (!topology_smt_supported())
 375                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 376 }
 377
 378 static int __init smt_cmdline_disable(char *str)
 379 {
 380         cpu_smt_disable(str && !strcmp(str, "force"));
 381         return 0;
 382 }
 383 early_param("nosmt", smt_cmdline_disable);
 384
 385 static inline bool cpu_smt_allowed(unsigned int cpu)
 386 {
 387         if (cpu_smt_control == CPU_SMT_ENABLED)
 388                 return true;
 389
 390         if (topology_is_primary_thread(cpu))
 391                 return true;
 392
 393         /*
 394          * On x86 it's required to boot all logical CPUs at least once so
 395          * that the init code can get a chance to set CR4.MCE on each
 396          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 397          * core will shutdown the machine.
 398          */
 399         return !per_cpu(cpuhp_state, cpu).booted_once;
 400 }
 401 #else
 402 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 403 #endif
 404
 405 static inline enum cpuhp_state
 406 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 407 {
 408         enum cpuhp_state prev_state = st->state;
 409
 410         st->rollback = false;
 411         st->last = NULL;
 412
 413         st->target = target;
 414         st->single = false;
 415         st->bringup = st->state < target;
 416
 417         return prev_state;
 418 }
 419
 420 static inline void
 421 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 422 {
 423         st->rollback = true;
 424
 425         /*
 426          * If we have st->last we need to undo partial multi_instance of this
 427          * state first. Otherwise start undo at the previous state.
 428          */
 429         if (!st->last) {
 430                 if (st->bringup)
 431                         st->state--;
 432                 else
 433                         st->state++;
 434         }
 435
 436         st->target = prev_state;
 437         st->bringup = !st->bringup;
 438 }
 439
 440 /* Regular hotplug invocation of the AP hotplug thread */
 441 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 442 {
 443         if (!st->single && st->state == st->target)
 444                 return;
 445
 446         st->result = 0;
 447         /*
 448          * Make sure the above stores are visible before should_run becomes
 449          * true. Paired with the mb() above in cpuhp_thread_fun()
 450          */
 451         smp_mb();
 452         st->should_run = true;
 453         wake_up_process(st->thread);
 454         wait_for_ap_thread(st, st->bringup);
 455 }
 456
 457 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 458 {
 459         enum cpuhp_state prev_state;
 460         int ret;
 461
 462         prev_state = cpuhp_set_state(st, target);
 463         __cpuhp_kick_ap(st);
 464         if ((ret = st->result)) {
 465                 cpuhp_reset_state(st, prev_state);
 466                 __cpuhp_kick_ap(st);
 467         }
 468
 469         return ret;
 470 }
 471
 472 static int bringup_wait_for_ap(unsigned int cpu)
 473 {
 474         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 475
 476         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 477         wait_for_ap_thread(st, true);
 478         if (WARN_ON_ONCE((!cpu_online(cpu))))
 479                 return -ECANCELED;
 480
 481         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 482         stop_machine_unpark(cpu);
 483         kthread_unpark(st->thread);
 484
 485         /*
 486          * SMT soft disabling on X86 requires to bring the CPU out of the
 487          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 488          * CPU marked itself as booted_once in cpu_notify_starting() so the
 489          * cpu_smt_allowed() check will now return false if this is not the
 490          * primary sibling.
 491          */
 492         if (!cpu_smt_allowed(cpu))
 493                 return -ECANCELED;
 494
 495         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 496                 return 0;
 497
 498         return cpuhp_kick_ap(st, st->target);
 499 }
 500
 501 static int bringup_cpu(unsigned int cpu)
 502 {
 503         struct task_struct *idle = idle_thread_get(cpu);
 504         int ret;
 505
 506         /*
 507          * Some architectures have to walk the irq descriptors to
 508          * setup the vector space for the cpu which comes online.
 509          * Prevent irq alloc/free across the bringup.
 510          */
 511         irq_lock_sparse();
 512
 513         /* Arch-specific enabling code. */
 514         ret = __cpu_up(cpu, idle);
 515         irq_unlock_sparse();
 516         if (ret)
 517                 return ret;
 518         return bringup_wait_for_ap(cpu);
 519 }
 520
 521 /*
 522  * Hotplug state machine related functions
 523  */
 524
 525 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 526 {
 527         for (st->state--; st->state > st->target; st->state--) {
 528                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 529
 530                 if (!step->skip_onerr)
 531                         cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 532         }
 533 }
 534
 535 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 536                               enum cpuhp_state target)
 537 {
 538         enum cpuhp_state prev_state = st->state;
 539         int ret = 0;
 540
 541         while (st->state < target) {
 542                 st->state++;
 543                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 544                 if (ret) {
 545                         st->target = prev_state;
 546                         undo_cpu_up(cpu, st);
 547                         break;
 548                 }
 549         }
 550         return ret;
 551 }
 552
 553 /*
 554  * The cpu hotplug threads manage the bringup and teardown of the cpus
 555  */
 556 static void cpuhp_create(unsigned int cpu)
 557 {
 558         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 559
 560         init_completion(&st->done_up);
 561         init_completion(&st->done_down);
 562 }
 563
 564 static int cpuhp_should_run(unsigned int cpu)
 565 {
 566         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 567
 568         return st->should_run;
 569 }
 570
 571 /*
 572  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 573  * callbacks when a state gets [un]installed at runtime.
 574  *
 575  * Each invocation of this function by the smpboot thread does a single AP
 576  * state callback.
 577  *
 578  * It has 3 modes of operation:
 579  *  - single: runs st->cb_state
 580  *  - up:     runs ++st->state, while st->state < st->target
 581  *  - down:   runs st->state--, while st->state > st->target
 582  *
 583  * When complete or on error, should_run is cleared and the completion is fired.
 584  */
 585 static void cpuhp_thread_fun(unsigned int cpu)
 586 {
 587         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 588         bool bringup = st->bringup;
 589         enum cpuhp_state state;
 590
 591         /*
 592          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 593          * that if we see ->should_run we also see the rest of the state.
 594          */
 595         smp_mb();
 596
 597         if (WARN_ON_ONCE(!st->should_run))
 598                 return;
 599
 600         cpuhp_lock_acquire(bringup);
 601
 602         if (st->single) {
 603                 state = st->cb_state;
 604                 st->should_run = false;
 605         } else {
 606                 if (bringup) {
 607                         st->state++;
 608                         state = st->state;
 609                         st->should_run = (st->state < st->target);
 610                         WARN_ON_ONCE(st->state > st->target);
 611                 } else {
 612                         state = st->state;
 613                         st->state--;
 614                         st->should_run = (st->state > st->target);
 615                         WARN_ON_ONCE(st->state < st->target);
 616                 }
 617         }
 618
 619         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 620
 621         if (st->rollback) {
 622                 struct cpuhp_step *step = cpuhp_get_step(state);
 623                 if (step->skip_onerr)
 624                         goto next;
 625         }
 626
 627         if (cpuhp_is_atomic_state(state)) {
 628                 local_irq_disable();
 629                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 630                 local_irq_enable();
 631
 632                 /*
 633                  * STARTING/DYING must not fail!
 634                  */
 635                 WARN_ON_ONCE(st->result);
 636         } else {
 637                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 638         }
 639
 640         if (st->result) {
 641                 /*
 642                  * If we fail on a rollback, we're up a creek without no
 643                  * paddle, no way forward, no way back. We loose, thanks for
 644                  * playing.
 645                  */
 646                 WARN_ON_ONCE(st->rollback);
 647                 st->should_run = false;
 648         }
 649
 650 next:
 651         cpuhp_lock_release(bringup);
 652
 653         if (!st->should_run)
 654                 complete_ap_thread(st, bringup);
 655 }
 656
 657 /* Invoke a single callback on a remote cpu */
 658 static int
 659 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 660                          struct hlist_node *node)
 661 {
 662         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 663         int ret;
 664
 665         if (!cpu_online(cpu))
 666                 return 0;
 667
 668         cpuhp_lock_acquire(false);
 669         cpuhp_lock_release(false);
 670
 671         cpuhp_lock_acquire(true);
 672         cpuhp_lock_release(true);
 673
 674         /*
 675          * If we are up and running, use the hotplug thread. For early calls
 676          * we invoke the thread function directly.
 677          */
 678         if (!st->thread)
 679                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 680
 681         st->rollback = false;
 682         st->last = NULL;
 683
 684         st->node = node;
 685         st->bringup = bringup;
 686         st->cb_state = state;
 687         st->single = true;
 688
 689         __cpuhp_kick_ap(st);
 690
 691         /*
 692          * If we failed and did a partial, do a rollback.
 693          */
 694         if ((ret = st->result) && st->last) {
 695                 st->rollback = true;
 696                 st->bringup = !bringup;
 697
 698                 __cpuhp_kick_ap(st);
 699         }
 700
 701         /*
 702          * Clean up the leftovers so the next hotplug operation wont use stale
 703          * data.
 704          */
 705         st->node = st->last = NULL;
 706         return ret;
 707 }
 708
 709 static int cpuhp_kick_ap_work(unsigned int cpu)
 710 {
 711         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 712         enum cpuhp_state prev_state = st->state;
 713         int ret;
 714
 715         cpuhp_lock_acquire(false);
 716         cpuhp_lock_release(false);
 717
 718         cpuhp_lock_acquire(true);
 719         cpuhp_lock_release(true);
 720
 721         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 722         ret = cpuhp_kick_ap(st, st->target);
 723         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 724
 725         return ret;
 726 }
 727
 728 static struct smp_hotplug_thread cpuhp_threads = {
 729         .store                  = &cpuhp_state.thread,
 730         .create                 = &cpuhp_create,
 731         .thread_should_run      = cpuhp_should_run,
 732         .thread_fn              = cpuhp_thread_fun,
 733         .thread_comm            = "cpuhp/%u",
 734         .selfparking            = true,
 735 };
 736
 737 void __init cpuhp_threads_init(void)
 738 {
 739         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 740         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 741 }
 742
 743 #ifdef CONFIG_HOTPLUG_CPU
 744 /**
 745  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 746  * @cpu: a CPU id
 747  *
 748  * This function walks all processes, finds a valid mm struct for each one and
 749  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 750  * trivial, there are various non-obvious corner cases, which this function
 751  * tries to solve in a safe manner.
 752  *
 753  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 754  * be called only for an already offlined CPU.
 755  */
 756 void clear_tasks_mm_cpumask(int cpu)
 757 {
 758         struct task_struct *p;
 759
 760         /*
 761          * This function is called after the cpu is taken down and marked
 762          * offline, so its not like new tasks will ever get this cpu set in
 763          * their mm mask. -- Peter Zijlstra
 764          * Thus, we may use rcu_read_lock() here, instead of grabbing
 765          * full-fledged tasklist_lock.
 766          */
 767         WARN_ON(cpu_online(cpu));
 768         rcu_read_lock();
 769         for_each_process(p) {
 770                 struct task_struct *t;
 771
 772                 /*
 773                  * Main thread might exit, but other threads may still have
 774                  * a valid mm. Find one.
 775                  */
 776                 t = find_lock_task_mm(p);
 777                 if (!t)
 778                         continue;
 779                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 780                 task_unlock(t);
 781         }
 782         rcu_read_unlock();
 783 }
 784
 785 /* Take this CPU down. */
 786 static int take_cpu_down(void *_param)
 787 {
 788         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 789         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 790         int err, cpu = smp_processor_id();
 791         int ret;
 792
 793         /* Ensure this CPU doesn't handle any more interrupts. */
 794         err = __cpu_disable();
 795         if (err < 0)
 796                 return err;
 797
 798         /*
 799          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 800          * do this step again.
 801          */
 802         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 803         st->state--;
 804         /* Invoke the former CPU_DYING callbacks */
 805         for (; st->state > target; st->state--) {
 806                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 807                 /*
 808                  * DYING must not fail!
 809                  */
 810                 WARN_ON_ONCE(ret);
 811         }
 812
 813         /* Give up timekeeping duties */
 814         tick_handover_do_timer();
 815         /* Park the stopper thread */
 816         stop_machine_park(cpu);
 817         return 0;
 818 }
 819
 820 static int takedown_cpu(unsigned int cpu)
 821 {
 822         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 823         int err;
 824
 825         /* Park the smpboot threads */
 826         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 827
 828         /*
 829          * Prevent irq alloc/free while the dying cpu reorganizes the
 830          * interrupt affinities.
 831          */
 832         irq_lock_sparse();
 833
 834         /*
 835          * So now all preempt/rcu users must observe !cpu_active().
 836          */
 837         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 838         if (err) {
 839                 /* CPU refused to die */
 840                 irq_unlock_sparse();
 841                 /* Unpark the hotplug thread so we can rollback there */
 842                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 843                 return err;
 844         }
 845         BUG_ON(cpu_online(cpu));
 846
 847         /*
 848          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 849          * runnable tasks from the cpu, there's only the idle task left now
 850          * that the migration thread is done doing the stop_machine thing.
 851          *
 852          * Wait for the stop thread to go away.
 853          */
 854         wait_for_ap_thread(st, false);
 855         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 856
 857         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 858         irq_unlock_sparse();
 859
 860         hotplug_cpu__broadcast_tick_pull(cpu);
 861         /* This actually kills the CPU. */
 862         __cpu_die(cpu);
 863
 864         tick_cleanup_dead_cpu(cpu);
 865         rcutree_migrate_callbacks(cpu);
 866         return 0;
 867 }
 868
 869 static void cpuhp_complete_idle_dead(void *arg)
 870 {
 871         struct cpuhp_cpu_state *st = arg;
 872
 873         complete_ap_thread(st, false);
 874 }
 875
 876 void cpuhp_report_idle_dead(void)
 877 {
 878         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 879
 880         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 881         rcu_report_dead(smp_processor_id());
 882         st->state = CPUHP_AP_IDLE_DEAD;
 883         /*
 884          * We cannot call complete after rcu_report_dead() so we delegate it
 885          * to an online cpu.
 886          */
 887         smp_call_function_single(cpumask_first(cpu_online_mask),
 888                                  cpuhp_complete_idle_dead, st, 0);
 889 }
 890
 891 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 892 {
 893         for (st->state++; st->state < st->target; st->state++) {
 894                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 895
 896                 if (!step->skip_onerr)
 897                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 898         }
 899 }
 900
 901 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 902                                 enum cpuhp_state target)
 903 {
 904         enum cpuhp_state prev_state = st->state;
 905         int ret = 0;
 906
 907         for (; st->state > target; st->state--) {
 908                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 909                 if (ret) {
 910                         st->target = prev_state;
 911                         undo_cpu_down(cpu, st);
 912                         break;
 913                 }
 914         }
 915         return ret;
 916 }
 917
 918 /* Requires cpu_add_remove_lock to be held */
 919 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 920                            enum cpuhp_state target)
 921 {
 922         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 923         int prev_state, ret = 0;
 924
 925         if (num_online_cpus() == 1)
 926                 return -EBUSY;
 927
 928         if (!cpu_present(cpu))
 929                 return -EINVAL;
 930
 931         cpus_write_lock();
 932
 933         cpuhp_tasks_frozen = tasks_frozen;
 934
 935         prev_state = cpuhp_set_state(st, target);
 936         /*
 937          * If the current CPU state is in the range of the AP hotplug thread,
 938          * then we need to kick the thread.
 939          */
 940         if (st->state > CPUHP_TEARDOWN_CPU) {
 941                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 942                 ret = cpuhp_kick_ap_work(cpu);
 943                 /*
 944                  * The AP side has done the error rollback already. Just
 945                  * return the error code..
 946                  */
 947                 if (ret)
 948                         goto out;
 949
 950                 /*
 951                  * We might have stopped still in the range of the AP hotplug
 952                  * thread. Nothing to do anymore.
 953                  */
 954                 if (st->state > CPUHP_TEARDOWN_CPU)
 955                         goto out;
 956
 957                 st->target = target;
 958         }
 959         /*
 960          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 961          * to do the further cleanups.
 962          */
 963         ret = cpuhp_down_callbacks(cpu, st, target);
 964         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 965                 cpuhp_reset_state(st, prev_state);
 966                 __cpuhp_kick_ap(st);
 967         }
 968
 969 out:
 970         cpus_write_unlock();
 971         /*
 972          * Do post unplug cleanup. This is still protected against
 973          * concurrent CPU hotplug via cpu_add_remove_lock.
 974          */
 975         lockup_detector_cleanup();
 976         return ret;
 977 }
 978
 979 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
 980 {
 981         if (cpu_hotplug_disabled)
 982                 return -EBUSY;
 983         return _cpu_down(cpu, 0, target);
 984 }
 985
 986 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 987 {
 988         int err;
 989
 990         cpu_maps_update_begin();
 991         err = cpu_down_maps_locked(cpu, target);
 992         cpu_maps_update_done();
 993         return err;
 994 }
 995
 996 int cpu_down(unsigned int cpu)
 997 {
 998         return do_cpu_down(cpu, CPUHP_OFFLINE);
 999 }
1000 EXPORT_SYMBOL(cpu_down);
1001
1002 #else
1003 #define takedown_cpu            NULL
1004 #endif /*CONFIG_HOTPLUG_CPU*/
1005
1006 /**
1007  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1008  * @cpu: cpu that just started
1009  *
1010  * It must be called by the arch code on the new cpu, before the new cpu
1011  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1012  */
1013 void notify_cpu_starting(unsigned int cpu)
1014 {
1015         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1016         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1017         int ret;
1018
1019         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1020         st->booted_once = true;
1021         while (st->state < target) {
1022                 st->state++;
1023                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1024                 /*
1025                  * STARTING must not fail!
1026                  */
1027                 WARN_ON_ONCE(ret);
1028         }
1029 }
1030
1031 /*
1032  * Called from the idle task. Wake up the controlling task which brings the
1033  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1034  * the rest of the online bringup to the hotplug thread.
1035  */
1036 void cpuhp_online_idle(enum cpuhp_state state)
1037 {
1038         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1039
1040         /* Happens for the boot cpu */
1041         if (state != CPUHP_AP_ONLINE_IDLE)
1042                 return;
1043
1044         st->state = CPUHP_AP_ONLINE_IDLE;
1045         complete_ap_thread(st, true);
1046 }
1047
1048 /* Requires cpu_add_remove_lock to be held */
1049 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1050 {
1051         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1052         struct task_struct *idle;
1053         int ret = 0;
1054
1055         cpus_write_lock();
1056
1057         if (!cpu_present(cpu)) {
1058                 ret = -EINVAL;
1059                 goto out;
1060         }
1061
1062         /*
1063          * The caller of do_cpu_up might have raced with another
1064          * caller. Ignore it for now.
1065          */
1066         if (st->state >= target)
1067                 goto out;
1068
1069         if (st->state == CPUHP_OFFLINE) {
1070                 /* Let it fail before we try to bring the cpu up */
1071                 idle = idle_thread_get(cpu);
1072                 if (IS_ERR(idle)) {
1073                         ret = PTR_ERR(idle);
1074                         goto out;
1075                 }
1076         }
1077
1078         cpuhp_tasks_frozen = tasks_frozen;
1079
1080         cpuhp_set_state(st, target);
1081         /*
1082          * If the current CPU state is in the range of the AP hotplug thread,
1083          * then we need to kick the thread once more.
1084          */
1085         if (st->state > CPUHP_BRINGUP_CPU) {
1086                 ret = cpuhp_kick_ap_work(cpu);
1087                 /*
1088                  * The AP side has done the error rollback already. Just
1089                  * return the error code..
1090                  */
1091                 if (ret)
1092                         goto out;
1093         }
1094
1095         /*
1096          * Try to reach the target state. We max out on the BP at
1097          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1098          * responsible for bringing it up to the target state.
1099          */
1100         target = min((int)target, CPUHP_BRINGUP_CPU);
1101         ret = cpuhp_up_callbacks(cpu, st, target);
1102 out:
1103         cpus_write_unlock();
1104         return ret;
1105 }
1106
1107 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1108 {
1109         int err = 0;
1110
1111         if (!cpu_possible(cpu)) {
1112                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1113                        cpu);
1114 #if defined(CONFIG_IA64)
1115                 pr_err("please check additional_cpus= boot parameter\n");
1116 #endif
1117                 return -EINVAL;
1118         }
1119
1120         err = try_online_node(cpu_to_node(cpu));
1121         if (err)
1122                 return err;
1123
1124         cpu_maps_update_begin();
1125
1126         if (cpu_hotplug_disabled) {
1127                 err = -EBUSY;
1128                 goto out;
1129         }
1130         if (!cpu_smt_allowed(cpu)) {
1131                 err = -EPERM;
1132                 goto out;
1133         }
1134
1135         err = _cpu_up(cpu, 0, target);
1136 out:
1137         cpu_maps_update_done();
1138         return err;
1139 }
1140
1141 int cpu_up(unsigned int cpu)
1142 {
1143         return do_cpu_up(cpu, CPUHP_ONLINE);
1144 }
1145 EXPORT_SYMBOL_GPL(cpu_up);
1146
1147 #ifdef CONFIG_PM_SLEEP_SMP
1148 static cpumask_var_t frozen_cpus;
1149
1150 int freeze_secondary_cpus(int primary)
1151 {
1152         int cpu, error = 0;
1153
1154         cpu_maps_update_begin();
1155         if (!cpu_online(primary))
1156                 primary = cpumask_first(cpu_online_mask);
1157         /*
1158          * We take down all of the non-boot CPUs in one shot to avoid races
1159          * with the userspace trying to use the CPU hotplug at the same time
1160          */
1161         cpumask_clear(frozen_cpus);
1162
1163         pr_info("Disabling non-boot CPUs ...\n");
1164         for_each_online_cpu(cpu) {
1165                 if (cpu == primary)
1166                         continue;
1167                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1168                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1169                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1170                 if (!error)
1171                         cpumask_set_cpu(cpu, frozen_cpus);
1172                 else {
1173                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1174                         break;
1175                 }
1176         }
1177
1178         if (!error)
1179                 BUG_ON(num_online_cpus() > 1);
1180         else
1181                 pr_err("Non-boot CPUs are not disabled\n");
1182
1183         /*
1184          * Make sure the CPUs won't be enabled by someone else. We need to do
1185          * this even in case of failure as all disable_nonboot_cpus() users are
1186          * supposed to do enable_nonboot_cpus() on the failure path.
1187          */
1188         cpu_hotplug_disabled++;
1189
1190         cpu_maps_update_done();
1191         return error;
1192 }
1193
1194 void __weak arch_enable_nonboot_cpus_begin(void)
1195 {
1196 }
1197
1198 void __weak arch_enable_nonboot_cpus_end(void)
1199 {
1200 }
1201
1202 void enable_nonboot_cpus(void)
1203 {
1204         int cpu, error;
1205
1206         /* Allow everyone to use the CPU hotplug again */
1207         cpu_maps_update_begin();
1208         __cpu_hotplug_enable();
1209         if (cpumask_empty(frozen_cpus))
1210                 goto out;
1211
1212         pr_info("Enabling non-boot CPUs ...\n");
1213
1214         arch_enable_nonboot_cpus_begin();
1215
1216         for_each_cpu(cpu, frozen_cpus) {
1217                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1218                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1219                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1220                 if (!error) {
1221                         pr_info("CPU%d is up\n", cpu);
1222                         continue;
1223                 }
1224                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1225         }
1226
1227         arch_enable_nonboot_cpus_end();
1228
1229         cpumask_clear(frozen_cpus);
1230 out:
1231         cpu_maps_update_done();
1232 }
1233
1234 static int __init alloc_frozen_cpus(void)
1235 {
1236         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1237                 return -ENOMEM;
1238         return 0;
1239 }
1240 core_initcall(alloc_frozen_cpus);
1241
1242 /*
1243  * When callbacks for CPU hotplug notifications are being executed, we must
1244  * ensure that the state of the system with respect to the tasks being frozen
1245  * or not, as reported by the notification, remains unchanged *throughout the
1246  * duration* of the execution of the callbacks.
1247  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1248  *
1249  * This synchronization is implemented by mutually excluding regular CPU
1250  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1251  * Hibernate notifications.
1252  */
1253 static int
1254 cpu_hotplug_pm_callback(struct notifier_block *nb,
1255                         unsigned long action, void *ptr)
1256 {
1257         switch (action) {
1258
1259         case PM_SUSPEND_PREPARE:
1260         case PM_HIBERNATION_PREPARE:
1261                 cpu_hotplug_disable();
1262                 break;
1263
1264         case PM_POST_SUSPEND:
1265         case PM_POST_HIBERNATION:
1266                 cpu_hotplug_enable();
1267                 break;
1268
1269         default:
1270                 return NOTIFY_DONE;
1271         }
1272
1273         return NOTIFY_OK;
1274 }
1275
1276
1277 static int __init cpu_hotplug_pm_sync_init(void)
1278 {
1279         /*
1280          * cpu_hotplug_pm_callback has higher priority than x86
1281          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1282          * to disable cpu hotplug to avoid cpu hotplug race.
1283          */
1284         pm_notifier(cpu_hotplug_pm_callback, 0);
1285         return 0;
1286 }
1287 core_initcall(cpu_hotplug_pm_sync_init);
1288
1289 #endif /* CONFIG_PM_SLEEP_SMP */
1290
1291 int __boot_cpu_id;
1292
1293 #endif /* CONFIG_SMP */
1294
1295 /* Boot processor state steps */
1296 static struct cpuhp_step cpuhp_bp_states[] = {
1297         [CPUHP_OFFLINE] = {
1298                 .name                   = "offline",
1299                 .startup.single         = NULL,
1300                 .teardown.single        = NULL,
1301         },
1302 #ifdef CONFIG_SMP
1303         [CPUHP_CREATE_THREADS]= {
1304                 .name                   = "threads:prepare",
1305                 .startup.single         = smpboot_create_threads,
1306                 .teardown.single        = NULL,
1307                 .cant_stop              = true,
1308         },
1309         [CPUHP_PERF_PREPARE] = {
1310                 .name                   = "perf:prepare",
1311                 .startup.single         = perf_event_init_cpu,
1312                 .teardown.single        = perf_event_exit_cpu,
1313         },
1314         [CPUHP_WORKQUEUE_PREP] = {
1315                 .name                   = "workqueue:prepare",
1316                 .startup.single         = workqueue_prepare_cpu,
1317                 .teardown.single        = NULL,
1318         },
1319         [CPUHP_HRTIMERS_PREPARE] = {
1320                 .name                   = "hrtimers:prepare",
1321                 .startup.single         = hrtimers_prepare_cpu,
1322                 .teardown.single        = hrtimers_dead_cpu,
1323         },
1324         [CPUHP_SMPCFD_PREPARE] = {
1325                 .name                   = "smpcfd:prepare",
1326                 .startup.single         = smpcfd_prepare_cpu,
1327                 .teardown.single        = smpcfd_dead_cpu,
1328         },
1329         [CPUHP_RELAY_PREPARE] = {
1330                 .name                   = "relay:prepare",
1331                 .startup.single         = relay_prepare_cpu,
1332                 .teardown.single        = NULL,
1333         },
1334         [CPUHP_SLAB_PREPARE] = {
1335                 .name                   = "slab:prepare",
1336                 .startup.single         = slab_prepare_cpu,
1337                 .teardown.single        = slab_dead_cpu,
1338         },
1339         [CPUHP_RCUTREE_PREP] = {
1340                 .name                   = "RCU/tree:prepare",
1341                 .startup.single         = rcutree_prepare_cpu,
1342                 .teardown.single        = rcutree_dead_cpu,
1343         },
1344         /*
1345          * On the tear-down path, timers_dead_cpu() must be invoked
1346          * before blk_mq_queue_reinit_notify() from notify_dead(),
1347          * otherwise a RCU stall occurs.
1348          */
1349         [CPUHP_TIMERS_PREPARE] = {
1350                 .name                   = "timers:dead",
1351                 .startup.single         = timers_prepare_cpu,
1352                 .teardown.single        = timers_dead_cpu,
1353         },
1354         /* Kicks the plugged cpu into life */
1355         [CPUHP_BRINGUP_CPU] = {
1356                 .name                   = "cpu:bringup",
1357                 .startup.single         = bringup_cpu,
1358                 .teardown.single        = NULL,
1359                 .cant_stop              = true,
1360         },
1361         /*
1362          * Handled on controll processor until the plugged processor manages
1363          * this itself.
1364          */
1365         [CPUHP_TEARDOWN_CPU] = {
1366                 .name                   = "cpu:teardown",
1367                 .startup.single         = NULL,
1368                 .teardown.single        = takedown_cpu,
1369                 .cant_stop              = true,
1370         },
1371 #else
1372         [CPUHP_BRINGUP_CPU] = { },
1373 #endif
1374 };
1375
1376 /* Application processor state steps */
1377 static struct cpuhp_step cpuhp_ap_states[] = {
1378 #ifdef CONFIG_SMP
1379         /* Final state before CPU kills itself */
1380         [CPUHP_AP_IDLE_DEAD] = {
1381                 .name                   = "idle:dead",
1382         },
1383         /*
1384          * Last state before CPU enters the idle loop to die. Transient state
1385          * for synchronization.
1386          */
1387         [CPUHP_AP_OFFLINE] = {
1388                 .name                   = "ap:offline",
1389                 .cant_stop              = true,
1390         },
1391         /* First state is scheduler control. Interrupts are disabled */
1392         [CPUHP_AP_SCHED_STARTING] = {
1393                 .name                   = "sched:starting",
1394                 .startup.single         = sched_cpu_starting,
1395                 .teardown.single        = sched_cpu_dying,
1396         },
1397         [CPUHP_AP_RCUTREE_DYING] = {
1398                 .name                   = "RCU/tree:dying",
1399                 .startup.single         = NULL,
1400                 .teardown.single        = rcutree_dying_cpu,
1401         },
1402         [CPUHP_AP_SMPCFD_DYING] = {
1403                 .name                   = "smpcfd:dying",
1404                 .startup.single         = NULL,
1405                 .teardown.single        = smpcfd_dying_cpu,
1406         },
1407         /* Entry state on starting. Interrupts enabled from here on. Transient
1408          * state for synchronsization */
1409         [CPUHP_AP_ONLINE] = {
1410                 .name                   = "ap:online",
1411         },
1412         /* Handle smpboot threads park/unpark */
1413         [CPUHP_AP_SMPBOOT_THREADS] = {
1414                 .name                   = "smpboot/threads:online",
1415                 .startup.single         = smpboot_unpark_threads,
1416                 .teardown.single        = smpboot_park_threads,
1417         },
1418         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1419                 .name                   = "irq/affinity:online",
1420                 .startup.single         = irq_affinity_online_cpu,
1421                 .teardown.single        = NULL,
1422         },
1423         [CPUHP_AP_PERF_ONLINE] = {
1424                 .name                   = "perf:online",
1425                 .startup.single         = perf_event_init_cpu,
1426                 .teardown.single        = perf_event_exit_cpu,
1427         },
1428         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1429                 .name                   = "workqueue:online",
1430                 .startup.single         = workqueue_online_cpu,
1431                 .teardown.single        = workqueue_offline_cpu,
1432         },
1433         [CPUHP_AP_RCUTREE_ONLINE] = {
1434                 .name                   = "RCU/tree:online",
1435                 .startup.single         = rcutree_online_cpu,
1436                 .teardown.single        = rcutree_offline_cpu,
1437         },
1438 #endif
1439         /*
1440          * The dynamically registered state space is here
1441          */
1442
1443 #ifdef CONFIG_SMP
1444         /* Last state is scheduler control setting the cpu active */
1445         [CPUHP_AP_ACTIVE] = {
1446                 .name                   = "sched:active",
1447                 .startup.single         = sched_cpu_activate,
1448                 .teardown.single        = sched_cpu_deactivate,
1449         },
1450 #endif
1451
1452         /* CPU is fully up and running. */
1453         [CPUHP_ONLINE] = {
1454                 .name                   = "online",
1455                 .startup.single         = NULL,
1456                 .teardown.single        = NULL,
1457         },
1458 };
1459
1460 /* Sanity check for callbacks */
1461 static int cpuhp_cb_check(enum cpuhp_state state)
1462 {
1463         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1464                 return -EINVAL;
1465         return 0;
1466 }
1467
1468 /*
1469  * Returns a free for dynamic slot assignment of the Online state. The states
1470  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1471  * by having no name assigned.
1472  */
1473 static int cpuhp_reserve_state(enum cpuhp_state state)
1474 {
1475         enum cpuhp_state i, end;
1476         struct cpuhp_step *step;
1477
1478         switch (state) {
1479         case CPUHP_AP_ONLINE_DYN:
1480                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1481                 end = CPUHP_AP_ONLINE_DYN_END;
1482                 break;
1483         case CPUHP_BP_PREPARE_DYN:
1484                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1485                 end = CPUHP_BP_PREPARE_DYN_END;
1486                 break;
1487         default:
1488                 return -EINVAL;
1489         }
1490
1491         for (i = state; i <= end; i++, step++) {
1492                 if (!step->name)
1493                         return i;
1494         }
1495         WARN(1, "No more dynamic states available for CPU hotplug\n");
1496         return -ENOSPC;
1497 }
1498
1499 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1500                                  int (*startup)(unsigned int cpu),
1501                                  int (*teardown)(unsigned int cpu),
1502                                  bool multi_instance)
1503 {
1504         /* (Un)Install the callbacks for further cpu hotplug operations */
1505         struct cpuhp_step *sp;
1506         int ret = 0;
1507
1508         /*
1509          * If name is NULL, then the state gets removed.
1510          *
1511          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1512          * the first allocation from these dynamic ranges, so the removal
1513          * would trigger a new allocation and clear the wrong (already
1514          * empty) state, leaving the callbacks of the to be cleared state
1515          * dangling, which causes wreckage on the next hotplug operation.
1516          */
1517         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1518                      state == CPUHP_BP_PREPARE_DYN)) {
1519                 ret = cpuhp_reserve_state(state);
1520                 if (ret < 0)
1521                         return ret;
1522                 state = ret;
1523         }
1524         sp = cpuhp_get_step(state);
1525         if (name && sp->name)
1526                 return -EBUSY;
1527
1528         sp->startup.single = startup;
1529         sp->teardown.single = teardown;
1530         sp->name = name;
1531         sp->multi_instance = multi_instance;
1532         INIT_HLIST_HEAD(&sp->list);
1533         return ret;
1534 }
1535
1536 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1537 {
1538         return cpuhp_get_step(state)->teardown.single;
1539 }
1540
1541 /*
1542  * Call the startup/teardown function for a step either on the AP or
1543  * on the current CPU.
1544  */
1545 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1546                             struct hlist_node *node)
1547 {
1548         struct cpuhp_step *sp = cpuhp_get_step(state);
1549         int ret;
1550
1551         /*
1552          * If there's nothing to do, we done.
1553          * Relies on the union for multi_instance.
1554          */
1555         if ((bringup && !sp->startup.single) ||
1556             (!bringup && !sp->teardown.single))
1557                 return 0;
1558         /*
1559          * The non AP bound callbacks can fail on bringup. On teardown
1560          * e.g. module removal we crash for now.
1561          */
1562 #ifdef CONFIG_SMP
1563         if (cpuhp_is_ap_state(state))
1564                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1565         else
1566                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1567 #else
1568         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1569 #endif
1570         BUG_ON(ret && !bringup);
1571         return ret;
1572 }
1573
1574 /*
1575  * Called from __cpuhp_setup_state on a recoverable failure.
1576  *
1577  * Note: The teardown callbacks for rollback are not allowed to fail!
1578  */
1579 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1580                                    struct hlist_node *node)
1581 {
1582         int cpu;
1583
1584         /* Roll back the already executed steps on the other cpus */
1585         for_each_present_cpu(cpu) {
1586                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1587                 int cpustate = st->state;
1588
1589                 if (cpu >= failedcpu)
1590                         break;
1591
1592                 /* Did we invoke the startup call on that cpu ? */
1593                 if (cpustate >= state)
1594                         cpuhp_issue_call(cpu, state, false, node);
1595         }
1596 }
1597
1598 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1599                                           struct hlist_node *node,
1600                                           bool invoke)
1601 {
1602         struct cpuhp_step *sp;
1603         int cpu;
1604         int ret;
1605
1606         lockdep_assert_cpus_held();
1607
1608         sp = cpuhp_get_step(state);
1609         if (sp->multi_instance == false)
1610                 return -EINVAL;
1611
1612         mutex_lock(&cpuhp_state_mutex);
1613
1614         if (!invoke || !sp->startup.multi)
1615                 goto add_node;
1616
1617         /*
1618          * Try to call the startup callback for each present cpu
1619          * depending on the hotplug state of the cpu.
1620          */
1621         for_each_present_cpu(cpu) {
1622                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1623                 int cpustate = st->state;
1624
1625                 if (cpustate < state)
1626                         continue;
1627
1628                 ret = cpuhp_issue_call(cpu, state, true, node);
1629                 if (ret) {
1630                         if (sp->teardown.multi)
1631                                 cpuhp_rollback_install(cpu, state, node);
1632                         goto unlock;
1633                 }
1634         }
1635 add_node:
1636         ret = 0;
1637         hlist_add_head(node, &sp->list);
1638 unlock:
1639         mutex_unlock(&cpuhp_state_mutex);
1640         return ret;
1641 }
1642
1643 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1644                                bool invoke)
1645 {
1646         int ret;
1647
1648         cpus_read_lock();
1649         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1650         cpus_read_unlock();
1651         return ret;
1652 }
1653 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1654
1655 /**
1656  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1657  * @state:              The state to setup
1658  * @invoke:             If true, the startup function is invoked for cpus where
1659  *                      cpu state >= @state
1660  * @startup:            startup callback function
1661  * @teardown:           teardown callback function
1662  * @multi_instance:     State is set up for multiple instances which get
1663  *                      added afterwards.
1664  *
1665  * The caller needs to hold cpus read locked while calling this function.
1666  * Returns:
1667  *   On success:
1668  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1669  *      0 for all other states
1670  *   On failure: proper (negative) error code
1671  */
1672 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1673                                    const char *name, bool invoke,
1674                                    int (*startup)(unsigned int cpu),
1675                                    int (*teardown)(unsigned int cpu),
1676                                    bool multi_instance)
1677 {
1678         int cpu, ret = 0;
1679         bool dynstate;
1680
1681         lockdep_assert_cpus_held();
1682
1683         if (cpuhp_cb_check(state) || !name)
1684                 return -EINVAL;
1685
1686         mutex_lock(&cpuhp_state_mutex);
1687
1688         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1689                                     multi_instance);
1690
1691         dynstate = state == CPUHP_AP_ONLINE_DYN;
1692         if (ret > 0 && dynstate) {
1693                 state = ret;
1694                 ret = 0;
1695         }
1696
1697         if (ret || !invoke || !startup)
1698                 goto out;
1699
1700         /*
1701          * Try to call the startup callback for each present cpu
1702          * depending on the hotplug state of the cpu.
1703          */
1704         for_each_present_cpu(cpu) {
1705                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1706                 int cpustate = st->state;
1707
1708                 if (cpustate < state)
1709                         continue;
1710
1711                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1712                 if (ret) {
1713                         if (teardown)
1714                                 cpuhp_rollback_install(cpu, state, NULL);
1715                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1716                         goto out;
1717                 }
1718         }
1719 out:
1720         mutex_unlock(&cpuhp_state_mutex);
1721         /*
1722          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1723          * dynamically allocated state in case of success.
1724          */
1725         if (!ret && dynstate)
1726                 return state;
1727         return ret;
1728 }
1729 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1730
1731 int __cpuhp_setup_state(enum cpuhp_state state,
1732                         const char *name, bool invoke,
1733                         int (*startup)(unsigned int cpu),
1734                         int (*teardown)(unsigned int cpu),
1735                         bool multi_instance)
1736 {
1737         int ret;
1738
1739         cpus_read_lock();
1740         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1741                                              teardown, multi_instance);
1742         cpus_read_unlock();
1743         return ret;
1744 }
1745 EXPORT_SYMBOL(__cpuhp_setup_state);
1746
1747 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1748                                   struct hlist_node *node, bool invoke)
1749 {
1750         struct cpuhp_step *sp = cpuhp_get_step(state);
1751         int cpu;
1752
1753         BUG_ON(cpuhp_cb_check(state));
1754
1755         if (!sp->multi_instance)
1756                 return -EINVAL;
1757
1758         cpus_read_lock();
1759         mutex_lock(&cpuhp_state_mutex);
1760
1761         if (!invoke || !cpuhp_get_teardown_cb(state))
1762                 goto remove;
1763         /*
1764          * Call the teardown callback for each present cpu depending
1765          * on the hotplug state of the cpu. This function is not
1766          * allowed to fail currently!
1767          */
1768         for_each_present_cpu(cpu) {
1769                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1770                 int cpustate = st->state;
1771
1772                 if (cpustate >= state)
1773                         cpuhp_issue_call(cpu, state, false, node);
1774         }
1775
1776 remove:
1777         hlist_del(node);
1778         mutex_unlock(&cpuhp_state_mutex);
1779         cpus_read_unlock();
1780
1781         return 0;
1782 }
1783 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1784
1785 /**
1786  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1787  * @state:      The state to remove
1788  * @invoke:     If true, the teardown function is invoked for cpus where
1789  *              cpu state >= @state
1790  *
1791  * The caller needs to hold cpus read locked while calling this function.
1792  * The teardown callback is currently not allowed to fail. Think
1793  * about module removal!
1794  */
1795 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1796 {
1797         struct cpuhp_step *sp = cpuhp_get_step(state);
1798         int cpu;
1799
1800         BUG_ON(cpuhp_cb_check(state));
1801
1802         lockdep_assert_cpus_held();
1803
1804         mutex_lock(&cpuhp_state_mutex);
1805         if (sp->multi_instance) {
1806                 WARN(!hlist_empty(&sp->list),
1807                      "Error: Removing state %d which has instances left.\n",
1808                      state);
1809                 goto remove;
1810         }
1811
1812         if (!invoke || !cpuhp_get_teardown_cb(state))
1813                 goto remove;
1814
1815         /*
1816          * Call the teardown callback for each present cpu depending
1817          * on the hotplug state of the cpu. This function is not
1818          * allowed to fail currently!
1819          */
1820         for_each_present_cpu(cpu) {
1821                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1822                 int cpustate = st->state;
1823
1824                 if (cpustate >= state)
1825                         cpuhp_issue_call(cpu, state, false, NULL);
1826         }
1827 remove:
1828         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1829         mutex_unlock(&cpuhp_state_mutex);
1830 }
1831 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1832
1833 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1834 {
1835         cpus_read_lock();
1836         __cpuhp_remove_state_cpuslocked(state, invoke);
1837         cpus_read_unlock();
1838 }
1839 EXPORT_SYMBOL(__cpuhp_remove_state);
1840
1841 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1842 static ssize_t show_cpuhp_state(struct device *dev,
1843                                 struct device_attribute *attr, char *buf)
1844 {
1845         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1846
1847         return sprintf(buf, "%d\n", st->state);
1848 }
1849 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1850
1851 static ssize_t write_cpuhp_target(struct device *dev,
1852                                   struct device_attribute *attr,
1853                                   const char *buf, size_t count)
1854 {
1855         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1856         struct cpuhp_step *sp;
1857         int target, ret;
1858
1859         ret = kstrtoint(buf, 10, &target);
1860         if (ret)
1861                 return ret;
1862
1863 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1864         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1865                 return -EINVAL;
1866 #else
1867         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1868                 return -EINVAL;
1869 #endif
1870
1871         ret = lock_device_hotplug_sysfs();
1872         if (ret)
1873                 return ret;
1874
1875         mutex_lock(&cpuhp_state_mutex);
1876         sp = cpuhp_get_step(target);
1877         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1878         mutex_unlock(&cpuhp_state_mutex);
1879         if (ret)
1880                 goto out;
1881
1882         if (st->state < target)
1883                 ret = do_cpu_up(dev->id, target);
1884         else
1885                 ret = do_cpu_down(dev->id, target);
1886 out:
1887         unlock_device_hotplug();
1888         return ret ? ret : count;
1889 }
1890
1891 static ssize_t show_cpuhp_target(struct device *dev,
1892                                  struct device_attribute *attr, char *buf)
1893 {
1894         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1895
1896         return sprintf(buf, "%d\n", st->target);
1897 }
1898 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1899
1900
1901 static ssize_t write_cpuhp_fail(struct device *dev,
1902                                 struct device_attribute *attr,
1903                                 const char *buf, size_t count)
1904 {
1905         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1906         struct cpuhp_step *sp;
1907         int fail, ret;
1908
1909         ret = kstrtoint(buf, 10, &fail);
1910         if (ret)
1911                 return ret;
1912
1913         /*
1914          * Cannot fail STARTING/DYING callbacks.
1915          */
1916         if (cpuhp_is_atomic_state(fail))
1917                 return -EINVAL;
1918
1919         /*
1920          * Cannot fail anything that doesn't have callbacks.
1921          */
1922         mutex_lock(&cpuhp_state_mutex);
1923         sp = cpuhp_get_step(fail);
1924         if (!sp->startup.single && !sp->teardown.single)
1925                 ret = -EINVAL;
1926         mutex_unlock(&cpuhp_state_mutex);
1927         if (ret)
1928                 return ret;
1929
1930         st->fail = fail;
1931
1932         return count;
1933 }
1934
1935 static ssize_t show_cpuhp_fail(struct device *dev,
1936                                struct device_attribute *attr, char *buf)
1937 {
1938         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1939
1940         return sprintf(buf, "%d\n", st->fail);
1941 }
1942
1943 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1944
1945 static struct attribute *cpuhp_cpu_attrs[] = {
1946         &dev_attr_state.attr,
1947         &dev_attr_target.attr,
1948         &dev_attr_fail.attr,
1949         NULL
1950 };
1951
1952 static const struct attribute_group cpuhp_cpu_attr_group = {
1953         .attrs = cpuhp_cpu_attrs,
1954         .name = "hotplug",
1955         NULL
1956 };
1957
1958 static ssize_t show_cpuhp_states(struct device *dev,
1959                                  struct device_attribute *attr, char *buf)
1960 {
1961         ssize_t cur, res = 0;
1962         int i;
1963
1964         mutex_lock(&cpuhp_state_mutex);
1965         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1966                 struct cpuhp_step *sp = cpuhp_get_step(i);
1967
1968                 if (sp->name) {
1969                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1970                         buf += cur;
1971                         res += cur;
1972                 }
1973         }
1974         mutex_unlock(&cpuhp_state_mutex);
1975         return res;
1976 }
1977 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1978
1979 static struct attribute *cpuhp_cpu_root_attrs[] = {
1980         &dev_attr_states.attr,
1981         NULL
1982 };
1983
1984 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1985         .attrs = cpuhp_cpu_root_attrs,
1986         .name = "hotplug",
1987         NULL
1988 };
1989
1990 #ifdef CONFIG_HOTPLUG_SMT
1991
1992 static const char *smt_states[] = {
1993         [CPU_SMT_ENABLED]               = "on",
1994         [CPU_SMT_DISABLED]              = "off",
1995         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
1996         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
1997 };
1998
1999 static ssize_t
2000 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2001 {
2002         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2003 }
2004
2005 static void cpuhp_offline_cpu_device(unsigned int cpu)
2006 {
2007         struct device *dev = get_cpu_device(cpu);
2008
2009         dev->offline = true;
2010         /* Tell user space about the state change */
2011         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2012 }
2013
2014 static void cpuhp_online_cpu_device(unsigned int cpu)
2015 {
2016         struct device *dev = get_cpu_device(cpu);
2017
2018         dev->offline = false;
2019         /* Tell user space about the state change */
2020         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2021 }
2022
2023 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2024 {
2025         int cpu, ret = 0;
2026
2027         cpu_maps_update_begin();
2028         for_each_online_cpu(cpu) {
2029                 if (topology_is_primary_thread(cpu))
2030                         continue;
2031                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2032                 if (ret)
2033                         break;
2034                 /*
2035                  * As this needs to hold the cpu maps lock it's impossible
2036                  * to call device_offline() because that ends up calling
2037                  * cpu_down() which takes cpu maps lock. cpu maps lock
2038                  * needs to be held as this might race against in kernel
2039                  * abusers of the hotplug machinery (thermal management).
2040                  *
2041                  * So nothing would update device:offline state. That would
2042                  * leave the sysfs entry stale and prevent onlining after
2043                  * smt control has been changed to 'off' again. This is
2044                  * called under the sysfs hotplug lock, so it is properly
2045                  * serialized against the regular offline usage.
2046                  */
2047                 cpuhp_offline_cpu_device(cpu);
2048         }
2049         if (!ret)
2050                 cpu_smt_control = ctrlval;
2051         cpu_maps_update_done();
2052         return ret;
2053 }
2054
2055 static int cpuhp_smt_enable(void)
2056 {
2057         int cpu, ret = 0;
2058
2059         cpu_maps_update_begin();
2060         cpu_smt_control = CPU_SMT_ENABLED;
2061         for_each_present_cpu(cpu) {
2062                 /* Skip online CPUs and CPUs on offline nodes */
2063                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2064                         continue;
2065                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2066                 if (ret)
2067                         break;
2068                 /* See comment in cpuhp_smt_disable() */
2069                 cpuhp_online_cpu_device(cpu);
2070         }
2071         cpu_maps_update_done();
2072         return ret;
2073 }
2074
2075 static ssize_t
2076 store_smt_control(struct device *dev, struct device_attribute *attr,
2077                   const char *buf, size_t count)
2078 {
2079         int ctrlval, ret;
2080
2081         if (sysfs_streq(buf, "on"))
2082                 ctrlval = CPU_SMT_ENABLED;
2083         else if (sysfs_streq(buf, "off"))
2084                 ctrlval = CPU_SMT_DISABLED;
2085         else if (sysfs_streq(buf, "forceoff"))
2086                 ctrlval = CPU_SMT_FORCE_DISABLED;
2087         else
2088                 return -EINVAL;
2089
2090         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2091                 return -EPERM;
2092
2093         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2094                 return -ENODEV;
2095
2096         ret = lock_device_hotplug_sysfs();
2097         if (ret)
2098                 return ret;
2099
2100         if (ctrlval != cpu_smt_control) {
2101                 switch (ctrlval) {
2102                 case CPU_SMT_ENABLED:
2103                         ret = cpuhp_smt_enable();
2104                         break;
2105                 case CPU_SMT_DISABLED:
2106                 case CPU_SMT_FORCE_DISABLED:
2107                         ret = cpuhp_smt_disable(ctrlval);
2108                         break;
2109                 }
2110         }
2111
2112         unlock_device_hotplug();
2113         return ret ? ret : count;
2114 }
2115 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2116
2117 static ssize_t
2118 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2119 {
2120         bool active = topology_max_smt_threads() > 1;
2121
2122         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2123 }
2124 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2125
2126 static struct attribute *cpuhp_smt_attrs[] = {
2127         &dev_attr_control.attr,
2128         &dev_attr_active.attr,
2129         NULL
2130 };
2131
2132 static const struct attribute_group cpuhp_smt_attr_group = {
2133         .attrs = cpuhp_smt_attrs,
2134         .name = "smt",
2135         NULL
2136 };
2137
2138 static int __init cpu_smt_state_init(void)
2139 {
2140         /*
2141          * If SMT was disabled by BIOS, detect it here, after the CPUs have
2142          * been brought online.  This ensures the smt/l1tf sysfs entries are
2143          * consistent with reality.  Note this may overwrite cpu_smt_control's
2144          * previous setting.
2145          */
2146         if (topology_max_smt_threads() == 1)
2147                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
2148
2149         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2150                                   &cpuhp_smt_attr_group);
2151 }
2152
2153 #else
2154 static inline int cpu_smt_state_init(void) { return 0; }
2155 #endif
2156
2157 static int __init cpuhp_sysfs_init(void)
2158 {
2159         int cpu, ret;
2160
2161         ret = cpu_smt_state_init();
2162         if (ret)
2163                 return ret;
2164
2165         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2166                                  &cpuhp_cpu_root_attr_group);
2167         if (ret)
2168                 return ret;
2169
2170         for_each_possible_cpu(cpu) {
2171                 struct device *dev = get_cpu_device(cpu);
2172
2173                 if (!dev)
2174                         continue;
2175                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2176                 if (ret)
2177                         return ret;
2178         }
2179         return 0;
2180 }
2181 device_initcall(cpuhp_sysfs_init);
2182 #endif
2183
2184 /*
2185  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2186  * represents all NR_CPUS bits binary values of 1<<nr.
2187  *
2188  * It is used by cpumask_of() to get a constant address to a CPU
2189  * mask value that has a single bit set only.
2190  */
2191
2192 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2193 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2194 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2195 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2196 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2197
2198 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2199
2200         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2201         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2202 #if BITS_PER_LONG > 32
2203         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2204         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2205 #endif
2206 };
2207 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2208
2209 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2210 EXPORT_SYMBOL(cpu_all_bits);
2211
2212 #ifdef CONFIG_INIT_ALL_POSSIBLE
2213 struct cpumask __cpu_possible_mask __read_mostly
2214         = {CPU_BITS_ALL};
2215 #else
2216 struct cpumask __cpu_possible_mask __read_mostly;
2217 #endif
2218 EXPORT_SYMBOL(__cpu_possible_mask);
2219
2220 struct cpumask __cpu_online_mask __read_mostly;
2221 EXPORT_SYMBOL(__cpu_online_mask);
2222
2223 struct cpumask __cpu_present_mask __read_mostly;
2224 EXPORT_SYMBOL(__cpu_present_mask);
2225
2226 struct cpumask __cpu_active_mask __read_mostly;
2227 EXPORT_SYMBOL(__cpu_active_mask);
2228
2229 void init_cpu_present(const struct cpumask *src)
2230 {
2231         cpumask_copy(&__cpu_present_mask, src);
2232 }
2233
2234 void init_cpu_possible(const struct cpumask *src)
2235 {
2236         cpumask_copy(&__cpu_possible_mask, src);
2237 }
2238
2239 void init_cpu_online(const struct cpumask *src)
2240 {
2241         cpumask_copy(&__cpu_online_mask, src);
2242 }
2243
2244 /*
2245  * Activate the first processor.
2246  */
2247 void __init boot_cpu_init(void)
2248 {
2249         int cpu = smp_processor_id();
2250
2251         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2252         set_cpu_online(cpu, true);
2253         set_cpu_active(cpu, true);
2254         set_cpu_present(cpu, true);
2255         set_cpu_possible(cpu, true);
2256
2257 #ifdef CONFIG_SMP
2258         __boot_cpu_id = cpu;
2259 #endif
2260 }
2261
2262 /*
2263  * Must be called _AFTER_ setting up the per_cpu areas
2264  */
2265 void __init boot_cpu_hotplug_init(void)
2266 {
2267         this_cpu_write(cpuhp_state.booted_once, true);
2268         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2269 }