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 /* Serializes the updates to cpu_online_map, cpu_present_map */
  19 static DEFINE_MUTEX(cpu_add_remove_lock);
  20
  21 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
  22
  23 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  24  * Should always be manipulated under cpu_add_remove_lock
  25  */
  26 static int cpu_hotplug_disabled;
  27
  28 static struct {
  29         struct task_struct *active_writer;
  30         struct mutex lock; /* Synchronizes accesses to refcount, */
  31         /*
  32          * Also blocks the new readers during
  33          * an ongoing cpu hotplug operation.
  34          */
  35         int refcount;
  36         wait_queue_head_t writer_queue;
  37 } cpu_hotplug;
  38
  39 #define writer_exists() (cpu_hotplug.active_writer != NULL)
  40
  41 void __init cpu_hotplug_init(void)
  42 {
  43         cpu_hotplug.active_writer = NULL;
  44         mutex_init(&cpu_hotplug.lock);
  45         cpu_hotplug.refcount = 0;
  46         init_waitqueue_head(&cpu_hotplug.writer_queue);
  47 }
  48
  49 #ifdef CONFIG_HOTPLUG_CPU
  50
  51 void get_online_cpus(void)
  52 {
  53         might_sleep();
  54         if (cpu_hotplug.active_writer == current)
  55                 return;
  56         mutex_lock(&cpu_hotplug.lock);
  57         cpu_hotplug.refcount++;
  58         mutex_unlock(&cpu_hotplug.lock);
  59
  60 }
  61 EXPORT_SYMBOL_GPL(get_online_cpus);
  62
  63 void put_online_cpus(void)
  64 {
  65         if (cpu_hotplug.active_writer == current)
  66                 return;
  67         mutex_lock(&cpu_hotplug.lock);
  68         cpu_hotplug.refcount--;
  69
  70         if (unlikely(writer_exists()) && !cpu_hotplug.refcount)
  71                 wake_up(&cpu_hotplug.writer_queue);
  72
  73         mutex_unlock(&cpu_hotplug.lock);
  74
  75 }
  76 EXPORT_SYMBOL_GPL(put_online_cpus);
  77
  78 #endif  /* CONFIG_HOTPLUG_CPU */
  79
  80 /*
  81  * The following two API's must be used when attempting
  82  * to serialize the updates to cpu_online_map, cpu_present_map.
  83  */
  84 void cpu_maps_update_begin(void)
  85 {
  86         mutex_lock(&cpu_add_remove_lock);
  87 }
  88
  89 void cpu_maps_update_done(void)
  90 {
  91         mutex_unlock(&cpu_add_remove_lock);
  92 }
  93
  94 /*
  95  * This ensures that the hotplug operation can begin only when the
  96  * refcount goes to zero.
  97  *
  98  * Note that during a cpu-hotplug operation, the new readers, if any,
  99  * will be blocked by the cpu_hotplug.lock
 100  *
 101  * Since cpu_maps_update_begin is always called after invoking
 102  * cpu_maps_update_begin, we can be sure that only one writer is active.
 103  *
 104  * Note that theoretically, there is a possibility of a livelock:
 105  * - Refcount goes to zero, last reader wakes up the sleeping
 106  *   writer.
 107  * - Last reader unlocks the cpu_hotplug.lock.
 108  * - A new reader arrives at this moment, bumps up the refcount.
 109  * - The writer acquires the cpu_hotplug.lock finds the refcount
 110  *   non zero and goes to sleep again.
 111  *
 112  * However, this is very difficult to achieve in practice since
 113  * get_online_cpus() not an api which is called all that often.
 114  *
 115  */
 116 static void cpu_hotplug_begin(void)
 117 {
 118         DECLARE_WAITQUEUE(wait, current);
 119
 120         mutex_lock(&cpu_hotplug.lock);
 121
 122         cpu_hotplug.active_writer = current;
 123         add_wait_queue_exclusive(&cpu_hotplug.writer_queue, &wait);
 124         while (cpu_hotplug.refcount) {
 125                 set_current_state(TASK_UNINTERRUPTIBLE);
 126                 mutex_unlock(&cpu_hotplug.lock);
 127                 schedule();
 128                 mutex_lock(&cpu_hotplug.lock);
 129         }
 130         remove_wait_queue_locked(&cpu_hotplug.writer_queue, &wait);
 131 }
 132
 133 static void cpu_hotplug_done(void)
 134 {
 135         cpu_hotplug.active_writer = NULL;
 136         mutex_unlock(&cpu_hotplug.lock);
 137 }
 138 /* Need to know about CPUs going up/down? */
 139 int __cpuinit register_cpu_notifier(struct notifier_block *nb)
 140 {
 141         int ret;
 142         cpu_maps_update_begin();
 143         ret = raw_notifier_chain_register(&cpu_chain, nb);
 144         cpu_maps_update_done();
 145         return ret;
 146 }
 147
 148 #ifdef CONFIG_HOTPLUG_CPU
 149
 150 EXPORT_SYMBOL(register_cpu_notifier);
 151
 152 void unregister_cpu_notifier(struct notifier_block *nb)
 153 {
 154         cpu_maps_update_begin();
 155         raw_notifier_chain_unregister(&cpu_chain, nb);
 156         cpu_maps_update_done();
 157 }
 158 EXPORT_SYMBOL(unregister_cpu_notifier);
 159
 160 static inline void check_for_tasks(int cpu)
 161 {
 162         struct task_struct *p;
 163
 164         write_lock_irq(&tasklist_lock);
 165         for_each_process(p) {
 166                 if (task_cpu(p) == cpu &&
 167                     (!cputime_eq(p->utime, cputime_zero) ||
 168                      !cputime_eq(p->stime, cputime_zero)))
 169                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
 170                                 (state = %ld, flags = %x) \n",
 171                                  p->comm, task_pid_nr(p), cpu,
 172                                  p->state, p->flags);
 173         }
 174         write_unlock_irq(&tasklist_lock);
 175 }
 176
 177 struct take_cpu_down_param {
 178         unsigned long mod;
 179         void *hcpu;
 180 };
 181
 182 /* Take this CPU down. */
 183 static int take_cpu_down(void *_param)
 184 {
 185         struct take_cpu_down_param *param = _param;
 186         int err;
 187
 188         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
 189                                 param->hcpu);
 190         /* Ensure this CPU doesn't handle any more interrupts. */
 191         err = __cpu_disable();
 192         if (err < 0)
 193                 return err;
 194
 195         /* Force idle task to run as soon as we yield: it should
 196            immediately notice cpu is offline and die quickly. */
 197         sched_idle_next();
 198         return 0;
 199 }
 200
 201 /* Requires cpu_add_remove_lock to be held */
 202 static int _cpu_down(unsigned int cpu, int tasks_frozen)
 203 {
 204         int err, nr_calls = 0;
 205         struct task_struct *p;
 206         cpumask_t old_allowed, tmp;
 207         void *hcpu = (void *)(long)cpu;
 208         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 209         struct take_cpu_down_param tcd_param = {
 210                 .mod = mod,
 211                 .hcpu = hcpu,
 212         };
 213
 214         if (num_online_cpus() == 1)
 215                 return -EBUSY;
 216
 217         if (!cpu_online(cpu))
 218                 return -EINVAL;
 219
 220         cpu_hotplug_begin();
 221         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
 222                                         hcpu, -1, &nr_calls);
 223         if (err == NOTIFY_BAD) {
 224                 nr_calls--;
 225                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 226                                           hcpu, nr_calls, NULL);
 227                 printk("%s: attempt to take down CPU %u failed\n",
 228                                 __FUNCTION__, cpu);
 229                 err = -EINVAL;
 230                 goto out_release;
 231         }
 232
 233         /* Ensure that we are not runnable on dying cpu */
 234         old_allowed = current->cpus_allowed;
 235         tmp = CPU_MASK_ALL;
 236         cpu_clear(cpu, tmp);
 237         set_cpus_allowed(current, tmp);
 238
 239         p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
 240
 241         if (IS_ERR(p) || cpu_online(cpu)) {
 242                 /* CPU didn't die: tell everyone.  Can't complain. */
 243                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 244                                             hcpu) == NOTIFY_BAD)
 245                         BUG();
 246
 247                 if (IS_ERR(p)) {
 248                         err = PTR_ERR(p);
 249                         goto out_allowed;
 250                 }
 251                 goto out_thread;
 252         }
 253
 254         /* Wait for it to sleep (leaving idle task). */
 255         while (!idle_cpu(cpu))
 256                 yield();
 257
 258         /* This actually kills the CPU. */
 259         __cpu_die(cpu);
 260
 261         /* CPU is completely dead: tell everyone.  Too late to complain. */
 262         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
 263                                     hcpu) == NOTIFY_BAD)
 264                 BUG();
 265
 266         check_for_tasks(cpu);
 267
 268 out_thread:
 269         err = kthread_stop(p);
 270 out_allowed:
 271         set_cpus_allowed(current, old_allowed);
 272 out_release:
 273         cpu_hotplug_done();
 274         return err;
 275 }
 276
 277 int cpu_down(unsigned int cpu)
 278 {
 279         int err = 0;
 280
 281         cpu_maps_update_begin();
 282         if (cpu_hotplug_disabled)
 283                 err = -EBUSY;
 284         else
 285                 err = _cpu_down(cpu, 0);
 286
 287         cpu_maps_update_done();
 288         return err;
 289 }
 290 #endif /*CONFIG_HOTPLUG_CPU*/
 291
 292 /* Requires cpu_add_remove_lock to be held */
 293 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 294 {
 295         int ret, nr_calls = 0;
 296         void *hcpu = (void *)(long)cpu;
 297         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 298
 299         if (cpu_online(cpu) || !cpu_present(cpu))
 300                 return -EINVAL;
 301
 302         cpu_hotplug_begin();
 303         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
 304                                                         -1, &nr_calls);
 305         if (ret == NOTIFY_BAD) {
 306                 nr_calls--;
 307                 printk("%s: attempt to bring up CPU %u failed\n",
 308                                 __FUNCTION__, cpu);
 309                 ret = -EINVAL;
 310                 goto out_notify;
 311         }
 312
 313         /* Arch-specific enabling code. */
 314         ret = __cpu_up(cpu);
 315         if (ret != 0)
 316                 goto out_notify;
 317         BUG_ON(!cpu_online(cpu));
 318
 319         /* Now call notifier in preparation. */
 320         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 321
 322 out_notify:
 323         if (ret != 0)
 324                 __raw_notifier_call_chain(&cpu_chain,
 325                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 326         cpu_hotplug_done();
 327
 328         return ret;
 329 }
 330
 331 int __cpuinit cpu_up(unsigned int cpu)
 332 {
 333         int err = 0;
 334         if (!cpu_isset(cpu, cpu_possible_map)) {
 335                 printk(KERN_ERR "can't online cpu %d because it is not "
 336                         "configured as may-hotadd at boot time\n", cpu);
 337 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64) || defined(CONFIG_S390)
 338                 printk(KERN_ERR "please check additional_cpus= boot "
 339                                 "parameter\n");
 340 #endif
 341                 return -EINVAL;
 342         }
 343
 344         cpu_maps_update_begin();
 345         if (cpu_hotplug_disabled)
 346                 err = -EBUSY;
 347         else
 348                 err = _cpu_up(cpu, 0);
 349
 350         cpu_maps_update_done();
 351         return err;
 352 }
 353
 354 #ifdef CONFIG_PM_SLEEP_SMP
 355 static cpumask_t frozen_cpus;
 356
 357 int disable_nonboot_cpus(void)
 358 {
 359         int cpu, first_cpu, error = 0;
 360
 361         cpu_maps_update_begin();
 362         first_cpu = first_cpu(cpu_online_map);
 363         /* We take down all of the non-boot CPUs in one shot to avoid races
 364          * with the userspace trying to use the CPU hotplug at the same time
 365          */
 366         cpus_clear(frozen_cpus);
 367         printk("Disabling non-boot CPUs ...\n");
 368         for_each_online_cpu(cpu) {
 369                 if (cpu == first_cpu)
 370                         continue;
 371                 error = _cpu_down(cpu, 1);
 372                 if (!error) {
 373                         cpu_set(cpu, frozen_cpus);
 374                         printk("CPU%d is down\n", cpu);
 375                 } else {
 376                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 377                                 cpu, error);
 378                         break;
 379                 }
 380         }
 381         if (!error) {
 382                 BUG_ON(num_online_cpus() > 1);
 383                 /* Make sure the CPUs won't be enabled by someone else */
 384                 cpu_hotplug_disabled = 1;
 385         } else {
 386                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 387         }
 388         cpu_maps_update_done();
 389         return error;
 390 }
 391
 392 void enable_nonboot_cpus(void)
 393 {
 394         int cpu, error;
 395
 396         /* Allow everyone to use the CPU hotplug again */
 397         cpu_maps_update_begin();
 398         cpu_hotplug_disabled = 0;
 399         if (cpus_empty(frozen_cpus))
 400                 goto out;
 401
 402         printk("Enabling non-boot CPUs ...\n");
 403         for_each_cpu_mask(cpu, frozen_cpus) {
 404                 error = _cpu_up(cpu, 1);
 405                 if (!error) {
 406                         printk("CPU%d is up\n", cpu);
 407                         continue;
 408                 }
 409                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 410         }
 411         cpus_clear(frozen_cpus);
 412 out:
 413         cpu_maps_update_done();
 414 }
 415 #endif /* CONFIG_PM_SLEEP_SMP */