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 __cpuinitdata 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 static void cpu_notify_nofail(unsigned long val, void *v)
 158 {
 159         int err;
 160
 161         err = cpu_notify(val, v);
 162         BUG_ON(err);
 163 }
 164
 165 #ifdef CONFIG_HOTPLUG_CPU
 166
 167 EXPORT_SYMBOL(register_cpu_notifier);
 168
 169 void __ref unregister_cpu_notifier(struct notifier_block *nb)
 170 {
 171         cpu_maps_update_begin();
 172         raw_notifier_chain_unregister(&cpu_chain, nb);
 173         cpu_maps_update_done();
 174 }
 175 EXPORT_SYMBOL(unregister_cpu_notifier);
 176
 177 static inline void check_for_tasks(int cpu)
 178 {
 179         struct task_struct *p;
 180
 181         write_lock_irq(&tasklist_lock);
 182         for_each_process(p) {
 183                 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
 184                     (!cputime_eq(p->utime, cputime_zero) ||
 185                      !cputime_eq(p->stime, cputime_zero)))
 186                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
 187                                 "(state = %ld, flags = %x)\n",
 188                                 p->comm, task_pid_nr(p), cpu,
 189                                 p->state, p->flags);
 190         }
 191         write_unlock_irq(&tasklist_lock);
 192 }
 193
 194 struct take_cpu_down_param {
 195         struct task_struct *caller;
 196         unsigned long mod;
 197         void *hcpu;
 198 };
 199
 200 /* Take this CPU down. */
 201 static int __ref take_cpu_down(void *_param)
 202 {
 203         struct take_cpu_down_param *param = _param;
 204         unsigned int cpu = (unsigned long)param->hcpu;
 205         int err;
 206
 207         /* Ensure this CPU doesn't handle any more interrupts. */
 208         err = __cpu_disable();
 209         if (err < 0)
 210                 return err;
 211
 212         cpu_notify(CPU_DYING | param->mod, param->hcpu);
 213
 214         if (task_cpu(param->caller) == cpu)
 215                 move_task_off_dead_cpu(cpu, param->caller);
 216         /* Force idle task to run as soon as we yield: it should
 217            immediately notice cpu is offline and die quickly. */
 218         sched_idle_next();
 219         return 0;
 220 }
 221
 222 /* Requires cpu_add_remove_lock to be held */
 223 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 224 {
 225         int err, nr_calls = 0;
 226         void *hcpu = (void *)(long)cpu;
 227         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 228         struct take_cpu_down_param tcd_param = {
 229                 .caller = current,
 230                 .mod = mod,
 231                 .hcpu = hcpu,
 232         };
 233
 234         if (num_online_cpus() == 1)
 235                 return -EBUSY;
 236
 237         if (!cpu_online(cpu))
 238                 return -EINVAL;
 239
 240         cpu_hotplug_begin();
 241         set_cpu_active(cpu, false);
 242         err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
 243         if (err) {
 244                 set_cpu_active(cpu, true);
 245
 246                 nr_calls--;
 247                 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
 248                 printk("%s: attempt to take down CPU %u failed\n",
 249                                 __func__, cpu);
 250                 goto out_release;
 251         }
 252
 253         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
 254         if (err) {
 255                 set_cpu_active(cpu, true);
 256                 /* CPU didn't die: tell everyone.  Can't complain. */
 257                 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
 258
 259                 goto out_release;
 260         }
 261         BUG_ON(cpu_online(cpu));
 262
 263         /* Wait for it to sleep (leaving idle task). */
 264         while (!idle_cpu(cpu))
 265                 yield();
 266
 267         /* This actually kills the CPU. */
 268         __cpu_die(cpu);
 269
 270         /* CPU is completely dead: tell everyone.  Too late to complain. */
 271         cpu_notify_nofail(CPU_DEAD | mod, hcpu);
 272
 273         check_for_tasks(cpu);
 274
 275 out_release:
 276         cpu_hotplug_done();
 277         if (!err)
 278                 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
 279         return err;
 280 }
 281
 282 int __ref cpu_down(unsigned int cpu)
 283 {
 284         int err;
 285
 286         cpu_maps_update_begin();
 287
 288         if (cpu_hotplug_disabled) {
 289                 err = -EBUSY;
 290                 goto out;
 291         }
 292
 293         err = _cpu_down(cpu, 0);
 294
 295 out:
 296         cpu_maps_update_done();
 297         return err;
 298 }
 299 EXPORT_SYMBOL(cpu_down);
 300 #endif /*CONFIG_HOTPLUG_CPU*/
 301
 302 /* Requires cpu_add_remove_lock to be held */
 303 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 304 {
 305         int ret, nr_calls = 0;
 306         void *hcpu = (void *)(long)cpu;
 307         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 308
 309         if (cpu_online(cpu) || !cpu_present(cpu))
 310                 return -EINVAL;
 311
 312         cpu_hotplug_begin();
 313         ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
 314         if (ret) {
 315                 nr_calls--;
 316                 printk("%s: attempt to bring up CPU %u failed\n",
 317                                 __func__, cpu);
 318                 goto out_notify;
 319         }
 320
 321         /* Arch-specific enabling code. */
 322         ret = __cpu_up(cpu);
 323         if (ret != 0)
 324                 goto out_notify;
 325         BUG_ON(!cpu_online(cpu));
 326
 327         set_cpu_active(cpu, true);
 328
 329         /* Now call notifier in preparation. */
 330         cpu_notify(CPU_ONLINE | mod, hcpu);
 331
 332 out_notify:
 333         if (ret != 0)
 334                 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 335         cpu_hotplug_done();
 336
 337         return ret;
 338 }
 339
 340 int __cpuinit cpu_up(unsigned int cpu)
 341 {
 342         int err = 0;
 343
 344 #ifdef  CONFIG_MEMORY_HOTPLUG
 345         int nid;
 346         pg_data_t       *pgdat;
 347 #endif
 348
 349         if (!cpu_possible(cpu)) {
 350                 printk(KERN_ERR "can't online cpu %d because it is not "
 351                         "configured as may-hotadd at boot time\n", cpu);
 352 #if defined(CONFIG_IA64)
 353                 printk(KERN_ERR "please check additional_cpus= boot "
 354                                 "parameter\n");
 355 #endif
 356                 return -EINVAL;
 357         }
 358
 359 #ifdef  CONFIG_MEMORY_HOTPLUG
 360         nid = cpu_to_node(cpu);
 361         if (!node_online(nid)) {
 362                 err = mem_online_node(nid);
 363                 if (err)
 364                         return err;
 365         }
 366
 367         pgdat = NODE_DATA(nid);
 368         if (!pgdat) {
 369                 printk(KERN_ERR
 370                         "Can't online cpu %d due to NULL pgdat\n", cpu);
 371                 return -ENOMEM;
 372         }
 373
 374         if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
 375                 mutex_lock(&zonelists_mutex);
 376                 build_all_zonelists(NULL);
 377                 mutex_unlock(&zonelists_mutex);
 378         }
 379 #endif
 380
 381         cpu_maps_update_begin();
 382
 383         if (cpu_hotplug_disabled) {
 384                 err = -EBUSY;
 385                 goto out;
 386         }
 387
 388         err = _cpu_up(cpu, 0);
 389
 390 out:
 391         cpu_maps_update_done();
 392         return err;
 393 }
 394
 395 #ifdef CONFIG_PM_SLEEP_SMP
 396 static cpumask_var_t frozen_cpus;
 397
 398 int disable_nonboot_cpus(void)
 399 {
 400         int cpu, first_cpu, error;
 401
 402         cpu_maps_update_begin();
 403         first_cpu = cpumask_first(cpu_online_mask);
 404         /*
 405          * We take down all of the non-boot CPUs in one shot to avoid races
 406          * with the userspace trying to use the CPU hotplug at the same time
 407          */
 408         cpumask_clear(frozen_cpus);
 409
 410         printk("Disabling non-boot CPUs ...\n");
 411         for_each_online_cpu(cpu) {
 412                 if (cpu == first_cpu)
 413                         continue;
 414                 error = _cpu_down(cpu, 1);
 415                 if (!error)
 416                         cpumask_set_cpu(cpu, frozen_cpus);
 417                 else {
 418                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 419                                 cpu, error);
 420                         break;
 421                 }
 422         }
 423
 424         if (!error) {
 425                 BUG_ON(num_online_cpus() > 1);
 426                 /* Make sure the CPUs won't be enabled by someone else */
 427                 cpu_hotplug_disabled = 1;
 428         } else {
 429                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 430         }
 431         cpu_maps_update_done();
 432         return error;
 433 }
 434
 435 void __weak arch_enable_nonboot_cpus_begin(void)
 436 {
 437 }
 438
 439 void __weak arch_enable_nonboot_cpus_end(void)
 440 {
 441 }
 442
 443 void __ref enable_nonboot_cpus(void)
 444 {
 445         int cpu, error;
 446
 447         /* Allow everyone to use the CPU hotplug again */
 448         cpu_maps_update_begin();
 449         cpu_hotplug_disabled = 0;
 450         if (cpumask_empty(frozen_cpus))
 451                 goto out;
 452
 453         printk("Enabling non-boot CPUs ...\n");
 454
 455         arch_enable_nonboot_cpus_begin();
 456
 457         for_each_cpu(cpu, frozen_cpus) {
 458                 error = _cpu_up(cpu, 1);
 459                 if (!error) {
 460                         printk("CPU%d is up\n", cpu);
 461                         continue;
 462                 }
 463                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 464         }
 465
 466         arch_enable_nonboot_cpus_end();
 467
 468         cpumask_clear(frozen_cpus);
 469 out:
 470         cpu_maps_update_done();
 471 }
 472
 473 static int alloc_frozen_cpus(void)
 474 {
 475         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 476                 return -ENOMEM;
 477         return 0;
 478 }
 479 core_initcall(alloc_frozen_cpus);
 480 #endif /* CONFIG_PM_SLEEP_SMP */
 481
 482 /**
 483  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 484  * @cpu: cpu that just started
 485  *
 486  * This function calls the cpu_chain notifiers with CPU_STARTING.
 487  * It must be called by the arch code on the new cpu, before the new cpu
 488  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 489  */
 490 void __cpuinit notify_cpu_starting(unsigned int cpu)
 491 {
 492         unsigned long val = CPU_STARTING;
 493
 494 #ifdef CONFIG_PM_SLEEP_SMP
 495         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
 496                 val = CPU_STARTING_FROZEN;
 497 #endif /* CONFIG_PM_SLEEP_SMP */
 498         cpu_notify(val, (void *)(long)cpu);
 499 }
 500
 501 #endif /* CONFIG_SMP */
 502
 503 /*
 504  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 505  * represents all NR_CPUS bits binary values of 1<<nr.
 506  *
 507  * It is used by cpumask_of() to get a constant address to a CPU
 508  * mask value that has a single bit set only.
 509  */
 510
 511 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 512 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
 513 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 514 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 515 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 516
 517 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 518
 519         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 520         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 521 #if BITS_PER_LONG > 32
 522         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 523         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 524 #endif
 525 };
 526 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 527
 528 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 529 EXPORT_SYMBOL(cpu_all_bits);
 530
 531 #ifdef CONFIG_INIT_ALL_POSSIBLE
 532 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
 533         = CPU_BITS_ALL;
 534 #else
 535 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
 536 #endif
 537 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
 538 EXPORT_SYMBOL(cpu_possible_mask);
 539
 540 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
 541 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
 542 EXPORT_SYMBOL(cpu_online_mask);
 543
 544 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
 545 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
 546 EXPORT_SYMBOL(cpu_present_mask);
 547
 548 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
 549 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
 550 EXPORT_SYMBOL(cpu_active_mask);
 551
 552 void set_cpu_possible(unsigned int cpu, bool possible)
 553 {
 554         if (possible)
 555                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
 556         else
 557                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
 558 }
 559
 560 void set_cpu_present(unsigned int cpu, bool present)
 561 {
 562         if (present)
 563                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
 564         else
 565                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
 566 }
 567
 568 void set_cpu_online(unsigned int cpu, bool online)
 569 {
 570         if (online)
 571                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
 572         else
 573                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
 574 }
 575
 576 void set_cpu_active(unsigned int cpu, bool active)
 577 {
 578         if (active)
 579                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
 580         else
 581                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
 582 }
 583
 584 void init_cpu_present(const struct cpumask *src)
 585 {
 586         cpumask_copy(to_cpumask(cpu_present_bits), src);
 587 }
 588
 589 void init_cpu_possible(const struct cpumask *src)
 590 {
 591         cpumask_copy(to_cpumask(cpu_possible_bits), src);
 592 }
 593
 594 void init_cpu_online(const struct cpumask *src)
 595 {
 596         cpumask_copy(to_cpumask(cpu_online_bits), src);
 597 }