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