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/module.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
36int sysctl_panic_on_oom;
37int sysctl_oom_kill_allocating_task;
38int sysctl_oom_dump_tasks = 1;
39static DEFINE_SPINLOCK(zone_scan_lock);
40
41/**
42 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
43 * @new_val: new oom_score_adj value
44 *
45 * Sets the oom_score_adj value for current to @new_val with proper
46 * synchronization and returns the old value. Usually used to temporarily
47 * set a value, save the old value in the caller, and then reinstate it later.
48 */
49int test_set_oom_score_adj(int new_val)
50{
51 struct sighand_struct *sighand = current->sighand;
52 int old_val;
53
54 spin_lock_irq(&sighand->siglock);
55 old_val = current->signal->oom_score_adj;
56 if (new_val != old_val) {
57 if (new_val == OOM_SCORE_ADJ_MIN)
58 atomic_inc(¤t->mm->oom_disable_count);
59 else if (old_val == OOM_SCORE_ADJ_MIN)
60 atomic_dec(¤t->mm->oom_disable_count);
61 current->signal->oom_score_adj = new_val;
62 }
63 spin_unlock_irq(&sighand->siglock);
64
65 return old_val;
66}
67
68#ifdef CONFIG_NUMA
69/**
70 * has_intersects_mems_allowed() - check task eligiblity for kill
71 * @tsk: task struct of which task to consider
72 * @mask: nodemask passed to page allocator for mempolicy ooms
73 *
74 * Task eligibility is determined by whether or not a candidate task, @tsk,
75 * shares the same mempolicy nodes as current if it is bound by such a policy
76 * and whether or not it has the same set of allowed cpuset nodes.
77 */
78static bool has_intersects_mems_allowed(struct task_struct *tsk,
79 const nodemask_t *mask)
80{
81 struct task_struct *start = tsk;
82
83 do {
84 if (mask) {
85 /*
86 * If this is a mempolicy constrained oom, tsk's
87 * cpuset is irrelevant. Only return true if its
88 * mempolicy intersects current, otherwise it may be
89 * needlessly killed.
90 */
91 if (mempolicy_nodemask_intersects(tsk, mask))
92 return true;
93 } else {
94 /*
95 * This is not a mempolicy constrained oom, so only
96 * check the mems of tsk's cpuset.
97 */
98 if (cpuset_mems_allowed_intersects(current, tsk))
99 return true;
100 }
101 } while_each_thread(start, tsk);
102
103 return false;
104}
105#else
106static bool has_intersects_mems_allowed(struct task_struct *tsk,
107 const nodemask_t *mask)
108{
109 return true;
110}
111#endif /* CONFIG_NUMA */
112
113/*
114 * The process p may have detached its own ->mm while exiting or through
115 * use_mm(), but one or more of its subthreads may still have a valid
116 * pointer. Return p, or any of its subthreads with a valid ->mm, with
117 * task_lock() held.
118 */
119struct task_struct *find_lock_task_mm(struct task_struct *p)
120{
121 struct task_struct *t = p;
122
123 do {
124 task_lock(t);
125 if (likely(t->mm))
126 return t;
127 task_unlock(t);
128 } while_each_thread(p, t);
129
130 return NULL;
131}
132
133/* return true if the task is not adequate as candidate victim task. */
134static bool oom_unkillable_task(struct task_struct *p,
135 const struct mem_cgroup *mem, const nodemask_t *nodemask)
136{
137 if (is_global_init(p))
138 return true;
139 if (p->flags & PF_KTHREAD)
140 return true;
141
142 /* When mem_cgroup_out_of_memory() and p is not member of the group */
143 if (mem && !task_in_mem_cgroup(p, mem))
144 return true;
145
146 /* p may not have freeable memory in nodemask */
147 if (!has_intersects_mems_allowed(p, nodemask))
148 return true;
149
150 return false;
151}
152
153/**
154 * oom_badness - heuristic function to determine which candidate task to kill
155 * @p: task struct of which task we should calculate
156 * @totalpages: total present RAM allowed for page allocation
157 *
158 * The heuristic for determining which task to kill is made to be as simple and
159 * predictable as possible. The goal is to return the highest value for the
160 * task consuming the most memory to avoid subsequent oom failures.
161 */
162unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
163 const nodemask_t *nodemask, unsigned long totalpages)
164{
165 int points;
166
167 if (oom_unkillable_task(p, mem, nodemask))
168 return 0;
169
170 p = find_lock_task_mm(p);
171 if (!p)
172 return 0;
173
174 /*
175 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
176 * so the entire heuristic doesn't need to be executed for something
177 * that cannot be killed.
178 */
179 if (atomic_read(&p->mm->oom_disable_count)) {
180 task_unlock(p);
181 return 0;
182 }
183
184 /*
185 * The memory controller may have a limit of 0 bytes, so avoid a divide
186 * by zero, if necessary.
187 */
188 if (!totalpages)
189 totalpages = 1;
190
191 /*
192 * The baseline for the badness score is the proportion of RAM that each
193 * task's rss, pagetable and swap space use.
194 */
195 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
196 points += get_mm_counter(p->mm, MM_SWAPENTS);
197
198 points *= 1000;
199 points /= totalpages;
200 task_unlock(p);
201
202 /*
203 * Root processes get 3% bonus, just like the __vm_enough_memory()
204 * implementation used by LSMs.
205 */
206 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
207 points -= 30;
208
209 /*
210 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
211 * either completely disable oom killing or always prefer a certain
212 * task.
213 */
214 points += p->signal->oom_score_adj;
215
216 /*
217 * Never return 0 for an eligible task that may be killed since it's
218 * possible that no single user task uses more than 0.1% of memory and
219 * no single admin tasks uses more than 3.0%.
220 */
221 if (points <= 0)
222 return 1;
223 return (points < 1000) ? points : 1000;
224}
225
226/*
227 * Determine the type of allocation constraint.
228 */
229#ifdef CONFIG_NUMA
230static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
231 gfp_t gfp_mask, nodemask_t *nodemask,
232 unsigned long *totalpages)
233{
234 struct zone *zone;
235 struct zoneref *z;
236 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
237 bool cpuset_limited = false;
238 int nid;
239
240 /* Default to all available memory */
241 *totalpages = totalram_pages + total_swap_pages;
242
243 if (!zonelist)
244 return CONSTRAINT_NONE;
245 /*
246 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
247 * to kill current.We have to random task kill in this case.
248 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
249 */
250 if (gfp_mask & __GFP_THISNODE)
251 return CONSTRAINT_NONE;
252
253 /*
254 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
255 * the page allocator means a mempolicy is in effect. Cpuset policy
256 * is enforced in get_page_from_freelist().
257 */
258 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
259 *totalpages = total_swap_pages;
260 for_each_node_mask(nid, *nodemask)
261 *totalpages += node_spanned_pages(nid);
262 return CONSTRAINT_MEMORY_POLICY;
263 }
264
265 /* Check this allocation failure is caused by cpuset's wall function */
266 for_each_zone_zonelist_nodemask(zone, z, zonelist,
267 high_zoneidx, nodemask)
268 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
269 cpuset_limited = true;
270
271 if (cpuset_limited) {
272 *totalpages = total_swap_pages;
273 for_each_node_mask(nid, cpuset_current_mems_allowed)
274 *totalpages += node_spanned_pages(nid);
275 return CONSTRAINT_CPUSET;
276 }
277 return CONSTRAINT_NONE;
278}
279#else
280static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
281 gfp_t gfp_mask, nodemask_t *nodemask,
282 unsigned long *totalpages)
283{
284 *totalpages = totalram_pages + total_swap_pages;
285 return CONSTRAINT_NONE;
286}
287#endif
288
289/*
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. We expect the caller will lock the tasklist.
292 *
293 * (not docbooked, we don't want this one cluttering up the manual)
294 */
295static struct task_struct *select_bad_process(unsigned int *ppoints,
296 unsigned long totalpages, struct mem_cgroup *mem,
297 const nodemask_t *nodemask)
298{
299 struct task_struct *g, *p;
300 struct task_struct *chosen = NULL;
301 *ppoints = 0;
302
303 do_each_thread(g, p) {
304 unsigned int points;
305
306 if (p->exit_state)
307 continue;
308 if (oom_unkillable_task(p, mem, nodemask))
309 continue;
310
311 /*
312 * This task already has access to memory reserves and is
313 * being killed. Don't allow any other task access to the
314 * memory reserve.
315 *
316 * Note: this may have a chance of deadlock if it gets
317 * blocked waiting for another task which itself is waiting
318 * for memory. Is there a better alternative?
319 */
320 if (test_tsk_thread_flag(p, TIF_MEMDIE))
321 return ERR_PTR(-1UL);
322 if (!p->mm)
323 continue;
324
325 if (p->flags & PF_EXITING) {
326 /*
327 * If p is the current task and is in the process of
328 * releasing memory, we allow the "kill" to set
329 * TIF_MEMDIE, which will allow it to gain access to
330 * memory reserves. Otherwise, it may stall forever.
331 *
332 * The loop isn't broken here, however, in case other
333 * threads are found to have already been oom killed.
334 */
335 if (p == current) {
336 chosen = p;
337 *ppoints = 1000;
338 } else {
339 /*
340 * If this task is not being ptraced on exit,
341 * then wait for it to finish before killing
342 * some other task unnecessarily.
343 */
344 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
345 return ERR_PTR(-1UL);
346 }
347 }
348
349 points = oom_badness(p, mem, nodemask, totalpages);
350 if (points > *ppoints) {
351 chosen = p;
352 *ppoints = points;
353 }
354 } while_each_thread(g, p);
355
356 return chosen;
357}
358
359/**
360 * dump_tasks - dump current memory state of all system tasks
361 * @mem: current's memory controller, if constrained
362 * @nodemask: nodemask passed to page allocator for mempolicy ooms
363 *
364 * Dumps the current memory state of all eligible tasks. Tasks not in the same
365 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
366 * are not shown.
367 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
368 * value, oom_score_adj value, and name.
369 *
370 * Call with tasklist_lock read-locked.
371 */
372static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
373{
374 struct task_struct *p;
375 struct task_struct *task;
376
377 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
378 for_each_process(p) {
379 if (oom_unkillable_task(p, mem, nodemask))
380 continue;
381
382 task = find_lock_task_mm(p);
383 if (!task) {
384 /*
385 * This is a kthread or all of p's threads have already
386 * detached their mm's. There's no need to report
387 * them; they can't be oom killed anyway.
388 */
389 continue;
390 }
391
392 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
393 task->pid, task_uid(task), task->tgid,
394 task->mm->total_vm, get_mm_rss(task->mm),
395 task_cpu(task), task->signal->oom_adj,
396 task->signal->oom_score_adj, task->comm);
397 task_unlock(task);
398 }
399}
400
401static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
402 struct mem_cgroup *mem, const nodemask_t *nodemask)
403{
404 task_lock(current);
405 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
406 "oom_adj=%d, oom_score_adj=%d\n",
407 current->comm, gfp_mask, order, current->signal->oom_adj,
408 current->signal->oom_score_adj);
409 cpuset_print_task_mems_allowed(current);
410 task_unlock(current);
411 dump_stack();
412 mem_cgroup_print_oom_info(mem, p);
413 show_mem(SHOW_MEM_FILTER_NODES);
414 if (sysctl_oom_dump_tasks)
415 dump_tasks(mem, nodemask);
416}
417
418#define K(x) ((x) << (PAGE_SHIFT-10))
419static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
420{
421 struct task_struct *q;
422 struct mm_struct *mm;
423
424 p = find_lock_task_mm(p);
425 if (!p)
426 return 1;
427
428 /* mm cannot be safely dereferenced after task_unlock(p) */
429 mm = p->mm;
430
431 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
432 task_pid_nr(p), p->comm, K(p->mm->total_vm),
433 K(get_mm_counter(p->mm, MM_ANONPAGES)),
434 K(get_mm_counter(p->mm, MM_FILEPAGES)));
435 task_unlock(p);
436
437 /*
438 * Kill all processes sharing p->mm in other thread groups, if any.
439 * They don't get access to memory reserves or a higher scheduler
440 * priority, though, to avoid depletion of all memory or task
441 * starvation. This prevents mm->mmap_sem livelock when an oom killed
442 * task cannot exit because it requires the semaphore and its contended
443 * by another thread trying to allocate memory itself. That thread will
444 * now get access to memory reserves since it has a pending fatal
445 * signal.
446 */
447 for_each_process(q)
448 if (q->mm == mm && !same_thread_group(q, p)) {
449 task_lock(q); /* Protect ->comm from prctl() */
450 pr_err("Kill process %d (%s) sharing same memory\n",
451 task_pid_nr(q), q->comm);
452 task_unlock(q);
453 force_sig(SIGKILL, q);
454 }
455
456 set_tsk_thread_flag(p, TIF_MEMDIE);
457 force_sig(SIGKILL, p);
458
459 return 0;
460}
461#undef K
462
463static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
464 unsigned int points, unsigned long totalpages,
465 struct mem_cgroup *mem, nodemask_t *nodemask,
466 const char *message)
467{
468 struct task_struct *victim = p;
469 struct task_struct *child;
470 struct task_struct *t = p;
471 unsigned int victim_points = 0;
472
473 if (printk_ratelimit())
474 dump_header(p, gfp_mask, order, mem, nodemask);
475
476 /*
477 * If the task is already exiting, don't alarm the sysadmin or kill
478 * its children or threads, just set TIF_MEMDIE so it can die quickly
479 */
480 if (p->flags & PF_EXITING) {
481 set_tsk_thread_flag(p, TIF_MEMDIE);
482 return 0;
483 }
484
485 task_lock(p);
486 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
487 message, task_pid_nr(p), p->comm, points);
488 task_unlock(p);
489
490 /*
491 * If any of p's children has a different mm and is eligible for kill,
492 * the one with the highest oom_badness() score is sacrificed for its
493 * parent. This attempts to lose the minimal amount of work done while
494 * still freeing memory.
495 */
496 do {
497 list_for_each_entry(child, &t->children, sibling) {
498 unsigned int child_points;
499
500 if (child->mm == p->mm)
501 continue;
502 /*
503 * oom_badness() returns 0 if the thread is unkillable
504 */
505 child_points = oom_badness(child, mem, nodemask,
506 totalpages);
507 if (child_points > victim_points) {
508 victim = child;
509 victim_points = child_points;
510 }
511 }
512 } while_each_thread(p, t);
513
514 return oom_kill_task(victim, mem);
515}
516
517/*
518 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
519 */
520static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
521 int order, const nodemask_t *nodemask)
522{
523 if (likely(!sysctl_panic_on_oom))
524 return;
525 if (sysctl_panic_on_oom != 2) {
526 /*
527 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
528 * does not panic for cpuset, mempolicy, or memcg allocation
529 * failures.
530 */
531 if (constraint != CONSTRAINT_NONE)
532 return;
533 }
534 read_lock(&tasklist_lock);
535 dump_header(NULL, gfp_mask, order, NULL, nodemask);
536 read_unlock(&tasklist_lock);
537 panic("Out of memory: %s panic_on_oom is enabled\n",
538 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
539}
540
541#ifdef CONFIG_CGROUP_MEM_RES_CTLR
542void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
543{
544 unsigned long limit;
545 unsigned int points = 0;
546 struct task_struct *p;
547
548 /*
549 * If current has a pending SIGKILL, then automatically select it. The
550 * goal is to allow it to allocate so that it may quickly exit and free
551 * its memory.
552 */
553 if (fatal_signal_pending(current)) {
554 set_thread_flag(TIF_MEMDIE);
555 return;
556 }
557
558 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
559 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
560 read_lock(&tasklist_lock);
561retry:
562 p = select_bad_process(&points, limit, mem, NULL);
563 if (!p || PTR_ERR(p) == -1UL)
564 goto out;
565
566 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
567 "Memory cgroup out of memory"))
568 goto retry;
569out:
570 read_unlock(&tasklist_lock);
571}
572#endif
573
574static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
575
576int register_oom_notifier(struct notifier_block *nb)
577{
578 return blocking_notifier_chain_register(&oom_notify_list, nb);
579}
580EXPORT_SYMBOL_GPL(register_oom_notifier);
581
582int unregister_oom_notifier(struct notifier_block *nb)
583{
584 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
585}
586EXPORT_SYMBOL_GPL(unregister_oom_notifier);
587
588/*
589 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
590 * if a parallel OOM killing is already taking place that includes a zone in
591 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
592 */
593int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
594{
595 struct zoneref *z;
596 struct zone *zone;
597 int ret = 1;
598
599 spin_lock(&zone_scan_lock);
600 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
601 if (zone_is_oom_locked(zone)) {
602 ret = 0;
603 goto out;
604 }
605 }
606
607 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
608 /*
609 * Lock each zone in the zonelist under zone_scan_lock so a
610 * parallel invocation of try_set_zonelist_oom() doesn't succeed
611 * when it shouldn't.
612 */
613 zone_set_flag(zone, ZONE_OOM_LOCKED);
614 }
615
616out:
617 spin_unlock(&zone_scan_lock);
618 return ret;
619}
620
621/*
622 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
623 * allocation attempts with zonelists containing them may now recall the OOM
624 * killer, if necessary.
625 */
626void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
627{
628 struct zoneref *z;
629 struct zone *zone;
630
631 spin_lock(&zone_scan_lock);
632 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
633 zone_clear_flag(zone, ZONE_OOM_LOCKED);
634 }
635 spin_unlock(&zone_scan_lock);
636}
637
638/*
639 * Try to acquire the oom killer lock for all system zones. Returns zero if a
640 * parallel oom killing is taking place, otherwise locks all zones and returns
641 * non-zero.
642 */
643static int try_set_system_oom(void)
644{
645 struct zone *zone;
646 int ret = 1;
647
648 spin_lock(&zone_scan_lock);
649 for_each_populated_zone(zone)
650 if (zone_is_oom_locked(zone)) {
651 ret = 0;
652 goto out;
653 }
654 for_each_populated_zone(zone)
655 zone_set_flag(zone, ZONE_OOM_LOCKED);
656out:
657 spin_unlock(&zone_scan_lock);
658 return ret;
659}
660
661/*
662 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
663 * attempts or page faults may now recall the oom killer, if necessary.
664 */
665static void clear_system_oom(void)
666{
667 struct zone *zone;
668
669 spin_lock(&zone_scan_lock);
670 for_each_populated_zone(zone)
671 zone_clear_flag(zone, ZONE_OOM_LOCKED);
672 spin_unlock(&zone_scan_lock);
673}
674
675/**
676 * out_of_memory - kill the "best" process when we run out of memory
677 * @zonelist: zonelist pointer
678 * @gfp_mask: memory allocation flags
679 * @order: amount of memory being requested as a power of 2
680 * @nodemask: nodemask passed to page allocator
681 *
682 * If we run out of memory, we have the choice between either
683 * killing a random task (bad), letting the system crash (worse)
684 * OR try to be smart about which process to kill. Note that we
685 * don't have to be perfect here, we just have to be good.
686 */
687void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
688 int order, nodemask_t *nodemask)
689{
690 const nodemask_t *mpol_mask;
691 struct task_struct *p;
692 unsigned long totalpages;
693 unsigned long freed = 0;
694 unsigned int points;
695 enum oom_constraint constraint = CONSTRAINT_NONE;
696 int killed = 0;
697
698 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
699 if (freed > 0)
700 /* Got some memory back in the last second. */
701 return;
702
703 /*
704 * If current has a pending SIGKILL, then automatically select it. The
705 * goal is to allow it to allocate so that it may quickly exit and free
706 * its memory.
707 */
708 if (fatal_signal_pending(current)) {
709 set_thread_flag(TIF_MEMDIE);
710 return;
711 }
712
713 /*
714 * Check if there were limitations on the allocation (only relevant for
715 * NUMA) that may require different handling.
716 */
717 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
718 &totalpages);
719 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
720 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
721
722 read_lock(&tasklist_lock);
723 if (sysctl_oom_kill_allocating_task &&
724 !oom_unkillable_task(current, NULL, nodemask) &&
725 current->mm && !atomic_read(¤t->mm->oom_disable_count)) {
726 /*
727 * oom_kill_process() needs tasklist_lock held. If it returns
728 * non-zero, current could not be killed so we must fallback to
729 * the tasklist scan.
730 */
731 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
732 NULL, nodemask,
733 "Out of memory (oom_kill_allocating_task)"))
734 goto out;
735 }
736
737retry:
738 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
739 if (PTR_ERR(p) == -1UL)
740 goto out;
741
742 /* Found nothing?!?! Either we hang forever, or we panic. */
743 if (!p) {
744 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
745 read_unlock(&tasklist_lock);
746 panic("Out of memory and no killable processes...\n");
747 }
748
749 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
750 nodemask, "Out of memory"))
751 goto retry;
752 killed = 1;
753out:
754 read_unlock(&tasklist_lock);
755
756 /*
757 * Give "p" a good chance of killing itself before we
758 * retry to allocate memory unless "p" is current
759 */
760 if (killed && !test_thread_flag(TIF_MEMDIE))
761 schedule_timeout_uninterruptible(1);
762}
763
764/*
765 * The pagefault handler calls here because it is out of memory, so kill a
766 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
767 * oom killing is already in progress so do nothing. If a task is found with
768 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
769 */
770void pagefault_out_of_memory(void)
771{
772 if (try_set_system_oom()) {
773 out_of_memory(NULL, 0, 0, NULL);
774 clear_system_oom();
775 }
776 if (!test_thread_flag(TIF_MEMDIE))
777 schedule_timeout_uninterruptible(1);
778}
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}