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 /* This protects CPUs going up and down... */
  19 static DEFINE_MUTEX(cpu_add_remove_lock);
  20 static DEFINE_MUTEX(cpu_bitmask_lock);
  21
  22 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
  23
  24 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  25  * Should always be manipulated under cpu_add_remove_lock
  26  */
  27 static int cpu_hotplug_disabled;
  28
  29 #ifdef CONFIG_HOTPLUG_CPU
  30
  31 /* Crappy recursive lock-takers in cpufreq! Complain loudly about idiots */
  32 static struct task_struct *recursive;
  33 static int recursive_depth;
  34
  35 void lock_cpu_hotplug(void)
  36 {
  37         struct task_struct *tsk = current;
  38
  39         if (tsk == recursive) {
  40                 static int warnings = 10;
  41                 if (warnings) {
  42                         printk(KERN_ERR "Lukewarm IQ detected in hotplug locking\n");
  43                         WARN_ON(1);
  44                         warnings--;
  45                 }
  46                 recursive_depth++;
  47                 return;
  48         }
  49         mutex_lock(&cpu_bitmask_lock);
  50         recursive = tsk;
  51 }
  52 EXPORT_SYMBOL_GPL(lock_cpu_hotplug);
  53
  54 void unlock_cpu_hotplug(void)
  55 {
  56         WARN_ON(recursive != current);
  57         if (recursive_depth) {
  58                 recursive_depth--;
  59                 return;
  60         }
  61         recursive = NULL;
  62         mutex_unlock(&cpu_bitmask_lock);
  63 }
  64 EXPORT_SYMBOL_GPL(unlock_cpu_hotplug);
  65
  66 #endif  /* CONFIG_HOTPLUG_CPU */
  67
  68 /* Need to know about CPUs going up/down? */
  69 int __cpuinit register_cpu_notifier(struct notifier_block *nb)
  70 {
  71         int ret;
  72         mutex_lock(&cpu_add_remove_lock);
  73         ret = raw_notifier_chain_register(&cpu_chain, nb);
  74         mutex_unlock(&cpu_add_remove_lock);
  75         return ret;
  76 }
  77
  78 #ifdef CONFIG_HOTPLUG_CPU
  79
  80 EXPORT_SYMBOL(register_cpu_notifier);
  81
  82 void unregister_cpu_notifier(struct notifier_block *nb)
  83 {
  84         mutex_lock(&cpu_add_remove_lock);
  85         raw_notifier_chain_unregister(&cpu_chain, nb);
  86         mutex_unlock(&cpu_add_remove_lock);
  87 }
  88 EXPORT_SYMBOL(unregister_cpu_notifier);
  89
  90 static inline void check_for_tasks(int cpu)
  91 {
  92         struct task_struct *p;
  93
  94         write_lock_irq(&tasklist_lock);
  95         for_each_process(p) {
  96                 if (task_cpu(p) == cpu &&
  97                     (!cputime_eq(p->utime, cputime_zero) ||
  98                      !cputime_eq(p->stime, cputime_zero)))
  99                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
 100                                 (state = %ld, flags = %x) \n",
 101                                  p->comm, task_pid_nr(p), cpu,
 102                                  p->state, p->flags);
 103         }
 104         write_unlock_irq(&tasklist_lock);
 105 }
 106
 107 struct take_cpu_down_param {
 108         unsigned long mod;
 109         void *hcpu;
 110 };
 111
 112 /* Take this CPU down. */
 113 static int take_cpu_down(void *_param)
 114 {
 115         struct take_cpu_down_param *param = _param;
 116         int err;
 117
 118         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
 119                                 param->hcpu);
 120         /* Ensure this CPU doesn't handle any more interrupts. */
 121         err = __cpu_disable();
 122         if (err < 0)
 123                 return err;
 124
 125         /* Force idle task to run as soon as we yield: it should
 126            immediately notice cpu is offline and die quickly. */
 127         sched_idle_next();
 128         return 0;
 129 }
 130
 131 /* Requires cpu_add_remove_lock to be held */
 132 static int _cpu_down(unsigned int cpu, int tasks_frozen)
 133 {
 134         int err, nr_calls = 0;
 135         struct task_struct *p;
 136         cpumask_t old_allowed, tmp;
 137         void *hcpu = (void *)(long)cpu;
 138         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 139         struct take_cpu_down_param tcd_param = {
 140                 .mod = mod,
 141                 .hcpu = hcpu,
 142         };
 143
 144         if (num_online_cpus() == 1)
 145                 return -EBUSY;
 146
 147         if (!cpu_online(cpu))
 148                 return -EINVAL;
 149
 150         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
 151         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
 152                                         hcpu, -1, &nr_calls);
 153         if (err == NOTIFY_BAD) {
 154                 nr_calls--;
 155                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 156                                           hcpu, nr_calls, NULL);
 157                 printk("%s: attempt to take down CPU %u failed\n",
 158                                 __FUNCTION__, cpu);
 159                 err = -EINVAL;
 160                 goto out_release;
 161         }
 162
 163         /* Ensure that we are not runnable on dying cpu */
 164         old_allowed = current->cpus_allowed;
 165         tmp = CPU_MASK_ALL;
 166         cpu_clear(cpu, tmp);
 167         set_cpus_allowed(current, tmp);
 168
 169         mutex_lock(&cpu_bitmask_lock);
 170         p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
 171         mutex_unlock(&cpu_bitmask_lock);
 172
 173         if (IS_ERR(p) || cpu_online(cpu)) {
 174                 /* CPU didn't die: tell everyone.  Can't complain. */
 175                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
 176                                             hcpu) == NOTIFY_BAD)
 177                         BUG();
 178
 179                 if (IS_ERR(p)) {
 180                         err = PTR_ERR(p);
 181                         goto out_allowed;
 182                 }
 183                 goto out_thread;
 184         }
 185
 186         /* Wait for it to sleep (leaving idle task). */
 187         while (!idle_cpu(cpu))
 188                 yield();
 189
 190         /* This actually kills the CPU. */
 191         __cpu_die(cpu);
 192
 193         /* CPU is completely dead: tell everyone.  Too late to complain. */
 194         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
 195                                     hcpu) == NOTIFY_BAD)
 196                 BUG();
 197
 198         check_for_tasks(cpu);
 199
 200 out_thread:
 201         err = kthread_stop(p);
 202 out_allowed:
 203         set_cpus_allowed(current, old_allowed);
 204 out_release:
 205         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
 206         return err;
 207 }
 208
 209 int cpu_down(unsigned int cpu)
 210 {
 211         int err = 0;
 212
 213         mutex_lock(&cpu_add_remove_lock);
 214         if (cpu_hotplug_disabled)
 215                 err = -EBUSY;
 216         else
 217                 err = _cpu_down(cpu, 0);
 218
 219         mutex_unlock(&cpu_add_remove_lock);
 220         return err;
 221 }
 222 #endif /*CONFIG_HOTPLUG_CPU*/
 223
 224 /* Requires cpu_add_remove_lock to be held */
 225 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 226 {
 227         int ret, nr_calls = 0;
 228         void *hcpu = (void *)(long)cpu;
 229         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 230
 231         if (cpu_online(cpu) || !cpu_present(cpu))
 232                 return -EINVAL;
 233
 234         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
 235         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
 236                                                         -1, &nr_calls);
 237         if (ret == NOTIFY_BAD) {
 238                 nr_calls--;
 239                 printk("%s: attempt to bring up CPU %u failed\n",
 240                                 __FUNCTION__, cpu);
 241                 ret = -EINVAL;
 242                 goto out_notify;
 243         }
 244
 245         /* Arch-specific enabling code. */
 246         mutex_lock(&cpu_bitmask_lock);
 247         ret = __cpu_up(cpu);
 248         mutex_unlock(&cpu_bitmask_lock);
 249         if (ret != 0)
 250                 goto out_notify;
 251         BUG_ON(!cpu_online(cpu));
 252
 253         /* Now call notifier in preparation. */
 254         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 255
 256 out_notify:
 257         if (ret != 0)
 258                 __raw_notifier_call_chain(&cpu_chain,
 259                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
 260         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
 261
 262         return ret;
 263 }
 264
 265 int __cpuinit cpu_up(unsigned int cpu)
 266 {
 267         int err = 0;
 268         if (!cpu_isset(cpu, cpu_possible_map)) {
 269                 printk(KERN_ERR "can't online cpu %d because it is not "
 270                         "configured as may-hotadd at boot time\n", cpu);
 271 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64) || defined(CONFIG_S390)
 272                 printk(KERN_ERR "please check additional_cpus= boot "
 273                                 "parameter\n");
 274 #endif
 275                 return -EINVAL;
 276         }
 277
 278         mutex_lock(&cpu_add_remove_lock);
 279         if (cpu_hotplug_disabled)
 280                 err = -EBUSY;
 281         else
 282                 err = _cpu_up(cpu, 0);
 283
 284         mutex_unlock(&cpu_add_remove_lock);
 285         return err;
 286 }
 287
 288 #ifdef CONFIG_PM_SLEEP_SMP
 289 static cpumask_t frozen_cpus;
 290
 291 int disable_nonboot_cpus(void)
 292 {
 293         int cpu, first_cpu, error = 0;
 294
 295         mutex_lock(&cpu_add_remove_lock);
 296         first_cpu = first_cpu(cpu_online_map);
 297         /* We take down all of the non-boot CPUs in one shot to avoid races
 298          * with the userspace trying to use the CPU hotplug at the same time
 299          */
 300         cpus_clear(frozen_cpus);
 301         printk("Disabling non-boot CPUs ...\n");
 302         for_each_online_cpu(cpu) {
 303                 if (cpu == first_cpu)
 304                         continue;
 305                 error = _cpu_down(cpu, 1);
 306                 if (!error) {
 307                         cpu_set(cpu, frozen_cpus);
 308                         printk("CPU%d is down\n", cpu);
 309                 } else {
 310                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
 311                                 cpu, error);
 312                         break;
 313                 }
 314         }
 315         if (!error) {
 316                 BUG_ON(num_online_cpus() > 1);
 317                 /* Make sure the CPUs won't be enabled by someone else */
 318                 cpu_hotplug_disabled = 1;
 319         } else {
 320                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 321         }
 322         mutex_unlock(&cpu_add_remove_lock);
 323         return error;
 324 }
 325
 326 void enable_nonboot_cpus(void)
 327 {
 328         int cpu, error;
 329
 330         /* Allow everyone to use the CPU hotplug again */
 331         mutex_lock(&cpu_add_remove_lock);
 332         cpu_hotplug_disabled = 0;
 333         if (cpus_empty(frozen_cpus))
 334                 goto out;
 335
 336         printk("Enabling non-boot CPUs ...\n");
 337         for_each_cpu_mask(cpu, frozen_cpus) {
 338                 error = _cpu_up(cpu, 1);
 339                 if (!error) {
 340                         printk("CPU%d is up\n", cpu);
 341                         continue;
 342                 }
 343                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
 344         }
 345         cpus_clear(frozen_cpus);
 346 out:
 347         mutex_unlock(&cpu_add_remove_lock);
 348 }
 349 #endif /* CONFIG_PM_SLEEP_SMP */