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 #ifdef CONFIG_HOTPLUG_CPU
  68
  69 void get_online_cpus(void)
  70 {
  71         might_sleep();
  72         if (cpu_hotplug.active_writer == current)
  73                 return;
  74         mutex_lock(&cpu_hotplug.lock);
  75         cpu_hotplug.refcount++;
  76         mutex_unlock(&cpu_hotplug.lock);
  77
  78 }
  79 EXPORT_SYMBOL_GPL(get_online_cpus);
  80
  81 void put_online_cpus(void)
  82 {
  83         if (cpu_hotplug.active_writer == current)
  84                 return;
  85         mutex_lock(&cpu_hotplug.lock);
  86         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
  87                 wake_up_process(cpu_hotplug.active_writer);
  88         mutex_unlock(&cpu_hotplug.lock);
  89
  90 }
  91 EXPORT_SYMBOL_GPL(put_online_cpus);
  92
  93 #endif  /* CONFIG_HOTPLUG_CPU */
  94
  95 /*
  96  * The following two API's must be used when attempting
  97  * to serialize the updates to cpu_online_map, cpu_present_map.
  98  */
  99 void cpu_maps_update_begin(void)
 100 {
 101         mutex_lock(&cpu_add_remove_lock);
 102 }
 103
 104 void cpu_maps_update_done(void)
 105 {
 106         mutex_unlock(&cpu_add_remove_lock);
 107 }
 108
 109 /*
 110  * This ensures that the hotplug operation can begin only when the
 111  * refcount goes to zero.
 112  *
 113  * Note that during a cpu-hotplug operation, the new readers, if any,
 114  * will be blocked by the cpu_hotplug.lock
 115  *
 116  * Since cpu_hotplug_begin() is always called after invoking
 117  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 118  *
 119  * Note that theoretically, there is a possibility of a livelock:
 120  * - Refcount goes to zero, last reader wakes up the sleeping
 121  *   writer.
 122  * - Last reader unlocks the cpu_hotplug.lock.
 123  * - A new reader arrives at this moment, bumps up the refcount.
 124  * - The writer acquires the cpu_hotplug.lock finds the refcount
 125  *   non zero and goes to sleep again.
 126  *
 127  * However, this is very difficult to achieve in practice since
 128  * get_online_cpus() not an api which is called all that often.
 129  *
 130  */
 131 static void cpu_hotplug_begin(void)
 132 {
 133         cpu_hotplug.active_writer = current;
 134
 135         for (;;) {
 136                 mutex_lock(&cpu_hotplug.lock);
 137                 if (likely(!cpu_hotplug.refcount))
 138                         break;
 139                 __set_current_state(TASK_UNINTERRUPTIBLE);
 140                 mutex_unlock(&cpu_hotplug.lock);
 141                 schedule();
 142         }
 143 }
 144
 145 static void cpu_hotplug_done(void)
 146 {
 147         cpu_hotplug.active_writer = NULL;
 148         mutex_unlock(&cpu_hotplug.lock);
 149 }
 150 /* Need to know about CPUs going up/down? */
 151 int __ref register_cpu_notifier(struct notifier_block *nb)
 152 {
 153         int ret;
 154         cpu_maps_update_begin();
 155         ret = raw_notifier_chain_register(&cpu_chain, nb);
 156         cpu_maps_update_done();
 157         return ret;
 158 }
 159
 160 #ifdef CONFIG_HOTPLUG_CPU
 161
 162 EXPORT_SYMBOL(register_cpu_notifier);
 163
 164 void __ref unregister_cpu_notifier(struct notifier_block *nb)
 165 {
 166         cpu_maps_update_begin();
 167         raw_notifier_chain_unregister(&cpu_chain, nb);
 168         cpu_maps_update_done();
 169 }
 170 EXPORT_SYMBOL(unregister_cpu_notifier);
 171
 172 static inline void check_for_tasks(int cpu)
 173 {
 174         struct task_struct *p;
 175
 176         write_lock_irq(&tasklist_lock);
 177         for_each_process(p) {
 178                 if (task_cpu(p) == cpu &&
 179                     (!cputime_eq(p->utime, cputime_zero) ||
 180                      !cputime_eq(p->stime, cputime_zero)))
 181                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
 182                                 (state = %ld, flags = %x) \n",
 183                                  p->comm, task_pid_nr(p), cpu,
 184                                  p->state, p->flags);
 185         }
 186         write_unlock_irq(&tasklist_lock);
 187 }
 188
 189 struct take_cpu_down_param {
 190         unsigned long mod;
 191         void *hcpu;
 192 };
 193
 194 /* Take this CPU down. */
 195 static int __ref take_cpu_down(void *_param)
 196 {
 197         struct take_cpu_down_param *param = _param;
 198         int err;
 199
 200         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
 201                                 param->hcpu);
 202         /* Ensure this CPU doesn't handle any more interrupts. */
 203         err = __cpu_disable();
 204         if (err < 0)
 205                 return err;
 206
 207         /* Force idle task to run as soon as we yield: it should
 208            immediately notice cpu is offline and die quickly. */
 209         sched_idle_next();
 210         return 0;
 211 }
 212
 213 /* Requires cpu_add_remove_lock to be held */
 214 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 215 {
 216         int err, nr_calls = 0;
 217         struct task_struct *p;
 218         cpumask_t old_allowed, tmp;
 219         void *hcpu = (void *)(long)cpu;
 220         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 221         struct take_cpu_down_param tcd_param = {
 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         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
 234                                         hcpu, -1, &nr_calls);
 235         if (err == NOTIFY_BAD) {
 236                 nr_calls--;
 237                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 238                                           hcpu, nr_calls, NULL);
 239                 printk("%s: attempt to take down CPU %u failed\n",
 240                                 __func__, cpu);
 241                 err = -EINVAL;
 242                 goto out_release;
 243         }
 244
 245         /* Ensure that we are not runnable on dying cpu */
 246         old_allowed = current->cpus_allowed;
 247         cpus_setall(tmp);
 248         cpu_clear(cpu, tmp);
 249         set_cpus_allowed_ptr(current, &tmp);
 250
 251         p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
 252
 253         if (IS_ERR(p) || cpu_online(cpu)) {
 254                 /* CPU didn't die: tell everyone.  Can't complain. */
 255                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 256                                             hcpu) == NOTIFY_BAD)
 257                         BUG();
 258
 259                 if (IS_ERR(p)) {
 260                         err = PTR_ERR(p);
 261                         goto out_allowed;
 262                 }
 263                 goto out_thread;
 264         }
 265
 266         /* Wait for it to sleep (leaving idle task). */
 267         while (!idle_cpu(cpu))
 268                 yield();
 269
 270         /* This actually kills the CPU. */
 271         __cpu_die(cpu);
 272
 273         /* CPU is completely dead: tell everyone.  Too late to complain. */
 274         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
 275                                     hcpu) == NOTIFY_BAD)
 276                 BUG();
 277
 278         check_for_tasks(cpu);
 279
 280 out_thread:
 281         err = kthread_stop(p);
 282 out_allowed:
 283         set_cpus_allowed_ptr(current, &old_allowed);
 284 out_release:
 285         cpu_hotplug_done();
 286         return err;
 287 }
 288
 289 int __ref cpu_down(unsigned int cpu)
 290 {
 291         int err = 0;
 292
 293         cpu_maps_update_begin();
 294         if (cpu_hotplug_disabled)
 295                 err = -EBUSY;
 296         else
 297                 err = _cpu_down(cpu, 0);
 298
 299         cpu_maps_update_done();
 300         return err;
 301 }
 302 EXPORT_SYMBOL(cpu_down);
 303 #endif /*CONFIG_HOTPLUG_CPU*/
 304
 305 /* Requires cpu_add_remove_lock to be held */
 306 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 307 {
 308         int ret, nr_calls = 0;
 309         void *hcpu = (void *)(long)cpu;
 310         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 311
 312         if (cpu_online(cpu) || !cpu_present(cpu))
 313                 return -EINVAL;
 314
 315         cpu_hotplug_begin();
 316         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
 317                                                         -1, &nr_calls);
 318         if (ret == NOTIFY_BAD) {
 319                 nr_calls--;
 320                 printk("%s: attempt to bring up CPU %u failed\n",
 321                                 __func__, cpu);
 322                 ret = -EINVAL;
 323                 goto out_notify;
 324         }
 325
 326         /* Arch-specific enabling code. */
 327         ret = __cpu_up(cpu);
 328         if (ret != 0)
 329                 goto out_notify;
 330         BUG_ON(!cpu_online(cpu));
 331
 332         /* Now call notifier in preparation. */
 333         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 334
 335 out_notify:
 336         if (ret != 0)
 337                 __raw_notifier_call_chain(&cpu_chain,
 338                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 339         cpu_hotplug_done();
 340
 341         return ret;
 342 }
 343
 344 int __cpuinit cpu_up(unsigned int cpu)
 345 {
 346         int err = 0;
 347         if (!cpu_isset(cpu, cpu_possible_map)) {
 348                 printk(KERN_ERR "can't online cpu %d because it is not "
 349                         "configured as may-hotadd at boot time\n", cpu);
 350 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64) || defined(CONFIG_S390)
 351                 printk(KERN_ERR "please check additional_cpus= boot "
 352                                 "parameter\n");
 353 #endif
 354                 return -EINVAL;
 355         }
 356
 357         cpu_maps_update_begin();
 358         if (cpu_hotplug_disabled)
 359                 err = -EBUSY;
 360         else
 361                 err = _cpu_up(cpu, 0);
 362
 363         cpu_maps_update_done();
 364         return err;
 365 }
 366
 367 #ifdef CONFIG_PM_SLEEP_SMP
 368 static cpumask_t frozen_cpus;
 369
 370 int disable_nonboot_cpus(void)
 371 {
 372         int cpu, first_cpu, error = 0;
 373
 374         cpu_maps_update_begin();
 375         first_cpu = first_cpu(cpu_online_map);
 376         /* We take down all of the non-boot CPUs in one shot to avoid races
 377          * with the userspace trying to use the CPU hotplug at the same time
 378          */
 379         cpus_clear(frozen_cpus);
 380         printk("Disabling non-boot CPUs ...\n");
 381         for_each_online_cpu(cpu) {
 382                 if (cpu == first_cpu)
 383                         continue;
 384                 error = _cpu_down(cpu, 1);
 385                 if (!error) {
 386                         cpu_set(cpu, frozen_cpus);
 387                         printk("CPU%d is down\n", cpu);
 388                 } else {
 389                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 390                                 cpu, error);
 391                         break;
 392                 }
 393         }
 394         if (!error) {
 395                 BUG_ON(num_online_cpus() > 1);
 396                 /* Make sure the CPUs won't be enabled by someone else */
 397                 cpu_hotplug_disabled = 1;
 398         } else {
 399                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 400         }
 401         cpu_maps_update_done();
 402         return error;
 403 }
 404
 405 void __ref enable_nonboot_cpus(void)
 406 {
 407         int cpu, error;
 408
 409         /* Allow everyone to use the CPU hotplug again */
 410         cpu_maps_update_begin();
 411         cpu_hotplug_disabled = 0;
 412         if (cpus_empty(frozen_cpus))
 413                 goto out;
 414
 415         printk("Enabling non-boot CPUs ...\n");
 416         for_each_cpu_mask_nr(cpu, frozen_cpus) {
 417                 error = _cpu_up(cpu, 1);
 418                 if (!error) {
 419                         printk("CPU%d is up\n", cpu);
 420                         continue;
 421                 }
 422                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 423         }
 424         cpus_clear(frozen_cpus);
 425 out:
 426         cpu_maps_update_done();
 427 }
 428 #endif /* CONFIG_PM_SLEEP_SMP */
 429
 430 #endif /* CONFIG_SMP */