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