Loading...
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 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
9 *
10 * The routines in this file are used to kill a process when
11 * we're seriously out of memory. This gets called from __alloc_pages()
12 * in mm/page_alloc.c when we really run out of memory.
13 *
14 * Since we won't call these routines often (on a well-configured
15 * machine) this file will double as a 'coding guide' and a signpost
16 * for newbie kernel hackers. It features several pointers to major
17 * kernel subsystems and hints as to where to find out what things do.
18 */
19
20#include <linux/oom.h>
21#include <linux/mm.h>
22#include <linux/err.h>
23#include <linux/gfp.h>
24#include <linux/sched.h>
25#include <linux/swap.h>
26#include <linux/timex.h>
27#include <linux/jiffies.h>
28#include <linux/cpuset.h>
29#include <linux/export.h>
30#include <linux/notifier.h>
31#include <linux/memcontrol.h>
32#include <linux/mempolicy.h>
33#include <linux/security.h>
34#include <linux/ptrace.h>
35#include <linux/freezer.h>
36#include <linux/ftrace.h>
37#include <linux/ratelimit.h>
38
39#define CREATE_TRACE_POINTS
40#include <trace/events/oom.h>
41
42int sysctl_panic_on_oom;
43int sysctl_oom_kill_allocating_task;
44int sysctl_oom_dump_tasks = 1;
45static DEFINE_SPINLOCK(zone_scan_lock);
46
47#ifdef CONFIG_NUMA
48/**
49 * has_intersects_mems_allowed() - check task eligiblity for kill
50 * @start: task struct of which task to consider
51 * @mask: nodemask passed to page allocator for mempolicy ooms
52 *
53 * Task eligibility is determined by whether or not a candidate task, @tsk,
54 * shares the same mempolicy nodes as current if it is bound by such a policy
55 * and whether or not it has the same set of allowed cpuset nodes.
56 */
57static bool has_intersects_mems_allowed(struct task_struct *start,
58 const nodemask_t *mask)
59{
60 struct task_struct *tsk;
61 bool ret = false;
62
63 rcu_read_lock();
64 for_each_thread(start, tsk) {
65 if (mask) {
66 /*
67 * If this is a mempolicy constrained oom, tsk's
68 * cpuset is irrelevant. Only return true if its
69 * mempolicy intersects current, otherwise it may be
70 * needlessly killed.
71 */
72 ret = mempolicy_nodemask_intersects(tsk, mask);
73 } else {
74 /*
75 * This is not a mempolicy constrained oom, so only
76 * check the mems of tsk's cpuset.
77 */
78 ret = cpuset_mems_allowed_intersects(current, tsk);
79 }
80 if (ret)
81 break;
82 }
83 rcu_read_unlock();
84
85 return ret;
86}
87#else
88static bool has_intersects_mems_allowed(struct task_struct *tsk,
89 const nodemask_t *mask)
90{
91 return true;
92}
93#endif /* CONFIG_NUMA */
94
95/*
96 * The process p may have detached its own ->mm while exiting or through
97 * use_mm(), but one or more of its subthreads may still have a valid
98 * pointer. Return p, or any of its subthreads with a valid ->mm, with
99 * task_lock() held.
100 */
101struct task_struct *find_lock_task_mm(struct task_struct *p)
102{
103 struct task_struct *t;
104
105 rcu_read_lock();
106
107 for_each_thread(p, t) {
108 task_lock(t);
109 if (likely(t->mm))
110 goto found;
111 task_unlock(t);
112 }
113 t = NULL;
114found:
115 rcu_read_unlock();
116
117 return t;
118}
119
120/* return true if the task is not adequate as candidate victim task. */
121static bool oom_unkillable_task(struct task_struct *p,
122 const struct mem_cgroup *memcg, const nodemask_t *nodemask)
123{
124 if (is_global_init(p))
125 return true;
126 if (p->flags & PF_KTHREAD)
127 return true;
128
129 /* When mem_cgroup_out_of_memory() and p is not member of the group */
130 if (memcg && !task_in_mem_cgroup(p, memcg))
131 return true;
132
133 /* p may not have freeable memory in nodemask */
134 if (!has_intersects_mems_allowed(p, nodemask))
135 return true;
136
137 return false;
138}
139
140/**
141 * oom_badness - heuristic function to determine which candidate task to kill
142 * @p: task struct of which task we should calculate
143 * @totalpages: total present RAM allowed for page allocation
144 *
145 * The heuristic for determining which task to kill is made to be as simple and
146 * predictable as possible. The goal is to return the highest value for the
147 * task consuming the most memory to avoid subsequent oom failures.
148 */
149unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
150 const nodemask_t *nodemask, unsigned long totalpages)
151{
152 long points;
153 long adj;
154
155 if (oom_unkillable_task(p, memcg, nodemask))
156 return 0;
157
158 p = find_lock_task_mm(p);
159 if (!p)
160 return 0;
161
162 adj = (long)p->signal->oom_score_adj;
163 if (adj == OOM_SCORE_ADJ_MIN) {
164 task_unlock(p);
165 return 0;
166 }
167
168 /*
169 * The baseline for the badness score is the proportion of RAM that each
170 * task's rss, pagetable and swap space use.
171 */
172 points = get_mm_rss(p->mm) + atomic_long_read(&p->mm->nr_ptes) +
173 get_mm_counter(p->mm, MM_SWAPENTS);
174 task_unlock(p);
175
176 /*
177 * Root processes get 3% bonus, just like the __vm_enough_memory()
178 * implementation used by LSMs.
179 */
180 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
181 points -= (points * 3) / 100;
182
183 /* Normalize to oom_score_adj units */
184 adj *= totalpages / 1000;
185 points += adj;
186
187 /*
188 * Never return 0 for an eligible task regardless of the root bonus and
189 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
190 */
191 return points > 0 ? points : 1;
192}
193
194/*
195 * Determine the type of allocation constraint.
196 */
197#ifdef CONFIG_NUMA
198static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
199 gfp_t gfp_mask, nodemask_t *nodemask,
200 unsigned long *totalpages)
201{
202 struct zone *zone;
203 struct zoneref *z;
204 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
205 bool cpuset_limited = false;
206 int nid;
207
208 /* Default to all available memory */
209 *totalpages = totalram_pages + total_swap_pages;
210
211 if (!zonelist)
212 return CONSTRAINT_NONE;
213 /*
214 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
215 * to kill current.We have to random task kill in this case.
216 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
217 */
218 if (gfp_mask & __GFP_THISNODE)
219 return CONSTRAINT_NONE;
220
221 /*
222 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
223 * the page allocator means a mempolicy is in effect. Cpuset policy
224 * is enforced in get_page_from_freelist().
225 */
226 if (nodemask && !nodes_subset(node_states[N_MEMORY], *nodemask)) {
227 *totalpages = total_swap_pages;
228 for_each_node_mask(nid, *nodemask)
229 *totalpages += node_spanned_pages(nid);
230 return CONSTRAINT_MEMORY_POLICY;
231 }
232
233 /* Check this allocation failure is caused by cpuset's wall function */
234 for_each_zone_zonelist_nodemask(zone, z, zonelist,
235 high_zoneidx, nodemask)
236 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
237 cpuset_limited = true;
238
239 if (cpuset_limited) {
240 *totalpages = total_swap_pages;
241 for_each_node_mask(nid, cpuset_current_mems_allowed)
242 *totalpages += node_spanned_pages(nid);
243 return CONSTRAINT_CPUSET;
244 }
245 return CONSTRAINT_NONE;
246}
247#else
248static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
249 gfp_t gfp_mask, nodemask_t *nodemask,
250 unsigned long *totalpages)
251{
252 *totalpages = totalram_pages + total_swap_pages;
253 return CONSTRAINT_NONE;
254}
255#endif
256
257enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
258 unsigned long totalpages, const nodemask_t *nodemask,
259 bool force_kill)
260{
261 if (task->exit_state)
262 return OOM_SCAN_CONTINUE;
263 if (oom_unkillable_task(task, NULL, nodemask))
264 return OOM_SCAN_CONTINUE;
265
266 /*
267 * This task already has access to memory reserves and is being killed.
268 * Don't allow any other task to have access to the reserves.
269 */
270 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
271 if (unlikely(frozen(task)))
272 __thaw_task(task);
273 if (!force_kill)
274 return OOM_SCAN_ABORT;
275 }
276 if (!task->mm)
277 return OOM_SCAN_CONTINUE;
278
279 /*
280 * If task is allocating a lot of memory and has been marked to be
281 * killed first if it triggers an oom, then select it.
282 */
283 if (oom_task_origin(task))
284 return OOM_SCAN_SELECT;
285
286 if (task->flags & PF_EXITING && !force_kill) {
287 /*
288 * If this task is not being ptraced on exit, then wait for it
289 * to finish before killing some other task unnecessarily.
290 */
291 if (!(task->group_leader->ptrace & PT_TRACE_EXIT))
292 return OOM_SCAN_ABORT;
293 }
294 return OOM_SCAN_OK;
295}
296
297/*
298 * Simple selection loop. We chose the process with the highest
299 * number of 'points'. Returns -1 on scan abort.
300 *
301 * (not docbooked, we don't want this one cluttering up the manual)
302 */
303static struct task_struct *select_bad_process(unsigned int *ppoints,
304 unsigned long totalpages, const nodemask_t *nodemask,
305 bool force_kill)
306{
307 struct task_struct *g, *p;
308 struct task_struct *chosen = NULL;
309 unsigned long chosen_points = 0;
310
311 rcu_read_lock();
312 for_each_process_thread(g, p) {
313 unsigned int points;
314
315 switch (oom_scan_process_thread(p, totalpages, nodemask,
316 force_kill)) {
317 case OOM_SCAN_SELECT:
318 chosen = p;
319 chosen_points = ULONG_MAX;
320 /* fall through */
321 case OOM_SCAN_CONTINUE:
322 continue;
323 case OOM_SCAN_ABORT:
324 rcu_read_unlock();
325 return (struct task_struct *)(-1UL);
326 case OOM_SCAN_OK:
327 break;
328 };
329 points = oom_badness(p, NULL, nodemask, totalpages);
330 if (!points || points < chosen_points)
331 continue;
332 /* Prefer thread group leaders for display purposes */
333 if (points == chosen_points && thread_group_leader(chosen))
334 continue;
335
336 chosen = p;
337 chosen_points = points;
338 }
339 if (chosen)
340 get_task_struct(chosen);
341 rcu_read_unlock();
342
343 *ppoints = chosen_points * 1000 / totalpages;
344 return chosen;
345}
346
347/**
348 * dump_tasks - dump current memory state of all system tasks
349 * @memcg: current's memory controller, if constrained
350 * @nodemask: nodemask passed to page allocator for mempolicy ooms
351 *
352 * Dumps the current memory state of all eligible tasks. Tasks not in the same
353 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
354 * are not shown.
355 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
356 * swapents, oom_score_adj value, and name.
357 */
358static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
359{
360 struct task_struct *p;
361 struct task_struct *task;
362
363 pr_info("[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name\n");
364 rcu_read_lock();
365 for_each_process(p) {
366 if (oom_unkillable_task(p, memcg, nodemask))
367 continue;
368
369 task = find_lock_task_mm(p);
370 if (!task) {
371 /*
372 * This is a kthread or all of p's threads have already
373 * detached their mm's. There's no need to report
374 * them; they can't be oom killed anyway.
375 */
376 continue;
377 }
378
379 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %8lu %5hd %s\n",
380 task->pid, from_kuid(&init_user_ns, task_uid(task)),
381 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
382 atomic_long_read(&task->mm->nr_ptes),
383 get_mm_counter(task->mm, MM_SWAPENTS),
384 task->signal->oom_score_adj, task->comm);
385 task_unlock(task);
386 }
387 rcu_read_unlock();
388}
389
390static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
391 struct mem_cgroup *memcg, const nodemask_t *nodemask)
392{
393 task_lock(current);
394 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
395 "oom_score_adj=%hd\n",
396 current->comm, gfp_mask, order,
397 current->signal->oom_score_adj);
398 cpuset_print_task_mems_allowed(current);
399 task_unlock(current);
400 dump_stack();
401 if (memcg)
402 mem_cgroup_print_oom_info(memcg, p);
403 else
404 show_mem(SHOW_MEM_FILTER_NODES);
405 if (sysctl_oom_dump_tasks)
406 dump_tasks(memcg, nodemask);
407}
408
409#define K(x) ((x) << (PAGE_SHIFT-10))
410/*
411 * Must be called while holding a reference to p, which will be released upon
412 * returning.
413 */
414void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
415 unsigned int points, unsigned long totalpages,
416 struct mem_cgroup *memcg, nodemask_t *nodemask,
417 const char *message)
418{
419 struct task_struct *victim = p;
420 struct task_struct *child;
421 struct task_struct *t;
422 struct mm_struct *mm;
423 unsigned int victim_points = 0;
424 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
425 DEFAULT_RATELIMIT_BURST);
426
427 /*
428 * If the task is already exiting, don't alarm the sysadmin or kill
429 * its children or threads, just set TIF_MEMDIE so it can die quickly
430 */
431 if (p->flags & PF_EXITING) {
432 set_tsk_thread_flag(p, TIF_MEMDIE);
433 put_task_struct(p);
434 return;
435 }
436
437 if (__ratelimit(&oom_rs))
438 dump_header(p, gfp_mask, order, memcg, nodemask);
439
440 task_lock(p);
441 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
442 message, task_pid_nr(p), p->comm, points);
443 task_unlock(p);
444
445 /*
446 * If any of p's children has a different mm and is eligible for kill,
447 * the one with the highest oom_badness() score is sacrificed for its
448 * parent. This attempts to lose the minimal amount of work done while
449 * still freeing memory.
450 */
451 read_lock(&tasklist_lock);
452 for_each_thread(p, t) {
453 list_for_each_entry(child, &t->children, sibling) {
454 unsigned int child_points;
455
456 if (child->mm == p->mm)
457 continue;
458 /*
459 * oom_badness() returns 0 if the thread is unkillable
460 */
461 child_points = oom_badness(child, memcg, nodemask,
462 totalpages);
463 if (child_points > victim_points) {
464 put_task_struct(victim);
465 victim = child;
466 victim_points = child_points;
467 get_task_struct(victim);
468 }
469 }
470 }
471 read_unlock(&tasklist_lock);
472
473 p = find_lock_task_mm(victim);
474 if (!p) {
475 put_task_struct(victim);
476 return;
477 } else if (victim != p) {
478 get_task_struct(p);
479 put_task_struct(victim);
480 victim = p;
481 }
482
483 /* mm cannot safely be dereferenced after task_unlock(victim) */
484 mm = victim->mm;
485 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
486 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
487 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
488 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
489 task_unlock(victim);
490
491 /*
492 * Kill all user processes sharing victim->mm in other thread groups, if
493 * any. They don't get access to memory reserves, though, to avoid
494 * depletion of all memory. This prevents mm->mmap_sem livelock when an
495 * oom killed thread cannot exit because it requires the semaphore and
496 * its contended by another thread trying to allocate memory itself.
497 * That thread will now get access to memory reserves since it has a
498 * pending fatal signal.
499 */
500 rcu_read_lock();
501 for_each_process(p)
502 if (p->mm == mm && !same_thread_group(p, victim) &&
503 !(p->flags & PF_KTHREAD)) {
504 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
505 continue;
506
507 task_lock(p); /* Protect ->comm from prctl() */
508 pr_err("Kill process %d (%s) sharing same memory\n",
509 task_pid_nr(p), p->comm);
510 task_unlock(p);
511 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
512 }
513 rcu_read_unlock();
514
515 set_tsk_thread_flag(victim, TIF_MEMDIE);
516 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
517 put_task_struct(victim);
518}
519#undef K
520
521/*
522 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
523 */
524void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
525 int order, const nodemask_t *nodemask)
526{
527 if (likely(!sysctl_panic_on_oom))
528 return;
529 if (sysctl_panic_on_oom != 2) {
530 /*
531 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
532 * does not panic for cpuset, mempolicy, or memcg allocation
533 * failures.
534 */
535 if (constraint != CONSTRAINT_NONE)
536 return;
537 }
538 dump_header(NULL, gfp_mask, order, NULL, nodemask);
539 panic("Out of memory: %s panic_on_oom is enabled\n",
540 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
541}
542
543static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
544
545int register_oom_notifier(struct notifier_block *nb)
546{
547 return blocking_notifier_chain_register(&oom_notify_list, nb);
548}
549EXPORT_SYMBOL_GPL(register_oom_notifier);
550
551int unregister_oom_notifier(struct notifier_block *nb)
552{
553 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
554}
555EXPORT_SYMBOL_GPL(unregister_oom_notifier);
556
557/*
558 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
559 * if a parallel OOM killing is already taking place that includes a zone in
560 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
561 */
562int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
563{
564 struct zoneref *z;
565 struct zone *zone;
566 int ret = 1;
567
568 spin_lock(&zone_scan_lock);
569 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
570 if (zone_is_oom_locked(zone)) {
571 ret = 0;
572 goto out;
573 }
574 }
575
576 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
577 /*
578 * Lock each zone in the zonelist under zone_scan_lock so a
579 * parallel invocation of try_set_zonelist_oom() doesn't succeed
580 * when it shouldn't.
581 */
582 zone_set_flag(zone, ZONE_OOM_LOCKED);
583 }
584
585out:
586 spin_unlock(&zone_scan_lock);
587 return ret;
588}
589
590/*
591 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
592 * allocation attempts with zonelists containing them may now recall the OOM
593 * killer, if necessary.
594 */
595void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
596{
597 struct zoneref *z;
598 struct zone *zone;
599
600 spin_lock(&zone_scan_lock);
601 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
602 zone_clear_flag(zone, ZONE_OOM_LOCKED);
603 }
604 spin_unlock(&zone_scan_lock);
605}
606
607/**
608 * out_of_memory - kill the "best" process when we run out of memory
609 * @zonelist: zonelist pointer
610 * @gfp_mask: memory allocation flags
611 * @order: amount of memory being requested as a power of 2
612 * @nodemask: nodemask passed to page allocator
613 * @force_kill: true if a task must be killed, even if others are exiting
614 *
615 * If we run out of memory, we have the choice between either
616 * killing a random task (bad), letting the system crash (worse)
617 * OR try to be smart about which process to kill. Note that we
618 * don't have to be perfect here, we just have to be good.
619 */
620void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
621 int order, nodemask_t *nodemask, bool force_kill)
622{
623 const nodemask_t *mpol_mask;
624 struct task_struct *p;
625 unsigned long totalpages;
626 unsigned long freed = 0;
627 unsigned int uninitialized_var(points);
628 enum oom_constraint constraint = CONSTRAINT_NONE;
629 int killed = 0;
630
631 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
632 if (freed > 0)
633 /* Got some memory back in the last second. */
634 return;
635
636 /*
637 * If current has a pending SIGKILL or is exiting, then automatically
638 * select it. The goal is to allow it to allocate so that it may
639 * quickly exit and free its memory.
640 */
641 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
642 set_thread_flag(TIF_MEMDIE);
643 return;
644 }
645
646 /*
647 * Check if there were limitations on the allocation (only relevant for
648 * NUMA) that may require different handling.
649 */
650 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
651 &totalpages);
652 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
653 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
654
655 if (sysctl_oom_kill_allocating_task && current->mm &&
656 !oom_unkillable_task(current, NULL, nodemask) &&
657 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
658 get_task_struct(current);
659 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
660 nodemask,
661 "Out of memory (oom_kill_allocating_task)");
662 goto out;
663 }
664
665 p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
666 /* Found nothing?!?! Either we hang forever, or we panic. */
667 if (!p) {
668 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
669 panic("Out of memory and no killable processes...\n");
670 }
671 if (p != (void *)-1UL) {
672 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
673 nodemask, "Out of memory");
674 killed = 1;
675 }
676out:
677 /*
678 * Give the killed threads a good chance of exiting before trying to
679 * allocate memory again.
680 */
681 if (killed)
682 schedule_timeout_killable(1);
683}
684
685/*
686 * The pagefault handler calls here because it is out of memory, so kill a
687 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
688 * parallel oom killing is already in progress so do nothing.
689 */
690void pagefault_out_of_memory(void)
691{
692 struct zonelist *zonelist;
693
694 if (mem_cgroup_oom_synchronize(true))
695 return;
696
697 zonelist = node_zonelist(first_online_node, GFP_KERNEL);
698 if (try_set_zonelist_oom(zonelist, GFP_KERNEL)) {
699 out_of_memory(NULL, 0, 0, NULL, false);
700 clear_zonelist_oom(zonelist, GFP_KERNEL);
701 }
702}
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 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
9 *
10 * The routines in this file are used to kill a process when
11 * we're seriously out of memory. This gets called from __alloc_pages()
12 * in mm/page_alloc.c when we really run out of memory.
13 *
14 * Since we won't call these routines often (on a well-configured
15 * machine) this file will double as a 'coding guide' and a signpost
16 * for newbie kernel hackers. It features several pointers to major
17 * kernel subsystems and hints as to where to find out what things do.
18 */
19
20#include <linux/oom.h>
21#include <linux/mm.h>
22#include <linux/err.h>
23#include <linux/gfp.h>
24#include <linux/sched.h>
25#include <linux/swap.h>
26#include <linux/timex.h>
27#include <linux/jiffies.h>
28#include <linux/cpuset.h>
29#include <linux/export.h>
30#include <linux/notifier.h>
31#include <linux/memcontrol.h>
32#include <linux/mempolicy.h>
33#include <linux/security.h>
34#include <linux/ptrace.h>
35#include <linux/freezer.h>
36#include <linux/ftrace.h>
37#include <linux/ratelimit.h>
38#include <linux/kthread.h>
39#include <linux/init.h>
40
41#include <asm/tlb.h>
42#include "internal.h"
43
44#define CREATE_TRACE_POINTS
45#include <trace/events/oom.h>
46
47int sysctl_panic_on_oom;
48int sysctl_oom_kill_allocating_task;
49int sysctl_oom_dump_tasks = 1;
50
51DEFINE_MUTEX(oom_lock);
52
53#ifdef CONFIG_NUMA
54/**
55 * has_intersects_mems_allowed() - check task eligiblity for kill
56 * @start: task struct of which task to consider
57 * @mask: nodemask passed to page allocator for mempolicy ooms
58 *
59 * Task eligibility is determined by whether or not a candidate task, @tsk,
60 * shares the same mempolicy nodes as current if it is bound by such a policy
61 * and whether or not it has the same set of allowed cpuset nodes.
62 */
63static bool has_intersects_mems_allowed(struct task_struct *start,
64 const nodemask_t *mask)
65{
66 struct task_struct *tsk;
67 bool ret = false;
68
69 rcu_read_lock();
70 for_each_thread(start, tsk) {
71 if (mask) {
72 /*
73 * If this is a mempolicy constrained oom, tsk's
74 * cpuset is irrelevant. Only return true if its
75 * mempolicy intersects current, otherwise it may be
76 * needlessly killed.
77 */
78 ret = mempolicy_nodemask_intersects(tsk, mask);
79 } else {
80 /*
81 * This is not a mempolicy constrained oom, so only
82 * check the mems of tsk's cpuset.
83 */
84 ret = cpuset_mems_allowed_intersects(current, tsk);
85 }
86 if (ret)
87 break;
88 }
89 rcu_read_unlock();
90
91 return ret;
92}
93#else
94static bool has_intersects_mems_allowed(struct task_struct *tsk,
95 const nodemask_t *mask)
96{
97 return true;
98}
99#endif /* CONFIG_NUMA */
100
101/*
102 * The process p may have detached its own ->mm while exiting or through
103 * use_mm(), but one or more of its subthreads may still have a valid
104 * pointer. Return p, or any of its subthreads with a valid ->mm, with
105 * task_lock() held.
106 */
107struct task_struct *find_lock_task_mm(struct task_struct *p)
108{
109 struct task_struct *t;
110
111 rcu_read_lock();
112
113 for_each_thread(p, t) {
114 task_lock(t);
115 if (likely(t->mm))
116 goto found;
117 task_unlock(t);
118 }
119 t = NULL;
120found:
121 rcu_read_unlock();
122
123 return t;
124}
125
126/*
127 * order == -1 means the oom kill is required by sysrq, otherwise only
128 * for display purposes.
129 */
130static inline bool is_sysrq_oom(struct oom_control *oc)
131{
132 return oc->order == -1;
133}
134
135/* return true if the task is not adequate as candidate victim task. */
136static bool oom_unkillable_task(struct task_struct *p,
137 struct mem_cgroup *memcg, const nodemask_t *nodemask)
138{
139 if (is_global_init(p))
140 return true;
141 if (p->flags & PF_KTHREAD)
142 return true;
143
144 /* When mem_cgroup_out_of_memory() and p is not member of the group */
145 if (memcg && !task_in_mem_cgroup(p, memcg))
146 return true;
147
148 /* p may not have freeable memory in nodemask */
149 if (!has_intersects_mems_allowed(p, nodemask))
150 return true;
151
152 return false;
153}
154
155/**
156 * oom_badness - heuristic function to determine which candidate task to kill
157 * @p: task struct of which task we should calculate
158 * @totalpages: total present RAM allowed for page allocation
159 *
160 * The heuristic for determining which task to kill is made to be as simple and
161 * predictable as possible. The goal is to return the highest value for the
162 * task consuming the most memory to avoid subsequent oom failures.
163 */
164unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
165 const nodemask_t *nodemask, unsigned long totalpages)
166{
167 long points;
168 long adj;
169
170 if (oom_unkillable_task(p, memcg, nodemask))
171 return 0;
172
173 p = find_lock_task_mm(p);
174 if (!p)
175 return 0;
176
177 adj = (long)p->signal->oom_score_adj;
178 if (adj == OOM_SCORE_ADJ_MIN) {
179 task_unlock(p);
180 return 0;
181 }
182
183 /*
184 * The baseline for the badness score is the proportion of RAM that each
185 * task's rss, pagetable and swap space use.
186 */
187 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
188 atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm);
189 task_unlock(p);
190
191 /*
192 * Root processes get 3% bonus, just like the __vm_enough_memory()
193 * implementation used by LSMs.
194 */
195 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
196 points -= (points * 3) / 100;
197
198 /* Normalize to oom_score_adj units */
199 adj *= totalpages / 1000;
200 points += adj;
201
202 /*
203 * Never return 0 for an eligible task regardless of the root bonus and
204 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
205 */
206 return points > 0 ? points : 1;
207}
208
209/*
210 * Determine the type of allocation constraint.
211 */
212#ifdef CONFIG_NUMA
213static enum oom_constraint constrained_alloc(struct oom_control *oc,
214 unsigned long *totalpages)
215{
216 struct zone *zone;
217 struct zoneref *z;
218 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
219 bool cpuset_limited = false;
220 int nid;
221
222 /* Default to all available memory */
223 *totalpages = totalram_pages + total_swap_pages;
224
225 if (!oc->zonelist)
226 return CONSTRAINT_NONE;
227 /*
228 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
229 * to kill current.We have to random task kill in this case.
230 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
231 */
232 if (oc->gfp_mask & __GFP_THISNODE)
233 return CONSTRAINT_NONE;
234
235 /*
236 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
237 * the page allocator means a mempolicy is in effect. Cpuset policy
238 * is enforced in get_page_from_freelist().
239 */
240 if (oc->nodemask &&
241 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
242 *totalpages = total_swap_pages;
243 for_each_node_mask(nid, *oc->nodemask)
244 *totalpages += node_spanned_pages(nid);
245 return CONSTRAINT_MEMORY_POLICY;
246 }
247
248 /* Check this allocation failure is caused by cpuset's wall function */
249 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
250 high_zoneidx, oc->nodemask)
251 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
252 cpuset_limited = true;
253
254 if (cpuset_limited) {
255 *totalpages = total_swap_pages;
256 for_each_node_mask(nid, cpuset_current_mems_allowed)
257 *totalpages += node_spanned_pages(nid);
258 return CONSTRAINT_CPUSET;
259 }
260 return CONSTRAINT_NONE;
261}
262#else
263static enum oom_constraint constrained_alloc(struct oom_control *oc,
264 unsigned long *totalpages)
265{
266 *totalpages = totalram_pages + total_swap_pages;
267 return CONSTRAINT_NONE;
268}
269#endif
270
271enum oom_scan_t oom_scan_process_thread(struct oom_control *oc,
272 struct task_struct *task, unsigned long totalpages)
273{
274 if (oom_unkillable_task(task, NULL, oc->nodemask))
275 return OOM_SCAN_CONTINUE;
276
277 /*
278 * This task already has access to memory reserves and is being killed.
279 * Don't allow any other task to have access to the reserves.
280 */
281 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
282 if (!is_sysrq_oom(oc))
283 return OOM_SCAN_ABORT;
284 }
285 if (!task->mm)
286 return OOM_SCAN_CONTINUE;
287
288 /*
289 * If task is allocating a lot of memory and has been marked to be
290 * killed first if it triggers an oom, then select it.
291 */
292 if (oom_task_origin(task))
293 return OOM_SCAN_SELECT;
294
295 return OOM_SCAN_OK;
296}
297
298/*
299 * Simple selection loop. We chose the process with the highest
300 * number of 'points'. Returns -1 on scan abort.
301 */
302static struct task_struct *select_bad_process(struct oom_control *oc,
303 unsigned int *ppoints, unsigned long totalpages)
304{
305 struct task_struct *g, *p;
306 struct task_struct *chosen = NULL;
307 unsigned long chosen_points = 0;
308
309 rcu_read_lock();
310 for_each_process_thread(g, p) {
311 unsigned int points;
312
313 switch (oom_scan_process_thread(oc, p, totalpages)) {
314 case OOM_SCAN_SELECT:
315 chosen = p;
316 chosen_points = ULONG_MAX;
317 /* fall through */
318 case OOM_SCAN_CONTINUE:
319 continue;
320 case OOM_SCAN_ABORT:
321 rcu_read_unlock();
322 return (struct task_struct *)(-1UL);
323 case OOM_SCAN_OK:
324 break;
325 };
326 points = oom_badness(p, NULL, oc->nodemask, totalpages);
327 if (!points || points < chosen_points)
328 continue;
329 /* Prefer thread group leaders for display purposes */
330 if (points == chosen_points && thread_group_leader(chosen))
331 continue;
332
333 chosen = p;
334 chosen_points = points;
335 }
336 if (chosen)
337 get_task_struct(chosen);
338 rcu_read_unlock();
339
340 *ppoints = chosen_points * 1000 / totalpages;
341 return chosen;
342}
343
344/**
345 * dump_tasks - dump current memory state of all system tasks
346 * @memcg: current's memory controller, if constrained
347 * @nodemask: nodemask passed to page allocator for mempolicy ooms
348 *
349 * Dumps the current memory state of all eligible tasks. Tasks not in the same
350 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
351 * are not shown.
352 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
353 * swapents, oom_score_adj value, and name.
354 */
355static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
356{
357 struct task_struct *p;
358 struct task_struct *task;
359
360 pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n");
361 rcu_read_lock();
362 for_each_process(p) {
363 if (oom_unkillable_task(p, memcg, nodemask))
364 continue;
365
366 task = find_lock_task_mm(p);
367 if (!task) {
368 /*
369 * This is a kthread or all of p's threads have already
370 * detached their mm's. There's no need to report
371 * them; they can't be oom killed anyway.
372 */
373 continue;
374 }
375
376 pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n",
377 task->pid, from_kuid(&init_user_ns, task_uid(task)),
378 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
379 atomic_long_read(&task->mm->nr_ptes),
380 mm_nr_pmds(task->mm),
381 get_mm_counter(task->mm, MM_SWAPENTS),
382 task->signal->oom_score_adj, task->comm);
383 task_unlock(task);
384 }
385 rcu_read_unlock();
386}
387
388static void dump_header(struct oom_control *oc, struct task_struct *p,
389 struct mem_cgroup *memcg)
390{
391 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
392 current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
393 current->signal->oom_score_adj);
394
395 cpuset_print_current_mems_allowed();
396 dump_stack();
397 if (memcg)
398 mem_cgroup_print_oom_info(memcg, p);
399 else
400 show_mem(SHOW_MEM_FILTER_NODES);
401 if (sysctl_oom_dump_tasks)
402 dump_tasks(memcg, oc->nodemask);
403}
404
405/*
406 * Number of OOM victims in flight
407 */
408static atomic_t oom_victims = ATOMIC_INIT(0);
409static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
410
411bool oom_killer_disabled __read_mostly;
412
413#define K(x) ((x) << (PAGE_SHIFT-10))
414
415#ifdef CONFIG_MMU
416/*
417 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
418 * victim (if that is possible) to help the OOM killer to move on.
419 */
420static struct task_struct *oom_reaper_th;
421static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
422static struct task_struct *oom_reaper_list;
423static DEFINE_SPINLOCK(oom_reaper_lock);
424
425
426static bool __oom_reap_task(struct task_struct *tsk)
427{
428 struct mmu_gather tlb;
429 struct vm_area_struct *vma;
430 struct mm_struct *mm;
431 struct task_struct *p;
432 struct zap_details details = {.check_swap_entries = true,
433 .ignore_dirty = true};
434 bool ret = true;
435
436 /*
437 * Make sure we find the associated mm_struct even when the particular
438 * thread has already terminated and cleared its mm.
439 * We might have race with exit path so consider our work done if there
440 * is no mm.
441 */
442 p = find_lock_task_mm(tsk);
443 if (!p)
444 return true;
445
446 mm = p->mm;
447 if (!atomic_inc_not_zero(&mm->mm_users)) {
448 task_unlock(p);
449 return true;
450 }
451
452 task_unlock(p);
453
454 if (!down_read_trylock(&mm->mmap_sem)) {
455 ret = false;
456 goto out;
457 }
458
459 tlb_gather_mmu(&tlb, mm, 0, -1);
460 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
461 if (is_vm_hugetlb_page(vma))
462 continue;
463
464 /*
465 * mlocked VMAs require explicit munlocking before unmap.
466 * Let's keep it simple here and skip such VMAs.
467 */
468 if (vma->vm_flags & VM_LOCKED)
469 continue;
470
471 /*
472 * Only anonymous pages have a good chance to be dropped
473 * without additional steps which we cannot afford as we
474 * are OOM already.
475 *
476 * We do not even care about fs backed pages because all
477 * which are reclaimable have already been reclaimed and
478 * we do not want to block exit_mmap by keeping mm ref
479 * count elevated without a good reason.
480 */
481 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED))
482 unmap_page_range(&tlb, vma, vma->vm_start, vma->vm_end,
483 &details);
484 }
485 tlb_finish_mmu(&tlb, 0, -1);
486 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
487 task_pid_nr(tsk), tsk->comm,
488 K(get_mm_counter(mm, MM_ANONPAGES)),
489 K(get_mm_counter(mm, MM_FILEPAGES)),
490 K(get_mm_counter(mm, MM_SHMEMPAGES)));
491 up_read(&mm->mmap_sem);
492
493 /*
494 * Clear TIF_MEMDIE because the task shouldn't be sitting on a
495 * reasonably reclaimable memory anymore. OOM killer can continue
496 * by selecting other victim if unmapping hasn't led to any
497 * improvements. This also means that selecting this task doesn't
498 * make any sense.
499 */
500 tsk->signal->oom_score_adj = OOM_SCORE_ADJ_MIN;
501 exit_oom_victim(tsk);
502out:
503 mmput(mm);
504 return ret;
505}
506
507#define MAX_OOM_REAP_RETRIES 10
508static void oom_reap_task(struct task_struct *tsk)
509{
510 int attempts = 0;
511
512 /* Retry the down_read_trylock(mmap_sem) a few times */
513 while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task(tsk))
514 schedule_timeout_idle(HZ/10);
515
516 if (attempts > MAX_OOM_REAP_RETRIES) {
517 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
518 task_pid_nr(tsk), tsk->comm);
519 debug_show_all_locks();
520 }
521
522 /* Drop a reference taken by wake_oom_reaper */
523 put_task_struct(tsk);
524}
525
526static int oom_reaper(void *unused)
527{
528 set_freezable();
529
530 while (true) {
531 struct task_struct *tsk = NULL;
532
533 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
534 spin_lock(&oom_reaper_lock);
535 if (oom_reaper_list != NULL) {
536 tsk = oom_reaper_list;
537 oom_reaper_list = tsk->oom_reaper_list;
538 }
539 spin_unlock(&oom_reaper_lock);
540
541 if (tsk)
542 oom_reap_task(tsk);
543 }
544
545 return 0;
546}
547
548static void wake_oom_reaper(struct task_struct *tsk)
549{
550 if (!oom_reaper_th)
551 return;
552
553 /* tsk is already queued? */
554 if (tsk == oom_reaper_list || tsk->oom_reaper_list)
555 return;
556
557 get_task_struct(tsk);
558
559 spin_lock(&oom_reaper_lock);
560 tsk->oom_reaper_list = oom_reaper_list;
561 oom_reaper_list = tsk;
562 spin_unlock(&oom_reaper_lock);
563 wake_up(&oom_reaper_wait);
564}
565
566static int __init oom_init(void)
567{
568 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
569 if (IS_ERR(oom_reaper_th)) {
570 pr_err("Unable to start OOM reaper %ld. Continuing regardless\n",
571 PTR_ERR(oom_reaper_th));
572 oom_reaper_th = NULL;
573 }
574 return 0;
575}
576subsys_initcall(oom_init)
577#else
578static void wake_oom_reaper(struct task_struct *tsk)
579{
580}
581#endif
582
583/**
584 * mark_oom_victim - mark the given task as OOM victim
585 * @tsk: task to mark
586 *
587 * Has to be called with oom_lock held and never after
588 * oom has been disabled already.
589 */
590void mark_oom_victim(struct task_struct *tsk)
591{
592 WARN_ON(oom_killer_disabled);
593 /* OOM killer might race with memcg OOM */
594 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
595 return;
596 /*
597 * Make sure that the task is woken up from uninterruptible sleep
598 * if it is frozen because OOM killer wouldn't be able to free
599 * any memory and livelock. freezing_slow_path will tell the freezer
600 * that TIF_MEMDIE tasks should be ignored.
601 */
602 __thaw_task(tsk);
603 atomic_inc(&oom_victims);
604}
605
606/**
607 * exit_oom_victim - note the exit of an OOM victim
608 */
609void exit_oom_victim(struct task_struct *tsk)
610{
611 if (!test_and_clear_tsk_thread_flag(tsk, TIF_MEMDIE))
612 return;
613
614 if (!atomic_dec_return(&oom_victims))
615 wake_up_all(&oom_victims_wait);
616}
617
618/**
619 * oom_killer_disable - disable OOM killer
620 *
621 * Forces all page allocations to fail rather than trigger OOM killer.
622 * Will block and wait until all OOM victims are killed.
623 *
624 * The function cannot be called when there are runnable user tasks because
625 * the userspace would see unexpected allocation failures as a result. Any
626 * new usage of this function should be consulted with MM people.
627 *
628 * Returns true if successful and false if the OOM killer cannot be
629 * disabled.
630 */
631bool oom_killer_disable(void)
632{
633 /*
634 * Make sure to not race with an ongoing OOM killer. Check that the
635 * current is not killed (possibly due to sharing the victim's memory).
636 */
637 if (mutex_lock_killable(&oom_lock))
638 return false;
639 oom_killer_disabled = true;
640 mutex_unlock(&oom_lock);
641
642 wait_event(oom_victims_wait, !atomic_read(&oom_victims));
643
644 return true;
645}
646
647/**
648 * oom_killer_enable - enable OOM killer
649 */
650void oom_killer_enable(void)
651{
652 oom_killer_disabled = false;
653}
654
655/*
656 * task->mm can be NULL if the task is the exited group leader. So to
657 * determine whether the task is using a particular mm, we examine all the
658 * task's threads: if one of those is using this mm then this task was also
659 * using it.
660 */
661static bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
662{
663 struct task_struct *t;
664
665 for_each_thread(p, t) {
666 struct mm_struct *t_mm = READ_ONCE(t->mm);
667 if (t_mm)
668 return t_mm == mm;
669 }
670 return false;
671}
672
673/*
674 * Must be called while holding a reference to p, which will be released upon
675 * returning.
676 */
677void oom_kill_process(struct oom_control *oc, struct task_struct *p,
678 unsigned int points, unsigned long totalpages,
679 struct mem_cgroup *memcg, const char *message)
680{
681 struct task_struct *victim = p;
682 struct task_struct *child;
683 struct task_struct *t;
684 struct mm_struct *mm;
685 unsigned int victim_points = 0;
686 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
687 DEFAULT_RATELIMIT_BURST);
688 bool can_oom_reap = true;
689
690 /*
691 * If the task is already exiting, don't alarm the sysadmin or kill
692 * its children or threads, just set TIF_MEMDIE so it can die quickly
693 */
694 task_lock(p);
695 if (p->mm && task_will_free_mem(p)) {
696 mark_oom_victim(p);
697 task_unlock(p);
698 put_task_struct(p);
699 return;
700 }
701 task_unlock(p);
702
703 if (__ratelimit(&oom_rs))
704 dump_header(oc, p, memcg);
705
706 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
707 message, task_pid_nr(p), p->comm, points);
708
709 /*
710 * If any of p's children has a different mm and is eligible for kill,
711 * the one with the highest oom_badness() score is sacrificed for its
712 * parent. This attempts to lose the minimal amount of work done while
713 * still freeing memory.
714 */
715 read_lock(&tasklist_lock);
716 for_each_thread(p, t) {
717 list_for_each_entry(child, &t->children, sibling) {
718 unsigned int child_points;
719
720 if (process_shares_mm(child, p->mm))
721 continue;
722 /*
723 * oom_badness() returns 0 if the thread is unkillable
724 */
725 child_points = oom_badness(child, memcg, oc->nodemask,
726 totalpages);
727 if (child_points > victim_points) {
728 put_task_struct(victim);
729 victim = child;
730 victim_points = child_points;
731 get_task_struct(victim);
732 }
733 }
734 }
735 read_unlock(&tasklist_lock);
736
737 p = find_lock_task_mm(victim);
738 if (!p) {
739 put_task_struct(victim);
740 return;
741 } else if (victim != p) {
742 get_task_struct(p);
743 put_task_struct(victim);
744 victim = p;
745 }
746
747 /* Get a reference to safely compare mm after task_unlock(victim) */
748 mm = victim->mm;
749 atomic_inc(&mm->mm_count);
750 /*
751 * We should send SIGKILL before setting TIF_MEMDIE in order to prevent
752 * the OOM victim from depleting the memory reserves from the user
753 * space under its control.
754 */
755 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
756 mark_oom_victim(victim);
757 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
758 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
759 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
760 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
761 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
762 task_unlock(victim);
763
764 /*
765 * Kill all user processes sharing victim->mm in other thread groups, if
766 * any. They don't get access to memory reserves, though, to avoid
767 * depletion of all memory. This prevents mm->mmap_sem livelock when an
768 * oom killed thread cannot exit because it requires the semaphore and
769 * its contended by another thread trying to allocate memory itself.
770 * That thread will now get access to memory reserves since it has a
771 * pending fatal signal.
772 */
773 rcu_read_lock();
774 for_each_process(p) {
775 if (!process_shares_mm(p, mm))
776 continue;
777 if (same_thread_group(p, victim))
778 continue;
779 if (unlikely(p->flags & PF_KTHREAD) || is_global_init(p) ||
780 p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
781 /*
782 * We cannot use oom_reaper for the mm shared by this
783 * process because it wouldn't get killed and so the
784 * memory might be still used.
785 */
786 can_oom_reap = false;
787 continue;
788 }
789 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
790 }
791 rcu_read_unlock();
792
793 if (can_oom_reap)
794 wake_oom_reaper(victim);
795
796 mmdrop(mm);
797 put_task_struct(victim);
798}
799#undef K
800
801/*
802 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
803 */
804void check_panic_on_oom(struct oom_control *oc, enum oom_constraint constraint,
805 struct mem_cgroup *memcg)
806{
807 if (likely(!sysctl_panic_on_oom))
808 return;
809 if (sysctl_panic_on_oom != 2) {
810 /*
811 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
812 * does not panic for cpuset, mempolicy, or memcg allocation
813 * failures.
814 */
815 if (constraint != CONSTRAINT_NONE)
816 return;
817 }
818 /* Do not panic for oom kills triggered by sysrq */
819 if (is_sysrq_oom(oc))
820 return;
821 dump_header(oc, NULL, memcg);
822 panic("Out of memory: %s panic_on_oom is enabled\n",
823 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
824}
825
826static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
827
828int register_oom_notifier(struct notifier_block *nb)
829{
830 return blocking_notifier_chain_register(&oom_notify_list, nb);
831}
832EXPORT_SYMBOL_GPL(register_oom_notifier);
833
834int unregister_oom_notifier(struct notifier_block *nb)
835{
836 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
837}
838EXPORT_SYMBOL_GPL(unregister_oom_notifier);
839
840/**
841 * out_of_memory - kill the "best" process when we run out of memory
842 * @oc: pointer to struct oom_control
843 *
844 * If we run out of memory, we have the choice between either
845 * killing a random task (bad), letting the system crash (worse)
846 * OR try to be smart about which process to kill. Note that we
847 * don't have to be perfect here, we just have to be good.
848 */
849bool out_of_memory(struct oom_control *oc)
850{
851 struct task_struct *p;
852 unsigned long totalpages;
853 unsigned long freed = 0;
854 unsigned int uninitialized_var(points);
855 enum oom_constraint constraint = CONSTRAINT_NONE;
856
857 if (oom_killer_disabled)
858 return false;
859
860 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
861 if (freed > 0)
862 /* Got some memory back in the last second. */
863 return true;
864
865 /*
866 * If current has a pending SIGKILL or is exiting, then automatically
867 * select it. The goal is to allow it to allocate so that it may
868 * quickly exit and free its memory.
869 *
870 * But don't select if current has already released its mm and cleared
871 * TIF_MEMDIE flag at exit_mm(), otherwise an OOM livelock may occur.
872 */
873 if (current->mm &&
874 (fatal_signal_pending(current) || task_will_free_mem(current))) {
875 mark_oom_victim(current);
876 return true;
877 }
878
879 /*
880 * Check if there were limitations on the allocation (only relevant for
881 * NUMA) that may require different handling.
882 */
883 constraint = constrained_alloc(oc, &totalpages);
884 if (constraint != CONSTRAINT_MEMORY_POLICY)
885 oc->nodemask = NULL;
886 check_panic_on_oom(oc, constraint, NULL);
887
888 if (sysctl_oom_kill_allocating_task && current->mm &&
889 !oom_unkillable_task(current, NULL, oc->nodemask) &&
890 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
891 get_task_struct(current);
892 oom_kill_process(oc, current, 0, totalpages, NULL,
893 "Out of memory (oom_kill_allocating_task)");
894 return true;
895 }
896
897 p = select_bad_process(oc, &points, totalpages);
898 /* Found nothing?!?! Either we hang forever, or we panic. */
899 if (!p && !is_sysrq_oom(oc)) {
900 dump_header(oc, NULL, NULL);
901 panic("Out of memory and no killable processes...\n");
902 }
903 if (p && p != (void *)-1UL) {
904 oom_kill_process(oc, p, points, totalpages, NULL,
905 "Out of memory");
906 /*
907 * Give the killed process a good chance to exit before trying
908 * to allocate memory again.
909 */
910 schedule_timeout_killable(1);
911 }
912 return true;
913}
914
915/*
916 * The pagefault handler calls here because it is out of memory, so kill a
917 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
918 * parallel oom killing is already in progress so do nothing.
919 */
920void pagefault_out_of_memory(void)
921{
922 struct oom_control oc = {
923 .zonelist = NULL,
924 .nodemask = NULL,
925 .gfp_mask = 0,
926 .order = 0,
927 };
928
929 if (mem_cgroup_oom_synchronize(true))
930 return;
931
932 if (!mutex_trylock(&oom_lock))
933 return;
934
935 if (!out_of_memory(&oc)) {
936 /*
937 * There shouldn't be any user tasks runnable while the
938 * OOM killer is disabled, so the current task has to
939 * be a racing OOM victim for which oom_killer_disable()
940 * is waiting for.
941 */
942 WARN_ON(test_thread_flag(TIF_MEMDIE));
943 }
944
945 mutex_unlock(&oom_lock);
946}