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
  18 #ifdef CONFIG_SMP
  19 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
  20 static DEFINE_MUTEX(cpu_add_remove_lock);
  21
  22 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
  23
  24 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  25  * Should always be manipulated under cpu_add_remove_lock
  26  */
  27 static int cpu_hotplug_disabled;
  28
  29 static struct {
  30         struct task_struct *active_writer;
  31         struct mutex lock; /* Synchronizes accesses to refcount, */
  32         /*
  33          * Also blocks the new readers during
  34          * an ongoing cpu hotplug operation.
  35          */
  36         int refcount;
  37 } cpu_hotplug = {
  38         .active_writer = NULL,
  39         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
  40         .refcount = 0,
  41 };
  42
  43 #ifdef CONFIG_HOTPLUG_CPU
  44
  45 void get_online_cpus(void)
  46 {
  47         might_sleep();
  48         if (cpu_hotplug.active_writer == current)
  49                 return;
  50         mutex_lock(&cpu_hotplug.lock);
  51         cpu_hotplug.refcount++;
  52         mutex_unlock(&cpu_hotplug.lock);
  53
  54 }
  55 EXPORT_SYMBOL_GPL(get_online_cpus);
  56
  57 void put_online_cpus(void)
  58 {
  59         if (cpu_hotplug.active_writer == current)
  60                 return;
  61         mutex_lock(&cpu_hotplug.lock);
  62         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
  63                 wake_up_process(cpu_hotplug.active_writer);
  64         mutex_unlock(&cpu_hotplug.lock);
  65
  66 }
  67 EXPORT_SYMBOL_GPL(put_online_cpus);
  68
  69 #endif  /* CONFIG_HOTPLUG_CPU */
  70
  71 /*
  72  * The following two API's must be used when attempting
  73  * to serialize the updates to cpu_online_mask, cpu_present_mask.
  74  */
  75 void cpu_maps_update_begin(void)
  76 {
  77         mutex_lock(&cpu_add_remove_lock);
  78 }
  79
  80 void cpu_maps_update_done(void)
  81 {
  82         mutex_unlock(&cpu_add_remove_lock);
  83 }
  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 /* Need to know about CPUs going up/down? */
 127 int __ref register_cpu_notifier(struct notifier_block *nb)
 128 {
 129         int ret;
 130         cpu_maps_update_begin();
 131         ret = raw_notifier_chain_register(&cpu_chain, nb);
 132         cpu_maps_update_done();
 133         return ret;
 134 }
 135
 136 #ifdef CONFIG_HOTPLUG_CPU
 137
 138 EXPORT_SYMBOL(register_cpu_notifier);
 139
 140 void __ref unregister_cpu_notifier(struct notifier_block *nb)
 141 {
 142         cpu_maps_update_begin();
 143         raw_notifier_chain_unregister(&cpu_chain, nb);
 144         cpu_maps_update_done();
 145 }
 146 EXPORT_SYMBOL(unregister_cpu_notifier);
 147
 148 static inline void check_for_tasks(int cpu)
 149 {
 150         struct task_struct *p;
 151
 152         write_lock_irq(&tasklist_lock);
 153         for_each_process(p) {
 154                 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
 155                     (!cputime_eq(p->utime, cputime_zero) ||
 156                      !cputime_eq(p->stime, cputime_zero)))
 157                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
 158                                 "(state = %ld, flags = %x)\n",
 159                                 p->comm, task_pid_nr(p), cpu,
 160                                 p->state, p->flags);
 161         }
 162         write_unlock_irq(&tasklist_lock);
 163 }
 164
 165 struct take_cpu_down_param {
 166         unsigned long mod;
 167         void *hcpu;
 168 };
 169
 170 /* Take this CPU down. */
 171 static int __ref take_cpu_down(void *_param)
 172 {
 173         struct take_cpu_down_param *param = _param;
 174         int err;
 175
 176         /* Ensure this CPU doesn't handle any more interrupts. */
 177         err = __cpu_disable();
 178         if (err < 0)
 179                 return err;
 180
 181         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
 182                                 param->hcpu);
 183
 184         /* Force idle task to run as soon as we yield: it should
 185            immediately notice cpu is offline and die quickly. */
 186         sched_idle_next();
 187         return 0;
 188 }
 189
 190 /* Requires cpu_add_remove_lock to be held */
 191 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 192 {
 193         int err, nr_calls = 0;
 194         cpumask_var_t old_allowed;
 195         void *hcpu = (void *)(long)cpu;
 196         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 197         struct take_cpu_down_param tcd_param = {
 198                 .mod = mod,
 199                 .hcpu = hcpu,
 200         };
 201
 202         if (num_online_cpus() == 1)
 203                 return -EBUSY;
 204
 205         if (!cpu_online(cpu))
 206                 return -EINVAL;
 207
 208         if (!alloc_cpumask_var(&old_allowed, GFP_KERNEL))
 209                 return -ENOMEM;
 210
 211         cpu_hotplug_begin();
 212         set_cpu_active(cpu, false);
 213         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
 214                                         hcpu, -1, &nr_calls);
 215         if (err == NOTIFY_BAD) {
 216                 set_cpu_active(cpu, true);
 217
 218                 nr_calls--;
 219                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 220                                           hcpu, nr_calls, NULL);
 221                 printk("%s: attempt to take down CPU %u failed\n",
 222                                 __func__, cpu);
 223                 err = -EINVAL;
 224                 goto out_release;
 225         }
 226
 227         /* Ensure that we are not runnable on dying cpu */
 228         cpumask_copy(old_allowed, &current->cpus_allowed);
 229         set_cpus_allowed_ptr(current, cpu_active_mask);
 230
 231         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
 232         if (err) {
 233                 set_cpu_active(cpu, true);
 234                 /* CPU didn't die: tell everyone.  Can't complain. */
 235                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 236                                             hcpu) == NOTIFY_BAD)
 237                         BUG();
 238
 239                 goto out_allowed;
 240         }
 241         BUG_ON(cpu_online(cpu));
 242
 243         /* Wait for it to sleep (leaving idle task). */
 244         while (!idle_cpu(cpu))
 245                 yield();
 246
 247         /* This actually kills the CPU. */
 248         __cpu_die(cpu);
 249
 250         /* CPU is completely dead: tell everyone.  Too late to complain. */
 251         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
 252                                     hcpu) == NOTIFY_BAD)
 253                 BUG();
 254
 255         check_for_tasks(cpu);
 256
 257 out_allowed:
 258         set_cpus_allowed_ptr(current, old_allowed);
 259 out_release:
 260         cpu_hotplug_done();
 261         if (!err) {
 262                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
 263                                             hcpu) == NOTIFY_BAD)
 264                         BUG();
 265         }
 266         free_cpumask_var(old_allowed);
 267         return err;
 268 }
 269
 270 int __ref cpu_down(unsigned int cpu)
 271 {
 272         int err;
 273
 274         err = stop_machine_create();
 275         if (err)
 276                 return err;
 277         cpu_maps_update_begin();
 278
 279         if (cpu_hotplug_disabled) {
 280                 err = -EBUSY;
 281                 goto out;
 282         }
 283
 284         err = _cpu_down(cpu, 0);
 285
 286 out:
 287         cpu_maps_update_done();
 288         stop_machine_destroy();
 289         return err;
 290 }
 291 EXPORT_SYMBOL(cpu_down);
 292 #endif /*CONFIG_HOTPLUG_CPU*/
 293
 294 /* Requires cpu_add_remove_lock to be held */
 295 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 296 {
 297         int ret, nr_calls = 0;
 298         void *hcpu = (void *)(long)cpu;
 299         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 300
 301         if (cpu_online(cpu) || !cpu_present(cpu))
 302                 return -EINVAL;
 303
 304         cpu_hotplug_begin();
 305         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
 306                                                         -1, &nr_calls);
 307         if (ret == NOTIFY_BAD) {
 308                 nr_calls--;
 309                 printk("%s: attempt to bring up CPU %u failed\n",
 310                                 __func__, cpu);
 311                 ret = -EINVAL;
 312                 goto out_notify;
 313         }
 314
 315         /* Arch-specific enabling code. */
 316         ret = __cpu_up(cpu);
 317         if (ret != 0)
 318                 goto out_notify;
 319         BUG_ON(!cpu_online(cpu));
 320
 321         set_cpu_active(cpu, true);
 322
 323         /* Now call notifier in preparation. */
 324         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 325
 326 out_notify:
 327         if (ret != 0)
 328                 __raw_notifier_call_chain(&cpu_chain,
 329                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 330         cpu_hotplug_done();
 331
 332         return ret;
 333 }
 334
 335 int __cpuinit cpu_up(unsigned int cpu)
 336 {
 337         int err = 0;
 338         if (!cpu_possible(cpu)) {
 339                 printk(KERN_ERR "can't online cpu %d because it is not "
 340                         "configured as may-hotadd at boot time\n", cpu);
 341 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
 342                 printk(KERN_ERR "please check additional_cpus= boot "
 343                                 "parameter\n");
 344 #endif
 345                 return -EINVAL;
 346         }
 347
 348         cpu_maps_update_begin();
 349
 350         if (cpu_hotplug_disabled) {
 351                 err = -EBUSY;
 352                 goto out;
 353         }
 354
 355         err = _cpu_up(cpu, 0);
 356
 357 out:
 358         cpu_maps_update_done();
 359         return err;
 360 }
 361
 362 #ifdef CONFIG_PM_SLEEP_SMP
 363 static cpumask_var_t frozen_cpus;
 364
 365 int disable_nonboot_cpus(void)
 366 {
 367         int cpu, first_cpu, error;
 368
 369         error = stop_machine_create();
 370         if (error)
 371                 return error;
 372         cpu_maps_update_begin();
 373         first_cpu = cpumask_first(cpu_online_mask);
 374         /*
 375          * We take down all of the non-boot CPUs in one shot to avoid races
 376          * with the userspace trying to use the CPU hotplug at the same time
 377          */
 378         cpumask_clear(frozen_cpus);
 379
 380         printk("Disabling non-boot CPUs ...\n");
 381         for_each_online_cpu(cpu) {
 382                 if (cpu == first_cpu)
 383                         continue;
 384                 error = _cpu_down(cpu, 1);
 385                 if (!error)
 386                         cpumask_set_cpu(cpu, frozen_cpus);
 387                 else {
 388                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 389                                 cpu, error);
 390                         break;
 391                 }
 392         }
 393
 394         if (!error) {
 395                 BUG_ON(num_online_cpus() > 1);
 396                 /* Make sure the CPUs won't be enabled by someone else */
 397                 cpu_hotplug_disabled = 1;
 398         } else {
 399                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 400         }
 401         cpu_maps_update_done();
 402         stop_machine_destroy();
 403         return error;
 404 }
 405
 406 void __weak arch_enable_nonboot_cpus_begin(void)
 407 {
 408 }
 409
 410 void __weak arch_enable_nonboot_cpus_end(void)
 411 {
 412 }
 413
 414 void __ref enable_nonboot_cpus(void)
 415 {
 416         int cpu, error;
 417
 418         /* Allow everyone to use the CPU hotplug again */
 419         cpu_maps_update_begin();
 420         cpu_hotplug_disabled = 0;
 421         if (cpumask_empty(frozen_cpus))
 422                 goto out;
 423
 424         printk("Enabling non-boot CPUs ...\n");
 425
 426         arch_enable_nonboot_cpus_begin();
 427
 428         for_each_cpu(cpu, frozen_cpus) {
 429                 error = _cpu_up(cpu, 1);
 430                 if (!error) {
 431                         printk("CPU%d is up\n", cpu);
 432                         continue;
 433                 }
 434                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 435         }
 436
 437         arch_enable_nonboot_cpus_end();
 438
 439         cpumask_clear(frozen_cpus);
 440 out:
 441         cpu_maps_update_done();
 442 }
 443
 444 static int alloc_frozen_cpus(void)
 445 {
 446         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 447                 return -ENOMEM;
 448         return 0;
 449 }
 450 core_initcall(alloc_frozen_cpus);
 451 #endif /* CONFIG_PM_SLEEP_SMP */
 452
 453 /**
 454  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 455  * @cpu: cpu that just started
 456  *
 457  * This function calls the cpu_chain notifiers with CPU_STARTING.
 458  * It must be called by the arch code on the new cpu, before the new cpu
 459  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 460  */
 461 void __cpuinit notify_cpu_starting(unsigned int cpu)
 462 {
 463         unsigned long val = CPU_STARTING;
 464
 465 #ifdef CONFIG_PM_SLEEP_SMP
 466         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
 467                 val = CPU_STARTING_FROZEN;
 468 #endif /* CONFIG_PM_SLEEP_SMP */
 469         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
 470 }
 471
 472 #endif /* CONFIG_SMP */
 473
 474 /*
 475  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 476  * represents all NR_CPUS bits binary values of 1<<nr.
 477  *
 478  * It is used by cpumask_of() to get a constant address to a CPU
 479  * mask value that has a single bit set only.
 480  */
 481
 482 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 483 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
 484 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 485 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 486 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 487
 488 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 489
 490         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 491         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 492 #if BITS_PER_LONG > 32
 493         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 494         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 495 #endif
 496 };
 497 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 498
 499 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 500 EXPORT_SYMBOL(cpu_all_bits);
 501
 502 #ifdef CONFIG_INIT_ALL_POSSIBLE
 503 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
 504         = CPU_BITS_ALL;
 505 #else
 506 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
 507 #endif
 508 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
 509 EXPORT_SYMBOL(cpu_possible_mask);
 510
 511 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
 512 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
 513 EXPORT_SYMBOL(cpu_online_mask);
 514
 515 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
 516 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
 517 EXPORT_SYMBOL(cpu_present_mask);
 518
 519 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
 520 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
 521 EXPORT_SYMBOL(cpu_active_mask);
 522
 523 void set_cpu_possible(unsigned int cpu, bool possible)
 524 {
 525         if (possible)
 526                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
 527         else
 528                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
 529 }
 530
 531 void set_cpu_present(unsigned int cpu, bool present)
 532 {
 533         if (present)
 534                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
 535         else
 536                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
 537 }
 538
 539 void set_cpu_online(unsigned int cpu, bool online)
 540 {
 541         if (online)
 542                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
 543         else
 544                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
 545 }
 546
 547 void set_cpu_active(unsigned int cpu, bool active)
 548 {
 549         if (active)
 550                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
 551         else
 552                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
 553 }
 554
 555 void init_cpu_present(const struct cpumask *src)
 556 {
 557         cpumask_copy(to_cpumask(cpu_present_bits), src);
 558 }
 559
 560 void init_cpu_possible(const struct cpumask *src)
 561 {
 562         cpumask_copy(to_cpumask(cpu_possible_bits), src);
 563 }
 564
 565 void init_cpu_online(const struct cpumask *src)
 566 {
 567         cpumask_copy(to_cpumask(cpu_online_bits), src);
 568 }