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/module.h>
  14 #include <linux/kthread.h>
  15 #include <linux/stop_machine.h>
  16 #include <linux/mutex.h>
  17 #include <linux/suspend.h>
  18
  19 #ifdef CONFIG_SMP
  20 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
  21 static DEFINE_MUTEX(cpu_add_remove_lock);
  22
  23 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
  24
  25 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  26  * Should always be manipulated under cpu_add_remove_lock
  27  */
  28 static int cpu_hotplug_disabled;
  29
  30 static struct {
  31         struct task_struct *active_writer;
  32         struct mutex lock; /* Synchronizes accesses to refcount, */
  33         /*
  34          * Also blocks the new readers during
  35          * an ongoing cpu hotplug operation.
  36          */
  37         int refcount;
  38 } cpu_hotplug = {
  39         .active_writer = NULL,
  40         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
  41         .refcount = 0,
  42 };
  43
  44 #ifdef CONFIG_HOTPLUG_CPU
  45
  46 void get_online_cpus(void)
  47 {
  48         might_sleep();
  49         if (cpu_hotplug.active_writer == current)
  50                 return;
  51         mutex_lock(&cpu_hotplug.lock);
  52         cpu_hotplug.refcount++;
  53         mutex_unlock(&cpu_hotplug.lock);
  54
  55 }
  56 EXPORT_SYMBOL_GPL(get_online_cpus);
  57
  58 void put_online_cpus(void)
  59 {
  60         if (cpu_hotplug.active_writer == current)
  61                 return;
  62         mutex_lock(&cpu_hotplug.lock);
  63         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
  64                 wake_up_process(cpu_hotplug.active_writer);
  65         mutex_unlock(&cpu_hotplug.lock);
  66
  67 }
  68 EXPORT_SYMBOL_GPL(put_online_cpus);
  69
  70 #endif  /* CONFIG_HOTPLUG_CPU */
  71
  72 /*
  73  * The following two API's must be used when attempting
  74  * to serialize the updates to cpu_online_mask, cpu_present_mask.
  75  */
  76 void cpu_maps_update_begin(void)
  77 {
  78         mutex_lock(&cpu_add_remove_lock);
  79 }
  80
  81 void cpu_maps_update_done(void)
  82 {
  83         mutex_unlock(&cpu_add_remove_lock);
  84 }
  85
  86 /*
  87  * This ensures that the hotplug operation can begin only when the
  88  * refcount goes to zero.
  89  *
  90  * Note that during a cpu-hotplug operation, the new readers, if any,
  91  * will be blocked by the cpu_hotplug.lock
  92  *
  93  * Since cpu_hotplug_begin() is always called after invoking
  94  * cpu_maps_update_begin(), we can be sure that only one writer is active.
  95  *
  96  * Note that theoretically, there is a possibility of a livelock:
  97  * - Refcount goes to zero, last reader wakes up the sleeping
  98  *   writer.
  99  * - Last reader unlocks the cpu_hotplug.lock.
 100  * - A new reader arrives at this moment, bumps up the refcount.
 101  * - The writer acquires the cpu_hotplug.lock finds the refcount
 102  *   non zero and goes to sleep again.
 103  *
 104  * However, this is very difficult to achieve in practice since
 105  * get_online_cpus() not an api which is called all that often.
 106  *
 107  */
 108 static void cpu_hotplug_begin(void)
 109 {
 110         cpu_hotplug.active_writer = current;
 111
 112         for (;;) {
 113                 mutex_lock(&cpu_hotplug.lock);
 114                 if (likely(!cpu_hotplug.refcount))
 115                         break;
 116                 __set_current_state(TASK_UNINTERRUPTIBLE);
 117                 mutex_unlock(&cpu_hotplug.lock);
 118                 schedule();
 119         }
 120 }
 121
 122 static void cpu_hotplug_done(void)
 123 {
 124         cpu_hotplug.active_writer = NULL;
 125         mutex_unlock(&cpu_hotplug.lock);
 126 }
 127 /* Need to know about CPUs going up/down? */
 128 int __ref register_cpu_notifier(struct notifier_block *nb)
 129 {
 130         int ret;
 131         cpu_maps_update_begin();
 132         ret = raw_notifier_chain_register(&cpu_chain, nb);
 133         cpu_maps_update_done();
 134         return ret;
 135 }
 136
 137 #ifdef CONFIG_HOTPLUG_CPU
 138
 139 EXPORT_SYMBOL(register_cpu_notifier);
 140
 141 void __ref unregister_cpu_notifier(struct notifier_block *nb)
 142 {
 143         cpu_maps_update_begin();
 144         raw_notifier_chain_unregister(&cpu_chain, nb);
 145         cpu_maps_update_done();
 146 }
 147 EXPORT_SYMBOL(unregister_cpu_notifier);
 148
 149 static inline void check_for_tasks(int cpu)
 150 {
 151         struct task_struct *p;
 152
 153         write_lock_irq(&tasklist_lock);
 154         for_each_process(p) {
 155                 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
 156                     (!cputime_eq(p->utime, cputime_zero) ||
 157                      !cputime_eq(p->stime, cputime_zero)))
 158                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
 159                                 (state = %ld, flags = %x) \n",
 160                                  p->comm, task_pid_nr(p), cpu,
 161                                  p->state, p->flags);
 162         }
 163         write_unlock_irq(&tasklist_lock);
 164 }
 165
 166 struct take_cpu_down_param {
 167         struct task_struct *caller;
 168         unsigned long mod;
 169         void *hcpu;
 170 };
 171
 172 /* Take this CPU down. */
 173 static int __ref take_cpu_down(void *_param)
 174 {
 175         struct take_cpu_down_param *param = _param;
 176         unsigned int cpu = (unsigned long)param->hcpu;
 177         int err;
 178
 179         /* Ensure this CPU doesn't handle any more interrupts. */
 180         err = __cpu_disable();
 181         if (err < 0)
 182                 return err;
 183
 184         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
 185                                 param->hcpu);
 186
 187         if (task_cpu(param->caller) == cpu)
 188                 move_task_off_dead_cpu(cpu, param->caller);
 189         /* Force idle task to run as soon as we yield: it should
 190            immediately notice cpu is offline and die quickly. */
 191         sched_idle_next();
 192         return 0;
 193 }
 194
 195 /* Requires cpu_add_remove_lock to be held */
 196 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 197 {
 198         int err, nr_calls = 0;
 199         void *hcpu = (void *)(long)cpu;
 200         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 201         struct take_cpu_down_param tcd_param = {
 202                 .caller = current,
 203                 .mod = mod,
 204                 .hcpu = hcpu,
 205         };
 206
 207         if (num_online_cpus() == 1)
 208                 return -EBUSY;
 209
 210         if (!cpu_online(cpu))
 211                 return -EINVAL;
 212
 213         cpu_hotplug_begin();
 214         set_cpu_active(cpu, false);
 215         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
 216                                         hcpu, -1, &nr_calls);
 217         if (err == NOTIFY_BAD) {
 218                 set_cpu_active(cpu, true);
 219
 220                 nr_calls--;
 221                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 222                                           hcpu, nr_calls, NULL);
 223                 printk("%s: attempt to take down CPU %u failed\n",
 224                                 __func__, cpu);
 225                 err = -EINVAL;
 226                 goto out_release;
 227         }
 228
 229         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
 230         if (err) {
 231                 set_cpu_active(cpu, true);
 232                 /* CPU didn't die: tell everyone.  Can't complain. */
 233                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 234                                             hcpu) == NOTIFY_BAD)
 235                         BUG();
 236
 237                 goto out_release;
 238         }
 239         BUG_ON(cpu_online(cpu));
 240
 241         /* Wait for it to sleep (leaving idle task). */
 242         while (!idle_cpu(cpu))
 243                 yield();
 244
 245         /* This actually kills the CPU. */
 246         __cpu_die(cpu);
 247
 248         /* CPU is completely dead: tell everyone.  Too late to complain. */
 249         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
 250                                     hcpu) == NOTIFY_BAD)
 251                 BUG();
 252
 253         check_for_tasks(cpu);
 254
 255 out_release:
 256         cpu_hotplug_done();
 257         if (!err) {
 258                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
 259                                             hcpu) == NOTIFY_BAD)
 260                         BUG();
 261         }
 262         return err;
 263 }
 264
 265 int __ref cpu_down(unsigned int cpu)
 266 {
 267         int err;
 268
 269         err = stop_machine_create();
 270         if (err)
 271                 return err;
 272         cpu_maps_update_begin();
 273
 274         if (cpu_hotplug_disabled) {
 275                 err = -EBUSY;
 276                 goto out;
 277         }
 278
 279         err = _cpu_down(cpu, 0);
 280
 281 out:
 282         cpu_maps_update_done();
 283         stop_machine_destroy();
 284         return err;
 285 }
 286 EXPORT_SYMBOL(cpu_down);
 287 #endif /*CONFIG_HOTPLUG_CPU*/
 288
 289 /* Requires cpu_add_remove_lock to be held */
 290 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 291 {
 292         int ret, nr_calls = 0;
 293         void *hcpu = (void *)(long)cpu;
 294         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 295
 296         if (cpu_online(cpu) || !cpu_present(cpu))
 297                 return -EINVAL;
 298
 299         cpu_hotplug_begin();
 300         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
 301                                                         -1, &nr_calls);
 302         if (ret == NOTIFY_BAD) {
 303                 nr_calls--;
 304                 printk("%s: attempt to bring up CPU %u failed\n",
 305                                 __func__, cpu);
 306                 ret = -EINVAL;
 307                 goto out_notify;
 308         }
 309
 310         /* Arch-specific enabling code. */
 311         ret = __cpu_up(cpu);
 312         if (ret != 0)
 313                 goto out_notify;
 314         BUG_ON(!cpu_online(cpu));
 315
 316         set_cpu_active(cpu, true);
 317
 318         /* Now call notifier in preparation. */
 319         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 320
 321 out_notify:
 322         if (ret != 0)
 323                 __raw_notifier_call_chain(&cpu_chain,
 324                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 325         cpu_hotplug_done();
 326
 327         return ret;
 328 }
 329
 330 int __cpuinit cpu_up(unsigned int cpu)
 331 {
 332         int err = 0;
 333         if (!cpu_possible(cpu)) {
 334                 printk(KERN_ERR "can't online cpu %d because it is not "
 335                         "configured as may-hotadd at boot time\n", cpu);
 336 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
 337                 printk(KERN_ERR "please check additional_cpus= boot "
 338                                 "parameter\n");
 339 #endif
 340                 return -EINVAL;
 341         }
 342
 343         cpu_maps_update_begin();
 344
 345         if (cpu_hotplug_disabled) {
 346                 err = -EBUSY;
 347                 goto out;
 348         }
 349
 350         err = _cpu_up(cpu, 0);
 351
 352 out:
 353         cpu_maps_update_done();
 354         return err;
 355 }
 356
 357 #ifdef CONFIG_PM_SLEEP_SMP
 358 static cpumask_var_t frozen_cpus;
 359
 360 int disable_nonboot_cpus(void)
 361 {
 362         int cpu, first_cpu, error;
 363
 364         error = stop_machine_create();
 365         if (error)
 366                 return error;
 367         cpu_maps_update_begin();
 368         first_cpu = cpumask_first(cpu_online_mask);
 369         /*
 370          * We take down all of the non-boot CPUs in one shot to avoid races
 371          * with the userspace trying to use the CPU hotplug at the same time
 372          */
 373         cpumask_clear(frozen_cpus);
 374
 375         printk("Disabling non-boot CPUs ...\n");
 376         for_each_online_cpu(cpu) {
 377                 if (cpu == first_cpu)
 378                         continue;
 379                 error = _cpu_down(cpu, 1);
 380                 if (!error)
 381                         cpumask_set_cpu(cpu, frozen_cpus);
 382                 else {
 383                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 384                                 cpu, error);
 385                         break;
 386                 }
 387         }
 388
 389         if (!error) {
 390                 BUG_ON(num_online_cpus() > 1);
 391                 /* Make sure the CPUs won't be enabled by someone else */
 392                 cpu_hotplug_disabled = 1;
 393         } else {
 394                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 395         }
 396         cpu_maps_update_done();
 397         stop_machine_destroy();
 398         return error;
 399 }
 400
 401 void __weak arch_enable_nonboot_cpus_begin(void)
 402 {
 403 }
 404
 405 void __weak arch_enable_nonboot_cpus_end(void)
 406 {
 407 }
 408
 409 void __ref enable_nonboot_cpus(void)
 410 {
 411         int cpu, error;
 412
 413         /* Allow everyone to use the CPU hotplug again */
 414         cpu_maps_update_begin();
 415         cpu_hotplug_disabled = 0;
 416         if (cpumask_empty(frozen_cpus))
 417                 goto out;
 418
 419         printk("Enabling non-boot CPUs ...\n");
 420
 421         arch_enable_nonboot_cpus_begin();
 422
 423         for_each_cpu(cpu, frozen_cpus) {
 424                 error = _cpu_up(cpu, 1);
 425                 if (!error) {
 426                         printk("CPU%d is up\n", cpu);
 427                         continue;
 428                 }
 429                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 430         }
 431
 432         arch_enable_nonboot_cpus_end();
 433
 434         cpumask_clear(frozen_cpus);
 435 out:
 436         cpu_maps_update_done();
 437 }
 438
 439 static int alloc_frozen_cpus(void)
 440 {
 441         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 442                 return -ENOMEM;
 443         return 0;
 444 }
 445 core_initcall(alloc_frozen_cpus);
 446
 447 /*
 448  * Prevent regular CPU hotplug from racing with the freezer, by disabling CPU
 449  * hotplug when tasks are about to be frozen. Also, don't allow the freezer
 450  * to continue until any currently running CPU hotplug operation gets
 451  * completed.
 452  * To modify the 'cpu_hotplug_disabled' flag, we need to acquire the
 453  * 'cpu_add_remove_lock'. And this same lock is also taken by the regular
 454  * CPU hotplug path and released only after it is complete. Thus, we
 455  * (and hence the freezer) will block here until any currently running CPU
 456  * hotplug operation gets completed.
 457  */
 458 void cpu_hotplug_disable_before_freeze(void)
 459 {
 460         cpu_maps_update_begin();
 461         cpu_hotplug_disabled = 1;
 462         cpu_maps_update_done();
 463 }
 464
 465
 466 /*
 467  * When tasks have been thawed, re-enable regular CPU hotplug (which had been
 468  * disabled while beginning to freeze tasks).
 469  */
 470 void cpu_hotplug_enable_after_thaw(void)
 471 {
 472         cpu_maps_update_begin();
 473         cpu_hotplug_disabled = 0;
 474         cpu_maps_update_done();
 475 }
 476
 477 /*
 478  * When callbacks for CPU hotplug notifications are being executed, we must
 479  * ensure that the state of the system with respect to the tasks being frozen
 480  * or not, as reported by the notification, remains unchanged *throughout the
 481  * duration* of the execution of the callbacks.
 482  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
 483  *
 484  * This synchronization is implemented by mutually excluding regular CPU
 485  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
 486  * Hibernate notifications.
 487  */
 488 static int
 489 cpu_hotplug_pm_callback(struct notifier_block *nb,
 490                         unsigned long action, void *ptr)
 491 {
 492         switch (action) {
 493
 494         case PM_SUSPEND_PREPARE:
 495         case PM_HIBERNATION_PREPARE:
 496                 cpu_hotplug_disable_before_freeze();
 497                 break;
 498
 499         case PM_POST_SUSPEND:
 500         case PM_POST_HIBERNATION:
 501                 cpu_hotplug_enable_after_thaw();
 502                 break;
 503
 504         default:
 505                 return NOTIFY_DONE;
 506         }
 507
 508         return NOTIFY_OK;
 509 }
 510
 511
 512 int cpu_hotplug_pm_sync_init(void)
 513 {
 514         pm_notifier(cpu_hotplug_pm_callback, 0);
 515         return 0;
 516 }
 517 core_initcall(cpu_hotplug_pm_sync_init);
 518
 519 #endif /* CONFIG_PM_SLEEP_SMP */
 520
 521 /**
 522  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 523  * @cpu: cpu that just started
 524  *
 525  * This function calls the cpu_chain notifiers with CPU_STARTING.
 526  * It must be called by the arch code on the new cpu, before the new cpu
 527  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 528  */
 529 void __cpuinit notify_cpu_starting(unsigned int cpu)
 530 {
 531         unsigned long val = CPU_STARTING;
 532
 533 #ifdef CONFIG_PM_SLEEP_SMP
 534         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
 535                 val = CPU_STARTING_FROZEN;
 536 #endif /* CONFIG_PM_SLEEP_SMP */
 537         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
 538 }
 539
 540 #endif /* CONFIG_SMP */
 541
 542 /*
 543  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 544  * represents all NR_CPUS bits binary values of 1<<nr.
 545  *
 546  * It is used by cpumask_of() to get a constant address to a CPU
 547  * mask value that has a single bit set only.
 548  */
 549
 550 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 551 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
 552 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 553 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 554 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 555
 556 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 557
 558         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 559         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 560 #if BITS_PER_LONG > 32
 561         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 562         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 563 #endif
 564 };
 565 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 566
 567 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 568 EXPORT_SYMBOL(cpu_all_bits);
 569
 570 #ifdef CONFIG_INIT_ALL_POSSIBLE
 571 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
 572         = CPU_BITS_ALL;
 573 #else
 574 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
 575 #endif
 576 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
 577 EXPORT_SYMBOL(cpu_possible_mask);
 578
 579 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
 580 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
 581 EXPORT_SYMBOL(cpu_online_mask);
 582
 583 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
 584 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
 585 EXPORT_SYMBOL(cpu_present_mask);
 586
 587 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
 588 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
 589 EXPORT_SYMBOL(cpu_active_mask);
 590
 591 void set_cpu_possible(unsigned int cpu, bool possible)
 592 {
 593         if (possible)
 594                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
 595         else
 596                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
 597 }
 598
 599 void set_cpu_present(unsigned int cpu, bool present)
 600 {
 601         if (present)
 602                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
 603         else
 604                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
 605 }
 606
 607 void set_cpu_online(unsigned int cpu, bool online)
 608 {
 609         if (online)
 610                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
 611         else
 612                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
 613 }
 614
 615 void set_cpu_active(unsigned int cpu, bool active)
 616 {
 617         if (active)
 618                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
 619         else
 620                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
 621 }
 622
 623 void init_cpu_present(const struct cpumask *src)
 624 {
 625         cpumask_copy(to_cpumask(cpu_present_bits), src);
 626 }
 627
 628 void init_cpu_possible(const struct cpumask *src)
 629 {
 630         cpumask_copy(to_cpumask(cpu_possible_bits), src);
 631 }
 632
 633 void init_cpu_online(const struct cpumask *src)
 634 {
 635         cpumask_copy(to_cpumask(cpu_online_bits), src);
 636 }