oom_kill.c 16 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627
  1. /*
  2. * linux/mm/oom_kill.c
  3. *
  4. * Copyright (C) 1998,2000 Rik van Riel
  5. * Thanks go out to Claus Fischer for some serious inspiration and
  6. * for goading me into coding this file...
  7. *
  8. * The routines in this file are used to kill a process when
  9. * we're seriously out of memory. This gets called from __alloc_pages()
  10. * in mm/page_alloc.c when we really run out of memory.
  11. *
  12. * Since we won't call these routines often (on a well-configured
  13. * machine) this file will double as a 'coding guide' and a signpost
  14. * for newbie kernel hackers. It features several pointers to major
  15. * kernel subsystems and hints as to where to find out what things do.
  16. */
  17. #include <linux/oom.h>
  18. #include <linux/mm.h>
  19. #include <linux/err.h>
  20. #include <linux/sched.h>
  21. #include <linux/swap.h>
  22. #include <linux/timex.h>
  23. #include <linux/jiffies.h>
  24. #include <linux/cpuset.h>
  25. #include <linux/module.h>
  26. #include <linux/notifier.h>
  27. #include <linux/memcontrol.h>
  28. #include <linux/security.h>
  29. int sysctl_panic_on_oom;
  30. int sysctl_oom_kill_allocating_task;
  31. int sysctl_oom_dump_tasks;
  32. static DEFINE_SPINLOCK(zone_scan_lock);
  33. /* #define DEBUG */
  34. /**
  35. * badness - calculate a numeric value for how bad this task has been
  36. * @p: task struct of which task we should calculate
  37. * @uptime: current uptime in seconds
  38. *
  39. * The formula used is relatively simple and documented inline in the
  40. * function. The main rationale is that we want to select a good task
  41. * to kill when we run out of memory.
  42. *
  43. * Good in this context means that:
  44. * 1) we lose the minimum amount of work done
  45. * 2) we recover a large amount of memory
  46. * 3) we don't kill anything innocent of eating tons of memory
  47. * 4) we want to kill the minimum amount of processes (one)
  48. * 5) we try to kill the process the user expects us to kill, this
  49. * algorithm has been meticulously tuned to meet the principle
  50. * of least surprise ... (be careful when you change it)
  51. */
  52. unsigned long badness(struct task_struct *p, unsigned long uptime)
  53. {
  54. unsigned long points, cpu_time, run_time;
  55. struct mm_struct *mm;
  56. struct task_struct *child;
  57. int oom_adj;
  58. task_lock(p);
  59. mm = p->mm;
  60. if (!mm) {
  61. task_unlock(p);
  62. return 0;
  63. }
  64. oom_adj = mm->oom_adj;
  65. if (oom_adj == OOM_DISABLE) {
  66. task_unlock(p);
  67. return 0;
  68. }
  69. /*
  70. * The memory size of the process is the basis for the badness.
  71. */
  72. points = mm->total_vm;
  73. /*
  74. * After this unlock we can no longer dereference local variable `mm'
  75. */
  76. task_unlock(p);
  77. /*
  78. * swapoff can easily use up all memory, so kill those first.
  79. */
  80. if (p->flags & PF_SWAPOFF)
  81. return ULONG_MAX;
  82. /*
  83. * Processes which fork a lot of child processes are likely
  84. * a good choice. We add half the vmsize of the children if they
  85. * have an own mm. This prevents forking servers to flood the
  86. * machine with an endless amount of children. In case a single
  87. * child is eating the vast majority of memory, adding only half
  88. * to the parents will make the child our kill candidate of choice.
  89. */
  90. list_for_each_entry(child, &p->children, sibling) {
  91. task_lock(child);
  92. if (child->mm != mm && child->mm)
  93. points += child->mm->total_vm/2 + 1;
  94. task_unlock(child);
  95. }
  96. /*
  97. * CPU time is in tens of seconds and run time is in thousands
  98. * of seconds. There is no particular reason for this other than
  99. * that it turned out to work very well in practice.
  100. */
  101. cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
  102. >> (SHIFT_HZ + 3);
  103. if (uptime >= p->start_time.tv_sec)
  104. run_time = (uptime - p->start_time.tv_sec) >> 10;
  105. else
  106. run_time = 0;
  107. if (cpu_time)
  108. points /= int_sqrt(cpu_time);
  109. if (run_time)
  110. points /= int_sqrt(int_sqrt(run_time));
  111. /*
  112. * Niced processes are most likely less important, so double
  113. * their badness points.
  114. */
  115. if (task_nice(p) > 0)
  116. points *= 2;
  117. /*
  118. * Superuser processes are usually more important, so we make it
  119. * less likely that we kill those.
  120. */
  121. if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
  122. has_capability_noaudit(p, CAP_SYS_RESOURCE))
  123. points /= 4;
  124. /*
  125. * We don't want to kill a process with direct hardware access.
  126. * Not only could that mess up the hardware, but usually users
  127. * tend to only have this flag set on applications they think
  128. * of as important.
  129. */
  130. if (has_capability_noaudit(p, CAP_SYS_RAWIO))
  131. points /= 4;
  132. /*
  133. * If p's nodes don't overlap ours, it may still help to kill p
  134. * because p may have allocated or otherwise mapped memory on
  135. * this node before. However it will be less likely.
  136. */
  137. if (!cpuset_mems_allowed_intersects(current, p))
  138. points /= 8;
  139. /*
  140. * Adjust the score by oom_adj.
  141. */
  142. if (oom_adj) {
  143. if (oom_adj > 0) {
  144. if (!points)
  145. points = 1;
  146. points <<= oom_adj;
  147. } else
  148. points >>= -(oom_adj);
  149. }
  150. #ifdef DEBUG
  151. printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
  152. p->pid, p->comm, points);
  153. #endif
  154. return points;
  155. }
  156. /*
  157. * Determine the type of allocation constraint.
  158. */
  159. static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
  160. gfp_t gfp_mask)
  161. {
  162. #ifdef CONFIG_NUMA
  163. struct zone *zone;
  164. struct zoneref *z;
  165. enum zone_type high_zoneidx = gfp_zone(gfp_mask);
  166. nodemask_t nodes = node_states[N_HIGH_MEMORY];
  167. for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
  168. if (cpuset_zone_allowed_softwall(zone, gfp_mask))
  169. node_clear(zone_to_nid(zone), nodes);
  170. else
  171. return CONSTRAINT_CPUSET;
  172. if (!nodes_empty(nodes))
  173. return CONSTRAINT_MEMORY_POLICY;
  174. #endif
  175. return CONSTRAINT_NONE;
  176. }
  177. /*
  178. * Simple selection loop. We chose the process with the highest
  179. * number of 'points'. We expect the caller will lock the tasklist.
  180. *
  181. * (not docbooked, we don't want this one cluttering up the manual)
  182. */
  183. static struct task_struct *select_bad_process(unsigned long *ppoints,
  184. struct mem_cgroup *mem)
  185. {
  186. struct task_struct *g, *p;
  187. struct task_struct *chosen = NULL;
  188. struct timespec uptime;
  189. *ppoints = 0;
  190. do_posix_clock_monotonic_gettime(&uptime);
  191. do_each_thread(g, p) {
  192. unsigned long points;
  193. /*
  194. * skip kernel threads and tasks which have already released
  195. * their mm.
  196. */
  197. if (!p->mm)
  198. continue;
  199. /* skip the init task */
  200. if (is_global_init(p))
  201. continue;
  202. if (mem && !task_in_mem_cgroup(p, mem))
  203. continue;
  204. /*
  205. * This task already has access to memory reserves and is
  206. * being killed. Don't allow any other task access to the
  207. * memory reserve.
  208. *
  209. * Note: this may have a chance of deadlock if it gets
  210. * blocked waiting for another task which itself is waiting
  211. * for memory. Is there a better alternative?
  212. */
  213. if (test_tsk_thread_flag(p, TIF_MEMDIE))
  214. return ERR_PTR(-1UL);
  215. /*
  216. * This is in the process of releasing memory so wait for it
  217. * to finish before killing some other task by mistake.
  218. *
  219. * However, if p is the current task, we allow the 'kill' to
  220. * go ahead if it is exiting: this will simply set TIF_MEMDIE,
  221. * which will allow it to gain access to memory reserves in
  222. * the process of exiting and releasing its resources.
  223. * Otherwise we could get an easy OOM deadlock.
  224. */
  225. if (p->flags & PF_EXITING) {
  226. if (p != current)
  227. return ERR_PTR(-1UL);
  228. chosen = p;
  229. *ppoints = ULONG_MAX;
  230. }
  231. points = badness(p, uptime.tv_sec);
  232. if (points > *ppoints) {
  233. chosen = p;
  234. *ppoints = points;
  235. }
  236. } while_each_thread(g, p);
  237. return chosen;
  238. }
  239. /**
  240. * dump_tasks - dump current memory state of all system tasks
  241. * @mem: target memory controller
  242. *
  243. * Dumps the current memory state of all system tasks, excluding kernel threads.
  244. * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
  245. * score, and name.
  246. *
  247. * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
  248. * shown.
  249. *
  250. * Call with tasklist_lock read-locked.
  251. */
  252. static void dump_tasks(const struct mem_cgroup *mem)
  253. {
  254. struct task_struct *g, *p;
  255. printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
  256. "name\n");
  257. do_each_thread(g, p) {
  258. struct mm_struct *mm;
  259. if (mem && !task_in_mem_cgroup(p, mem))
  260. continue;
  261. if (!thread_group_leader(p))
  262. continue;
  263. task_lock(p);
  264. mm = p->mm;
  265. if (!mm) {
  266. /*
  267. * total_vm and rss sizes do not exist for tasks with no
  268. * mm so there's no need to report them; they can't be
  269. * oom killed anyway.
  270. */
  271. task_unlock(p);
  272. continue;
  273. }
  274. printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
  275. p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
  276. get_mm_rss(mm), (int)task_cpu(p), mm->oom_adj, p->comm);
  277. task_unlock(p);
  278. } while_each_thread(g, p);
  279. }
  280. /*
  281. * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
  282. * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
  283. * set.
  284. */
  285. static void __oom_kill_task(struct task_struct *p, int verbose)
  286. {
  287. if (is_global_init(p)) {
  288. WARN_ON(1);
  289. printk(KERN_WARNING "tried to kill init!\n");
  290. return;
  291. }
  292. if (!p->mm)
  293. return;
  294. if (verbose)
  295. printk(KERN_ERR "Killed process %d (%s)\n",
  296. task_pid_nr(p), p->comm);
  297. /*
  298. * We give our sacrificial lamb high priority and access to
  299. * all the memory it needs. That way it should be able to
  300. * exit() and clear out its resources quickly...
  301. */
  302. p->rt.time_slice = HZ;
  303. set_tsk_thread_flag(p, TIF_MEMDIE);
  304. force_sig(SIGKILL, p);
  305. }
  306. static int oom_kill_task(struct task_struct *p)
  307. {
  308. struct mm_struct *mm;
  309. struct task_struct *g, *q;
  310. task_lock(p);
  311. mm = p->mm;
  312. if (!mm || mm->oom_adj == OOM_DISABLE) {
  313. task_unlock(p);
  314. return 1;
  315. }
  316. task_unlock(p);
  317. __oom_kill_task(p, 1);
  318. /*
  319. * kill all processes that share the ->mm (i.e. all threads),
  320. * but are in a different thread group. Don't let them have access
  321. * to memory reserves though, otherwise we might deplete all memory.
  322. */
  323. do_each_thread(g, q) {
  324. if (q->mm == mm && !same_thread_group(q, p))
  325. force_sig(SIGKILL, q);
  326. } while_each_thread(g, q);
  327. return 0;
  328. }
  329. static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
  330. unsigned long points, struct mem_cgroup *mem,
  331. const char *message)
  332. {
  333. struct task_struct *c;
  334. if (printk_ratelimit()) {
  335. task_lock(current);
  336. printk(KERN_WARNING "%s invoked oom-killer: "
  337. "gfp_mask=0x%x, order=%d, oom_adj=%d\n",
  338. current->comm, gfp_mask, order,
  339. current->mm ? current->mm->oom_adj : OOM_DISABLE);
  340. cpuset_print_task_mems_allowed(current);
  341. task_unlock(current);
  342. dump_stack();
  343. mem_cgroup_print_oom_info(mem, current);
  344. show_mem();
  345. if (sysctl_oom_dump_tasks)
  346. dump_tasks(mem);
  347. }
  348. /*
  349. * If the task is already exiting, don't alarm the sysadmin or kill
  350. * its children or threads, just set TIF_MEMDIE so it can die quickly
  351. * if its mm is still attached.
  352. */
  353. if (p->mm && (p->flags & PF_EXITING)) {
  354. __oom_kill_task(p, 0);
  355. return 0;
  356. }
  357. printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
  358. message, task_pid_nr(p), p->comm, points);
  359. /* Try to kill a child first */
  360. list_for_each_entry(c, &p->children, sibling) {
  361. if (c->mm == p->mm)
  362. continue;
  363. if (!oom_kill_task(c))
  364. return 0;
  365. }
  366. return oom_kill_task(p);
  367. }
  368. #ifdef CONFIG_CGROUP_MEM_RES_CTLR
  369. void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
  370. {
  371. unsigned long points = 0;
  372. struct task_struct *p;
  373. read_lock(&tasklist_lock);
  374. retry:
  375. p = select_bad_process(&points, mem);
  376. if (PTR_ERR(p) == -1UL)
  377. goto out;
  378. if (!p)
  379. p = current;
  380. if (oom_kill_process(p, gfp_mask, 0, points, mem,
  381. "Memory cgroup out of memory"))
  382. goto retry;
  383. out:
  384. read_unlock(&tasklist_lock);
  385. }
  386. #endif
  387. static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
  388. int register_oom_notifier(struct notifier_block *nb)
  389. {
  390. return blocking_notifier_chain_register(&oom_notify_list, nb);
  391. }
  392. EXPORT_SYMBOL_GPL(register_oom_notifier);
  393. int unregister_oom_notifier(struct notifier_block *nb)
  394. {
  395. return blocking_notifier_chain_unregister(&oom_notify_list, nb);
  396. }
  397. EXPORT_SYMBOL_GPL(unregister_oom_notifier);
  398. /*
  399. * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
  400. * if a parallel OOM killing is already taking place that includes a zone in
  401. * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
  402. */
  403. int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
  404. {
  405. struct zoneref *z;
  406. struct zone *zone;
  407. int ret = 1;
  408. spin_lock(&zone_scan_lock);
  409. for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
  410. if (zone_is_oom_locked(zone)) {
  411. ret = 0;
  412. goto out;
  413. }
  414. }
  415. for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
  416. /*
  417. * Lock each zone in the zonelist under zone_scan_lock so a
  418. * parallel invocation of try_set_zone_oom() doesn't succeed
  419. * when it shouldn't.
  420. */
  421. zone_set_flag(zone, ZONE_OOM_LOCKED);
  422. }
  423. out:
  424. spin_unlock(&zone_scan_lock);
  425. return ret;
  426. }
  427. /*
  428. * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
  429. * allocation attempts with zonelists containing them may now recall the OOM
  430. * killer, if necessary.
  431. */
  432. void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
  433. {
  434. struct zoneref *z;
  435. struct zone *zone;
  436. spin_lock(&zone_scan_lock);
  437. for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
  438. zone_clear_flag(zone, ZONE_OOM_LOCKED);
  439. }
  440. spin_unlock(&zone_scan_lock);
  441. }
  442. /*
  443. * Must be called with tasklist_lock held for read.
  444. */
  445. static void __out_of_memory(gfp_t gfp_mask, int order)
  446. {
  447. struct task_struct *p;
  448. unsigned long points;
  449. if (sysctl_oom_kill_allocating_task)
  450. if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
  451. "Out of memory (oom_kill_allocating_task)"))
  452. return;
  453. retry:
  454. /*
  455. * Rambo mode: Shoot down a process and hope it solves whatever
  456. * issues we may have.
  457. */
  458. p = select_bad_process(&points, NULL);
  459. if (PTR_ERR(p) == -1UL)
  460. return;
  461. /* Found nothing?!?! Either we hang forever, or we panic. */
  462. if (!p) {
  463. read_unlock(&tasklist_lock);
  464. panic("Out of memory and no killable processes...\n");
  465. }
  466. if (oom_kill_process(p, gfp_mask, order, points, NULL,
  467. "Out of memory"))
  468. goto retry;
  469. }
  470. /*
  471. * pagefault handler calls into here because it is out of memory but
  472. * doesn't know exactly how or why.
  473. */
  474. void pagefault_out_of_memory(void)
  475. {
  476. unsigned long freed = 0;
  477. blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
  478. if (freed > 0)
  479. /* Got some memory back in the last second. */
  480. return;
  481. /*
  482. * If this is from memcg, oom-killer is already invoked.
  483. * and not worth to go system-wide-oom.
  484. */
  485. if (mem_cgroup_oom_called(current))
  486. goto rest_and_return;
  487. if (sysctl_panic_on_oom)
  488. panic("out of memory from page fault. panic_on_oom is selected.\n");
  489. read_lock(&tasklist_lock);
  490. __out_of_memory(0, 0); /* unknown gfp_mask and order */
  491. read_unlock(&tasklist_lock);
  492. /*
  493. * Give "p" a good chance of killing itself before we
  494. * retry to allocate memory.
  495. */
  496. rest_and_return:
  497. if (!test_thread_flag(TIF_MEMDIE))
  498. schedule_timeout_uninterruptible(1);
  499. }
  500. /**
  501. * out_of_memory - kill the "best" process when we run out of memory
  502. * @zonelist: zonelist pointer
  503. * @gfp_mask: memory allocation flags
  504. * @order: amount of memory being requested as a power of 2
  505. *
  506. * If we run out of memory, we have the choice between either
  507. * killing a random task (bad), letting the system crash (worse)
  508. * OR try to be smart about which process to kill. Note that we
  509. * don't have to be perfect here, we just have to be good.
  510. */
  511. void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
  512. {
  513. unsigned long freed = 0;
  514. enum oom_constraint constraint;
  515. blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
  516. if (freed > 0)
  517. /* Got some memory back in the last second. */
  518. return;
  519. if (sysctl_panic_on_oom == 2)
  520. panic("out of memory. Compulsory panic_on_oom is selected.\n");
  521. /*
  522. * Check if there were limitations on the allocation (only relevant for
  523. * NUMA) that may require different handling.
  524. */
  525. constraint = constrained_alloc(zonelist, gfp_mask);
  526. read_lock(&tasklist_lock);
  527. switch (constraint) {
  528. case CONSTRAINT_MEMORY_POLICY:
  529. oom_kill_process(current, gfp_mask, order, 0, NULL,
  530. "No available memory (MPOL_BIND)");
  531. break;
  532. case CONSTRAINT_NONE:
  533. if (sysctl_panic_on_oom)
  534. panic("out of memory. panic_on_oom is selected\n");
  535. /* Fall-through */
  536. case CONSTRAINT_CPUSET:
  537. __out_of_memory(gfp_mask, order);
  538. break;
  539. }
  540. read_unlock(&tasklist_lock);
  541. /*
  542. * Give "p" a good chance of killing itself before we
  543. * retry to allocate memory unless "p" is current
  544. */
  545. if (!test_thread_flag(TIF_MEMDIE))
  546. schedule_timeout_uninterruptible(1);
  547. }