kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/proc_fs.h>
   7 #include <linux/smp.h>
   8 #include <linux/init.h>
   9 #include <linux/notifier.h>
  10 #include <linux/sched/signal.h>
  11 #include <linux/sched/hotplug.h>
  12 #include <linux/sched/task.h>
  13 #include <linux/sched/smt.h>
  14 #include <linux/unistd.h>
  15 #include <linux/cpu.h>
  16 #include <linux/oom.h>
  17 #include <linux/rcupdate.h>
  18 #include <linux/export.h>
  19 #include <linux/bug.h>
  20 #include <linux/kthread.h>
  21 #include <linux/stop_machine.h>
  22 #include <linux/mutex.h>
  23 #include <linux/gfp.h>
  24 #include <linux/suspend.h>
  25 #include <linux/lockdep.h>
  26 #include <linux/tick.h>
  27 #include <linux/irq.h>
  28 #include <linux/nmi.h>
  29 #include <linux/smpboot.h>
  30 #include <linux/relay.h>
  31 #include <linux/slab.h>
  32 #include <linux/percpu-rwsem.h>
  33
  34 #include <trace/events/power.h>
  35 #define CREATE_TRACE_POINTS
  36 #include <trace/events/cpuhp.h>
  37
  38 #include "smpboot.h"
  39
  40 /**
  41  * cpuhp_cpu_state - Per cpu hotplug state storage
  42  * @state:      The current cpu state
  43  * @target:     The target state
  44  * @thread:     Pointer to the hotplug thread
  45  * @should_run: Thread should execute
  46  * @rollback:   Perform a rollback
  47  * @single:     Single callback invocation
  48  * @bringup:    Single callback bringup or teardown selector
  49  * @cb_state:   The state for a single callback (install/uninstall)
  50  * @result:     Result of the operation
  51  * @done_up:    Signal completion to the issuer of the task for cpu-up
  52  * @done_down:  Signal completion to the issuer of the task for cpu-down
  53  */
  54 struct cpuhp_cpu_state {
  55         enum cpuhp_state        state;
  56         enum cpuhp_state        target;
  57         enum cpuhp_state        fail;
  58 #ifdef CONFIG_SMP
  59         struct task_struct      *thread;
  60         bool                    should_run;
  61         bool                    rollback;
  62         bool                    single;
  63         bool                    bringup;
  64         bool                    booted_once;
  65         struct hlist_node       *node;
  66         struct hlist_node       *last;
  67         enum cpuhp_state        cb_state;
  68         int                     result;
  69         struct completion       done_up;
  70         struct completion       done_down;
  71 #endif
  72 };
  73
  74 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  75         .fail = CPUHP_INVALID,
  76 };
  77
  78 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  79 static struct lockdep_map cpuhp_state_up_map =
  80         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  81 static struct lockdep_map cpuhp_state_down_map =
  82         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  83
  84
  85 static inline void cpuhp_lock_acquire(bool bringup)
  86 {
  87         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  88 }
  89
  90 static inline void cpuhp_lock_release(bool bringup)
  91 {
  92         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  93 }
  94 #else
  95
  96 static inline void cpuhp_lock_acquire(bool bringup) { }
  97 static inline void cpuhp_lock_release(bool bringup) { }
  98
  99 #endif
 100
 101 /**
 102  * cpuhp_step - Hotplug state machine step
 103  * @name:       Name of the step
 104  * @startup:    Startup function of the step
 105  * @teardown:   Teardown function of the step
 106  * @cant_stop:  Bringup/teardown can't be stopped at this step
 107  */
 108 struct cpuhp_step {
 109         const char              *name;
 110         union {
 111                 int             (*single)(unsigned int cpu);
 112                 int             (*multi)(unsigned int cpu,
 113                                          struct hlist_node *node);
 114         } startup;
 115         union {
 116                 int             (*single)(unsigned int cpu);
 117                 int             (*multi)(unsigned int cpu,
 118                                          struct hlist_node *node);
 119         } teardown;
 120         struct hlist_head       list;
 121         bool                    cant_stop;
 122         bool                    multi_instance;
 123 };
 124
 125 static DEFINE_MUTEX(cpuhp_state_mutex);
 126 static struct cpuhp_step cpuhp_hp_states[];
 127
 128 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 129 {
 130         return cpuhp_hp_states + state;
 131 }
 132
 133 /**
 134  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 135  * @cpu:        The cpu for which the callback should be invoked
 136  * @state:      The state to do callbacks for
 137  * @bringup:    True if the bringup callback should be invoked
 138  * @node:       For multi-instance, do a single entry callback for install/remove
 139  * @lastp:      For multi-instance rollback, remember how far we got
 140  *
 141  * Called from cpu hotplug and from the state register machinery.
 142  */
 143 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 144                                  bool bringup, struct hlist_node *node,
 145                                  struct hlist_node **lastp)
 146 {
 147         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 148         struct cpuhp_step *step = cpuhp_get_step(state);
 149         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 150         int (*cb)(unsigned int cpu);
 151         int ret, cnt;
 152
 153         if (st->fail == state) {
 154                 st->fail = CPUHP_INVALID;
 155
 156                 if (!(bringup ? step->startup.single : step->teardown.single))
 157                         return 0;
 158
 159                 return -EAGAIN;
 160         }
 161
 162         if (!step->multi_instance) {
 163                 WARN_ON_ONCE(lastp && *lastp);
 164                 cb = bringup ? step->startup.single : step->teardown.single;
 165                 if (!cb)
 166                         return 0;
 167                 trace_cpuhp_enter(cpu, st->target, state, cb);
 168                 ret = cb(cpu);
 169                 trace_cpuhp_exit(cpu, st->state, state, ret);
 170                 return ret;
 171         }
 172         cbm = bringup ? step->startup.multi : step->teardown.multi;
 173         if (!cbm)
 174                 return 0;
 175
 176         /* Single invocation for instance add/remove */
 177         if (node) {
 178                 WARN_ON_ONCE(lastp && *lastp);
 179                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 180                 ret = cbm(cpu, node);
 181                 trace_cpuhp_exit(cpu, st->state, state, ret);
 182                 return ret;
 183         }
 184
 185         /* State transition. Invoke on all instances */
 186         cnt = 0;
 187         hlist_for_each(node, &step->list) {
 188                 if (lastp && node == *lastp)
 189                         break;
 190
 191                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 192                 ret = cbm(cpu, node);
 193                 trace_cpuhp_exit(cpu, st->state, state, ret);
 194                 if (ret) {
 195                         if (!lastp)
 196                                 goto err;
 197
 198                         *lastp = node;
 199                         return ret;
 200                 }
 201                 cnt++;
 202         }
 203         if (lastp)
 204                 *lastp = NULL;
 205         return 0;
 206 err:
 207         /* Rollback the instances if one failed */
 208         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 209         if (!cbm)
 210                 return ret;
 211
 212         hlist_for_each(node, &step->list) {
 213                 if (!cnt--)
 214                         break;
 215
 216                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 217                 ret = cbm(cpu, node);
 218                 trace_cpuhp_exit(cpu, st->state, state, ret);
 219                 /*
 220                  * Rollback must not fail,
 221                  */
 222                 WARN_ON_ONCE(ret);
 223         }
 224         return ret;
 225 }
 226
 227 #ifdef CONFIG_SMP
 228 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 229 {
 230         /*
 231          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 232          * purposes as that state is handled explicitly in cpu_down.
 233          */
 234         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 235 }
 236
 237 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 238 {
 239         struct completion *done = bringup ? &st->done_up : &st->done_down;
 240         wait_for_completion(done);
 241 }
 242
 243 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 244 {
 245         struct completion *done = bringup ? &st->done_up : &st->done_down;
 246         complete(done);
 247 }
 248
 249 /*
 250  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 251  */
 252 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 253 {
 254         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 255 }
 256
 257 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 258 static DEFINE_MUTEX(cpu_add_remove_lock);
 259 bool cpuhp_tasks_frozen;
 260 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 261
 262 /*
 263  * The following two APIs (cpu_maps_update_begin/done) must be used when
 264  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 265  */
 266 void cpu_maps_update_begin(void)
 267 {
 268         mutex_lock(&cpu_add_remove_lock);
 269 }
 270
 271 void cpu_maps_update_done(void)
 272 {
 273         mutex_unlock(&cpu_add_remove_lock);
 274 }
 275
 276 /*
 277  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 278  * Should always be manipulated under cpu_add_remove_lock
 279  */
 280 static int cpu_hotplug_disabled;
 281
 282 #ifdef CONFIG_HOTPLUG_CPU
 283
 284 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 285
 286 void cpus_read_lock(void)
 287 {
 288         percpu_down_read(&cpu_hotplug_lock);
 289 }
 290 EXPORT_SYMBOL_GPL(cpus_read_lock);
 291
 292 int cpus_read_trylock(void)
 293 {
 294         return percpu_down_read_trylock(&cpu_hotplug_lock);
 295 }
 296 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 297
 298 void cpus_read_unlock(void)
 299 {
 300         percpu_up_read(&cpu_hotplug_lock);
 301 }
 302 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 303
 304 void cpus_write_lock(void)
 305 {
 306         percpu_down_write(&cpu_hotplug_lock);
 307 }
 308
 309 void cpus_write_unlock(void)
 310 {
 311         percpu_up_write(&cpu_hotplug_lock);
 312 }
 313
 314 void lockdep_assert_cpus_held(void)
 315 {
 316         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 317 }
 318
 319 static void lockdep_acquire_cpus_lock(void)
 320 {
 321         rwsem_acquire(&cpu_hotplug_lock.rw_sem.dep_map, 0, 0, _THIS_IP_);
 322 }
 323
 324 static void lockdep_release_cpus_lock(void)
 325 {
 326         rwsem_release(&cpu_hotplug_lock.rw_sem.dep_map, 1, _THIS_IP_);
 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
 359 #else
 360
 361 static void lockdep_acquire_cpus_lock(void)
 362 {
 363 }
 364
 365 static void lockdep_release_cpus_lock(void)
 366 {
 367 }
 368
 369 #endif  /* CONFIG_HOTPLUG_CPU */
 370
 371 /*
 372  * Architectures that need SMT-specific errata handling during SMT hotplug
 373  * should override this.
 374  */
 375 void __weak arch_smt_update(void) { }
 376
 377 #ifdef CONFIG_HOTPLUG_SMT
 378 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 379 EXPORT_SYMBOL_GPL(cpu_smt_control);
 380
 381 static bool cpu_smt_available __read_mostly;
 382
 383 void __init cpu_smt_disable(bool force)
 384 {
 385         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED ||
 386                 cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
 387                 return;
 388
 389         if (force) {
 390                 pr_info("SMT: Force disabled\n");
 391                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 392         } else {
 393                 pr_info("SMT: disabled\n");
 394                 cpu_smt_control = CPU_SMT_DISABLED;
 395         }
 396 }
 397
 398 /*
 399  * The decision whether SMT is supported can only be done after the full
 400  * CPU identification. Called from architecture code before non boot CPUs
 401  * are brought up.
 402  */
 403 void __init cpu_smt_check_topology_early(void)
 404 {
 405         if (!topology_smt_supported())
 406                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 407 }
 408
 409 /*
 410  * If SMT was disabled by BIOS, detect it here, after the CPUs have been
 411  * brought online. This ensures the smt/l1tf sysfs entries are consistent
 412  * with reality. cpu_smt_available is set to true during the bringup of non
 413  * boot CPUs when a SMT sibling is detected. Note, this may overwrite
 414  * cpu_smt_control's previous setting.
 415  */
 416 void __init cpu_smt_check_topology(void)
 417 {
 418         if (!cpu_smt_available)
 419                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 420 }
 421
 422 static int __init smt_cmdline_disable(char *str)
 423 {
 424         cpu_smt_disable(str && !strcmp(str, "force"));
 425         return 0;
 426 }
 427 early_param("nosmt", smt_cmdline_disable);
 428
 429 static inline bool cpu_smt_allowed(unsigned int cpu)
 430 {
 431         if (topology_is_primary_thread(cpu))
 432                 return true;
 433
 434         /*
 435          * If the CPU is not a 'primary' thread and the booted_once bit is
 436          * set then the processor has SMT support. Store this information
 437          * for the late check of SMT support in cpu_smt_check_topology().
 438          */
 439         if (per_cpu(cpuhp_state, cpu).booted_once)
 440                 cpu_smt_available = true;
 441
 442         if (cpu_smt_control == CPU_SMT_ENABLED)
 443                 return true;
 444
 445         /*
 446          * On x86 it's required to boot all logical CPUs at least once so
 447          * that the init code can get a chance to set CR4.MCE on each
 448          * CPU. Otherwise, a broadacasted MCE observing CR4.MCE=0b on any
 449          * core will shutdown the machine.
 450          */
 451         return !per_cpu(cpuhp_state, cpu).booted_once;
 452 }
 453 #else
 454 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 455 #endif
 456
 457 static inline enum cpuhp_state
 458 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 459 {
 460         enum cpuhp_state prev_state = st->state;
 461
 462         st->rollback = false;
 463         st->last = NULL;
 464
 465         st->target = target;
 466         st->single = false;
 467         st->bringup = st->state < target;
 468
 469         return prev_state;
 470 }
 471
 472 static inline void
 473 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 474 {
 475         st->rollback = true;
 476
 477         /*
 478          * If we have st->last we need to undo partial multi_instance of this
 479          * state first. Otherwise start undo at the previous state.
 480          */
 481         if (!st->last) {
 482                 if (st->bringup)
 483                         st->state--;
 484                 else
 485                         st->state++;
 486         }
 487
 488         st->target = prev_state;
 489         st->bringup = !st->bringup;
 490 }
 491
 492 /* Regular hotplug invocation of the AP hotplug thread */
 493 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 494 {
 495         if (!st->single && st->state == st->target)
 496                 return;
 497
 498         st->result = 0;
 499         /*
 500          * Make sure the above stores are visible before should_run becomes
 501          * true. Paired with the mb() above in cpuhp_thread_fun()
 502          */
 503         smp_mb();
 504         st->should_run = true;
 505         wake_up_process(st->thread);
 506         wait_for_ap_thread(st, st->bringup);
 507 }
 508
 509 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 510 {
 511         enum cpuhp_state prev_state;
 512         int ret;
 513
 514         prev_state = cpuhp_set_state(st, target);
 515         __cpuhp_kick_ap(st);
 516         if ((ret = st->result)) {
 517                 cpuhp_reset_state(st, prev_state);
 518                 __cpuhp_kick_ap(st);
 519         }
 520
 521         return ret;
 522 }
 523
 524 static int bringup_wait_for_ap(unsigned int cpu)
 525 {
 526         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 527
 528         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 529         wait_for_ap_thread(st, true);
 530         if (WARN_ON_ONCE((!cpu_online(cpu))))
 531                 return -ECANCELED;
 532
 533         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 534         stop_machine_unpark(cpu);
 535         kthread_unpark(st->thread);
 536
 537         /*
 538          * SMT soft disabling on X86 requires to bring the CPU out of the
 539          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 540          * CPU marked itself as booted_once in cpu_notify_starting() so the
 541          * cpu_smt_allowed() check will now return false if this is not the
 542          * primary sibling.
 543          */
 544         if (!cpu_smt_allowed(cpu))
 545                 return -ECANCELED;
 546
 547         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 548                 return 0;
 549
 550         return cpuhp_kick_ap(st, st->target);
 551 }
 552
 553 static int bringup_cpu(unsigned int cpu)
 554 {
 555         struct task_struct *idle = idle_thread_get(cpu);
 556         int ret;
 557
 558         /*
 559          * Some architectures have to walk the irq descriptors to
 560          * setup the vector space for the cpu which comes online.
 561          * Prevent irq alloc/free across the bringup.
 562          */
 563         irq_lock_sparse();
 564
 565         /* Arch-specific enabling code. */
 566         ret = __cpu_up(cpu, idle);
 567         irq_unlock_sparse();
 568         if (ret)
 569                 return ret;
 570         return bringup_wait_for_ap(cpu);
 571 }
 572
 573 /*
 574  * Hotplug state machine related functions
 575  */
 576
 577 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 578 {
 579         for (st->state--; st->state > st->target; st->state--)
 580                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 581 }
 582
 583 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 584                               enum cpuhp_state target)
 585 {
 586         enum cpuhp_state prev_state = st->state;
 587         int ret = 0;
 588
 589         while (st->state < target) {
 590                 st->state++;
 591                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 592                 if (ret) {
 593                         st->target = prev_state;
 594                         undo_cpu_up(cpu, st);
 595                         break;
 596                 }
 597         }
 598         return ret;
 599 }
 600
 601 /*
 602  * The cpu hotplug threads manage the bringup and teardown of the cpus
 603  */
 604 static void cpuhp_create(unsigned int cpu)
 605 {
 606         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 607
 608         init_completion(&st->done_up);
 609         init_completion(&st->done_down);
 610 }
 611
 612 static int cpuhp_should_run(unsigned int cpu)
 613 {
 614         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 615
 616         return st->should_run;
 617 }
 618
 619 /*
 620  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 621  * callbacks when a state gets [un]installed at runtime.
 622  *
 623  * Each invocation of this function by the smpboot thread does a single AP
 624  * state callback.
 625  *
 626  * It has 3 modes of operation:
 627  *  - single: runs st->cb_state
 628  *  - up:     runs ++st->state, while st->state < st->target
 629  *  - down:   runs st->state--, while st->state > st->target
 630  *
 631  * When complete or on error, should_run is cleared and the completion is fired.
 632  */
 633 static void cpuhp_thread_fun(unsigned int cpu)
 634 {
 635         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 636         bool bringup = st->bringup;
 637         enum cpuhp_state state;
 638
 639         if (WARN_ON_ONCE(!st->should_run))
 640                 return;
 641
 642         /*
 643          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 644          * that if we see ->should_run we also see the rest of the state.
 645          */
 646         smp_mb();
 647
 648         /*
 649          * The BP holds the hotplug lock, but we're now running on the AP,
 650          * ensure that anybody asserting the lock is held, will actually find
 651          * it so.
 652          */
 653         lockdep_acquire_cpus_lock();
 654         cpuhp_lock_acquire(bringup);
 655
 656         if (st->single) {
 657                 state = st->cb_state;
 658                 st->should_run = false;
 659         } else {
 660                 if (bringup) {
 661                         st->state++;
 662                         state = st->state;
 663                         st->should_run = (st->state < st->target);
 664                         WARN_ON_ONCE(st->state > st->target);
 665                 } else {
 666                         state = st->state;
 667                         st->state--;
 668                         st->should_run = (st->state > st->target);
 669                         WARN_ON_ONCE(st->state < st->target);
 670                 }
 671         }
 672
 673         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 674
 675         if (cpuhp_is_atomic_state(state)) {
 676                 local_irq_disable();
 677                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 678                 local_irq_enable();
 679
 680                 /*
 681                  * STARTING/DYING must not fail!
 682                  */
 683                 WARN_ON_ONCE(st->result);
 684         } else {
 685                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 686         }
 687
 688         if (st->result) {
 689                 /*
 690                  * If we fail on a rollback, we're up a creek without no
 691                  * paddle, no way forward, no way back. We loose, thanks for
 692                  * playing.
 693                  */
 694                 WARN_ON_ONCE(st->rollback);
 695                 st->should_run = false;
 696         }
 697
 698         cpuhp_lock_release(bringup);
 699         lockdep_release_cpus_lock();
 700
 701         if (!st->should_run)
 702                 complete_ap_thread(st, bringup);
 703 }
 704
 705 /* Invoke a single callback on a remote cpu */
 706 static int
 707 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 708                          struct hlist_node *node)
 709 {
 710         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 711         int ret;
 712
 713         if (!cpu_online(cpu))
 714                 return 0;
 715
 716         cpuhp_lock_acquire(false);
 717         cpuhp_lock_release(false);
 718
 719         cpuhp_lock_acquire(true);
 720         cpuhp_lock_release(true);
 721
 722         /*
 723          * If we are up and running, use the hotplug thread. For early calls
 724          * we invoke the thread function directly.
 725          */
 726         if (!st->thread)
 727                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 728
 729         st->rollback = false;
 730         st->last = NULL;
 731
 732         st->node = node;
 733         st->bringup = bringup;
 734         st->cb_state = state;
 735         st->single = true;
 736
 737         __cpuhp_kick_ap(st);
 738
 739         /*
 740          * If we failed and did a partial, do a rollback.
 741          */
 742         if ((ret = st->result) && st->last) {
 743                 st->rollback = true;
 744                 st->bringup = !bringup;
 745
 746                 __cpuhp_kick_ap(st);
 747         }
 748
 749         /*
 750          * Clean up the leftovers so the next hotplug operation wont use stale
 751          * data.
 752          */
 753         st->node = st->last = NULL;
 754         return ret;
 755 }
 756
 757 static int cpuhp_kick_ap_work(unsigned int cpu)
 758 {
 759         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 760         enum cpuhp_state prev_state = st->state;
 761         int ret;
 762
 763         cpuhp_lock_acquire(false);
 764         cpuhp_lock_release(false);
 765
 766         cpuhp_lock_acquire(true);
 767         cpuhp_lock_release(true);
 768
 769         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 770         ret = cpuhp_kick_ap(st, st->target);
 771         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 772
 773         return ret;
 774 }
 775
 776 static struct smp_hotplug_thread cpuhp_threads = {
 777         .store                  = &cpuhp_state.thread,
 778         .create                 = &cpuhp_create,
 779         .thread_should_run      = cpuhp_should_run,
 780         .thread_fn              = cpuhp_thread_fun,
 781         .thread_comm            = "cpuhp/%u",
 782         .selfparking            = true,
 783 };
 784
 785 void __init cpuhp_threads_init(void)
 786 {
 787         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 788         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 789 }
 790
 791 #ifdef CONFIG_HOTPLUG_CPU
 792 /**
 793  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 794  * @cpu: a CPU id
 795  *
 796  * This function walks all processes, finds a valid mm struct for each one and
 797  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 798  * trivial, there are various non-obvious corner cases, which this function
 799  * tries to solve in a safe manner.
 800  *
 801  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 802  * be called only for an already offlined CPU.
 803  */
 804 void clear_tasks_mm_cpumask(int cpu)
 805 {
 806         struct task_struct *p;
 807
 808         /*
 809          * This function is called after the cpu is taken down and marked
 810          * offline, so its not like new tasks will ever get this cpu set in
 811          * their mm mask. -- Peter Zijlstra
 812          * Thus, we may use rcu_read_lock() here, instead of grabbing
 813          * full-fledged tasklist_lock.
 814          */
 815         WARN_ON(cpu_online(cpu));
 816         rcu_read_lock();
 817         for_each_process(p) {
 818                 struct task_struct *t;
 819
 820                 /*
 821                  * Main thread might exit, but other threads may still have
 822                  * a valid mm. Find one.
 823                  */
 824                 t = find_lock_task_mm(p);
 825                 if (!t)
 826                         continue;
 827                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 828                 task_unlock(t);
 829         }
 830         rcu_read_unlock();
 831 }
 832
 833 /* Take this CPU down. */
 834 static int take_cpu_down(void *_param)
 835 {
 836         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 837         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 838         int err, cpu = smp_processor_id();
 839         int ret;
 840
 841         /* Ensure this CPU doesn't handle any more interrupts. */
 842         err = __cpu_disable();
 843         if (err < 0)
 844                 return err;
 845
 846         /*
 847          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 848          * do this step again.
 849          */
 850         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 851         st->state--;
 852         /* Invoke the former CPU_DYING callbacks */
 853         for (; st->state > target; st->state--) {
 854                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 855                 /*
 856                  * DYING must not fail!
 857                  */
 858                 WARN_ON_ONCE(ret);
 859         }
 860
 861         /* Give up timekeeping duties */
 862         tick_handover_do_timer();
 863         /* Park the stopper thread */
 864         stop_machine_park(cpu);
 865         return 0;
 866 }
 867
 868 static int takedown_cpu(unsigned int cpu)
 869 {
 870         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 871         int err;
 872
 873         /* Park the smpboot threads */
 874         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 875
 876         /*
 877          * Prevent irq alloc/free while the dying cpu reorganizes the
 878          * interrupt affinities.
 879          */
 880         irq_lock_sparse();
 881
 882         /*
 883          * So now all preempt/rcu users must observe !cpu_active().
 884          */
 885         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 886         if (err) {
 887                 /* CPU refused to die */
 888                 irq_unlock_sparse();
 889                 /* Unpark the hotplug thread so we can rollback there */
 890                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 891                 return err;
 892         }
 893         BUG_ON(cpu_online(cpu));
 894
 895         /*
 896          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 897          * all runnable tasks from the CPU, there's only the idle task left now
 898          * that the migration thread is done doing the stop_machine thing.
 899          *
 900          * Wait for the stop thread to go away.
 901          */
 902         wait_for_ap_thread(st, false);
 903         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 904
 905         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 906         irq_unlock_sparse();
 907
 908         hotplug_cpu__broadcast_tick_pull(cpu);
 909         /* This actually kills the CPU. */
 910         __cpu_die(cpu);
 911
 912         tick_cleanup_dead_cpu(cpu);
 913         rcutree_migrate_callbacks(cpu);
 914         return 0;
 915 }
 916
 917 static void cpuhp_complete_idle_dead(void *arg)
 918 {
 919         struct cpuhp_cpu_state *st = arg;
 920
 921         complete_ap_thread(st, false);
 922 }
 923
 924 void cpuhp_report_idle_dead(void)
 925 {
 926         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 927
 928         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 929         rcu_report_dead(smp_processor_id());
 930         st->state = CPUHP_AP_IDLE_DEAD;
 931         /*
 932          * We cannot call complete after rcu_report_dead() so we delegate it
 933          * to an online cpu.
 934          */
 935         smp_call_function_single(cpumask_first(cpu_online_mask),
 936                                  cpuhp_complete_idle_dead, st, 0);
 937 }
 938
 939 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 940 {
 941         for (st->state++; st->state < st->target; st->state++)
 942                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 943 }
 944
 945 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 946                                 enum cpuhp_state target)
 947 {
 948         enum cpuhp_state prev_state = st->state;
 949         int ret = 0;
 950
 951         for (; st->state > target; st->state--) {
 952                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 953                 if (ret) {
 954                         st->target = prev_state;
 955                         if (st->state < prev_state)
 956                                 undo_cpu_down(cpu, st);
 957                         break;
 958                 }
 959         }
 960         return ret;
 961 }
 962
 963 /* Requires cpu_add_remove_lock to be held */
 964 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 965                            enum cpuhp_state target)
 966 {
 967         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 968         int prev_state, ret = 0;
 969
 970         if (num_online_cpus() == 1)
 971                 return -EBUSY;
 972
 973         if (!cpu_present(cpu))
 974                 return -EINVAL;
 975
 976         cpus_write_lock();
 977
 978         cpuhp_tasks_frozen = tasks_frozen;
 979
 980         prev_state = cpuhp_set_state(st, target);
 981         /*
 982          * If the current CPU state is in the range of the AP hotplug thread,
 983          * then we need to kick the thread.
 984          */
 985         if (st->state > CPUHP_TEARDOWN_CPU) {
 986                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 987                 ret = cpuhp_kick_ap_work(cpu);
 988                 /*
 989                  * The AP side has done the error rollback already. Just
 990                  * return the error code..
 991                  */
 992                 if (ret)
 993                         goto out;
 994
 995                 /*
 996                  * We might have stopped still in the range of the AP hotplug
 997                  * thread. Nothing to do anymore.
 998                  */
 999                 if (st->state > CPUHP_TEARDOWN_CPU)
1000                         goto out;
1001
1002                 st->target = target;
1003         }
1004         /*
1005          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1006          * to do the further cleanups.
1007          */
1008         ret = cpuhp_down_callbacks(cpu, st, target);
1009         if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1010                 cpuhp_reset_state(st, prev_state);
1011                 __cpuhp_kick_ap(st);
1012         }
1013
1014 out:
1015         cpus_write_unlock();
1016         /*
1017          * Do post unplug cleanup. This is still protected against
1018          * concurrent CPU hotplug via cpu_add_remove_lock.
1019          */
1020         lockup_detector_cleanup();
1021         arch_smt_update();
1022         return ret;
1023 }
1024
1025 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1026 {
1027         if (cpu_hotplug_disabled)
1028                 return -EBUSY;
1029         return _cpu_down(cpu, 0, target);
1030 }
1031
1032 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
1033 {
1034         int err;
1035
1036         cpu_maps_update_begin();
1037         err = cpu_down_maps_locked(cpu, target);
1038         cpu_maps_update_done();
1039         return err;
1040 }
1041
1042 int cpu_down(unsigned int cpu)
1043 {
1044         return do_cpu_down(cpu, CPUHP_OFFLINE);
1045 }
1046 EXPORT_SYMBOL(cpu_down);
1047
1048 #else
1049 #define takedown_cpu            NULL
1050 #endif /*CONFIG_HOTPLUG_CPU*/
1051
1052 /**
1053  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1054  * @cpu: cpu that just started
1055  *
1056  * It must be called by the arch code on the new cpu, before the new cpu
1057  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1058  */
1059 void notify_cpu_starting(unsigned int cpu)
1060 {
1061         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1062         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1063         int ret;
1064
1065         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1066         st->booted_once = true;
1067         while (st->state < target) {
1068                 st->state++;
1069                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1070                 /*
1071                  * STARTING must not fail!
1072                  */
1073                 WARN_ON_ONCE(ret);
1074         }
1075 }
1076
1077 /*
1078  * Called from the idle task. Wake up the controlling task which brings the
1079  * stopper and the hotplug thread of the upcoming CPU up and then delegates
1080  * the rest of the online bringup to the hotplug thread.
1081  */
1082 void cpuhp_online_idle(enum cpuhp_state state)
1083 {
1084         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1085
1086         /* Happens for the boot cpu */
1087         if (state != CPUHP_AP_ONLINE_IDLE)
1088                 return;
1089
1090         st->state = CPUHP_AP_ONLINE_IDLE;
1091         complete_ap_thread(st, true);
1092 }
1093
1094 /* Requires cpu_add_remove_lock to be held */
1095 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1096 {
1097         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1098         struct task_struct *idle;
1099         int ret = 0;
1100
1101         cpus_write_lock();
1102
1103         if (!cpu_present(cpu)) {
1104                 ret = -EINVAL;
1105                 goto out;
1106         }
1107
1108         /*
1109          * The caller of do_cpu_up might have raced with another
1110          * caller. Ignore it for now.
1111          */
1112         if (st->state >= target)
1113                 goto out;
1114
1115         if (st->state == CPUHP_OFFLINE) {
1116                 /* Let it fail before we try to bring the cpu up */
1117                 idle = idle_thread_get(cpu);
1118                 if (IS_ERR(idle)) {
1119                         ret = PTR_ERR(idle);
1120                         goto out;
1121                 }
1122         }
1123
1124         cpuhp_tasks_frozen = tasks_frozen;
1125
1126         cpuhp_set_state(st, target);
1127         /*
1128          * If the current CPU state is in the range of the AP hotplug thread,
1129          * then we need to kick the thread once more.
1130          */
1131         if (st->state > CPUHP_BRINGUP_CPU) {
1132                 ret = cpuhp_kick_ap_work(cpu);
1133                 /*
1134                  * The AP side has done the error rollback already. Just
1135                  * return the error code..
1136                  */
1137                 if (ret)
1138                         goto out;
1139         }
1140
1141         /*
1142          * Try to reach the target state. We max out on the BP at
1143          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1144          * responsible for bringing it up to the target state.
1145          */
1146         target = min((int)target, CPUHP_BRINGUP_CPU);
1147         ret = cpuhp_up_callbacks(cpu, st, target);
1148 out:
1149         cpus_write_unlock();
1150         arch_smt_update();
1151         return ret;
1152 }
1153
1154 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1155 {
1156         int err = 0;
1157
1158         if (!cpu_possible(cpu)) {
1159                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1160                        cpu);
1161 #if defined(CONFIG_IA64)
1162                 pr_err("please check additional_cpus= boot parameter\n");
1163 #endif
1164                 return -EINVAL;
1165         }
1166
1167         err = try_online_node(cpu_to_node(cpu));
1168         if (err)
1169                 return err;
1170
1171         cpu_maps_update_begin();
1172
1173         if (cpu_hotplug_disabled) {
1174                 err = -EBUSY;
1175                 goto out;
1176         }
1177         if (!cpu_smt_allowed(cpu)) {
1178                 err = -EPERM;
1179                 goto out;
1180         }
1181
1182         err = _cpu_up(cpu, 0, target);
1183 out:
1184         cpu_maps_update_done();
1185         return err;
1186 }
1187
1188 int cpu_up(unsigned int cpu)
1189 {
1190         return do_cpu_up(cpu, CPUHP_ONLINE);
1191 }
1192 EXPORT_SYMBOL_GPL(cpu_up);
1193
1194 #ifdef CONFIG_PM_SLEEP_SMP
1195 static cpumask_var_t frozen_cpus;
1196
1197 int freeze_secondary_cpus(int primary)
1198 {
1199         int cpu, error = 0;
1200
1201         cpu_maps_update_begin();
1202         if (!cpu_online(primary))
1203                 primary = cpumask_first(cpu_online_mask);
1204         /*
1205          * We take down all of the non-boot CPUs in one shot to avoid races
1206          * with the userspace trying to use the CPU hotplug at the same time
1207          */
1208         cpumask_clear(frozen_cpus);
1209
1210         pr_info("Disabling non-boot CPUs ...\n");
1211         for_each_online_cpu(cpu) {
1212                 if (cpu == primary)
1213                         continue;
1214                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1215                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1216                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1217                 if (!error)
1218                         cpumask_set_cpu(cpu, frozen_cpus);
1219                 else {
1220                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1221                         break;
1222                 }
1223         }
1224
1225         if (!error)
1226                 BUG_ON(num_online_cpus() > 1);
1227         else
1228                 pr_err("Non-boot CPUs are not disabled\n");
1229
1230         /*
1231          * Make sure the CPUs won't be enabled by someone else. We need to do
1232          * this even in case of failure as all disable_nonboot_cpus() users are
1233          * supposed to do enable_nonboot_cpus() on the failure path.
1234          */
1235         cpu_hotplug_disabled++;
1236
1237         cpu_maps_update_done();
1238         return error;
1239 }
1240
1241 void __weak arch_enable_nonboot_cpus_begin(void)
1242 {
1243 }
1244
1245 void __weak arch_enable_nonboot_cpus_end(void)
1246 {
1247 }
1248
1249 void enable_nonboot_cpus(void)
1250 {
1251         int cpu, error;
1252
1253         /* Allow everyone to use the CPU hotplug again */
1254         cpu_maps_update_begin();
1255         __cpu_hotplug_enable();
1256         if (cpumask_empty(frozen_cpus))
1257                 goto out;
1258
1259         pr_info("Enabling non-boot CPUs ...\n");
1260
1261         arch_enable_nonboot_cpus_begin();
1262
1263         for_each_cpu(cpu, frozen_cpus) {
1264                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1265                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1266                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1267                 if (!error) {
1268                         pr_info("CPU%d is up\n", cpu);
1269                         continue;
1270                 }
1271                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1272         }
1273
1274         arch_enable_nonboot_cpus_end();
1275
1276         cpumask_clear(frozen_cpus);
1277 out:
1278         cpu_maps_update_done();
1279 }
1280
1281 static int __init alloc_frozen_cpus(void)
1282 {
1283         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1284                 return -ENOMEM;
1285         return 0;
1286 }
1287 core_initcall(alloc_frozen_cpus);
1288
1289 /*
1290  * When callbacks for CPU hotplug notifications are being executed, we must
1291  * ensure that the state of the system with respect to the tasks being frozen
1292  * or not, as reported by the notification, remains unchanged *throughout the
1293  * duration* of the execution of the callbacks.
1294  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1295  *
1296  * This synchronization is implemented by mutually excluding regular CPU
1297  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1298  * Hibernate notifications.
1299  */
1300 static int
1301 cpu_hotplug_pm_callback(struct notifier_block *nb,
1302                         unsigned long action, void *ptr)
1303 {
1304         switch (action) {
1305
1306         case PM_SUSPEND_PREPARE:
1307         case PM_HIBERNATION_PREPARE:
1308                 cpu_hotplug_disable();
1309                 break;
1310
1311         case PM_POST_SUSPEND:
1312         case PM_POST_HIBERNATION:
1313                 cpu_hotplug_enable();
1314                 break;
1315
1316         default:
1317                 return NOTIFY_DONE;
1318         }
1319
1320         return NOTIFY_OK;
1321 }
1322
1323
1324 static int __init cpu_hotplug_pm_sync_init(void)
1325 {
1326         /*
1327          * cpu_hotplug_pm_callback has higher priority than x86
1328          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1329          * to disable cpu hotplug to avoid cpu hotplug race.
1330          */
1331         pm_notifier(cpu_hotplug_pm_callback, 0);
1332         return 0;
1333 }
1334 core_initcall(cpu_hotplug_pm_sync_init);
1335
1336 #endif /* CONFIG_PM_SLEEP_SMP */
1337
1338 int __boot_cpu_id;
1339
1340 #endif /* CONFIG_SMP */
1341
1342 /* Boot processor state steps */
1343 static struct cpuhp_step cpuhp_hp_states[] = {
1344         [CPUHP_OFFLINE] = {
1345                 .name                   = "offline",
1346                 .startup.single         = NULL,
1347                 .teardown.single        = NULL,
1348         },
1349 #ifdef CONFIG_SMP
1350         [CPUHP_CREATE_THREADS]= {
1351                 .name                   = "threads:prepare",
1352                 .startup.single         = smpboot_create_threads,
1353                 .teardown.single        = NULL,
1354                 .cant_stop              = true,
1355         },
1356         [CPUHP_PERF_PREPARE] = {
1357                 .name                   = "perf:prepare",
1358                 .startup.single         = perf_event_init_cpu,
1359                 .teardown.single        = perf_event_exit_cpu,
1360         },
1361         [CPUHP_WORKQUEUE_PREP] = {
1362                 .name                   = "workqueue:prepare",
1363                 .startup.single         = workqueue_prepare_cpu,
1364                 .teardown.single        = NULL,
1365         },
1366         [CPUHP_HRTIMERS_PREPARE] = {
1367                 .name                   = "hrtimers:prepare",
1368                 .startup.single         = hrtimers_prepare_cpu,
1369                 .teardown.single        = hrtimers_dead_cpu,
1370         },
1371         [CPUHP_SMPCFD_PREPARE] = {
1372                 .name                   = "smpcfd:prepare",
1373                 .startup.single         = smpcfd_prepare_cpu,
1374                 .teardown.single        = smpcfd_dead_cpu,
1375         },
1376         [CPUHP_RELAY_PREPARE] = {
1377                 .name                   = "relay:prepare",
1378                 .startup.single         = relay_prepare_cpu,
1379                 .teardown.single        = NULL,
1380         },
1381         [CPUHP_SLAB_PREPARE] = {
1382                 .name                   = "slab:prepare",
1383                 .startup.single         = slab_prepare_cpu,
1384                 .teardown.single        = slab_dead_cpu,
1385         },
1386         [CPUHP_RCUTREE_PREP] = {
1387                 .name                   = "RCU/tree:prepare",
1388                 .startup.single         = rcutree_prepare_cpu,
1389                 .teardown.single        = rcutree_dead_cpu,
1390         },
1391         /*
1392          * On the tear-down path, timers_dead_cpu() must be invoked
1393          * before blk_mq_queue_reinit_notify() from notify_dead(),
1394          * otherwise a RCU stall occurs.
1395          */
1396         [CPUHP_TIMERS_PREPARE] = {
1397                 .name                   = "timers:prepare",
1398                 .startup.single         = timers_prepare_cpu,
1399                 .teardown.single        = timers_dead_cpu,
1400         },
1401         /* Kicks the plugged cpu into life */
1402         [CPUHP_BRINGUP_CPU] = {
1403                 .name                   = "cpu:bringup",
1404                 .startup.single         = bringup_cpu,
1405                 .teardown.single        = NULL,
1406                 .cant_stop              = true,
1407         },
1408         /* Final state before CPU kills itself */
1409         [CPUHP_AP_IDLE_DEAD] = {
1410                 .name                   = "idle:dead",
1411         },
1412         /*
1413          * Last state before CPU enters the idle loop to die. Transient state
1414          * for synchronization.
1415          */
1416         [CPUHP_AP_OFFLINE] = {
1417                 .name                   = "ap:offline",
1418                 .cant_stop              = true,
1419         },
1420         /* First state is scheduler control. Interrupts are disabled */
1421         [CPUHP_AP_SCHED_STARTING] = {
1422                 .name                   = "sched:starting",
1423                 .startup.single         = sched_cpu_starting,
1424                 .teardown.single        = sched_cpu_dying,
1425         },
1426         [CPUHP_AP_RCUTREE_DYING] = {
1427                 .name                   = "RCU/tree:dying",
1428                 .startup.single         = NULL,
1429                 .teardown.single        = rcutree_dying_cpu,
1430         },
1431         [CPUHP_AP_SMPCFD_DYING] = {
1432                 .name                   = "smpcfd:dying",
1433                 .startup.single         = NULL,
1434                 .teardown.single        = smpcfd_dying_cpu,
1435         },
1436         /* Entry state on starting. Interrupts enabled from here on. Transient
1437          * state for synchronsization */
1438         [CPUHP_AP_ONLINE] = {
1439                 .name                   = "ap:online",
1440         },
1441         /*
1442          * Handled on controll processor until the plugged processor manages
1443          * this itself.
1444          */
1445         [CPUHP_TEARDOWN_CPU] = {
1446                 .name                   = "cpu:teardown",
1447                 .startup.single         = NULL,
1448                 .teardown.single        = takedown_cpu,
1449                 .cant_stop              = true,
1450         },
1451         /* Handle smpboot threads park/unpark */
1452         [CPUHP_AP_SMPBOOT_THREADS] = {
1453                 .name                   = "smpboot/threads:online",
1454                 .startup.single         = smpboot_unpark_threads,
1455                 .teardown.single        = smpboot_park_threads,
1456         },
1457         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1458                 .name                   = "irq/affinity:online",
1459                 .startup.single         = irq_affinity_online_cpu,
1460                 .teardown.single        = NULL,
1461         },
1462         [CPUHP_AP_PERF_ONLINE] = {
1463                 .name                   = "perf:online",
1464                 .startup.single         = perf_event_init_cpu,
1465                 .teardown.single        = perf_event_exit_cpu,
1466         },
1467         [CPUHP_AP_WATCHDOG_ONLINE] = {
1468                 .name                   = "lockup_detector:online",
1469                 .startup.single         = lockup_detector_online_cpu,
1470                 .teardown.single        = lockup_detector_offline_cpu,
1471         },
1472         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1473                 .name                   = "workqueue:online",
1474                 .startup.single         = workqueue_online_cpu,
1475                 .teardown.single        = workqueue_offline_cpu,
1476         },
1477         [CPUHP_AP_RCUTREE_ONLINE] = {
1478                 .name                   = "RCU/tree:online",
1479                 .startup.single         = rcutree_online_cpu,
1480                 .teardown.single        = rcutree_offline_cpu,
1481         },
1482 #endif
1483         /*
1484          * The dynamically registered state space is here
1485          */
1486
1487 #ifdef CONFIG_SMP
1488         /* Last state is scheduler control setting the cpu active */
1489         [CPUHP_AP_ACTIVE] = {
1490                 .name                   = "sched:active",
1491                 .startup.single         = sched_cpu_activate,
1492                 .teardown.single        = sched_cpu_deactivate,
1493         },
1494 #endif
1495
1496         /* CPU is fully up and running. */
1497         [CPUHP_ONLINE] = {
1498                 .name                   = "online",
1499                 .startup.single         = NULL,
1500                 .teardown.single        = NULL,
1501         },
1502 };
1503
1504 /* Sanity check for callbacks */
1505 static int cpuhp_cb_check(enum cpuhp_state state)
1506 {
1507         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1508                 return -EINVAL;
1509         return 0;
1510 }
1511
1512 /*
1513  * Returns a free for dynamic slot assignment of the Online state. The states
1514  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1515  * by having no name assigned.
1516  */
1517 static int cpuhp_reserve_state(enum cpuhp_state state)
1518 {
1519         enum cpuhp_state i, end;
1520         struct cpuhp_step *step;
1521
1522         switch (state) {
1523         case CPUHP_AP_ONLINE_DYN:
1524                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1525                 end = CPUHP_AP_ONLINE_DYN_END;
1526                 break;
1527         case CPUHP_BP_PREPARE_DYN:
1528                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1529                 end = CPUHP_BP_PREPARE_DYN_END;
1530                 break;
1531         default:
1532                 return -EINVAL;
1533         }
1534
1535         for (i = state; i <= end; i++, step++) {
1536                 if (!step->name)
1537                         return i;
1538         }
1539         WARN(1, "No more dynamic states available for CPU hotplug\n");
1540         return -ENOSPC;
1541 }
1542
1543 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1544                                  int (*startup)(unsigned int cpu),
1545                                  int (*teardown)(unsigned int cpu),
1546                                  bool multi_instance)
1547 {
1548         /* (Un)Install the callbacks for further cpu hotplug operations */
1549         struct cpuhp_step *sp;
1550         int ret = 0;
1551
1552         /*
1553          * If name is NULL, then the state gets removed.
1554          *
1555          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1556          * the first allocation from these dynamic ranges, so the removal
1557          * would trigger a new allocation and clear the wrong (already
1558          * empty) state, leaving the callbacks of the to be cleared state
1559          * dangling, which causes wreckage on the next hotplug operation.
1560          */
1561         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1562                      state == CPUHP_BP_PREPARE_DYN)) {
1563                 ret = cpuhp_reserve_state(state);
1564                 if (ret < 0)
1565                         return ret;
1566                 state = ret;
1567         }
1568         sp = cpuhp_get_step(state);
1569         if (name && sp->name)
1570                 return -EBUSY;
1571
1572         sp->startup.single = startup;
1573         sp->teardown.single = teardown;
1574         sp->name = name;
1575         sp->multi_instance = multi_instance;
1576         INIT_HLIST_HEAD(&sp->list);
1577         return ret;
1578 }
1579
1580 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1581 {
1582         return cpuhp_get_step(state)->teardown.single;
1583 }
1584
1585 /*
1586  * Call the startup/teardown function for a step either on the AP or
1587  * on the current CPU.
1588  */
1589 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1590                             struct hlist_node *node)
1591 {
1592         struct cpuhp_step *sp = cpuhp_get_step(state);
1593         int ret;
1594
1595         /*
1596          * If there's nothing to do, we done.
1597          * Relies on the union for multi_instance.
1598          */
1599         if ((bringup && !sp->startup.single) ||
1600             (!bringup && !sp->teardown.single))
1601                 return 0;
1602         /*
1603          * The non AP bound callbacks can fail on bringup. On teardown
1604          * e.g. module removal we crash for now.
1605          */
1606 #ifdef CONFIG_SMP
1607         if (cpuhp_is_ap_state(state))
1608                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1609         else
1610                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1611 #else
1612         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1613 #endif
1614         BUG_ON(ret && !bringup);
1615         return ret;
1616 }
1617
1618 /*
1619  * Called from __cpuhp_setup_state on a recoverable failure.
1620  *
1621  * Note: The teardown callbacks for rollback are not allowed to fail!
1622  */
1623 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1624                                    struct hlist_node *node)
1625 {
1626         int cpu;
1627
1628         /* Roll back the already executed steps on the other cpus */
1629         for_each_present_cpu(cpu) {
1630                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1631                 int cpustate = st->state;
1632
1633                 if (cpu >= failedcpu)
1634                         break;
1635
1636                 /* Did we invoke the startup call on that cpu ? */
1637                 if (cpustate >= state)
1638                         cpuhp_issue_call(cpu, state, false, node);
1639         }
1640 }
1641
1642 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1643                                           struct hlist_node *node,
1644                                           bool invoke)
1645 {
1646         struct cpuhp_step *sp;
1647         int cpu;
1648         int ret;
1649
1650         lockdep_assert_cpus_held();
1651
1652         sp = cpuhp_get_step(state);
1653         if (sp->multi_instance == false)
1654                 return -EINVAL;
1655
1656         mutex_lock(&cpuhp_state_mutex);
1657
1658         if (!invoke || !sp->startup.multi)
1659                 goto add_node;
1660
1661         /*
1662          * Try to call the startup callback for each present cpu
1663          * depending on the hotplug state of the cpu.
1664          */
1665         for_each_present_cpu(cpu) {
1666                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1667                 int cpustate = st->state;
1668
1669                 if (cpustate < state)
1670                         continue;
1671
1672                 ret = cpuhp_issue_call(cpu, state, true, node);
1673                 if (ret) {
1674                         if (sp->teardown.multi)
1675                                 cpuhp_rollback_install(cpu, state, node);
1676                         goto unlock;
1677                 }
1678         }
1679 add_node:
1680         ret = 0;
1681         hlist_add_head(node, &sp->list);
1682 unlock:
1683         mutex_unlock(&cpuhp_state_mutex);
1684         return ret;
1685 }
1686
1687 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1688                                bool invoke)
1689 {
1690         int ret;
1691
1692         cpus_read_lock();
1693         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1694         cpus_read_unlock();
1695         return ret;
1696 }
1697 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1698
1699 /**
1700  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1701  * @state:              The state to setup
1702  * @invoke:             If true, the startup function is invoked for cpus where
1703  *                      cpu state >= @state
1704  * @startup:            startup callback function
1705  * @teardown:           teardown callback function
1706  * @multi_instance:     State is set up for multiple instances which get
1707  *                      added afterwards.
1708  *
1709  * The caller needs to hold cpus read locked while calling this function.
1710  * Returns:
1711  *   On success:
1712  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1713  *      0 for all other states
1714  *   On failure: proper (negative) error code
1715  */
1716 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1717                                    const char *name, bool invoke,
1718                                    int (*startup)(unsigned int cpu),
1719                                    int (*teardown)(unsigned int cpu),
1720                                    bool multi_instance)
1721 {
1722         int cpu, ret = 0;
1723         bool dynstate;
1724
1725         lockdep_assert_cpus_held();
1726
1727         if (cpuhp_cb_check(state) || !name)
1728                 return -EINVAL;
1729
1730         mutex_lock(&cpuhp_state_mutex);
1731
1732         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1733                                     multi_instance);
1734
1735         dynstate = state == CPUHP_AP_ONLINE_DYN;
1736         if (ret > 0 && dynstate) {
1737                 state = ret;
1738                 ret = 0;
1739         }
1740
1741         if (ret || !invoke || !startup)
1742                 goto out;
1743
1744         /*
1745          * Try to call the startup callback for each present cpu
1746          * depending on the hotplug state of the cpu.
1747          */
1748         for_each_present_cpu(cpu) {
1749                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1750                 int cpustate = st->state;
1751
1752                 if (cpustate < state)
1753                         continue;
1754
1755                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1756                 if (ret) {
1757                         if (teardown)
1758                                 cpuhp_rollback_install(cpu, state, NULL);
1759                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1760                         goto out;
1761                 }
1762         }
1763 out:
1764         mutex_unlock(&cpuhp_state_mutex);
1765         /*
1766          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1767          * dynamically allocated state in case of success.
1768          */
1769         if (!ret && dynstate)
1770                 return state;
1771         return ret;
1772 }
1773 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1774
1775 int __cpuhp_setup_state(enum cpuhp_state state,
1776                         const char *name, bool invoke,
1777                         int (*startup)(unsigned int cpu),
1778                         int (*teardown)(unsigned int cpu),
1779                         bool multi_instance)
1780 {
1781         int ret;
1782
1783         cpus_read_lock();
1784         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1785                                              teardown, multi_instance);
1786         cpus_read_unlock();
1787         return ret;
1788 }
1789 EXPORT_SYMBOL(__cpuhp_setup_state);
1790
1791 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1792                                   struct hlist_node *node, bool invoke)
1793 {
1794         struct cpuhp_step *sp = cpuhp_get_step(state);
1795         int cpu;
1796
1797         BUG_ON(cpuhp_cb_check(state));
1798
1799         if (!sp->multi_instance)
1800                 return -EINVAL;
1801
1802         cpus_read_lock();
1803         mutex_lock(&cpuhp_state_mutex);
1804
1805         if (!invoke || !cpuhp_get_teardown_cb(state))
1806                 goto remove;
1807         /*
1808          * Call the teardown callback for each present cpu depending
1809          * on the hotplug state of the cpu. This function is not
1810          * allowed to fail currently!
1811          */
1812         for_each_present_cpu(cpu) {
1813                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1814                 int cpustate = st->state;
1815
1816                 if (cpustate >= state)
1817                         cpuhp_issue_call(cpu, state, false, node);
1818         }
1819
1820 remove:
1821         hlist_del(node);
1822         mutex_unlock(&cpuhp_state_mutex);
1823         cpus_read_unlock();
1824
1825         return 0;
1826 }
1827 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1828
1829 /**
1830  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1831  * @state:      The state to remove
1832  * @invoke:     If true, the teardown function is invoked for cpus where
1833  *              cpu state >= @state
1834  *
1835  * The caller needs to hold cpus read locked while calling this function.
1836  * The teardown callback is currently not allowed to fail. Think
1837  * about module removal!
1838  */
1839 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1840 {
1841         struct cpuhp_step *sp = cpuhp_get_step(state);
1842         int cpu;
1843
1844         BUG_ON(cpuhp_cb_check(state));
1845
1846         lockdep_assert_cpus_held();
1847
1848         mutex_lock(&cpuhp_state_mutex);
1849         if (sp->multi_instance) {
1850                 WARN(!hlist_empty(&sp->list),
1851                      "Error: Removing state %d which has instances left.\n",
1852                      state);
1853                 goto remove;
1854         }
1855
1856         if (!invoke || !cpuhp_get_teardown_cb(state))
1857                 goto remove;
1858
1859         /*
1860          * Call the teardown callback for each present cpu depending
1861          * on the hotplug state of the cpu. This function is not
1862          * allowed to fail currently!
1863          */
1864         for_each_present_cpu(cpu) {
1865                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1866                 int cpustate = st->state;
1867
1868                 if (cpustate >= state)
1869                         cpuhp_issue_call(cpu, state, false, NULL);
1870         }
1871 remove:
1872         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1873         mutex_unlock(&cpuhp_state_mutex);
1874 }
1875 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1876
1877 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1878 {
1879         cpus_read_lock();
1880         __cpuhp_remove_state_cpuslocked(state, invoke);
1881         cpus_read_unlock();
1882 }
1883 EXPORT_SYMBOL(__cpuhp_remove_state);
1884
1885 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1886 static ssize_t show_cpuhp_state(struct device *dev,
1887                                 struct device_attribute *attr, char *buf)
1888 {
1889         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1890
1891         return sprintf(buf, "%d\n", st->state);
1892 }
1893 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1894
1895 static ssize_t write_cpuhp_target(struct device *dev,
1896                                   struct device_attribute *attr,
1897                                   const char *buf, size_t count)
1898 {
1899         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1900         struct cpuhp_step *sp;
1901         int target, ret;
1902
1903         ret = kstrtoint(buf, 10, &target);
1904         if (ret)
1905                 return ret;
1906
1907 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1908         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1909                 return -EINVAL;
1910 #else
1911         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1912                 return -EINVAL;
1913 #endif
1914
1915         ret = lock_device_hotplug_sysfs();
1916         if (ret)
1917                 return ret;
1918
1919         mutex_lock(&cpuhp_state_mutex);
1920         sp = cpuhp_get_step(target);
1921         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1922         mutex_unlock(&cpuhp_state_mutex);
1923         if (ret)
1924                 goto out;
1925
1926         if (st->state < target)
1927                 ret = do_cpu_up(dev->id, target);
1928         else
1929                 ret = do_cpu_down(dev->id, target);
1930 out:
1931         unlock_device_hotplug();
1932         return ret ? ret : count;
1933 }
1934
1935 static ssize_t show_cpuhp_target(struct device *dev,
1936                                  struct device_attribute *attr, char *buf)
1937 {
1938         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1939
1940         return sprintf(buf, "%d\n", st->target);
1941 }
1942 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1943
1944
1945 static ssize_t write_cpuhp_fail(struct device *dev,
1946                                 struct device_attribute *attr,
1947                                 const char *buf, size_t count)
1948 {
1949         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1950         struct cpuhp_step *sp;
1951         int fail, ret;
1952
1953         ret = kstrtoint(buf, 10, &fail);
1954         if (ret)
1955                 return ret;
1956
1957         /*
1958          * Cannot fail STARTING/DYING callbacks.
1959          */
1960         if (cpuhp_is_atomic_state(fail))
1961                 return -EINVAL;
1962
1963         /*
1964          * Cannot fail anything that doesn't have callbacks.
1965          */
1966         mutex_lock(&cpuhp_state_mutex);
1967         sp = cpuhp_get_step(fail);
1968         if (!sp->startup.single && !sp->teardown.single)
1969                 ret = -EINVAL;
1970         mutex_unlock(&cpuhp_state_mutex);
1971         if (ret)
1972                 return ret;
1973
1974         st->fail = fail;
1975
1976         return count;
1977 }
1978
1979 static ssize_t show_cpuhp_fail(struct device *dev,
1980                                struct device_attribute *attr, char *buf)
1981 {
1982         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1983
1984         return sprintf(buf, "%d\n", st->fail);
1985 }
1986
1987 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1988
1989 static struct attribute *cpuhp_cpu_attrs[] = {
1990         &dev_attr_state.attr,
1991         &dev_attr_target.attr,
1992         &dev_attr_fail.attr,
1993         NULL
1994 };
1995
1996 static const struct attribute_group cpuhp_cpu_attr_group = {
1997         .attrs = cpuhp_cpu_attrs,
1998         .name = "hotplug",
1999         NULL
2000 };
2001
2002 static ssize_t show_cpuhp_states(struct device *dev,
2003                                  struct device_attribute *attr, char *buf)
2004 {
2005         ssize_t cur, res = 0;
2006         int i;
2007
2008         mutex_lock(&cpuhp_state_mutex);
2009         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2010                 struct cpuhp_step *sp = cpuhp_get_step(i);
2011
2012                 if (sp->name) {
2013                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2014                         buf += cur;
2015                         res += cur;
2016                 }
2017         }
2018         mutex_unlock(&cpuhp_state_mutex);
2019         return res;
2020 }
2021 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2022
2023 static struct attribute *cpuhp_cpu_root_attrs[] = {
2024         &dev_attr_states.attr,
2025         NULL
2026 };
2027
2028 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2029         .attrs = cpuhp_cpu_root_attrs,
2030         .name = "hotplug",
2031         NULL
2032 };
2033
2034 #ifdef CONFIG_HOTPLUG_SMT
2035
2036 static const char *smt_states[] = {
2037         [CPU_SMT_ENABLED]               = "on",
2038         [CPU_SMT_DISABLED]              = "off",
2039         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2040         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2041 };
2042
2043 static ssize_t
2044 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2045 {
2046         return snprintf(buf, PAGE_SIZE - 2, "%s\n", smt_states[cpu_smt_control]);
2047 }
2048
2049 static void cpuhp_offline_cpu_device(unsigned int cpu)
2050 {
2051         struct device *dev = get_cpu_device(cpu);
2052
2053         dev->offline = true;
2054         /* Tell user space about the state change */
2055         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2056 }
2057
2058 static void cpuhp_online_cpu_device(unsigned int cpu)
2059 {
2060         struct device *dev = get_cpu_device(cpu);
2061
2062         dev->offline = false;
2063         /* Tell user space about the state change */
2064         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2065 }
2066
2067 static int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2068 {
2069         int cpu, ret = 0;
2070
2071         cpu_maps_update_begin();
2072         for_each_online_cpu(cpu) {
2073                 if (topology_is_primary_thread(cpu))
2074                         continue;
2075                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2076                 if (ret)
2077                         break;
2078                 /*
2079                  * As this needs to hold the cpu maps lock it's impossible
2080                  * to call device_offline() because that ends up calling
2081                  * cpu_down() which takes cpu maps lock. cpu maps lock
2082                  * needs to be held as this might race against in kernel
2083                  * abusers of the hotplug machinery (thermal management).
2084                  *
2085                  * So nothing would update device:offline state. That would
2086                  * leave the sysfs entry stale and prevent onlining after
2087                  * smt control has been changed to 'off' again. This is
2088                  * called under the sysfs hotplug lock, so it is properly
2089                  * serialized against the regular offline usage.
2090                  */
2091                 cpuhp_offline_cpu_device(cpu);
2092         }
2093         if (!ret) {
2094                 cpu_smt_control = ctrlval;
2095                 arch_smt_update();
2096         }
2097         cpu_maps_update_done();
2098         return ret;
2099 }
2100
2101 static int cpuhp_smt_enable(void)
2102 {
2103         int cpu, ret = 0;
2104
2105         cpu_maps_update_begin();
2106         cpu_smt_control = CPU_SMT_ENABLED;
2107         arch_smt_update();
2108         for_each_present_cpu(cpu) {
2109                 /* Skip online CPUs and CPUs on offline nodes */
2110                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2111                         continue;
2112                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2113                 if (ret)
2114                         break;
2115                 /* See comment in cpuhp_smt_disable() */
2116                 cpuhp_online_cpu_device(cpu);
2117         }
2118         cpu_maps_update_done();
2119         return ret;
2120 }
2121
2122 static ssize_t
2123 store_smt_control(struct device *dev, struct device_attribute *attr,
2124                   const char *buf, size_t count)
2125 {
2126         int ctrlval, ret;
2127
2128         if (sysfs_streq(buf, "on"))
2129                 ctrlval = CPU_SMT_ENABLED;
2130         else if (sysfs_streq(buf, "off"))
2131                 ctrlval = CPU_SMT_DISABLED;
2132         else if (sysfs_streq(buf, "forceoff"))
2133                 ctrlval = CPU_SMT_FORCE_DISABLED;
2134         else
2135                 return -EINVAL;
2136
2137         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2138                 return -EPERM;
2139
2140         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2141                 return -ENODEV;
2142
2143         ret = lock_device_hotplug_sysfs();
2144         if (ret)
2145                 return ret;
2146
2147         if (ctrlval != cpu_smt_control) {
2148                 switch (ctrlval) {
2149                 case CPU_SMT_ENABLED:
2150                         ret = cpuhp_smt_enable();
2151                         break;
2152                 case CPU_SMT_DISABLED:
2153                 case CPU_SMT_FORCE_DISABLED:
2154                         ret = cpuhp_smt_disable(ctrlval);
2155                         break;
2156                 }
2157         }
2158
2159         unlock_device_hotplug();
2160         return ret ? ret : count;
2161 }
2162 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2163
2164 static ssize_t
2165 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2166 {
2167         bool active = topology_max_smt_threads() > 1;
2168
2169         return snprintf(buf, PAGE_SIZE - 2, "%d\n", active);
2170 }
2171 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2172
2173 static struct attribute *cpuhp_smt_attrs[] = {
2174         &dev_attr_control.attr,
2175         &dev_attr_active.attr,
2176         NULL
2177 };
2178
2179 static const struct attribute_group cpuhp_smt_attr_group = {
2180         .attrs = cpuhp_smt_attrs,
2181         .name = "smt",
2182         NULL
2183 };
2184
2185 static int __init cpu_smt_state_init(void)
2186 {
2187         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2188                                   &cpuhp_smt_attr_group);
2189 }
2190
2191 #else
2192 static inline int cpu_smt_state_init(void) { return 0; }
2193 #endif
2194
2195 static int __init cpuhp_sysfs_init(void)
2196 {
2197         int cpu, ret;
2198
2199         ret = cpu_smt_state_init();
2200         if (ret)
2201                 return ret;
2202
2203         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2204                                  &cpuhp_cpu_root_attr_group);
2205         if (ret)
2206                 return ret;
2207
2208         for_each_possible_cpu(cpu) {
2209                 struct device *dev = get_cpu_device(cpu);
2210
2211                 if (!dev)
2212                         continue;
2213                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2214                 if (ret)
2215                         return ret;
2216         }
2217         return 0;
2218 }
2219 device_initcall(cpuhp_sysfs_init);
2220 #endif
2221
2222 /*
2223  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2224  * represents all NR_CPUS bits binary values of 1<<nr.
2225  *
2226  * It is used by cpumask_of() to get a constant address to a CPU
2227  * mask value that has a single bit set only.
2228  */
2229
2230 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2231 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2232 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2233 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2234 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2235
2236 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2237
2238         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2239         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2240 #if BITS_PER_LONG > 32
2241         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2242         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2243 #endif
2244 };
2245 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2246
2247 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2248 EXPORT_SYMBOL(cpu_all_bits);
2249
2250 #ifdef CONFIG_INIT_ALL_POSSIBLE
2251 struct cpumask __cpu_possible_mask __read_mostly
2252         = {CPU_BITS_ALL};
2253 #else
2254 struct cpumask __cpu_possible_mask __read_mostly;
2255 #endif
2256 EXPORT_SYMBOL(__cpu_possible_mask);
2257
2258 struct cpumask __cpu_online_mask __read_mostly;
2259 EXPORT_SYMBOL(__cpu_online_mask);
2260
2261 struct cpumask __cpu_present_mask __read_mostly;
2262 EXPORT_SYMBOL(__cpu_present_mask);
2263
2264 struct cpumask __cpu_active_mask __read_mostly;
2265 EXPORT_SYMBOL(__cpu_active_mask);
2266
2267 void init_cpu_present(const struct cpumask *src)
2268 {
2269         cpumask_copy(&__cpu_present_mask, src);
2270 }
2271
2272 void init_cpu_possible(const struct cpumask *src)
2273 {
2274         cpumask_copy(&__cpu_possible_mask, src);
2275 }
2276
2277 void init_cpu_online(const struct cpumask *src)
2278 {
2279         cpumask_copy(&__cpu_online_mask, src);
2280 }
2281
2282 /*
2283  * Activate the first processor.
2284  */
2285 void __init boot_cpu_init(void)
2286 {
2287         int cpu = smp_processor_id();
2288
2289         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2290         set_cpu_online(cpu, true);
2291         set_cpu_active(cpu, true);
2292         set_cpu_present(cpu, true);
2293         set_cpu_possible(cpu, true);
2294
2295 #ifdef CONFIG_SMP
2296         __boot_cpu_id = cpu;
2297 #endif
2298 }
2299
2300 /*
2301  * Must be called _AFTER_ setting up the per_cpu areas
2302  */
2303 void __init boot_cpu_hotplug_init(void)
2304 {
2305 #ifdef CONFIG_SMP
2306         this_cpu_write(cpuhp_state.booted_once, true);
2307 #endif
2308         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2309 }