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/gfp.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 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
 138                         int *nr_calls)
 139 {
 140         return __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
 141                                         nr_calls);
 142 }
 143
 144 static int cpu_notify(unsigned long val, void *v)
 145 {
 146         return __cpu_notify(val, v, -1, NULL);
 147 }
 148
 149 static void cpu_notify_nofail(unsigned long val, void *v)
 150 {
 151         int err;
 152
 153         err = cpu_notify(val, v);
 154         BUG_ON(err == NOTIFY_BAD);
 155 }
 156
 157 #ifdef CONFIG_HOTPLUG_CPU
 158
 159 EXPORT_SYMBOL(register_cpu_notifier);
 160
 161 void __ref unregister_cpu_notifier(struct notifier_block *nb)
 162 {
 163         cpu_maps_update_begin();
 164         raw_notifier_chain_unregister(&cpu_chain, nb);
 165         cpu_maps_update_done();
 166 }
 167 EXPORT_SYMBOL(unregister_cpu_notifier);
 168
 169 static inline void check_for_tasks(int cpu)
 170 {
 171         struct task_struct *p;
 172
 173         write_lock_irq(&tasklist_lock);
 174         for_each_process(p) {
 175                 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
 176                     (!cputime_eq(p->utime, cputime_zero) ||
 177                      !cputime_eq(p->stime, cputime_zero)))
 178                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
 179                                 "(state = %ld, flags = %x)\n",
 180                                 p->comm, task_pid_nr(p), cpu,
 181                                 p->state, p->flags);
 182         }
 183         write_unlock_irq(&tasklist_lock);
 184 }
 185
 186 struct take_cpu_down_param {
 187         struct task_struct *caller;
 188         unsigned long mod;
 189         void *hcpu;
 190 };
 191
 192 /* Take this CPU down. */
 193 static int __ref take_cpu_down(void *_param)
 194 {
 195         struct take_cpu_down_param *param = _param;
 196         unsigned int cpu = (unsigned long)param->hcpu;
 197         int err;
 198
 199         /* Ensure this CPU doesn't handle any more interrupts. */
 200         err = __cpu_disable();
 201         if (err < 0)
 202                 return err;
 203
 204         cpu_notify(CPU_DYING | param->mod, param->hcpu);
 205
 206         if (task_cpu(param->caller) == cpu)
 207                 move_task_off_dead_cpu(cpu, param->caller);
 208         /* Force idle task to run as soon as we yield: it should
 209            immediately notice cpu is offline and die quickly. */
 210         sched_idle_next();
 211         return 0;
 212 }
 213
 214 /* Requires cpu_add_remove_lock to be held */
 215 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 216 {
 217         int err, nr_calls = 0;
 218         void *hcpu = (void *)(long)cpu;
 219         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 220         struct take_cpu_down_param tcd_param = {
 221                 .caller = current,
 222                 .mod = mod,
 223                 .hcpu = hcpu,
 224         };
 225
 226         if (num_online_cpus() == 1)
 227                 return -EBUSY;
 228
 229         if (!cpu_online(cpu))
 230                 return -EINVAL;
 231
 232         cpu_hotplug_begin();
 233         set_cpu_active(cpu, false);
 234         err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
 235         if (err == NOTIFY_BAD) {
 236                 set_cpu_active(cpu, true);
 237
 238                 nr_calls--;
 239                 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
 240                 printk("%s: attempt to take down CPU %u failed\n",
 241                                 __func__, cpu);
 242                 err = -EINVAL;
 243                 goto out_release;
 244         }
 245
 246         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
 247         if (err) {
 248                 set_cpu_active(cpu, true);
 249                 /* CPU didn't die: tell everyone.  Can't complain. */
 250                 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
 251
 252                 goto out_release;
 253         }
 254         BUG_ON(cpu_online(cpu));
 255
 256         /* Wait for it to sleep (leaving idle task). */
 257         while (!idle_cpu(cpu))
 258                 yield();
 259
 260         /* This actually kills the CPU. */
 261         __cpu_die(cpu);
 262
 263         /* CPU is completely dead: tell everyone.  Too late to complain. */
 264         cpu_notify_nofail(CPU_DEAD | mod, hcpu);
 265
 266         check_for_tasks(cpu);
 267
 268 out_release:
 269         cpu_hotplug_done();
 270         if (!err)
 271                 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
 272         return err;
 273 }
 274
 275 int __ref cpu_down(unsigned int cpu)
 276 {
 277         int err;
 278
 279         cpu_maps_update_begin();
 280
 281         if (cpu_hotplug_disabled) {
 282                 err = -EBUSY;
 283                 goto out;
 284         }
 285
 286         err = _cpu_down(cpu, 0);
 287
 288 out:
 289         cpu_maps_update_done();
 290         return err;
 291 }
 292 EXPORT_SYMBOL(cpu_down);
 293 #endif /*CONFIG_HOTPLUG_CPU*/
 294
 295 /* Requires cpu_add_remove_lock to be held */
 296 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 297 {
 298         int ret, nr_calls = 0;
 299         void *hcpu = (void *)(long)cpu;
 300         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 301
 302         if (cpu_online(cpu) || !cpu_present(cpu))
 303                 return -EINVAL;
 304
 305         cpu_hotplug_begin();
 306         ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -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         cpu_notify(CPU_ONLINE | mod, hcpu);
 325
 326 out_notify:
 327         if (ret != 0)
 328                 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 329         cpu_hotplug_done();
 330
 331         return ret;
 332 }
 333
 334 int __cpuinit cpu_up(unsigned int cpu)
 335 {
 336         int err = 0;
 337
 338 #ifdef  CONFIG_MEMORY_HOTPLUG
 339         int nid;
 340         pg_data_t       *pgdat;
 341 #endif
 342
 343         if (!cpu_possible(cpu)) {
 344                 printk(KERN_ERR "can't online cpu %d because it is not "
 345                         "configured as may-hotadd at boot time\n", cpu);
 346 #if defined(CONFIG_IA64)
 347                 printk(KERN_ERR "please check additional_cpus= boot "
 348                                 "parameter\n");
 349 #endif
 350                 return -EINVAL;
 351         }
 352
 353 #ifdef  CONFIG_MEMORY_HOTPLUG
 354         nid = cpu_to_node(cpu);
 355         if (!node_online(nid)) {
 356                 err = mem_online_node(nid);
 357                 if (err)
 358                         return err;
 359         }
 360
 361         pgdat = NODE_DATA(nid);
 362         if (!pgdat) {
 363                 printk(KERN_ERR
 364                         "Can't online cpu %d due to NULL pgdat\n", cpu);
 365                 return -ENOMEM;
 366         }
 367
 368         if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
 369                 mutex_lock(&zonelists_mutex);
 370                 build_all_zonelists(NULL);
 371                 mutex_unlock(&zonelists_mutex);
 372         }
 373 #endif
 374
 375         cpu_maps_update_begin();
 376
 377         if (cpu_hotplug_disabled) {
 378                 err = -EBUSY;
 379                 goto out;
 380         }
 381
 382         err = _cpu_up(cpu, 0);
 383
 384 out:
 385         cpu_maps_update_done();
 386         return err;
 387 }
 388
 389 #ifdef CONFIG_PM_SLEEP_SMP
 390 static cpumask_var_t frozen_cpus;
 391
 392 int disable_nonboot_cpus(void)
 393 {
 394         int cpu, first_cpu, error;
 395
 396         cpu_maps_update_begin();
 397         first_cpu = cpumask_first(cpu_online_mask);
 398         /*
 399          * We take down all of the non-boot CPUs in one shot to avoid races
 400          * with the userspace trying to use the CPU hotplug at the same time
 401          */
 402         cpumask_clear(frozen_cpus);
 403
 404         printk("Disabling non-boot CPUs ...\n");
 405         for_each_online_cpu(cpu) {
 406                 if (cpu == first_cpu)
 407                         continue;
 408                 error = _cpu_down(cpu, 1);
 409                 if (!error)
 410                         cpumask_set_cpu(cpu, frozen_cpus);
 411                 else {
 412                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 413                                 cpu, error);
 414                         break;
 415                 }
 416         }
 417
 418         if (!error) {
 419                 BUG_ON(num_online_cpus() > 1);
 420                 /* Make sure the CPUs won't be enabled by someone else */
 421                 cpu_hotplug_disabled = 1;
 422         } else {
 423                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 424         }
 425         cpu_maps_update_done();
 426         return error;
 427 }
 428
 429 void __weak arch_enable_nonboot_cpus_begin(void)
 430 {
 431 }
 432
 433 void __weak arch_enable_nonboot_cpus_end(void)
 434 {
 435 }
 436
 437 void __ref enable_nonboot_cpus(void)
 438 {
 439         int cpu, error;
 440
 441         /* Allow everyone to use the CPU hotplug again */
 442         cpu_maps_update_begin();
 443         cpu_hotplug_disabled = 0;
 444         if (cpumask_empty(frozen_cpus))
 445                 goto out;
 446
 447         printk("Enabling non-boot CPUs ...\n");
 448
 449         arch_enable_nonboot_cpus_begin();
 450
 451         for_each_cpu(cpu, frozen_cpus) {
 452                 error = _cpu_up(cpu, 1);
 453                 if (!error) {
 454                         printk("CPU%d is up\n", cpu);
 455                         continue;
 456                 }
 457                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 458         }
 459
 460         arch_enable_nonboot_cpus_end();
 461
 462         cpumask_clear(frozen_cpus);
 463 out:
 464         cpu_maps_update_done();
 465 }
 466
 467 static int alloc_frozen_cpus(void)
 468 {
 469         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 470                 return -ENOMEM;
 471         return 0;
 472 }
 473 core_initcall(alloc_frozen_cpus);
 474 #endif /* CONFIG_PM_SLEEP_SMP */
 475
 476 /**
 477  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 478  * @cpu: cpu that just started
 479  *
 480  * This function calls the cpu_chain notifiers with CPU_STARTING.
 481  * It must be called by the arch code on the new cpu, before the new cpu
 482  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 483  */
 484 void __cpuinit notify_cpu_starting(unsigned int cpu)
 485 {
 486         unsigned long val = CPU_STARTING;
 487
 488 #ifdef CONFIG_PM_SLEEP_SMP
 489         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
 490                 val = CPU_STARTING_FROZEN;
 491 #endif /* CONFIG_PM_SLEEP_SMP */
 492         cpu_notify(val, (void *)(long)cpu);
 493 }
 494
 495 #endif /* CONFIG_SMP */
 496
 497 /*
 498  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 499  * represents all NR_CPUS bits binary values of 1<<nr.
 500  *
 501  * It is used by cpumask_of() to get a constant address to a CPU
 502  * mask value that has a single bit set only.
 503  */
 504
 505 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 506 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
 507 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 508 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 509 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 510
 511 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 512
 513         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 514         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 515 #if BITS_PER_LONG > 32
 516         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 517         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 518 #endif
 519 };
 520 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 521
 522 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 523 EXPORT_SYMBOL(cpu_all_bits);
 524
 525 #ifdef CONFIG_INIT_ALL_POSSIBLE
 526 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
 527         = CPU_BITS_ALL;
 528 #else
 529 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
 530 #endif
 531 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
 532 EXPORT_SYMBOL(cpu_possible_mask);
 533
 534 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
 535 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
 536 EXPORT_SYMBOL(cpu_online_mask);
 537
 538 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
 539 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
 540 EXPORT_SYMBOL(cpu_present_mask);
 541
 542 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
 543 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
 544 EXPORT_SYMBOL(cpu_active_mask);
 545
 546 void set_cpu_possible(unsigned int cpu, bool possible)
 547 {
 548         if (possible)
 549                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
 550         else
 551                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
 552 }
 553
 554 void set_cpu_present(unsigned int cpu, bool present)
 555 {
 556         if (present)
 557                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
 558         else
 559                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
 560 }
 561
 562 void set_cpu_online(unsigned int cpu, bool online)
 563 {
 564         if (online)
 565                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
 566         else
 567                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
 568 }
 569
 570 void set_cpu_active(unsigned int cpu, bool active)
 571 {
 572         if (active)
 573                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
 574         else
 575                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
 576 }
 577
 578 void init_cpu_present(const struct cpumask *src)
 579 {
 580         cpumask_copy(to_cpumask(cpu_present_bits), src);
 581 }
 582
 583 void init_cpu_possible(const struct cpumask *src)
 584 {
 585         cpumask_copy(to_cpumask(cpu_possible_bits), src);
 586 }
 587
 588 void init_cpu_online(const struct cpumask *src)
 589 {
 590         cpumask_copy(to_cpumask(cpu_online_bits), src);
 591 }