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