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