kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/proc_fs.h>
   7 #include <linux/smp.h>
   8 #include <linux/init.h>
   9 #include <linux/notifier.h>
  10 #include <linux/sched.h>
  11 #include <linux/unistd.h>
  12 #include <linux/cpu.h>
  13 #include <linux/export.h>
  14 #include <linux/kthread.h>
  15 #include <linux/stop_machine.h>
  16 #include <linux/mutex.h>
  17 #include <linux/gfp.h>
  18 #include <linux/suspend.h>
  19
  20 #ifdef CONFIG_SMP
  21 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
  22 static DEFINE_MUTEX(cpu_add_remove_lock);
  23
  24 /*
  25  * The following two API's must be used when attempting
  26  * to serialize the updates to cpu_online_mask, cpu_present_mask.
  27  */
  28 void cpu_maps_update_begin(void)
  29 {
  30         mutex_lock(&cpu_add_remove_lock);
  31 }
  32
  33 void cpu_maps_update_done(void)
  34 {
  35         mutex_unlock(&cpu_add_remove_lock);
  36 }
  37
  38 static RAW_NOTIFIER_HEAD(cpu_chain);
  39
  40 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  41  * Should always be manipulated under cpu_add_remove_lock
  42  */
  43 static int cpu_hotplug_disabled;
  44
  45 #ifdef CONFIG_HOTPLUG_CPU
  46
  47 static struct {
  48         struct task_struct *active_writer;
  49         struct mutex lock; /* Synchronizes accesses to refcount, */
  50         /*
  51          * Also blocks the new readers during
  52          * an ongoing cpu hotplug operation.
  53          */
  54         int refcount;
  55 } cpu_hotplug = {
  56         .active_writer = NULL,
  57         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
  58         .refcount = 0,
  59 };
  60
  61 void get_online_cpus(void)
  62 {
  63         might_sleep();
  64         if (cpu_hotplug.active_writer == current)
  65                 return;
  66         mutex_lock(&cpu_hotplug.lock);
  67         cpu_hotplug.refcount++;
  68         mutex_unlock(&cpu_hotplug.lock);
  69
  70 }
  71 EXPORT_SYMBOL_GPL(get_online_cpus);
  72
  73 void put_online_cpus(void)
  74 {
  75         if (cpu_hotplug.active_writer == current)
  76                 return;
  77         mutex_lock(&cpu_hotplug.lock);
  78         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
  79                 wake_up_process(cpu_hotplug.active_writer);
  80         mutex_unlock(&cpu_hotplug.lock);
  81
  82 }
  83 EXPORT_SYMBOL_GPL(put_online_cpus);
  84
  85 /*
  86  * This ensures that the hotplug operation can begin only when the
  87  * refcount goes to zero.
  88  *
  89  * Note that during a cpu-hotplug operation, the new readers, if any,
  90  * will be blocked by the cpu_hotplug.lock
  91  *
  92  * Since cpu_hotplug_begin() is always called after invoking
  93  * cpu_maps_update_begin(), we can be sure that only one writer is active.
  94  *
  95  * Note that theoretically, there is a possibility of a livelock:
  96  * - Refcount goes to zero, last reader wakes up the sleeping
  97  *   writer.
  98  * - Last reader unlocks the cpu_hotplug.lock.
  99  * - A new reader arrives at this moment, bumps up the refcount.
 100  * - The writer acquires the cpu_hotplug.lock finds the refcount
 101  *   non zero and goes to sleep again.
 102  *
 103  * However, this is very difficult to achieve in practice since
 104  * get_online_cpus() not an api which is called all that often.
 105  *
 106  */
 107 static void cpu_hotplug_begin(void)
 108 {
 109         cpu_hotplug.active_writer = current;
 110
 111         for (;;) {
 112                 mutex_lock(&cpu_hotplug.lock);
 113                 if (likely(!cpu_hotplug.refcount))
 114                         break;
 115                 __set_current_state(TASK_UNINTERRUPTIBLE);
 116                 mutex_unlock(&cpu_hotplug.lock);
 117                 schedule();
 118         }
 119 }
 120
 121 static void cpu_hotplug_done(void)
 122 {
 123         cpu_hotplug.active_writer = NULL;
 124         mutex_unlock(&cpu_hotplug.lock);
 125 }
 126
 127 /*
 128  * Wait for currently running CPU hotplug operations to complete (if any) and
 129  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 130  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 131  * hotplug path before performing hotplug operations. So acquiring that lock
 132  * guarantees mutual exclusion from any currently running hotplug operations.
 133  */
 134 void cpu_hotplug_disable(void)
 135 {
 136         cpu_maps_update_begin();
 137         cpu_hotplug_disabled = 1;
 138         cpu_maps_update_done();
 139 }
 140
 141 void cpu_hotplug_enable(void)
 142 {
 143         cpu_maps_update_begin();
 144         cpu_hotplug_disabled = 0;
 145         cpu_maps_update_done();
 146 }
 147
 148 #else /* #if CONFIG_HOTPLUG_CPU */
 149 static void cpu_hotplug_begin(void) {}
 150 static void cpu_hotplug_done(void) {}
 151 #endif  /* #else #if CONFIG_HOTPLUG_CPU */
 152
 153 /* Need to know about CPUs going up/down? */
 154 int __ref register_cpu_notifier(struct notifier_block *nb)
 155 {
 156         int ret;
 157         cpu_maps_update_begin();
 158         ret = raw_notifier_chain_register(&cpu_chain, nb);
 159         cpu_maps_update_done();
 160         return ret;
 161 }
 162
 163 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
 164                         int *nr_calls)
 165 {
 166         int ret;
 167
 168         ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
 169                                         nr_calls);
 170
 171         return notifier_to_errno(ret);
 172 }
 173
 174 static int cpu_notify(unsigned long val, void *v)
 175 {
 176         return __cpu_notify(val, v, -1, NULL);
 177 }
 178
 179 #ifdef CONFIG_HOTPLUG_CPU
 180
 181 static void cpu_notify_nofail(unsigned long val, void *v)
 182 {
 183         BUG_ON(cpu_notify(val, v));
 184 }
 185 EXPORT_SYMBOL(register_cpu_notifier);
 186
 187 void __ref unregister_cpu_notifier(struct notifier_block *nb)
 188 {
 189         cpu_maps_update_begin();
 190         raw_notifier_chain_unregister(&cpu_chain, nb);
 191         cpu_maps_update_done();
 192 }
 193 EXPORT_SYMBOL(unregister_cpu_notifier);
 194
 195 static inline void check_for_tasks(int cpu)
 196 {
 197         struct task_struct *p;
 198
 199         write_lock_irq(&tasklist_lock);
 200         for_each_process(p) {
 201                 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
 202                     (p->utime || p->stime))
 203                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
 204                                 "(state = %ld, flags = %x)\n",
 205                                 p->comm, task_pid_nr(p), cpu,
 206                                 p->state, p->flags);
 207         }
 208         write_unlock_irq(&tasklist_lock);
 209 }
 210
 211 struct take_cpu_down_param {
 212         unsigned long mod;
 213         void *hcpu;
 214 };
 215
 216 /* Take this CPU down. */
 217 static int __ref take_cpu_down(void *_param)
 218 {
 219         struct take_cpu_down_param *param = _param;
 220         int err;
 221
 222         /* Ensure this CPU doesn't handle any more interrupts. */
 223         err = __cpu_disable();
 224         if (err < 0)
 225                 return err;
 226
 227         cpu_notify(CPU_DYING | param->mod, param->hcpu);
 228         return 0;
 229 }
 230
 231 /* Requires cpu_add_remove_lock to be held */
 232 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 233 {
 234         int err, nr_calls = 0;
 235         void *hcpu = (void *)(long)cpu;
 236         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 237         struct take_cpu_down_param tcd_param = {
 238                 .mod = mod,
 239                 .hcpu = hcpu,
 240         };
 241
 242         if (num_online_cpus() == 1)
 243                 return -EBUSY;
 244
 245         if (!cpu_online(cpu))
 246                 return -EINVAL;
 247
 248         cpu_hotplug_begin();
 249
 250         err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
 251         if (err) {
 252                 nr_calls--;
 253                 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
 254                 printk("%s: attempt to take down CPU %u failed\n",
 255                                 __func__, cpu);
 256                 goto out_release;
 257         }
 258
 259         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
 260         if (err) {
 261                 /* CPU didn't die: tell everyone.  Can't complain. */
 262                 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
 263
 264                 goto out_release;
 265         }
 266         BUG_ON(cpu_online(cpu));
 267
 268         /*
 269          * The migration_call() CPU_DYING callback will have removed all
 270          * runnable tasks from the cpu, there's only the idle task left now
 271          * that the migration thread is done doing the stop_machine thing.
 272          *
 273          * Wait for the stop thread to go away.
 274          */
 275         while (!idle_cpu(cpu))
 276                 cpu_relax();
 277
 278         /* This actually kills the CPU. */
 279         __cpu_die(cpu);
 280
 281         /* CPU is completely dead: tell everyone.  Too late to complain. */
 282         cpu_notify_nofail(CPU_DEAD | mod, hcpu);
 283
 284         check_for_tasks(cpu);
 285
 286 out_release:
 287         cpu_hotplug_done();
 288         if (!err)
 289                 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
 290         return err;
 291 }
 292
 293 int __ref cpu_down(unsigned int cpu)
 294 {
 295         int err;
 296
 297         cpu_maps_update_begin();
 298
 299         if (cpu_hotplug_disabled) {
 300                 err = -EBUSY;
 301                 goto out;
 302         }
 303
 304         err = _cpu_down(cpu, 0);
 305
 306 out:
 307         cpu_maps_update_done();
 308         return err;
 309 }
 310 EXPORT_SYMBOL(cpu_down);
 311 #endif /*CONFIG_HOTPLUG_CPU*/
 312
 313 /* Requires cpu_add_remove_lock to be held */
 314 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 315 {
 316         int ret, nr_calls = 0;
 317         void *hcpu = (void *)(long)cpu;
 318         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 319
 320         if (cpu_online(cpu) || !cpu_present(cpu))
 321                 return -EINVAL;
 322
 323         cpu_hotplug_begin();
 324         ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
 325         if (ret) {
 326                 nr_calls--;
 327                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 328                                 __func__, cpu);
 329                 goto out_notify;
 330         }
 331
 332         /* Arch-specific enabling code. */
 333         ret = __cpu_up(cpu);
 334         if (ret != 0)
 335                 goto out_notify;
 336         BUG_ON(!cpu_online(cpu));
 337
 338         /* Now call notifier in preparation. */
 339         cpu_notify(CPU_ONLINE | mod, hcpu);
 340
 341 out_notify:
 342         if (ret != 0)
 343                 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 344         cpu_hotplug_done();
 345
 346         return ret;
 347 }
 348
 349 int __cpuinit cpu_up(unsigned int cpu)
 350 {
 351         int err = 0;
 352
 353 #ifdef  CONFIG_MEMORY_HOTPLUG
 354         int nid;
 355         pg_data_t       *pgdat;
 356 #endif
 357
 358         if (!cpu_possible(cpu)) {
 359                 printk(KERN_ERR "can't online cpu %d because it is not "
 360                         "configured as may-hotadd at boot time\n", cpu);
 361 #if defined(CONFIG_IA64)
 362                 printk(KERN_ERR "please check additional_cpus= boot "
 363                                 "parameter\n");
 364 #endif
 365                 return -EINVAL;
 366         }
 367
 368 #ifdef  CONFIG_MEMORY_HOTPLUG
 369         nid = cpu_to_node(cpu);
 370         if (!node_online(nid)) {
 371                 err = mem_online_node(nid);
 372                 if (err)
 373                         return err;
 374         }
 375
 376         pgdat = NODE_DATA(nid);
 377         if (!pgdat) {
 378                 printk(KERN_ERR
 379                         "Can't online cpu %d due to NULL pgdat\n", cpu);
 380                 return -ENOMEM;
 381         }
 382
 383         if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
 384                 mutex_lock(&zonelists_mutex);
 385                 build_all_zonelists(NULL);
 386                 mutex_unlock(&zonelists_mutex);
 387         }
 388 #endif
 389
 390         cpu_maps_update_begin();
 391
 392         if (cpu_hotplug_disabled) {
 393                 err = -EBUSY;
 394                 goto out;
 395         }
 396
 397         err = _cpu_up(cpu, 0);
 398
 399 out:
 400         cpu_maps_update_done();
 401         return err;
 402 }
 403 EXPORT_SYMBOL_GPL(cpu_up);
 404
 405 #ifdef CONFIG_PM_SLEEP_SMP
 406 static cpumask_var_t frozen_cpus;
 407
 408 void __weak arch_disable_nonboot_cpus_begin(void)
 409 {
 410 }
 411
 412 void __weak arch_disable_nonboot_cpus_end(void)
 413 {
 414 }
 415
 416 int disable_nonboot_cpus(void)
 417 {
 418         int cpu, first_cpu, error = 0;
 419
 420         cpu_maps_update_begin();
 421         first_cpu = cpumask_first(cpu_online_mask);
 422         /*
 423          * We take down all of the non-boot CPUs in one shot to avoid races
 424          * with the userspace trying to use the CPU hotplug at the same time
 425          */
 426         cpumask_clear(frozen_cpus);
 427         arch_disable_nonboot_cpus_begin();
 428
 429         printk("Disabling non-boot CPUs ...\n");
 430         for_each_online_cpu(cpu) {
 431                 if (cpu == first_cpu)
 432                         continue;
 433                 error = _cpu_down(cpu, 1);
 434                 if (!error)
 435                         cpumask_set_cpu(cpu, frozen_cpus);
 436                 else {
 437                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 438                                 cpu, error);
 439                         break;
 440                 }
 441         }
 442
 443         arch_disable_nonboot_cpus_end();
 444
 445         if (!error) {
 446                 BUG_ON(num_online_cpus() > 1);
 447                 /* Make sure the CPUs won't be enabled by someone else */
 448                 cpu_hotplug_disabled = 1;
 449         } else {
 450                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 451         }
 452         cpu_maps_update_done();
 453         return error;
 454 }
 455
 456 void __weak arch_enable_nonboot_cpus_begin(void)
 457 {
 458 }
 459
 460 void __weak arch_enable_nonboot_cpus_end(void)
 461 {
 462 }
 463
 464 void __ref enable_nonboot_cpus(void)
 465 {
 466         int cpu, error;
 467
 468         /* Allow everyone to use the CPU hotplug again */
 469         cpu_maps_update_begin();
 470         cpu_hotplug_disabled = 0;
 471         if (cpumask_empty(frozen_cpus))
 472                 goto out;
 473
 474         printk(KERN_INFO "Enabling non-boot CPUs ...\n");
 475
 476         arch_enable_nonboot_cpus_begin();
 477
 478         for_each_cpu(cpu, frozen_cpus) {
 479                 error = _cpu_up(cpu, 1);
 480                 if (!error) {
 481                         printk(KERN_INFO "CPU%d is up\n", cpu);
 482                         continue;
 483                 }
 484                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 485         }
 486
 487         arch_enable_nonboot_cpus_end();
 488
 489         cpumask_clear(frozen_cpus);
 490 out:
 491         cpu_maps_update_done();
 492 }
 493
 494 static int __init alloc_frozen_cpus(void)
 495 {
 496         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 497                 return -ENOMEM;
 498         return 0;
 499 }
 500 core_initcall(alloc_frozen_cpus);
 501
 502 /*
 503  * When callbacks for CPU hotplug notifications are being executed, we must
 504  * ensure that the state of the system with respect to the tasks being frozen
 505  * or not, as reported by the notification, remains unchanged *throughout the
 506  * duration* of the execution of the callbacks.
 507  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
 508  *
 509  * This synchronization is implemented by mutually excluding regular CPU
 510  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
 511  * Hibernate notifications.
 512  */
 513 static int
 514 cpu_hotplug_pm_callback(struct notifier_block *nb,
 515                         unsigned long action, void *ptr)
 516 {
 517         switch (action) {
 518
 519         case PM_SUSPEND_PREPARE:
 520         case PM_HIBERNATION_PREPARE:
 521                 cpu_hotplug_disable();
 522                 break;
 523
 524         case PM_POST_SUSPEND:
 525         case PM_POST_HIBERNATION:
 526                 cpu_hotplug_enable();
 527                 break;
 528
 529         default:
 530                 return NOTIFY_DONE;
 531         }
 532
 533         return NOTIFY_OK;
 534 }
 535
 536
 537 static int __init cpu_hotplug_pm_sync_init(void)
 538 {
 539         pm_notifier(cpu_hotplug_pm_callback, 0);
 540         return 0;
 541 }
 542 core_initcall(cpu_hotplug_pm_sync_init);
 543
 544 #endif /* CONFIG_PM_SLEEP_SMP */
 545
 546 /**
 547  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 548  * @cpu: cpu that just started
 549  *
 550  * This function calls the cpu_chain notifiers with CPU_STARTING.
 551  * It must be called by the arch code on the new cpu, before the new cpu
 552  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 553  */
 554 void __cpuinit notify_cpu_starting(unsigned int cpu)
 555 {
 556         unsigned long val = CPU_STARTING;
 557
 558 #ifdef CONFIG_PM_SLEEP_SMP
 559         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
 560                 val = CPU_STARTING_FROZEN;
 561 #endif /* CONFIG_PM_SLEEP_SMP */
 562         cpu_notify(val, (void *)(long)cpu);
 563 }
 564
 565 #endif /* CONFIG_SMP */
 566
 567 /*
 568  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 569  * represents all NR_CPUS bits binary values of 1<<nr.
 570  *
 571  * It is used by cpumask_of() to get a constant address to a CPU
 572  * mask value that has a single bit set only.
 573  */
 574
 575 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 576 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
 577 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 578 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 579 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 580
 581 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 582
 583         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 584         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 585 #if BITS_PER_LONG > 32
 586         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 587         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 588 #endif
 589 };
 590 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 591
 592 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 593 EXPORT_SYMBOL(cpu_all_bits);
 594
 595 #ifdef CONFIG_INIT_ALL_POSSIBLE
 596 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
 597         = CPU_BITS_ALL;
 598 #else
 599 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
 600 #endif
 601 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
 602 EXPORT_SYMBOL(cpu_possible_mask);
 603
 604 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
 605 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
 606 EXPORT_SYMBOL(cpu_online_mask);
 607
 608 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
 609 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
 610 EXPORT_SYMBOL(cpu_present_mask);
 611
 612 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
 613 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
 614 EXPORT_SYMBOL(cpu_active_mask);
 615
 616 void set_cpu_possible(unsigned int cpu, bool possible)
 617 {
 618         if (possible)
 619                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
 620         else
 621                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
 622 }
 623
 624 void set_cpu_present(unsigned int cpu, bool present)
 625 {
 626         if (present)
 627                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
 628         else
 629                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
 630 }
 631
 632 void set_cpu_online(unsigned int cpu, bool online)
 633 {
 634         if (online)
 635                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
 636         else
 637                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
 638 }
 639
 640 void set_cpu_active(unsigned int cpu, bool active)
 641 {
 642         if (active)
 643                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
 644         else
 645                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
 646 }
 647
 648 void init_cpu_present(const struct cpumask *src)
 649 {
 650         cpumask_copy(to_cpumask(cpu_present_bits), src);
 651 }
 652
 653 void init_cpu_possible(const struct cpumask *src)
 654 {
 655         cpumask_copy(to_cpumask(cpu_possible_bits), src);
 656 }
 657
 658 void init_cpu_online(const struct cpumask *src)
 659 {
 660         cpumask_copy(to_cpumask(cpu_online_bits), src);
 661 }