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.h>
  11 #include <linux/sched/smt.h>
  12 #include <linux/unistd.h>
  13 #include <linux/cpu.h>
  14 #include <linux/oom.h>
  15 #include <linux/rcupdate.h>
  16 #include <linux/export.h>
  17 #include <linux/bug.h>
  18 #include <linux/kthread.h>
  19 #include <linux/stop_machine.h>
  20 #include <linux/mutex.h>
  21 #include <linux/gfp.h>
  22 #include <linux/suspend.h>
  23 #include <linux/lockdep.h>
  24 #include <linux/tick.h>
  25 #include <linux/irq.h>
  26 #include <linux/smpboot.h>
  27 #include <linux/relay.h>
  28 #include <linux/slab.h>
  29 #include <linux/random.h>
  30
  31 #include <trace/events/power.h>
  32 #define CREATE_TRACE_POINTS
  33 #include <trace/events/cpuhp.h>
  34
  35 #include "smpboot.h"
  36
  37 /**
  38  * cpuhp_cpu_state - Per cpu hotplug state storage
  39  * @state:      The current cpu state
  40  * @target:     The target state
  41  * @thread:     Pointer to the hotplug thread
  42  * @should_run: Thread should execute
  43  * @rollback:   Perform a rollback
  44  * @single:     Single callback invocation
  45  * @bringup:    Single callback bringup or teardown selector
  46  * @cb_state:   The state for a single callback (install/uninstall)
  47  * @result:     Result of the operation
  48  * @done:       Signal completion to the issuer of the task
  49  */
  50 struct cpuhp_cpu_state {
  51         enum cpuhp_state        state;
  52         enum cpuhp_state        target;
  53 #ifdef CONFIG_SMP
  54         struct task_struct      *thread;
  55         bool                    should_run;
  56         bool                    rollback;
  57         bool                    single;
  58         bool                    bringup;
  59         bool                    booted_once;
  60         struct hlist_node       *node;
  61         enum cpuhp_state        cb_state;
  62         int                     result;
  63         struct completion       done;
  64 #endif
  65 };
  66
  67 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  68
  69 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  70 static struct lock_class_key cpuhp_state_key;
  71 static struct lockdep_map cpuhp_state_lock_map =
  72         STATIC_LOCKDEP_MAP_INIT("cpuhp_state", &cpuhp_state_key);
  73 #endif
  74
  75 /**
  76  * cpuhp_step - Hotplug state machine step
  77  * @name:       Name of the step
  78  * @startup:    Startup function of the step
  79  * @teardown:   Teardown function of the step
  80  * @skip_onerr: Do not invoke the functions on error rollback
  81  *              Will go away once the notifiers are gone
  82  * @cant_stop:  Bringup/teardown can't be stopped at this step
  83  */
  84 struct cpuhp_step {
  85         const char              *name;
  86         union {
  87                 int             (*single)(unsigned int cpu);
  88                 int             (*multi)(unsigned int cpu,
  89                                          struct hlist_node *node);
  90         } startup;
  91         union {
  92                 int             (*single)(unsigned int cpu);
  93                 int             (*multi)(unsigned int cpu,
  94                                          struct hlist_node *node);
  95         } teardown;
  96         struct hlist_head       list;
  97         bool                    skip_onerr;
  98         bool                    cant_stop;
  99         bool                    multi_instance;
 100 };
 101
 102 static DEFINE_MUTEX(cpuhp_state_mutex);
 103 static struct cpuhp_step cpuhp_bp_states[];
 104 static struct cpuhp_step cpuhp_ap_states[];
 105
 106 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 107 {
 108         /*
 109          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 110          * purposes as that state is handled explicitly in cpu_down.
 111          */
 112         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 113 }
 114
 115 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 116 {
 117         struct cpuhp_step *sp;
 118
 119         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 120         return sp + state;
 121 }
 122
 123 /**
 124  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 125  * @cpu:        The cpu for which the callback should be invoked
 126  * @step:       The step in the state machine
 127  * @bringup:    True if the bringup callback should be invoked
 128  *
 129  * Called from cpu hotplug and from the state register machinery.
 130  */
 131 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 132                                  bool bringup, struct hlist_node *node)
 133 {
 134         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 135         struct cpuhp_step *step = cpuhp_get_step(state);
 136         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 137         int (*cb)(unsigned int cpu);
 138         int ret, cnt;
 139
 140         if (!step->multi_instance) {
 141                 cb = bringup ? step->startup.single : step->teardown.single;
 142                 if (!cb)
 143                         return 0;
 144                 trace_cpuhp_enter(cpu, st->target, state, cb);
 145                 ret = cb(cpu);
 146                 trace_cpuhp_exit(cpu, st->state, state, ret);
 147                 return ret;
 148         }
 149         cbm = bringup ? step->startup.multi : step->teardown.multi;
 150         if (!cbm)
 151                 return 0;
 152
 153         /* Single invocation for instance add/remove */
 154         if (node) {
 155                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 156                 ret = cbm(cpu, node);
 157                 trace_cpuhp_exit(cpu, st->state, state, ret);
 158                 return ret;
 159         }
 160
 161         /* State transition. Invoke on all instances */
 162         cnt = 0;
 163         hlist_for_each(node, &step->list) {
 164                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 165                 ret = cbm(cpu, node);
 166                 trace_cpuhp_exit(cpu, st->state, state, ret);
 167                 if (ret)
 168                         goto err;
 169                 cnt++;
 170         }
 171         return 0;
 172 err:
 173         /* Rollback the instances if one failed */
 174         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 175         if (!cbm)
 176                 return ret;
 177
 178         hlist_for_each(node, &step->list) {
 179                 if (!cnt--)
 180                         break;
 181                 cbm(cpu, node);
 182         }
 183         return ret;
 184 }
 185
 186 #ifdef CONFIG_SMP
 187 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 188 static DEFINE_MUTEX(cpu_add_remove_lock);
 189 bool cpuhp_tasks_frozen;
 190 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 191
 192 /*
 193  * The following two APIs (cpu_maps_update_begin/done) must be used when
 194  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 195  * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
 196  * hotplug callback (un)registration performed using __register_cpu_notifier()
 197  * or __unregister_cpu_notifier().
 198  */
 199 void cpu_maps_update_begin(void)
 200 {
 201         mutex_lock(&cpu_add_remove_lock);
 202 }
 203 EXPORT_SYMBOL(cpu_notifier_register_begin);
 204
 205 void cpu_maps_update_done(void)
 206 {
 207         mutex_unlock(&cpu_add_remove_lock);
 208 }
 209 EXPORT_SYMBOL(cpu_notifier_register_done);
 210
 211 static RAW_NOTIFIER_HEAD(cpu_chain);
 212
 213 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 214  * Should always be manipulated under cpu_add_remove_lock
 215  */
 216 static int cpu_hotplug_disabled;
 217
 218 #ifdef CONFIG_HOTPLUG_CPU
 219
 220 static struct {
 221         struct task_struct *active_writer;
 222         /* wait queue to wake up the active_writer */
 223         wait_queue_head_t wq;
 224         /* verifies that no writer will get active while readers are active */
 225         struct mutex lock;
 226         /*
 227          * Also blocks the new readers during
 228          * an ongoing cpu hotplug operation.
 229          */
 230         atomic_t refcount;
 231
 232 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 233         struct lockdep_map dep_map;
 234 #endif
 235 } cpu_hotplug = {
 236         .active_writer = NULL,
 237         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 238         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 239 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 240         .dep_map = STATIC_LOCKDEP_MAP_INIT("cpu_hotplug.dep_map", &cpu_hotplug.dep_map),
 241 #endif
 242 };
 243
 244 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 245 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 246 #define cpuhp_lock_acquire_tryread() \
 247                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 248 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 249 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 250
 251
 252 void get_online_cpus(void)
 253 {
 254         might_sleep();
 255         if (cpu_hotplug.active_writer == current)
 256                 return;
 257         cpuhp_lock_acquire_read();
 258         mutex_lock(&cpu_hotplug.lock);
 259         atomic_inc(&cpu_hotplug.refcount);
 260         mutex_unlock(&cpu_hotplug.lock);
 261 }
 262 EXPORT_SYMBOL_GPL(get_online_cpus);
 263
 264 void put_online_cpus(void)
 265 {
 266         int refcount;
 267
 268         if (cpu_hotplug.active_writer == current)
 269                 return;
 270
 271         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 272         if (WARN_ON(refcount < 0)) /* try to fix things up */
 273                 atomic_inc(&cpu_hotplug.refcount);
 274
 275         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 276                 wake_up(&cpu_hotplug.wq);
 277
 278         cpuhp_lock_release();
 279
 280 }
 281 EXPORT_SYMBOL_GPL(put_online_cpus);
 282
 283 /*
 284  * This ensures that the hotplug operation can begin only when the
 285  * refcount goes to zero.
 286  *
 287  * Note that during a cpu-hotplug operation, the new readers, if any,
 288  * will be blocked by the cpu_hotplug.lock
 289  *
 290  * Since cpu_hotplug_begin() is always called after invoking
 291  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 292  *
 293  * Note that theoretically, there is a possibility of a livelock:
 294  * - Refcount goes to zero, last reader wakes up the sleeping
 295  *   writer.
 296  * - Last reader unlocks the cpu_hotplug.lock.
 297  * - A new reader arrives at this moment, bumps up the refcount.
 298  * - The writer acquires the cpu_hotplug.lock finds the refcount
 299  *   non zero and goes to sleep again.
 300  *
 301  * However, this is very difficult to achieve in practice since
 302  * get_online_cpus() not an api which is called all that often.
 303  *
 304  */
 305 void cpu_hotplug_begin(void)
 306 {
 307         DEFINE_WAIT(wait);
 308
 309         cpu_hotplug.active_writer = current;
 310         cpuhp_lock_acquire();
 311
 312         for (;;) {
 313                 mutex_lock(&cpu_hotplug.lock);
 314                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 315                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 316                                 break;
 317                 mutex_unlock(&cpu_hotplug.lock);
 318                 schedule();
 319         }
 320         finish_wait(&cpu_hotplug.wq, &wait);
 321 }
 322
 323 void cpu_hotplug_done(void)
 324 {
 325         cpu_hotplug.active_writer = NULL;
 326         mutex_unlock(&cpu_hotplug.lock);
 327         cpuhp_lock_release();
 328 }
 329
 330 /*
 331  * Wait for currently running CPU hotplug operations to complete (if any) and
 332  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 333  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 334  * hotplug path before performing hotplug operations. So acquiring that lock
 335  * guarantees mutual exclusion from any currently running hotplug operations.
 336  */
 337 void cpu_hotplug_disable(void)
 338 {
 339         cpu_maps_update_begin();
 340         cpu_hotplug_disabled++;
 341         cpu_maps_update_done();
 342 }
 343 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 344
 345 static void __cpu_hotplug_enable(void)
 346 {
 347         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 348                 return;
 349         cpu_hotplug_disabled--;
 350 }
 351
 352 void cpu_hotplug_enable(void)
 353 {
 354         cpu_maps_update_begin();
 355         __cpu_hotplug_enable();
 356         cpu_maps_update_done();
 357 }
 358 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 359 #endif  /* CONFIG_HOTPLUG_CPU */
 360
 361 /*
 362  * Architectures that need SMT-specific errata handling during SMT hotplug
 363  * should override this.
 364  */
 365 void __weak arch_smt_update(void) { }
 366
 367 #ifdef CONFIG_HOTPLUG_SMT
 368 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 369 EXPORT_SYMBOL_GPL(cpu_smt_control);
 370
 371 static bool cpu_smt_available __read_mostly;
 372
 373 void __init cpu_smt_disable(bool force)
 374 {
 375         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 376                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 377                 return;
 378
 379         if (force) {
 380                 pr_info("SMT: Force disabled\n");
 381                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 382         } else {
 383                 pr_info("SMT: disabled\n");
 384                 cpu_smt_control = CPU_SMT_DISABLED;
 385         }
 386 }
 387
 388 /*
 389  * The decision whether SMT is supported can only be done after the full
 390  * CPU identification. Called from architecture code before non boot CPUs
 391  * are brought up.
 392  */
 393 void __init cpu_smt_check_topology_early(void)
 394 {
 395         if (!topology_smt_supported())
 396                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 397 }
 398
 399 /*
 400  * If SMT was disabled by BIOS, detect it here, after the CPUs have been
 401  * brought online. This ensures the smt/l1tf sysfs entries are consistent
 402  * with reality. cpu_smt_available is set to true during the bringup of non
 403  * boot CPUs when a SMT sibling is detected. Note, this may overwrite
 404  * cpu_smt_control's previous setting.
 405  */
 406 void __init cpu_smt_check_topology(void)
 407 {
 408         if (!cpu_smt_available)
 409                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 410 }
 411
 412 static int __init smt_cmdline_disable(char *str)
 413 {
 414         cpu_smt_disable(str && !strcmp(str, "force"));
 415         return 0;
 416 }
 417 early_param("nosmt", smt_cmdline_disable);
 418
 419 static inline bool cpu_smt_allowed(unsigned int cpu)
 420 {
 421         if (topology_is_primary_thread(cpu))
 422                 return true;
 423
 424         /*
 425          * If the CPU is not a 'primary' thread and the booted_once bit is
 426          * set then the processor has SMT support. Store this information
 427          * for the late check of SMT support in cpu_smt_check_topology().
 428          */
 429         if (per_cpu(cpuhp_state, cpu).booted_once)
 430                 cpu_smt_available = true;
 431
 432         if (cpu_smt_control == CPU_SMT_ENABLED)
 433                 return true;
 434
 435         /*
 436          * On x86 it's required to boot all logical CPUs at least once so
 437          * that the init code can get a chance to set CR4.MCE on each
 438          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 439          * core will shutdown the machine.
 440          */
 441         return !per_cpu(cpuhp_state, cpu).booted_once;
 442 }
 443 #else
 444 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 445 #endif
 446
 447 /* Need to know about CPUs going up/down? */
 448 int register_cpu_notifier(struct notifier_block *nb)
 449 {
 450         int ret;
 451         cpu_maps_update_begin();
 452         ret = raw_notifier_chain_register(&cpu_chain, nb);
 453         cpu_maps_update_done();
 454         return ret;
 455 }
 456
 457 int __register_cpu_notifier(struct notifier_block *nb)
 458 {
 459         return raw_notifier_chain_register(&cpu_chain, nb);
 460 }
 461
 462 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 463                         int *nr_calls)
 464 {
 465         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 466         void *hcpu = (void *)(long)cpu;
 467
 468         int ret;
 469
 470         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 471                                         nr_calls);
 472
 473         return notifier_to_errno(ret);
 474 }
 475
 476 static int cpu_notify(unsigned long val, unsigned int cpu)
 477 {
 478         return __cpu_notify(val, cpu, -1, NULL);
 479 }
 480
 481 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 482 {
 483         BUG_ON(cpu_notify(val, cpu));
 484 }
 485
 486 /* Notifier wrappers for transitioning to state machine */
 487 static int notify_prepare(unsigned int cpu)
 488 {
 489         int nr_calls = 0;
 490         int ret;
 491
 492         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 493         if (ret) {
 494                 nr_calls--;
 495                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 496                                 __func__, cpu);
 497                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 498         }
 499         return ret;
 500 }
 501
 502 static int notify_online(unsigned int cpu)
 503 {
 504         cpu_notify(CPU_ONLINE, cpu);
 505         return 0;
 506 }
 507
 508 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st);
 509
 510 static int bringup_wait_for_ap(unsigned int cpu)
 511 {
 512         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 513
 514         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 515         wait_for_completion(&st->done);
 516         if (WARN_ON_ONCE((!cpu_online(cpu))))
 517                 return -ECANCELED;
 518
 519         /* Unpark the hotplug thread of the target cpu */
 520         kthread_unpark(st->thread);
 521
 522         /*
 523          * SMT soft disabling on X86 requires to bring the CPU out of the
 524          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 525          * CPU marked itself as booted_once in cpu_notify_starting() so the
 526          * cpu_smt_allowed() check will now return false if this is not the
 527          * primary sibling.
 528          */
 529         if (!cpu_smt_allowed(cpu))
 530                 return -ECANCELED;
 531
 532         /* Should we go further up ? */
 533         if (st->target > CPUHP_AP_ONLINE_IDLE) {
 534                 __cpuhp_kick_ap_work(st);
 535                 wait_for_completion(&st->done);
 536         }
 537         return st->result;
 538 }
 539
 540 static int bringup_cpu(unsigned int cpu)
 541 {
 542         struct task_struct *idle = idle_thread_get(cpu);
 543         int ret;
 544
 545         /*
 546          * Some architectures have to walk the irq descriptors to
 547          * setup the vector space for the cpu which comes online.
 548          * Prevent irq alloc/free across the bringup.
 549          */
 550         irq_lock_sparse();
 551
 552         /* Arch-specific enabling code. */
 553         ret = __cpu_up(cpu, idle);
 554         irq_unlock_sparse();
 555         if (ret) {
 556                 cpu_notify(CPU_UP_CANCELED, cpu);
 557                 return ret;
 558         }
 559         return bringup_wait_for_ap(cpu);
 560 }
 561
 562 /*
 563  * Hotplug state machine related functions
 564  */
 565 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 566 {
 567         for (st->state++; st->state < st->target; st->state++) {
 568                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 569
 570                 if (!step->skip_onerr)
 571                         cpuhp_invoke_callback(cpu, st->state, true, NULL);
 572         }
 573 }
 574
 575 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 576                                 enum cpuhp_state target)
 577 {
 578         enum cpuhp_state prev_state = st->state;
 579         int ret = 0;
 580
 581         for (; st->state > target; st->state--) {
 582                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
 583                 if (ret) {
 584                         st->target = prev_state;
 585                         undo_cpu_down(cpu, st);
 586                         break;
 587                 }
 588         }
 589         return ret;
 590 }
 591
 592 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 593 {
 594         for (st->state--; st->state > st->target; st->state--) {
 595                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 596
 597                 if (!step->skip_onerr)
 598                         cpuhp_invoke_callback(cpu, st->state, false, NULL);
 599         }
 600 }
 601
 602 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 603 {
 604         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 605                 return true;
 606         /*
 607          * When CPU hotplug is disabled, then taking the CPU down is not
 608          * possible because takedown_cpu() and the architecture and
 609          * subsystem specific mechanisms are not available. So the CPU
 610          * which would be completely unplugged again needs to stay around
 611          * in the current state.
 612          */
 613         return st->state <= CPUHP_BRINGUP_CPU;
 614 }
 615
 616 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 617                               enum cpuhp_state target)
 618 {
 619         enum cpuhp_state prev_state = st->state;
 620         int ret = 0;
 621
 622         while (st->state < target) {
 623                 st->state++;
 624                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
 625                 if (ret) {
 626                         if (can_rollback_cpu(st)) {
 627                                 st->target = prev_state;
 628                                 undo_cpu_up(cpu, st);
 629                         }
 630                         break;
 631                 }
 632         }
 633         return ret;
 634 }
 635
 636 /*
 637  * The cpu hotplug threads manage the bringup and teardown of the cpus
 638  */
 639 static void cpuhp_create(unsigned int cpu)
 640 {
 641         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 642
 643         init_completion(&st->done);
 644 }
 645
 646 static int cpuhp_should_run(unsigned int cpu)
 647 {
 648         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 649
 650         return st->should_run;
 651 }
 652
 653 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 654 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 655 {
 656         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 657
 658         return cpuhp_down_callbacks(cpu, st, target);
 659 }
 660
 661 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 662 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 663 {
 664         return cpuhp_up_callbacks(cpu, st, st->target);
 665 }
 666
 667 /*
 668  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 669  * callbacks when a state gets [un]installed at runtime.
 670  */
 671 static void cpuhp_thread_fun(unsigned int cpu)
 672 {
 673         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 674         int ret = 0;
 675
 676         /*
 677          * Paired with the mb() in cpuhp_kick_ap_work and
 678          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 679          */
 680         smp_mb();
 681         if (!st->should_run)
 682                 return;
 683
 684         st->should_run = false;
 685
 686         lock_map_acquire(&cpuhp_state_lock_map);
 687         /* Single callback invocation for [un]install ? */
 688         if (st->single) {
 689                 if (st->cb_state < CPUHP_AP_ONLINE) {
 690                         local_irq_disable();
 691                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 692                                                     st->bringup, st->node);
 693                         local_irq_enable();
 694                 } else {
 695                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 696                                                     st->bringup, st->node);
 697                 }
 698         } else if (st->rollback) {
 699                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 700
 701                 undo_cpu_down(cpu, st);
 702                 /*
 703                  * This is a momentary workaround to keep the notifier users
 704                  * happy. Will go away once we got rid of the notifiers.
 705                  */
 706                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 707                 st->rollback = false;
 708         } else {
 709                 /* Cannot happen .... */
 710                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 711
 712                 /* Regular hotplug work */
 713                 if (st->state < st->target)
 714                         ret = cpuhp_ap_online(cpu, st);
 715                 else if (st->state > st->target)
 716                         ret = cpuhp_ap_offline(cpu, st);
 717         }
 718         lock_map_release(&cpuhp_state_lock_map);
 719         st->result = ret;
 720         complete(&st->done);
 721 }
 722
 723 /* Invoke a single callback on a remote cpu */
 724 static int
 725 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 726                          struct hlist_node *node)
 727 {
 728         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 729
 730         if (!cpu_online(cpu))
 731                 return 0;
 732
 733         lock_map_acquire(&cpuhp_state_lock_map);
 734         lock_map_release(&cpuhp_state_lock_map);
 735
 736         /*
 737          * If we are up and running, use the hotplug thread. For early calls
 738          * we invoke the thread function directly.
 739          */
 740         if (!st->thread)
 741                 return cpuhp_invoke_callback(cpu, state, bringup, node);
 742
 743         st->cb_state = state;
 744         st->single = true;
 745         st->bringup = bringup;
 746         st->node = node;
 747
 748         /*
 749          * Make sure the above stores are visible before should_run becomes
 750          * true. Paired with the mb() above in cpuhp_thread_fun()
 751          */
 752         smp_mb();
 753         st->should_run = true;
 754         wake_up_process(st->thread);
 755         wait_for_completion(&st->done);
 756         return st->result;
 757 }
 758
 759 /* Regular hotplug invocation of the AP hotplug thread */
 760 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 761 {
 762         st->result = 0;
 763         st->single = false;
 764         /*
 765          * Make sure the above stores are visible before should_run becomes
 766          * true. Paired with the mb() above in cpuhp_thread_fun()
 767          */
 768         smp_mb();
 769         st->should_run = true;
 770         wake_up_process(st->thread);
 771 }
 772
 773 static int cpuhp_kick_ap_work(unsigned int cpu)
 774 {
 775         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 776         enum cpuhp_state state = st->state;
 777
 778         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 779         lock_map_acquire(&cpuhp_state_lock_map);
 780         lock_map_release(&cpuhp_state_lock_map);
 781         __cpuhp_kick_ap_work(st);
 782         wait_for_completion(&st->done);
 783         trace_cpuhp_exit(cpu, st->state, state, st->result);
 784         return st->result;
 785 }
 786
 787 static struct smp_hotplug_thread cpuhp_threads = {
 788         .store                  = &cpuhp_state.thread,
 789         .create                 = &cpuhp_create,
 790         .thread_should_run      = cpuhp_should_run,
 791         .thread_fn              = cpuhp_thread_fun,
 792         .thread_comm            = "cpuhp/%u",
 793         .selfparking            = true,
 794 };
 795
 796 void __init cpuhp_threads_init(void)
 797 {
 798         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 799         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 800 }
 801
 802 EXPORT_SYMBOL(register_cpu_notifier);
 803 EXPORT_SYMBOL(__register_cpu_notifier);
 804 void unregister_cpu_notifier(struct notifier_block *nb)
 805 {
 806         cpu_maps_update_begin();
 807         raw_notifier_chain_unregister(&cpu_chain, nb);
 808         cpu_maps_update_done();
 809 }
 810 EXPORT_SYMBOL(unregister_cpu_notifier);
 811
 812 void __unregister_cpu_notifier(struct notifier_block *nb)
 813 {
 814         raw_notifier_chain_unregister(&cpu_chain, nb);
 815 }
 816 EXPORT_SYMBOL(__unregister_cpu_notifier);
 817
 818 #ifdef CONFIG_HOTPLUG_CPU
 819 #ifndef arch_clear_mm_cpumask_cpu
 820 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
 821 #endif
 822
 823 /**
 824  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 825  * @cpu: a CPU id
 826  *
 827  * This function walks all processes, finds a valid mm struct for each one and
 828  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 829  * trivial, there are various non-obvious corner cases, which this function
 830  * tries to solve in a safe manner.
 831  *
 832  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 833  * be called only for an already offlined CPU.
 834  */
 835 void clear_tasks_mm_cpumask(int cpu)
 836 {
 837         struct task_struct *p;
 838
 839         /*
 840          * This function is called after the cpu is taken down and marked
 841          * offline, so its not like new tasks will ever get this cpu set in
 842          * their mm mask. -- Peter Zijlstra
 843          * Thus, we may use rcu_read_lock() here, instead of grabbing
 844          * full-fledged tasklist_lock.
 845          */
 846         WARN_ON(cpu_online(cpu));
 847         rcu_read_lock();
 848         for_each_process(p) {
 849                 struct task_struct *t;
 850
 851                 /*
 852                  * Main thread might exit, but other threads may still have
 853                  * a valid mm. Find one.
 854                  */
 855                 t = find_lock_task_mm(p);
 856                 if (!t)
 857                         continue;
 858                 arch_clear_mm_cpumask_cpu(cpu, t->mm);
 859                 task_unlock(t);
 860         }
 861         rcu_read_unlock();
 862 }
 863
 864 static inline void check_for_tasks(int dead_cpu)
 865 {
 866         struct task_struct *g, *p;
 867
 868         read_lock(&tasklist_lock);
 869         for_each_process_thread(g, p) {
 870                 if (!p->on_rq)
 871                         continue;
 872                 /*
 873                  * We do the check with unlocked task_rq(p)->lock.
 874                  * Order the reading to do not warn about a task,
 875                  * which was running on this cpu in the past, and
 876                  * it's just been woken on another cpu.
 877                  */
 878                 rmb();
 879                 if (task_cpu(p) != dead_cpu)
 880                         continue;
 881
 882                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 883                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 884         }
 885         read_unlock(&tasklist_lock);
 886 }
 887
 888 static int notify_down_prepare(unsigned int cpu)
 889 {
 890         int err, nr_calls = 0;
 891
 892         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 893         if (err) {
 894                 nr_calls--;
 895                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 896                 pr_warn("%s: attempt to take down CPU %u failed\n",
 897                                 __func__, cpu);
 898         }
 899         return err;
 900 }
 901
 902 /* Take this CPU down. */
 903 static int take_cpu_down(void *_param)
 904 {
 905         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 906         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 907         int err, cpu = smp_processor_id();
 908
 909         /* Ensure this CPU doesn't handle any more interrupts. */
 910         err = __cpu_disable();
 911         if (err < 0)
 912                 return err;
 913
 914         /*
 915          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 916          * do this step again.
 917          */
 918         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 919         st->state--;
 920         /* Invoke the former CPU_DYING callbacks */
 921         for (; st->state > target; st->state--)
 922                 cpuhp_invoke_callback(cpu, st->state, false, NULL);
 923
 924         /* Give up timekeeping duties */
 925         tick_handover_do_timer();
 926         /* Park the stopper thread */
 927         stop_machine_park(cpu);
 928         return 0;
 929 }
 930
 931 static int takedown_cpu(unsigned int cpu)
 932 {
 933         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 934         int err;
 935
 936         /* Park the smpboot threads */
 937         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 938
 939         /*
 940          * Prevent irq alloc/free while the dying cpu reorganizes the
 941          * interrupt affinities.
 942          */
 943         irq_lock_sparse();
 944
 945         /*
 946          * So now all preempt/rcu users must observe !cpu_active().
 947          */
 948         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 949         if (err) {
 950                 /* CPU refused to die */
 951                 irq_unlock_sparse();
 952                 /* Unpark the hotplug thread so we can rollback there */
 953                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 954                 return err;
 955         }
 956         BUG_ON(cpu_online(cpu));
 957
 958         /*
 959          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 960          * runnable tasks from the cpu, there's only the idle task left now
 961          * that the migration thread is done doing the stop_machine thing.
 962          *
 963          * Wait for the stop thread to go away.
 964          */
 965         wait_for_completion(&st->done);
 966         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 967
 968         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 969         irq_unlock_sparse();
 970
 971         hotplug_cpu__broadcast_tick_pull(cpu);
 972         /* This actually kills the CPU. */
 973         __cpu_die(cpu);
 974
 975         tick_cleanup_dead_cpu(cpu);
 976         return 0;
 977 }
 978
 979 static int notify_dead(unsigned int cpu)
 980 {
 981         cpu_notify_nofail(CPU_DEAD, cpu);
 982         check_for_tasks(cpu);
 983         return 0;
 984 }
 985
 986 static void cpuhp_complete_idle_dead(void *arg)
 987 {
 988         struct cpuhp_cpu_state *st = arg;
 989
 990         complete(&st->done);
 991 }
 992
 993 void cpuhp_report_idle_dead(void)
 994 {
 995         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 996
 997         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 998         rcu_report_dead(smp_processor_id());
 999         st->state = CPUHP_AP_IDLE_DEAD;
1000         /*
1001          * We cannot call complete after rcu_report_dead() so we delegate it
1002          * to an online cpu.
1003          */
1004         smp_call_function_single(cpumask_first(cpu_online_mask),
1005                                  cpuhp_complete_idle_dead, st, 0);
1006 }
1007
1008 #else
1009 #define notify_down_prepare     NULL
1010 #define takedown_cpu            NULL
1011 #define notify_dead             NULL
1012 #endif
1013
1014 #ifdef CONFIG_HOTPLUG_CPU
1015
1016 /* Requires cpu_add_remove_lock to be held */
1017 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1018                            enum cpuhp_state target)
1019 {
1020         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1021         int prev_state, ret = 0;
1022         bool hasdied = false;
1023
1024         if (num_online_cpus() == 1)
1025                 return -EBUSY;
1026
1027         if (!cpu_present(cpu))
1028                 return -EINVAL;
1029
1030         cpu_hotplug_begin();
1031
1032         cpuhp_tasks_frozen = tasks_frozen;
1033
1034         prev_state = st->state;
1035         st->target = target;
1036         /*
1037          * If the current CPU state is in the range of the AP hotplug thread,
1038          * then we need to kick the thread.
1039          */
1040         if (st->state > CPUHP_TEARDOWN_CPU) {
1041                 ret = cpuhp_kick_ap_work(cpu);
1042                 /*
1043                  * The AP side has done the error rollback already. Just
1044                  * return the error code..
1045                  */
1046                 if (ret)
1047                         goto out;
1048
1049                 /*
1050                  * We might have stopped still in the range of the AP hotplug
1051                  * thread. Nothing to do anymore.
1052                  */
1053                 if (st->state > CPUHP_TEARDOWN_CPU)
1054                         goto out;
1055         }
1056         /*
1057          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1058          * to do the further cleanups.
1059          */
1060         ret = cpuhp_down_callbacks(cpu, st, target);
1061         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1062                 st->target = prev_state;
1063                 st->rollback = true;
1064                 cpuhp_kick_ap_work(cpu);
1065         }
1066
1067         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
1068 out:
1069         cpu_hotplug_done();
1070         /* This post dead nonsense must die */
1071         if (!ret && hasdied)
1072                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
1073         arch_smt_update();
1074         return ret;
1075 }
1076
1077 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1078 {
1079         if (cpu_hotplug_disabled)
1080                 return -EBUSY;
1081         return _cpu_down(cpu, 0, target);
1082 }
1083
1084 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1085 {
1086         int err;
1087
1088         cpu_maps_update_begin();
1089         err = cpu_down_maps_locked(cpu, target);
1090         cpu_maps_update_done();
1091         return err;
1092 }
1093 int cpu_down(unsigned int cpu)
1094 {
1095         return do_cpu_down(cpu, CPUHP_OFFLINE);
1096 }
1097 EXPORT_SYMBOL(cpu_down);
1098 #endif /*CONFIG_HOTPLUG_CPU*/
1099
1100 /**
1101  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1102  * @cpu: cpu that just started
1103  *
1104  * It must be called by the arch code on the new cpu, before the new cpu
1105  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1106  */
1107 void notify_cpu_starting(unsigned int cpu)
1108 {
1109         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1110         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1111
1112         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1113         st->booted_once = true;
1114         while (st->state < target) {
1115                 st->state++;
1116                 cpuhp_invoke_callback(cpu, st->state, true, NULL);
1117         }
1118 }
1119
1120 /*
1121  * Called from the idle task. Wake up the controlling task which brings the
1122  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1123  * online bringup to the hotplug thread.
1124  */
1125 void cpuhp_online_idle(enum cpuhp_state state)
1126 {
1127         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1128
1129         /* Happens for the boot cpu */
1130         if (state != CPUHP_AP_ONLINE_IDLE)
1131                 return;
1132
1133         /*
1134          * Unpart the stopper thread before we start the idle loop (and start
1135          * scheduling); this ensures the stopper task is always available.
1136          */
1137         stop_machine_unpark(smp_processor_id());
1138
1139         st->state = CPUHP_AP_ONLINE_IDLE;
1140         complete(&st->done);
1141 }
1142
1143 /* Requires cpu_add_remove_lock to be held */
1144 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1145 {
1146         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1147         struct task_struct *idle;
1148         int ret = 0;
1149
1150         cpu_hotplug_begin();
1151
1152         if (!cpu_present(cpu)) {
1153                 ret = -EINVAL;
1154                 goto out;
1155         }
1156
1157         /*
1158          * The caller of do_cpu_up might have raced with another
1159          * caller. Ignore it for now.
1160          */
1161         if (st->state >= target)
1162                 goto out;
1163
1164         if (st->state == CPUHP_OFFLINE) {
1165                 /* Let it fail before we try to bring the cpu up */
1166                 idle = idle_thread_get(cpu);
1167                 if (IS_ERR(idle)) {
1168                         ret = PTR_ERR(idle);
1169                         goto out;
1170                 }
1171         }
1172
1173         cpuhp_tasks_frozen = tasks_frozen;
1174
1175         st->target = target;
1176         /*
1177          * If the current CPU state is in the range of the AP hotplug thread,
1178          * then we need to kick the thread once more.
1179          */
1180         if (st->state > CPUHP_BRINGUP_CPU) {
1181                 ret = cpuhp_kick_ap_work(cpu);
1182                 /*
1183                  * The AP side has done the error rollback already. Just
1184                  * return the error code..
1185                  */
1186                 if (ret)
1187                         goto out;
1188         }
1189
1190         /*
1191          * Try to reach the target state. We max out on the BP at
1192          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1193          * responsible for bringing it up to the target state.
1194          */
1195         target = min((int)target, CPUHP_BRINGUP_CPU);
1196         ret = cpuhp_up_callbacks(cpu, st, target);
1197 out:
1198         cpu_hotplug_done();
1199         arch_smt_update();
1200         return ret;
1201 }
1202
1203 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1204 {
1205         int err = 0;
1206
1207         if (!cpu_possible(cpu)) {
1208                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1209                        cpu);
1210 #if defined(CONFIG_IA64)
1211                 pr_err("please check additional_cpus= boot parameter\n");
1212 #endif
1213                 return -EINVAL;
1214         }
1215
1216         err = try_online_node(cpu_to_node(cpu));
1217         if (err)
1218                 return err;
1219
1220         cpu_maps_update_begin();
1221
1222         if (cpu_hotplug_disabled) {
1223                 err = -EBUSY;
1224                 goto out;
1225         }
1226         if (!cpu_smt_allowed(cpu)) {
1227                 err = -EPERM;
1228                 goto out;
1229         }
1230
1231         err = _cpu_up(cpu, 0, target);
1232 out:
1233         cpu_maps_update_done();
1234         return err;
1235 }
1236
1237 int cpu_up(unsigned int cpu)
1238 {
1239         return do_cpu_up(cpu, CPUHP_ONLINE);
1240 }
1241 EXPORT_SYMBOL_GPL(cpu_up);
1242
1243 #ifdef CONFIG_PM_SLEEP_SMP
1244 static cpumask_var_t frozen_cpus;
1245
1246 int freeze_secondary_cpus(int primary)
1247 {
1248         int cpu, error = 0;
1249
1250         cpu_maps_update_begin();
1251         if (!cpu_online(primary))
1252                 primary = cpumask_first(cpu_online_mask);
1253         /*
1254          * We take down all of the non-boot CPUs in one shot to avoid races
1255          * with the userspace trying to use the CPU hotplug at the same time
1256          */
1257         cpumask_clear(frozen_cpus);
1258
1259         pr_info("Disabling non-boot CPUs ...\n");
1260         for_each_online_cpu(cpu) {
1261                 if (cpu == primary)
1262                         continue;
1263                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1264                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1265                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1266                 if (!error)
1267                         cpumask_set_cpu(cpu, frozen_cpus);
1268                 else {
1269                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1270                         break;
1271                 }
1272         }
1273
1274         if (!error)
1275                 BUG_ON(num_online_cpus() > 1);
1276         else
1277                 pr_err("Non-boot CPUs are not disabled\n");
1278
1279         /*
1280          * Make sure the CPUs won't be enabled by someone else. We need to do
1281          * this even in case of failure as all disable_nonboot_cpus() users are
1282          * supposed to do enable_nonboot_cpus() on the failure path.
1283          */
1284         cpu_hotplug_disabled++;
1285
1286         cpu_maps_update_done();
1287         return error;
1288 }
1289
1290 void __weak arch_enable_nonboot_cpus_begin(void)
1291 {
1292 }
1293
1294 void __weak arch_enable_nonboot_cpus_end(void)
1295 {
1296 }
1297
1298 void enable_nonboot_cpus(void)
1299 {
1300         int cpu, error;
1301
1302         /* Allow everyone to use the CPU hotplug again */
1303         cpu_maps_update_begin();
1304         __cpu_hotplug_enable();
1305         if (cpumask_empty(frozen_cpus))
1306                 goto out;
1307
1308         pr_info("Enabling non-boot CPUs ...\n");
1309
1310         arch_enable_nonboot_cpus_begin();
1311
1312         for_each_cpu(cpu, frozen_cpus) {
1313                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1314                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1315                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1316                 if (!error) {
1317                         pr_info("CPU%d is up\n", cpu);
1318                         continue;
1319                 }
1320                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1321         }
1322
1323         arch_enable_nonboot_cpus_end();
1324
1325         cpumask_clear(frozen_cpus);
1326 out:
1327         cpu_maps_update_done();
1328 }
1329
1330 static int __init alloc_frozen_cpus(void)
1331 {
1332         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1333                 return -ENOMEM;
1334         return 0;
1335 }
1336 core_initcall(alloc_frozen_cpus);
1337
1338 /*
1339  * When callbacks for CPU hotplug notifications are being executed, we must
1340  * ensure that the state of the system with respect to the tasks being frozen
1341  * or not, as reported by the notification, remains unchanged *throughout the
1342  * duration* of the execution of the callbacks.
1343  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1344  *
1345  * This synchronization is implemented by mutually excluding regular CPU
1346  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1347  * Hibernate notifications.
1348  */
1349 static int
1350 cpu_hotplug_pm_callback(struct notifier_block *nb,
1351                         unsigned long action, void *ptr)
1352 {
1353         switch (action) {
1354
1355         case PM_SUSPEND_PREPARE:
1356         case PM_HIBERNATION_PREPARE:
1357                 cpu_hotplug_disable();
1358                 break;
1359
1360         case PM_POST_SUSPEND:
1361         case PM_POST_HIBERNATION:
1362                 cpu_hotplug_enable();
1363                 break;
1364
1365         default:
1366                 return NOTIFY_DONE;
1367         }
1368
1369         return NOTIFY_OK;
1370 }
1371
1372
1373 static int __init cpu_hotplug_pm_sync_init(void)
1374 {
1375         /*
1376          * cpu_hotplug_pm_callback has higher priority than x86
1377          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1378          * to disable cpu hotplug to avoid cpu hotplug race.
1379          */
1380         pm_notifier(cpu_hotplug_pm_callback, 0);
1381         return 0;
1382 }
1383 core_initcall(cpu_hotplug_pm_sync_init);
1384
1385 #endif /* CONFIG_PM_SLEEP_SMP */
1386
1387 #endif /* CONFIG_SMP */
1388
1389 /* Boot processor state steps */
1390 static struct cpuhp_step cpuhp_bp_states[] = {
1391         [CPUHP_OFFLINE] = {
1392                 .name                   = "offline",
1393                 .startup.single         = NULL,
1394                 .teardown.single        = NULL,
1395         },
1396 #ifdef CONFIG_SMP
1397         [CPUHP_CREATE_THREADS]= {
1398                 .name                   = "threads:prepare",
1399                 .startup.single         = smpboot_create_threads,
1400                 .teardown.single        = NULL,
1401                 .cant_stop              = true,
1402         },
1403         [CPUHP_PERF_PREPARE] = {
1404                 .name                   = "perf:prepare",
1405                 .startup.single         = perf_event_init_cpu,
1406                 .teardown.single        = perf_event_exit_cpu,
1407         },
1408         [CPUHP_RANDOM_PREPARE] = {
1409                 .name                   = "random:prepare",
1410                 .startup.single         = random_prepare_cpu,
1411                 .teardown.single        = NULL,
1412         },
1413         [CPUHP_WORKQUEUE_PREP] = {
1414                 .name                   = "workqueue:prepare",
1415                 .startup.single         = workqueue_prepare_cpu,
1416                 .teardown.single        = NULL,
1417         },
1418         [CPUHP_HRTIMERS_PREPARE] = {
1419                 .name                   = "hrtimers:prepare",
1420                 .startup.single         = hrtimers_prepare_cpu,
1421                 .teardown.single        = hrtimers_dead_cpu,
1422         },
1423         [CPUHP_SMPCFD_PREPARE] = {
1424                 .name                   = "smpcfd:prepare",
1425                 .startup.single         = smpcfd_prepare_cpu,
1426                 .teardown.single        = smpcfd_dead_cpu,
1427         },
1428         [CPUHP_RELAY_PREPARE] = {
1429                 .name                   = "relay:prepare",
1430                 .startup.single         = relay_prepare_cpu,
1431                 .teardown.single        = NULL,
1432         },
1433         [CPUHP_SLAB_PREPARE] = {
1434                 .name                   = "slab:prepare",
1435                 .startup.single         = slab_prepare_cpu,
1436                 .teardown.single        = slab_dead_cpu,
1437         },
1438         [CPUHP_RCUTREE_PREP] = {
1439                 .name                   = "RCU/tree:prepare",
1440                 .startup.single         = rcutree_prepare_cpu,
1441                 .teardown.single        = rcutree_dead_cpu,
1442         },
1443         /*
1444          * Preparatory and dead notifiers. Will be replaced once the notifiers
1445          * are converted to states.
1446          */
1447         [CPUHP_NOTIFY_PREPARE] = {
1448                 .name                   = "notify:prepare",
1449                 .startup.single         = notify_prepare,
1450                 .teardown.single        = notify_dead,
1451                 .skip_onerr             = true,
1452                 .cant_stop              = true,
1453         },
1454         /*
1455          * On the tear-down path, timers_dead_cpu() must be invoked
1456          * before blk_mq_queue_reinit_notify() from notify_dead(),
1457          * otherwise a RCU stall occurs.
1458          */
1459         [CPUHP_TIMERS_PREPARE] = {
1460                 .name                   = "timers:dead",
1461                 .startup.single         = timers_prepare_cpu,
1462                 .teardown.single        = timers_dead_cpu,
1463         },
1464         /* Kicks the plugged cpu into life */
1465         [CPUHP_BRINGUP_CPU] = {
1466                 .name                   = "cpu:bringup",
1467                 .startup.single         = bringup_cpu,
1468                 .teardown.single        = NULL,
1469                 .cant_stop              = true,
1470         },
1471         /*
1472          * Handled on controll processor until the plugged processor manages
1473          * this itself.
1474          */
1475         [CPUHP_TEARDOWN_CPU] = {
1476                 .name                   = "cpu:teardown",
1477                 .startup.single         = NULL,
1478                 .teardown.single        = takedown_cpu,
1479                 .cant_stop              = true,
1480         },
1481 #else
1482         [CPUHP_BRINGUP_CPU] = { },
1483 #endif
1484 };
1485
1486 /* Application processor state steps */
1487 static struct cpuhp_step cpuhp_ap_states[] = {
1488 #ifdef CONFIG_SMP
1489         /* Final state before CPU kills itself */
1490         [CPUHP_AP_IDLE_DEAD] = {
1491                 .name                   = "idle:dead",
1492         },
1493         /*
1494          * Last state before CPU enters the idle loop to die. Transient state
1495          * for synchronization.
1496          */
1497         [CPUHP_AP_OFFLINE] = {
1498                 .name                   = "ap:offline",
1499                 .cant_stop              = true,
1500         },
1501         /* First state is scheduler control. Interrupts are disabled */
1502         [CPUHP_AP_SCHED_STARTING] = {
1503                 .name                   = "sched:starting",
1504                 .startup.single         = sched_cpu_starting,
1505                 .teardown.single        = sched_cpu_dying,
1506         },
1507         [CPUHP_AP_RCUTREE_DYING] = {
1508                 .name                   = "RCU/tree:dying",
1509                 .startup.single         = NULL,
1510                 .teardown.single        = rcutree_dying_cpu,
1511         },
1512         [CPUHP_AP_SMPCFD_DYING] = {
1513                 .name                   = "smpcfd:dying",
1514                 .startup.single         = NULL,
1515                 .teardown.single        = smpcfd_dying_cpu,
1516         },
1517         /* Entry state on starting. Interrupts enabled from here on. Transient
1518          * state for synchronsization */
1519         [CPUHP_AP_ONLINE] = {
1520                 .name                   = "ap:online",
1521         },
1522         /* Handle smpboot threads park/unpark */
1523         [CPUHP_AP_SMPBOOT_THREADS] = {
1524                 .name                   = "smpboot/threads:online",
1525                 .startup.single         = smpboot_unpark_threads,
1526                 .teardown.single        = smpboot_park_threads,
1527         },
1528         [CPUHP_AP_PERF_ONLINE] = {
1529                 .name                   = "perf:online",
1530                 .startup.single         = perf_event_init_cpu,
1531                 .teardown.single        = perf_event_exit_cpu,
1532         },
1533         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1534                 .name                   = "workqueue:online",
1535                 .startup.single         = workqueue_online_cpu,
1536                 .teardown.single        = workqueue_offline_cpu,
1537         },
1538         [CPUHP_AP_RANDOM_ONLINE] = {
1539                 .name                   = "random:online",
1540                 .startup.single         = random_online_cpu,
1541                 .teardown.single        = NULL,
1542         },
1543         [CPUHP_AP_RCUTREE_ONLINE] = {
1544                 .name                   = "RCU/tree:online",
1545                 .startup.single         = rcutree_online_cpu,
1546                 .teardown.single        = rcutree_offline_cpu,
1547         },
1548
1549         /*
1550          * Online/down_prepare notifiers. Will be removed once the notifiers
1551          * are converted to states.
1552          */
1553         [CPUHP_AP_NOTIFY_ONLINE] = {
1554                 .name                   = "notify:online",
1555                 .startup.single         = notify_online,
1556                 .teardown.single        = notify_down_prepare,
1557                 .skip_onerr             = true,
1558         },
1559 #endif
1560         /*
1561          * The dynamically registered state space is here
1562          */
1563
1564 #ifdef CONFIG_SMP
1565         /* Last state is scheduler control setting the cpu active */
1566         [CPUHP_AP_ACTIVE] = {
1567                 .name                   = "sched:active",
1568                 .startup.single         = sched_cpu_activate,
1569                 .teardown.single        = sched_cpu_deactivate,
1570         },
1571 #endif
1572
1573         /* CPU is fully up and running. */
1574         [CPUHP_ONLINE] = {
1575                 .name                   = "online",
1576                 .startup.single         = NULL,
1577                 .teardown.single        = NULL,
1578         },
1579 };
1580
1581 /* Sanity check for callbacks */
1582 static int cpuhp_cb_check(enum cpuhp_state state)
1583 {
1584         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1585                 return -EINVAL;
1586         return 0;
1587 }
1588
1589 static void cpuhp_store_callbacks(enum cpuhp_state state,
1590                                   const char *name,
1591                                   int (*startup)(unsigned int cpu),
1592                                   int (*teardown)(unsigned int cpu),
1593                                   bool multi_instance)
1594 {
1595         /* (Un)Install the callbacks for further cpu hotplug operations */
1596         struct cpuhp_step *sp;
1597
1598         sp = cpuhp_get_step(state);
1599         sp->startup.single = startup;
1600         sp->teardown.single = teardown;
1601         sp->name = name;
1602         sp->multi_instance = multi_instance;
1603         INIT_HLIST_HEAD(&sp->list);
1604 }
1605
1606 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1607 {
1608         return cpuhp_get_step(state)->teardown.single;
1609 }
1610
1611 /*
1612  * Call the startup/teardown function for a step either on the AP or
1613  * on the current CPU.
1614  */
1615 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1616                             struct hlist_node *node)
1617 {
1618         struct cpuhp_step *sp = cpuhp_get_step(state);
1619         int ret;
1620
1621         if ((bringup && !sp->startup.single) ||
1622             (!bringup && !sp->teardown.single))
1623                 return 0;
1624         /*
1625          * The non AP bound callbacks can fail on bringup. On teardown
1626          * e.g. module removal we crash for now.
1627          */
1628 #ifdef CONFIG_SMP
1629         if (cpuhp_is_ap_state(state))
1630                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1631         else
1632                 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1633 #else
1634         ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1635 #endif
1636         BUG_ON(ret && !bringup);
1637         return ret;
1638 }
1639
1640 /*
1641  * Called from __cpuhp_setup_state on a recoverable failure.
1642  *
1643  * Note: The teardown callbacks for rollback are not allowed to fail!
1644  */
1645 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1646                                    struct hlist_node *node)
1647 {
1648         int cpu;
1649
1650         /* Roll back the already executed steps on the other cpus */
1651         for_each_present_cpu(cpu) {
1652                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1653                 int cpustate = st->state;
1654
1655                 if (cpu >= failedcpu)
1656                         break;
1657
1658                 /* Did we invoke the startup call on that cpu ? */
1659                 if (cpustate >= state)
1660                         cpuhp_issue_call(cpu, state, false, node);
1661         }
1662 }
1663
1664 /*
1665  * Returns a free for dynamic slot assignment of the Online state. The states
1666  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1667  * by having no name assigned.
1668  */
1669 static int cpuhp_reserve_state(enum cpuhp_state state)
1670 {
1671         enum cpuhp_state i;
1672
1673         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1674                 if (cpuhp_ap_states[i].name)
1675                         continue;
1676
1677                 cpuhp_ap_states[i].name = "Reserved";
1678                 return i;
1679         }
1680         WARN(1, "No more dynamic states available for CPU hotplug\n");
1681         return -ENOSPC;
1682 }
1683
1684 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1685                                bool invoke)
1686 {
1687         struct cpuhp_step *sp;
1688         int cpu;
1689         int ret;
1690
1691         sp = cpuhp_get_step(state);
1692         if (sp->multi_instance == false)
1693                 return -EINVAL;
1694
1695         get_online_cpus();
1696         mutex_lock(&cpuhp_state_mutex);
1697
1698         if (!invoke || !sp->startup.multi)
1699                 goto add_node;
1700
1701         /*
1702          * Try to call the startup callback for each present cpu
1703          * depending on the hotplug state of the cpu.
1704          */
1705         for_each_present_cpu(cpu) {
1706                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1707                 int cpustate = st->state;
1708
1709                 if (cpustate < state)
1710                         continue;
1711
1712                 ret = cpuhp_issue_call(cpu, state, true, node);
1713                 if (ret) {
1714                         if (sp->teardown.multi)
1715                                 cpuhp_rollback_install(cpu, state, node);
1716                         goto err;
1717                 }
1718         }
1719 add_node:
1720         ret = 0;
1721         hlist_add_head(node, &sp->list);
1722
1723 err:
1724         mutex_unlock(&cpuhp_state_mutex);
1725         put_online_cpus();
1726         return ret;
1727 }
1728 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1729
1730 /**
1731  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1732  * @state:      The state to setup
1733  * @invoke:     If true, the startup function is invoked for cpus where
1734  *              cpu state >= @state
1735  * @startup:    startup callback function
1736  * @teardown:   teardown callback function
1737  *
1738  * Returns 0 if successful, otherwise a proper error code
1739  */
1740 int __cpuhp_setup_state(enum cpuhp_state state,
1741                         const char *name, bool invoke,
1742                         int (*startup)(unsigned int cpu),
1743                         int (*teardown)(unsigned int cpu),
1744                         bool multi_instance)
1745 {
1746         int cpu, ret = 0;
1747         int dyn_state = 0;
1748
1749         if (cpuhp_cb_check(state) || !name)
1750                 return -EINVAL;
1751
1752         get_online_cpus();
1753         mutex_lock(&cpuhp_state_mutex);
1754
1755         /* currently assignments for the ONLINE state are possible */
1756         if (state == CPUHP_AP_ONLINE_DYN) {
1757                 dyn_state = 1;
1758                 ret = cpuhp_reserve_state(state);
1759                 if (ret < 0)
1760                         goto out;
1761                 state = ret;
1762         }
1763
1764         cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1765
1766         if (!invoke || !startup)
1767                 goto out;
1768
1769         /*
1770          * Try to call the startup callback for each present cpu
1771          * depending on the hotplug state of the cpu.
1772          */
1773         for_each_present_cpu(cpu) {
1774                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1775                 int cpustate = st->state;
1776
1777                 if (cpustate < state)
1778                         continue;
1779
1780                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1781                 if (ret) {
1782                         if (teardown)
1783                                 cpuhp_rollback_install(cpu, state, NULL);
1784                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1785                         goto out;
1786                 }
1787         }
1788 out:
1789         mutex_unlock(&cpuhp_state_mutex);
1790
1791         put_online_cpus();
1792         if (!ret && dyn_state)
1793                 return state;
1794         return ret;
1795 }
1796 EXPORT_SYMBOL(__cpuhp_setup_state);
1797
1798 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1799                                   struct hlist_node *node, bool invoke)
1800 {
1801         struct cpuhp_step *sp = cpuhp_get_step(state);
1802         int cpu;
1803
1804         BUG_ON(cpuhp_cb_check(state));
1805
1806         if (!sp->multi_instance)
1807                 return -EINVAL;
1808
1809         get_online_cpus();
1810         mutex_lock(&cpuhp_state_mutex);
1811
1812         if (!invoke || !cpuhp_get_teardown_cb(state))
1813                 goto remove;
1814         /*
1815          * Call the teardown callback for each present cpu depending
1816          * on the hotplug state of the cpu. This function is not
1817          * allowed to fail currently!
1818          */
1819         for_each_present_cpu(cpu) {
1820                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1821                 int cpustate = st->state;
1822
1823                 if (cpustate >= state)
1824                         cpuhp_issue_call(cpu, state, false, node);
1825         }
1826
1827 remove:
1828         hlist_del(node);
1829         mutex_unlock(&cpuhp_state_mutex);
1830         put_online_cpus();
1831
1832         return 0;
1833 }
1834 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1835 /**
1836  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1837  * @state:      The state to remove
1838  * @invoke:     If true, the teardown function is invoked for cpus where
1839  *              cpu state >= @state
1840  *
1841  * The teardown callback is currently not allowed to fail. Think
1842  * about module removal!
1843  */
1844 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1845 {
1846         struct cpuhp_step *sp = cpuhp_get_step(state);
1847         int cpu;
1848
1849         BUG_ON(cpuhp_cb_check(state));
1850
1851         get_online_cpus();
1852         mutex_lock(&cpuhp_state_mutex);
1853
1854         if (sp->multi_instance) {
1855                 WARN(!hlist_empty(&sp->list),
1856                      "Error: Removing state %d which has instances left.\n",
1857                      state);
1858                 goto remove;
1859         }
1860
1861         if (!invoke || !cpuhp_get_teardown_cb(state))
1862                 goto remove;
1863
1864         /*
1865          * Call the teardown callback for each present cpu depending
1866          * on the hotplug state of the cpu. This function is not
1867          * allowed to fail currently!
1868          */
1869         for_each_present_cpu(cpu) {
1870                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1871                 int cpustate = st->state;
1872
1873                 if (cpustate >= state)
1874                         cpuhp_issue_call(cpu, state, false, NULL);
1875         }
1876 remove:
1877         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1878         mutex_unlock(&cpuhp_state_mutex);
1879         put_online_cpus();
1880 }
1881 EXPORT_SYMBOL(__cpuhp_remove_state);
1882
1883 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1884 static ssize_t show_cpuhp_state(struct device *dev,
1885                                 struct device_attribute *attr, char *buf)
1886 {
1887         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1888
1889         return sprintf(buf, "%d\n", st->state);
1890 }
1891 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1892
1893 static ssize_t write_cpuhp_target(struct device *dev,
1894                                   struct device_attribute *attr,
1895                                   const char *buf, size_t count)
1896 {
1897         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1898         struct cpuhp_step *sp;
1899         int target, ret;
1900
1901         ret = kstrtoint(buf, 10, &target);
1902         if (ret)
1903                 return ret;
1904
1905 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1906         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1907                 return -EINVAL;
1908 #else
1909         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1910                 return -EINVAL;
1911 #endif
1912
1913         ret = lock_device_hotplug_sysfs();
1914         if (ret)
1915                 return ret;
1916
1917         mutex_lock(&cpuhp_state_mutex);
1918         sp = cpuhp_get_step(target);
1919         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1920         mutex_unlock(&cpuhp_state_mutex);
1921         if (ret)
1922                 goto out;
1923
1924         if (st->state < target)
1925                 ret = do_cpu_up(dev->id, target);
1926         else
1927                 ret = do_cpu_down(dev->id, target);
1928 out:
1929         unlock_device_hotplug();
1930         return ret ? ret : count;
1931 }
1932
1933 static ssize_t show_cpuhp_target(struct device *dev,
1934                                  struct device_attribute *attr, char *buf)
1935 {
1936         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1937
1938         return sprintf(buf, "%d\n", st->target);
1939 }
1940 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1941
1942 static struct attribute *cpuhp_cpu_attrs[] = {
1943         &dev_attr_state.attr,
1944         &dev_attr_target.attr,
1945         NULL
1946 };
1947
1948 static struct attribute_group cpuhp_cpu_attr_group = {
1949         .attrs = cpuhp_cpu_attrs,
1950         .name = "hotplug",
1951         NULL
1952 };
1953
1954 static ssize_t show_cpuhp_states(struct device *dev,
1955                                  struct device_attribute *attr, char *buf)
1956 {
1957         ssize_t cur, res = 0;
1958         int i;
1959
1960         mutex_lock(&cpuhp_state_mutex);
1961         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1962                 struct cpuhp_step *sp = cpuhp_get_step(i);
1963
1964                 if (sp->name) {
1965                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1966                         buf += cur;
1967                         res += cur;
1968                 }
1969         }
1970         mutex_unlock(&cpuhp_state_mutex);
1971         return res;
1972 }
1973 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1974
1975 static struct attribute *cpuhp_cpu_root_attrs[] = {
1976         &dev_attr_states.attr,
1977         NULL
1978 };
1979
1980 static struct attribute_group cpuhp_cpu_root_attr_group = {
1981         .attrs = cpuhp_cpu_root_attrs,
1982         .name = "hotplug",
1983         NULL
1984 };
1985
1986 #ifdef CONFIG_HOTPLUG_SMT
1987
1988 static const char *smt_states[] = {
1989         [CPU_SMT_ENABLED]               = "on",
1990         [CPU_SMT_DISABLED]              = "off",
1991         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
1992         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
1993 };
1994
1995 static ssize_t
1996 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
1997 {
1998         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
1999 }
2000
2001 static void cpuhp_offline_cpu_device(unsigned int cpu)
2002 {
2003         struct device *dev = get_cpu_device(cpu);
2004
2005         dev->offline = true;
2006         /* Tell user space about the state change */
2007         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2008 }
2009
2010 static void cpuhp_online_cpu_device(unsigned int cpu)
2011 {
2012         struct device *dev = get_cpu_device(cpu);
2013
2014         dev->offline = false;
2015         /* Tell user space about the state change */
2016         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2017 }
2018
2019 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2020 {
2021         int cpu, ret = 0;
2022
2023         cpu_maps_update_begin();
2024         for_each_online_cpu(cpu) {
2025                 if (topology_is_primary_thread(cpu))
2026                         continue;
2027                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2028                 if (ret)
2029                         break;
2030                 /*
2031                  * As this needs to hold the cpu maps lock it's impossible
2032                  * to call device_offline() because that ends up calling
2033                  * cpu_down() which takes cpu maps lock. cpu maps lock
2034                  * needs to be held as this might race against in kernel
2035                  * abusers of the hotplug machinery (thermal management).
2036                  *
2037                  * So nothing would update device:offline state. That would
2038                  * leave the sysfs entry stale and prevent onlining after
2039                  * smt control has been changed to 'off' again. This is
2040                  * called under the sysfs hotplug lock, so it is properly
2041                  * serialized against the regular offline usage.
2042                  */
2043                 cpuhp_offline_cpu_device(cpu);
2044         }
2045         if (!ret)
2046                 cpu_smt_control = ctrlval;
2047         cpu_maps_update_done();
2048         return ret;
2049 }
2050
2051 int cpuhp_smt_enable(void)
2052 {
2053         int cpu, ret = 0;
2054
2055         cpu_maps_update_begin();
2056         cpu_smt_control = CPU_SMT_ENABLED;
2057         for_each_present_cpu(cpu) {
2058                 /* Skip online CPUs and CPUs on offline nodes */
2059                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2060                         continue;
2061                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2062                 if (ret)
2063                         break;
2064                 /* See comment in cpuhp_smt_disable() */
2065                 cpuhp_online_cpu_device(cpu);
2066         }
2067         cpu_maps_update_done();
2068         return ret;
2069 }
2070
2071 static ssize_t
2072 store_smt_control(struct device *dev, struct device_attribute *attr,
2073                   const char *buf, size_t count)
2074 {
2075         int ctrlval, ret;
2076
2077         if (sysfs_streq(buf, "on"))
2078                 ctrlval = CPU_SMT_ENABLED;
2079         else if (sysfs_streq(buf, "off"))
2080                 ctrlval = CPU_SMT_DISABLED;
2081         else if (sysfs_streq(buf, "forceoff"))
2082                 ctrlval = CPU_SMT_FORCE_DISABLED;
2083         else
2084                 return -EINVAL;
2085
2086         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2087                 return -EPERM;
2088
2089         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2090                 return -ENODEV;
2091
2092         ret = lock_device_hotplug_sysfs();
2093         if (ret)
2094                 return ret;
2095
2096         if (ctrlval != cpu_smt_control) {
2097                 switch (ctrlval) {
2098                 case CPU_SMT_ENABLED:
2099                         ret = cpuhp_smt_enable();
2100                         break;
2101                 case CPU_SMT_DISABLED:
2102                 case CPU_SMT_FORCE_DISABLED:
2103                         ret = cpuhp_smt_disable(ctrlval);
2104                         break;
2105                 }
2106         }
2107
2108         unlock_device_hotplug();
2109         return ret ? ret : count;
2110 }
2111 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2112
2113 static ssize_t
2114 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2115 {
2116         bool active = topology_max_smt_threads() > 1;
2117
2118         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2119 }
2120 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2121
2122 static struct attribute *cpuhp_smt_attrs[] = {
2123         &dev_attr_control.attr,
2124         &dev_attr_active.attr,
2125         NULL
2126 };
2127
2128 static const struct attribute_group cpuhp_smt_attr_group = {
2129         .attrs = cpuhp_smt_attrs,
2130         .name = "smt",
2131         NULL
2132 };
2133
2134 static int __init cpu_smt_state_init(void)
2135 {
2136         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2137                                   &cpuhp_smt_attr_group);
2138 }
2139
2140 #else
2141 static inline int cpu_smt_state_init(void) { return 0; }
2142 #endif
2143
2144 static int __init cpuhp_sysfs_init(void)
2145 {
2146         int cpu, ret;
2147
2148         ret = cpu_smt_state_init();
2149         if (ret)
2150                 return ret;
2151
2152         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2153                                  &cpuhp_cpu_root_attr_group);
2154         if (ret)
2155                 return ret;
2156
2157         for_each_possible_cpu(cpu) {
2158                 struct device *dev = get_cpu_device(cpu);
2159
2160                 if (!dev)
2161                         continue;
2162                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2163                 if (ret)
2164                         return ret;
2165         }
2166         return 0;
2167 }
2168 device_initcall(cpuhp_sysfs_init);
2169 #endif
2170
2171 /*
2172  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2173  * represents all NR_CPUS bits binary values of 1<<nr.
2174  *
2175  * It is used by cpumask_of() to get a constant address to a CPU
2176  * mask value that has a single bit set only.
2177  */
2178
2179 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2180 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2181 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2182 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2183 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2184
2185 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2186
2187         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2188         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2189 #if BITS_PER_LONG > 32
2190         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2191         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2192 #endif
2193 };
2194 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2195
2196 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2197 EXPORT_SYMBOL(cpu_all_bits);
2198
2199 #ifdef CONFIG_INIT_ALL_POSSIBLE
2200 struct cpumask __cpu_possible_mask __read_mostly
2201         = {CPU_BITS_ALL};
2202 #else
2203 struct cpumask __cpu_possible_mask __read_mostly;
2204 #endif
2205 EXPORT_SYMBOL(__cpu_possible_mask);
2206
2207 struct cpumask __cpu_online_mask __read_mostly;
2208 EXPORT_SYMBOL(__cpu_online_mask);
2209
2210 struct cpumask __cpu_present_mask __read_mostly;
2211 EXPORT_SYMBOL(__cpu_present_mask);
2212
2213 struct cpumask __cpu_active_mask __read_mostly;
2214 EXPORT_SYMBOL(__cpu_active_mask);
2215
2216 void init_cpu_present(const struct cpumask *src)
2217 {
2218         cpumask_copy(&__cpu_present_mask, src);
2219 }
2220
2221 void init_cpu_possible(const struct cpumask *src)
2222 {
2223         cpumask_copy(&__cpu_possible_mask, src);
2224 }
2225
2226 void init_cpu_online(const struct cpumask *src)
2227 {
2228         cpumask_copy(&__cpu_online_mask, src);
2229 }
2230
2231 /*
2232  * Activate the first processor.
2233  */
2234 void __init boot_cpu_init(void)
2235 {
2236         int cpu = smp_processor_id();
2237
2238         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2239         set_cpu_online(cpu, true);
2240         set_cpu_active(cpu, true);
2241         set_cpu_present(cpu, true);
2242         set_cpu_possible(cpu, true);
2243 }
2244
2245 /*
2246  * Must be called _AFTER_ setting up the per_cpu areas
2247  */
2248 void __init boot_cpu_hotplug_init(void)
2249 {
2250 #ifdef CONFIG_SMP
2251         this_cpu_write(cpuhp_state.booted_once, true);
2252 #endif
2253         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2254 }
2255
2256 /*
2257  * These are used for a global "mitigations=" cmdline option for toggling
2258  * optional CPU mitigations.
2259  */
2260 enum cpu_mitigations {
2261         CPU_MITIGATIONS_OFF,
2262         CPU_MITIGATIONS_AUTO,
2263         CPU_MITIGATIONS_AUTO_NOSMT,
2264 };
2265
2266 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2267         CPU_MITIGATIONS_AUTO;
2268
2269 static int __init mitigations_parse_cmdline(char *arg)
2270 {
2271         if (!strcmp(arg, "off"))
2272                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2273         else if (!strcmp(arg, "auto"))
2274                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2275         else if (!strcmp(arg, "auto,nosmt"))
2276                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2277         else
2278                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2279                         arg);
2280
2281         return 0;
2282 }
2283 early_param("mitigations", mitigations_parse_cmdline);
2284
2285 /* mitigations=off */
2286 bool cpu_mitigations_off(void)
2287 {
2288         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2289 }
2290 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2291
2292 /* mitigations=auto,nosmt */
2293 bool cpu_mitigations_auto_nosmt(void)
2294 {
2295         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2296 }
2297 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);