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 /*
  19  * Represents all cpu's present in the system
  20  * In systems capable of hotplug, this map could dynamically grow
  21  * as new cpu's are detected in the system via any platform specific
  22  * method, such as ACPI for e.g.
  23  */
  24 cpumask_t cpu_present_map __read_mostly;
  25 EXPORT_SYMBOL(cpu_present_map);
  26
  27 #ifndef CONFIG_SMP
  28
  29 /*
  30  * Represents all cpu's that are currently online.
  31  */
  32 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
  33 EXPORT_SYMBOL(cpu_online_map);
  34
  35 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
  36 EXPORT_SYMBOL(cpu_possible_map);
  37
  38 #else /* CONFIG_SMP */
  39
  40 /* Serializes the updates to cpu_online_map, cpu_present_map */
  41 static DEFINE_MUTEX(cpu_add_remove_lock);
  42
  43 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
  44
  45 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  46  * Should always be manipulated under cpu_add_remove_lock
  47  */
  48 static int cpu_hotplug_disabled;
  49
  50 static struct {
  51         struct task_struct *active_writer;
  52         struct mutex lock; /* Synchronizes accesses to refcount, */
  53         /*
  54          * Also blocks the new readers during
  55          * an ongoing cpu hotplug operation.
  56          */
  57         int refcount;
  58 } cpu_hotplug;
  59
  60 void __init cpu_hotplug_init(void)
  61 {
  62         cpu_hotplug.active_writer = NULL;
  63         mutex_init(&cpu_hotplug.lock);
  64         cpu_hotplug.refcount = 0;
  65 }
  66
  67 cpumask_t cpu_active_map;
  68
  69 #ifdef CONFIG_HOTPLUG_CPU
  70
  71 void get_online_cpus(void)
  72 {
  73         might_sleep();
  74         if (cpu_hotplug.active_writer == current)
  75                 return;
  76         mutex_lock(&cpu_hotplug.lock);
  77         cpu_hotplug.refcount++;
  78         mutex_unlock(&cpu_hotplug.lock);
  79
  80 }
  81 EXPORT_SYMBOL_GPL(get_online_cpus);
  82
  83 void put_online_cpus(void)
  84 {
  85         if (cpu_hotplug.active_writer == current)
  86                 return;
  87         mutex_lock(&cpu_hotplug.lock);
  88         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
  89                 wake_up_process(cpu_hotplug.active_writer);
  90         mutex_unlock(&cpu_hotplug.lock);
  91
  92 }
  93 EXPORT_SYMBOL_GPL(put_online_cpus);
  94
  95 #endif  /* CONFIG_HOTPLUG_CPU */
  96
  97 /*
  98  * The following two API's must be used when attempting
  99  * to serialize the updates to cpu_online_map, cpu_present_map.
 100  */
 101 void cpu_maps_update_begin(void)
 102 {
 103         mutex_lock(&cpu_add_remove_lock);
 104 }
 105
 106 void cpu_maps_update_done(void)
 107 {
 108         mutex_unlock(&cpu_add_remove_lock);
 109 }
 110
 111 /*
 112  * This ensures that the hotplug operation can begin only when the
 113  * refcount goes to zero.
 114  *
 115  * Note that during a cpu-hotplug operation, the new readers, if any,
 116  * will be blocked by the cpu_hotplug.lock
 117  *
 118  * Since cpu_hotplug_begin() is always called after invoking
 119  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 120  *
 121  * Note that theoretically, there is a possibility of a livelock:
 122  * - Refcount goes to zero, last reader wakes up the sleeping
 123  *   writer.
 124  * - Last reader unlocks the cpu_hotplug.lock.
 125  * - A new reader arrives at this moment, bumps up the refcount.
 126  * - The writer acquires the cpu_hotplug.lock finds the refcount
 127  *   non zero and goes to sleep again.
 128  *
 129  * However, this is very difficult to achieve in practice since
 130  * get_online_cpus() not an api which is called all that often.
 131  *
 132  */
 133 static void cpu_hotplug_begin(void)
 134 {
 135         cpu_hotplug.active_writer = current;
 136
 137         for (;;) {
 138                 mutex_lock(&cpu_hotplug.lock);
 139                 if (likely(!cpu_hotplug.refcount))
 140                         break;
 141                 __set_current_state(TASK_UNINTERRUPTIBLE);
 142                 mutex_unlock(&cpu_hotplug.lock);
 143                 schedule();
 144         }
 145 }
 146
 147 static void cpu_hotplug_done(void)
 148 {
 149         cpu_hotplug.active_writer = NULL;
 150         mutex_unlock(&cpu_hotplug.lock);
 151 }
 152 /* Need to know about CPUs going up/down? */
 153 int __ref register_cpu_notifier(struct notifier_block *nb)
 154 {
 155         int ret;
 156         cpu_maps_update_begin();
 157         ret = raw_notifier_chain_register(&cpu_chain, nb);
 158         cpu_maps_update_done();
 159         return ret;
 160 }
 161
 162 #ifdef CONFIG_HOTPLUG_CPU
 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 &&
 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         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         int err;
 201
 202         /* Ensure this CPU doesn't handle any more interrupts. */
 203         err = __cpu_disable();
 204         if (err < 0)
 205                 return err;
 206
 207         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
 208                                 param->hcpu);
 209
 210         /* Force idle task to run as soon as we yield: it should
 211            immediately notice cpu is offline and die quickly. */
 212         sched_idle_next();
 213         return 0;
 214 }
 215
 216 /* Requires cpu_add_remove_lock to be held */
 217 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 218 {
 219         int err, nr_calls = 0;
 220         cpumask_t old_allowed, tmp;
 221         void *hcpu = (void *)(long)cpu;
 222         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 223         struct take_cpu_down_param tcd_param = {
 224                 .mod = mod,
 225                 .hcpu = hcpu,
 226         };
 227
 228         if (num_online_cpus() == 1)
 229                 return -EBUSY;
 230
 231         if (!cpu_online(cpu))
 232                 return -EINVAL;
 233
 234         cpu_hotplug_begin();
 235         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
 236                                         hcpu, -1, &nr_calls);
 237         if (err == NOTIFY_BAD) {
 238                 nr_calls--;
 239                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 240                                           hcpu, nr_calls, NULL);
 241                 printk("%s: attempt to take down CPU %u failed\n",
 242                                 __func__, cpu);
 243                 err = -EINVAL;
 244                 goto out_release;
 245         }
 246
 247         /* Ensure that we are not runnable on dying cpu */
 248         old_allowed = current->cpus_allowed;
 249         cpus_setall(tmp);
 250         cpu_clear(cpu, tmp);
 251         set_cpus_allowed_ptr(current, &tmp);
 252         tmp = cpumask_of_cpu(cpu);
 253
 254         err = __stop_machine(take_cpu_down, &tcd_param, &tmp);
 255         if (err) {
 256                 /* CPU didn't die: tell everyone.  Can't complain. */
 257                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 258                                             hcpu) == NOTIFY_BAD)
 259                         BUG();
 260
 261                 goto out_allowed;
 262         }
 263         BUG_ON(cpu_online(cpu));
 264
 265         /* Wait for it to sleep (leaving idle task). */
 266         while (!idle_cpu(cpu))
 267                 yield();
 268
 269         /* This actually kills the CPU. */
 270         __cpu_die(cpu);
 271
 272         /* CPU is completely dead: tell everyone.  Too late to complain. */
 273         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
 274                                     hcpu) == NOTIFY_BAD)
 275                 BUG();
 276
 277         check_for_tasks(cpu);
 278
 279 out_allowed:
 280         set_cpus_allowed_ptr(current, &old_allowed);
 281 out_release:
 282         cpu_hotplug_done();
 283         if (!err) {
 284                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
 285                                             hcpu) == NOTIFY_BAD)
 286                         BUG();
 287         }
 288         return err;
 289 }
 290
 291 int __ref cpu_down(unsigned int cpu)
 292 {
 293         int err = 0;
 294
 295         cpu_maps_update_begin();
 296
 297         if (cpu_hotplug_disabled) {
 298                 err = -EBUSY;
 299                 goto out;
 300         }
 301
 302         cpu_clear(cpu, cpu_active_map);
 303
 304         /*
 305          * Make sure the all cpus did the reschedule and are not
 306          * using stale version of the cpu_active_map.
 307          * This is not strictly necessary becuase stop_machine()
 308          * that we run down the line already provides the required
 309          * synchronization. But it's really a side effect and we do not
 310          * want to depend on the innards of the stop_machine here.
 311          */
 312         synchronize_sched();
 313
 314         err = _cpu_down(cpu, 0);
 315
 316         if (cpu_online(cpu))
 317                 cpu_set(cpu, cpu_active_map);
 318
 319 out:
 320         cpu_maps_update_done();
 321         return err;
 322 }
 323 EXPORT_SYMBOL(cpu_down);
 324 #endif /*CONFIG_HOTPLUG_CPU*/
 325
 326 /* Requires cpu_add_remove_lock to be held */
 327 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 328 {
 329         int ret, nr_calls = 0;
 330         void *hcpu = (void *)(long)cpu;
 331         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 332
 333         if (cpu_online(cpu) || !cpu_present(cpu))
 334                 return -EINVAL;
 335
 336         cpu_hotplug_begin();
 337         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
 338                                                         -1, &nr_calls);
 339         if (ret == NOTIFY_BAD) {
 340                 nr_calls--;
 341                 printk("%s: attempt to bring up CPU %u failed\n",
 342                                 __func__, cpu);
 343                 ret = -EINVAL;
 344                 goto out_notify;
 345         }
 346
 347         /* Arch-specific enabling code. */
 348         ret = __cpu_up(cpu);
 349         if (ret != 0)
 350                 goto out_notify;
 351         BUG_ON(!cpu_online(cpu));
 352
 353         cpu_set(cpu, cpu_active_map);
 354
 355         /* Now call notifier in preparation. */
 356         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 357
 358 out_notify:
 359         if (ret != 0)
 360                 __raw_notifier_call_chain(&cpu_chain,
 361                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 362         cpu_hotplug_done();
 363
 364         return ret;
 365 }
 366
 367 int __cpuinit cpu_up(unsigned int cpu)
 368 {
 369         int err = 0;
 370         if (!cpu_isset(cpu, cpu_possible_map)) {
 371                 printk(KERN_ERR "can't online cpu %d because it is not "
 372                         "configured as may-hotadd at boot time\n", cpu);
 373 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
 374                 printk(KERN_ERR "please check additional_cpus= boot "
 375                                 "parameter\n");
 376 #endif
 377                 return -EINVAL;
 378         }
 379
 380         cpu_maps_update_begin();
 381
 382         if (cpu_hotplug_disabled) {
 383                 err = -EBUSY;
 384                 goto out;
 385         }
 386
 387         err = _cpu_up(cpu, 0);
 388
 389 out:
 390         cpu_maps_update_done();
 391         return err;
 392 }
 393
 394 #ifdef CONFIG_PM_SLEEP_SMP
 395 static cpumask_t frozen_cpus;
 396
 397 int disable_nonboot_cpus(void)
 398 {
 399         int cpu, first_cpu, error = 0;
 400
 401         cpu_maps_update_begin();
 402         first_cpu = first_cpu(cpu_online_map);
 403         /* We take down all of the non-boot CPUs in one shot to avoid races
 404          * with the userspace trying to use the CPU hotplug at the same time
 405          */
 406         cpus_clear(frozen_cpus);
 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                         cpu_set(cpu, frozen_cpus);
 414                         printk("CPU%d is down\n", cpu);
 415                 } else {
 416                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 417                                 cpu, error);
 418                         break;
 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 __ref enable_nonboot_cpus(void)
 433 {
 434         int cpu, error;
 435
 436         /* Allow everyone to use the CPU hotplug again */
 437         cpu_maps_update_begin();
 438         cpu_hotplug_disabled = 0;
 439         if (cpus_empty(frozen_cpus))
 440                 goto out;
 441
 442         printk("Enabling non-boot CPUs ...\n");
 443         for_each_cpu_mask_nr(cpu, frozen_cpus) {
 444                 error = _cpu_up(cpu, 1);
 445                 if (!error) {
 446                         printk("CPU%d is up\n", cpu);
 447                         continue;
 448                 }
 449                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 450         }
 451         cpus_clear(frozen_cpus);
 452 out:
 453         cpu_maps_update_done();
 454 }
 455 #endif /* CONFIG_PM_SLEEP_SMP */
 456
 457 /**
 458  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 459  * @cpu: cpu that just started
 460  *
 461  * This function calls the cpu_chain notifiers with CPU_STARTING.
 462  * It must be called by the arch code on the new cpu, before the new cpu
 463  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 464  */
 465 void __cpuinit notify_cpu_starting(unsigned int cpu)
 466 {
 467         unsigned long val = CPU_STARTING;
 468
 469 #ifdef CONFIG_PM_SLEEP_SMP
 470         if (cpu_isset(cpu, frozen_cpus))
 471                 val = CPU_STARTING_FROZEN;
 472 #endif /* CONFIG_PM_SLEEP_SMP */
 473         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
 474 }
 475
 476 #endif /* CONFIG_SMP */
 477
 478 /*
 479  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 480  * represents all NR_CPUS bits binary values of 1<<nr.
 481  *
 482  * It is used by cpumask_of_cpu() to get a constant address to a CPU
 483  * mask value that has a single bit set only.
 484  */
 485
 486 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 487 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
 488 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 489 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 490 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 491
 492 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 493
 494         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 495         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 496 #if BITS_PER_LONG > 32
 497         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 498         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 499 #endif
 500 };
 501 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 502
 503 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 504 EXPORT_SYMBOL(cpu_all_bits);