cpu.c 13 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562
  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. #ifdef CONFIG_SMP
  18. /* Serializes the updates to cpu_online_mask, cpu_present_mask */
  19. static DEFINE_MUTEX(cpu_add_remove_lock);
  20. static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
  21. /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  22. * Should always be manipulated under cpu_add_remove_lock
  23. */
  24. static int cpu_hotplug_disabled;
  25. static struct {
  26. struct task_struct *active_writer;
  27. struct mutex lock; /* Synchronizes accesses to refcount, */
  28. /*
  29. * Also blocks the new readers during
  30. * an ongoing cpu hotplug operation.
  31. */
  32. int refcount;
  33. } cpu_hotplug = {
  34. .active_writer = NULL,
  35. .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
  36. .refcount = 0,
  37. };
  38. #ifdef CONFIG_HOTPLUG_CPU
  39. void get_online_cpus(void)
  40. {
  41. might_sleep();
  42. if (cpu_hotplug.active_writer == current)
  43. return;
  44. mutex_lock(&cpu_hotplug.lock);
  45. cpu_hotplug.refcount++;
  46. mutex_unlock(&cpu_hotplug.lock);
  47. }
  48. EXPORT_SYMBOL_GPL(get_online_cpus);
  49. void put_online_cpus(void)
  50. {
  51. if (cpu_hotplug.active_writer == current)
  52. return;
  53. mutex_lock(&cpu_hotplug.lock);
  54. if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
  55. wake_up_process(cpu_hotplug.active_writer);
  56. mutex_unlock(&cpu_hotplug.lock);
  57. }
  58. EXPORT_SYMBOL_GPL(put_online_cpus);
  59. #endif /* CONFIG_HOTPLUG_CPU */
  60. /*
  61. * The following two API's must be used when attempting
  62. * to serialize the updates to cpu_online_mask, cpu_present_mask.
  63. */
  64. void cpu_maps_update_begin(void)
  65. {
  66. mutex_lock(&cpu_add_remove_lock);
  67. }
  68. void cpu_maps_update_done(void)
  69. {
  70. mutex_unlock(&cpu_add_remove_lock);
  71. }
  72. /*
  73. * This ensures that the hotplug operation can begin only when the
  74. * refcount goes to zero.
  75. *
  76. * Note that during a cpu-hotplug operation, the new readers, if any,
  77. * will be blocked by the cpu_hotplug.lock
  78. *
  79. * Since cpu_hotplug_begin() is always called after invoking
  80. * cpu_maps_update_begin(), we can be sure that only one writer is active.
  81. *
  82. * Note that theoretically, there is a possibility of a livelock:
  83. * - Refcount goes to zero, last reader wakes up the sleeping
  84. * writer.
  85. * - Last reader unlocks the cpu_hotplug.lock.
  86. * - A new reader arrives at this moment, bumps up the refcount.
  87. * - The writer acquires the cpu_hotplug.lock finds the refcount
  88. * non zero and goes to sleep again.
  89. *
  90. * However, this is very difficult to achieve in practice since
  91. * get_online_cpus() not an api which is called all that often.
  92. *
  93. */
  94. static void cpu_hotplug_begin(void)
  95. {
  96. cpu_hotplug.active_writer = current;
  97. for (;;) {
  98. mutex_lock(&cpu_hotplug.lock);
  99. if (likely(!cpu_hotplug.refcount))
  100. break;
  101. __set_current_state(TASK_UNINTERRUPTIBLE);
  102. mutex_unlock(&cpu_hotplug.lock);
  103. schedule();
  104. }
  105. }
  106. static void cpu_hotplug_done(void)
  107. {
  108. cpu_hotplug.active_writer = NULL;
  109. mutex_unlock(&cpu_hotplug.lock);
  110. }
  111. /* Need to know about CPUs going up/down? */
  112. int __ref register_cpu_notifier(struct notifier_block *nb)
  113. {
  114. int ret;
  115. cpu_maps_update_begin();
  116. ret = raw_notifier_chain_register(&cpu_chain, nb);
  117. cpu_maps_update_done();
  118. return ret;
  119. }
  120. #ifdef CONFIG_HOTPLUG_CPU
  121. EXPORT_SYMBOL(register_cpu_notifier);
  122. void __ref unregister_cpu_notifier(struct notifier_block *nb)
  123. {
  124. cpu_maps_update_begin();
  125. raw_notifier_chain_unregister(&cpu_chain, nb);
  126. cpu_maps_update_done();
  127. }
  128. EXPORT_SYMBOL(unregister_cpu_notifier);
  129. static inline void check_for_tasks(int cpu)
  130. {
  131. struct task_struct *p;
  132. write_lock_irq(&tasklist_lock);
  133. for_each_process(p) {
  134. if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
  135. (!cputime_eq(p->utime, cputime_zero) ||
  136. !cputime_eq(p->stime, cputime_zero)))
  137. printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
  138. "(state = %ld, flags = %x)\n",
  139. p->comm, task_pid_nr(p), cpu,
  140. p->state, p->flags);
  141. }
  142. write_unlock_irq(&tasklist_lock);
  143. }
  144. struct take_cpu_down_param {
  145. struct task_struct *caller;
  146. unsigned long mod;
  147. void *hcpu;
  148. };
  149. /* Take this CPU down. */
  150. static int __ref take_cpu_down(void *_param)
  151. {
  152. struct take_cpu_down_param *param = _param;
  153. unsigned int cpu = (unsigned long)param->hcpu;
  154. int err;
  155. /* Ensure this CPU doesn't handle any more interrupts. */
  156. err = __cpu_disable();
  157. if (err < 0)
  158. return err;
  159. raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
  160. param->hcpu);
  161. if (task_cpu(param->caller) == cpu)
  162. move_task_off_dead_cpu(cpu, param->caller);
  163. /* Force idle task to run as soon as we yield: it should
  164. immediately notice cpu is offline and die quickly. */
  165. sched_idle_next();
  166. return 0;
  167. }
  168. /* Requires cpu_add_remove_lock to be held */
  169. static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
  170. {
  171. int err, nr_calls = 0;
  172. void *hcpu = (void *)(long)cpu;
  173. unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
  174. struct take_cpu_down_param tcd_param = {
  175. .caller = current,
  176. .mod = mod,
  177. .hcpu = hcpu,
  178. };
  179. if (num_online_cpus() == 1)
  180. return -EBUSY;
  181. if (!cpu_online(cpu))
  182. return -EINVAL;
  183. cpu_hotplug_begin();
  184. set_cpu_active(cpu, false);
  185. err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
  186. hcpu, -1, &nr_calls);
  187. if (err == NOTIFY_BAD) {
  188. set_cpu_active(cpu, true);
  189. nr_calls--;
  190. __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
  191. hcpu, nr_calls, NULL);
  192. printk("%s: attempt to take down CPU %u failed\n",
  193. __func__, cpu);
  194. err = -EINVAL;
  195. goto out_release;
  196. }
  197. err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
  198. if (err) {
  199. set_cpu_active(cpu, true);
  200. /* CPU didn't die: tell everyone. Can't complain. */
  201. if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
  202. hcpu) == NOTIFY_BAD)
  203. BUG();
  204. goto out_release;
  205. }
  206. BUG_ON(cpu_online(cpu));
  207. /* Wait for it to sleep (leaving idle task). */
  208. while (!idle_cpu(cpu))
  209. yield();
  210. /* This actually kills the CPU. */
  211. __cpu_die(cpu);
  212. /* CPU is completely dead: tell everyone. Too late to complain. */
  213. if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
  214. hcpu) == NOTIFY_BAD)
  215. BUG();
  216. check_for_tasks(cpu);
  217. out_release:
  218. cpu_hotplug_done();
  219. if (!err) {
  220. if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
  221. hcpu) == NOTIFY_BAD)
  222. BUG();
  223. }
  224. return err;
  225. }
  226. int __ref cpu_down(unsigned int cpu)
  227. {
  228. int err;
  229. err = stop_machine_create();
  230. if (err)
  231. return err;
  232. cpu_maps_update_begin();
  233. if (cpu_hotplug_disabled) {
  234. err = -EBUSY;
  235. goto out;
  236. }
  237. err = _cpu_down(cpu, 0);
  238. out:
  239. cpu_maps_update_done();
  240. stop_machine_destroy();
  241. return err;
  242. }
  243. EXPORT_SYMBOL(cpu_down);
  244. #endif /*CONFIG_HOTPLUG_CPU*/
  245. /* Requires cpu_add_remove_lock to be held */
  246. static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
  247. {
  248. int ret, nr_calls = 0;
  249. void *hcpu = (void *)(long)cpu;
  250. unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
  251. if (cpu_online(cpu) || !cpu_present(cpu))
  252. return -EINVAL;
  253. cpu_hotplug_begin();
  254. ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
  255. -1, &nr_calls);
  256. if (ret == NOTIFY_BAD) {
  257. nr_calls--;
  258. printk("%s: attempt to bring up CPU %u failed\n",
  259. __func__, cpu);
  260. ret = -EINVAL;
  261. goto out_notify;
  262. }
  263. /* Arch-specific enabling code. */
  264. ret = __cpu_up(cpu);
  265. if (ret != 0)
  266. goto out_notify;
  267. BUG_ON(!cpu_online(cpu));
  268. set_cpu_active(cpu, true);
  269. /* Now call notifier in preparation. */
  270. raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
  271. out_notify:
  272. if (ret != 0)
  273. __raw_notifier_call_chain(&cpu_chain,
  274. CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
  275. cpu_hotplug_done();
  276. return ret;
  277. }
  278. int __cpuinit cpu_up(unsigned int cpu)
  279. {
  280. int err = 0;
  281. if (!cpu_possible(cpu)) {
  282. printk(KERN_ERR "can't online cpu %d because it is not "
  283. "configured as may-hotadd at boot time\n", cpu);
  284. #if defined(CONFIG_IA64)
  285. printk(KERN_ERR "please check additional_cpus= boot "
  286. "parameter\n");
  287. #endif
  288. return -EINVAL;
  289. }
  290. cpu_maps_update_begin();
  291. if (cpu_hotplug_disabled) {
  292. err = -EBUSY;
  293. goto out;
  294. }
  295. err = _cpu_up(cpu, 0);
  296. out:
  297. cpu_maps_update_done();
  298. return err;
  299. }
  300. #ifdef CONFIG_PM_SLEEP_SMP
  301. static cpumask_var_t frozen_cpus;
  302. int disable_nonboot_cpus(void)
  303. {
  304. int cpu, first_cpu, error;
  305. error = stop_machine_create();
  306. if (error)
  307. return error;
  308. cpu_maps_update_begin();
  309. first_cpu = cpumask_first(cpu_online_mask);
  310. /*
  311. * We take down all of the non-boot CPUs in one shot to avoid races
  312. * with the userspace trying to use the CPU hotplug at the same time
  313. */
  314. cpumask_clear(frozen_cpus);
  315. printk("Disabling non-boot CPUs ...\n");
  316. for_each_online_cpu(cpu) {
  317. if (cpu == first_cpu)
  318. continue;
  319. error = _cpu_down(cpu, 1);
  320. if (!error)
  321. cpumask_set_cpu(cpu, frozen_cpus);
  322. else {
  323. printk(KERN_ERR "Error taking CPU%d down: %d\n",
  324. cpu, error);
  325. break;
  326. }
  327. }
  328. if (!error) {
  329. BUG_ON(num_online_cpus() > 1);
  330. /* Make sure the CPUs won't be enabled by someone else */
  331. cpu_hotplug_disabled = 1;
  332. } else {
  333. printk(KERN_ERR "Non-boot CPUs are not disabled\n");
  334. }
  335. cpu_maps_update_done();
  336. stop_machine_destroy();
  337. return error;
  338. }
  339. void __weak arch_enable_nonboot_cpus_begin(void)
  340. {
  341. }
  342. void __weak arch_enable_nonboot_cpus_end(void)
  343. {
  344. }
  345. void __ref enable_nonboot_cpus(void)
  346. {
  347. int cpu, error;
  348. /* Allow everyone to use the CPU hotplug again */
  349. cpu_maps_update_begin();
  350. cpu_hotplug_disabled = 0;
  351. if (cpumask_empty(frozen_cpus))
  352. goto out;
  353. printk("Enabling non-boot CPUs ...\n");
  354. arch_enable_nonboot_cpus_begin();
  355. for_each_cpu(cpu, frozen_cpus) {
  356. error = _cpu_up(cpu, 1);
  357. if (!error) {
  358. printk("CPU%d is up\n", cpu);
  359. continue;
  360. }
  361. printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
  362. }
  363. arch_enable_nonboot_cpus_end();
  364. cpumask_clear(frozen_cpus);
  365. out:
  366. cpu_maps_update_done();
  367. }
  368. static int alloc_frozen_cpus(void)
  369. {
  370. if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
  371. return -ENOMEM;
  372. return 0;
  373. }
  374. core_initcall(alloc_frozen_cpus);
  375. #endif /* CONFIG_PM_SLEEP_SMP */
  376. /**
  377. * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
  378. * @cpu: cpu that just started
  379. *
  380. * This function calls the cpu_chain notifiers with CPU_STARTING.
  381. * It must be called by the arch code on the new cpu, before the new cpu
  382. * enables interrupts and before the "boot" cpu returns from __cpu_up().
  383. */
  384. void __cpuinit notify_cpu_starting(unsigned int cpu)
  385. {
  386. unsigned long val = CPU_STARTING;
  387. #ifdef CONFIG_PM_SLEEP_SMP
  388. if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
  389. val = CPU_STARTING_FROZEN;
  390. #endif /* CONFIG_PM_SLEEP_SMP */
  391. raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
  392. }
  393. #endif /* CONFIG_SMP */
  394. /*
  395. * cpu_bit_bitmap[] is a special, "compressed" data structure that
  396. * represents all NR_CPUS bits binary values of 1<<nr.
  397. *
  398. * It is used by cpumask_of() to get a constant address to a CPU
  399. * mask value that has a single bit set only.
  400. */
  401. /* cpu_bit_bitmap[0] is empty - so we can back into it */
  402. #define MASK_DECLARE_1(x) [x+1][0] = 1UL << (x)
  403. #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
  404. #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
  405. #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
  406. const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
  407. MASK_DECLARE_8(0), MASK_DECLARE_8(8),
  408. MASK_DECLARE_8(16), MASK_DECLARE_8(24),
  409. #if BITS_PER_LONG > 32
  410. MASK_DECLARE_8(32), MASK_DECLARE_8(40),
  411. MASK_DECLARE_8(48), MASK_DECLARE_8(56),
  412. #endif
  413. };
  414. EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
  415. const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
  416. EXPORT_SYMBOL(cpu_all_bits);
  417. #ifdef CONFIG_INIT_ALL_POSSIBLE
  418. static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
  419. = CPU_BITS_ALL;
  420. #else
  421. static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
  422. #endif
  423. const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
  424. EXPORT_SYMBOL(cpu_possible_mask);
  425. static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
  426. const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
  427. EXPORT_SYMBOL(cpu_online_mask);
  428. static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
  429. const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
  430. EXPORT_SYMBOL(cpu_present_mask);
  431. static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
  432. const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
  433. EXPORT_SYMBOL(cpu_active_mask);
  434. void set_cpu_possible(unsigned int cpu, bool possible)
  435. {
  436. if (possible)
  437. cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
  438. else
  439. cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
  440. }
  441. void set_cpu_present(unsigned int cpu, bool present)
  442. {
  443. if (present)
  444. cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
  445. else
  446. cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
  447. }
  448. void set_cpu_online(unsigned int cpu, bool online)
  449. {
  450. if (online)
  451. cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
  452. else
  453. cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
  454. }
  455. void set_cpu_active(unsigned int cpu, bool active)
  456. {
  457. if (active)
  458. cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
  459. else
  460. cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
  461. }
  462. void init_cpu_present(const struct cpumask *src)
  463. {
  464. cpumask_copy(to_cpumask(cpu_present_bits), src);
  465. }
  466. void init_cpu_possible(const struct cpumask *src)
  467. {
  468. cpumask_copy(to_cpumask(cpu_possible_bits), src);
  469. }
  470. void init_cpu_online(const struct cpumask *src)
  471. {
  472. cpumask_copy(to_cpumask(cpu_online_bits), src);
  473. }