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