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