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/sched/mm.h>
26#include <linux/sched/coredump.h>
27#include <linux/sched/task.h>
28#include <linux/swap.h>
29#include <linux/timex.h>
30#include <linux/jiffies.h>
31#include <linux/cpuset.h>
32#include <linux/export.h>
33#include <linux/notifier.h>
34#include <linux/memcontrol.h>
35#include <linux/mempolicy.h>
36#include <linux/security.h>
37#include <linux/ptrace.h>
38#include <linux/freezer.h>
39#include <linux/ftrace.h>
40#include <linux/ratelimit.h>
41#include <linux/kthread.h>
42#include <linux/init.h>
43#include <linux/mmu_notifier.h>
44
45#include <asm/tlb.h>
46#include "internal.h"
47#include "slab.h"
48
49#define CREATE_TRACE_POINTS
50#include <trace/events/oom.h>
51
52int sysctl_panic_on_oom;
53int sysctl_oom_kill_allocating_task;
54int sysctl_oom_dump_tasks = 1;
55
56DEFINE_MUTEX(oom_lock);
57
58#ifdef CONFIG_NUMA
59/**
60 * has_intersects_mems_allowed() - check task eligiblity for kill
61 * @start: task struct of which task to consider
62 * @mask: nodemask passed to page allocator for mempolicy ooms
63 *
64 * Task eligibility is determined by whether or not a candidate task, @tsk,
65 * shares the same mempolicy nodes as current if it is bound by such a policy
66 * and whether or not it has the same set of allowed cpuset nodes.
67 */
68static bool has_intersects_mems_allowed(struct task_struct *start,
69 const nodemask_t *mask)
70{
71 struct task_struct *tsk;
72 bool ret = false;
73
74 rcu_read_lock();
75 for_each_thread(start, tsk) {
76 if (mask) {
77 /*
78 * If this is a mempolicy constrained oom, tsk's
79 * cpuset is irrelevant. Only return true if its
80 * mempolicy intersects current, otherwise it may be
81 * needlessly killed.
82 */
83 ret = mempolicy_nodemask_intersects(tsk, mask);
84 } else {
85 /*
86 * This is not a mempolicy constrained oom, so only
87 * check the mems of tsk's cpuset.
88 */
89 ret = cpuset_mems_allowed_intersects(current, tsk);
90 }
91 if (ret)
92 break;
93 }
94 rcu_read_unlock();
95
96 return ret;
97}
98#else
99static bool has_intersects_mems_allowed(struct task_struct *tsk,
100 const nodemask_t *mask)
101{
102 return true;
103}
104#endif /* CONFIG_NUMA */
105
106/*
107 * The process p may have detached its own ->mm while exiting or through
108 * use_mm(), but one or more of its subthreads may still have a valid
109 * pointer. Return p, or any of its subthreads with a valid ->mm, with
110 * task_lock() held.
111 */
112struct task_struct *find_lock_task_mm(struct task_struct *p)
113{
114 struct task_struct *t;
115
116 rcu_read_lock();
117
118 for_each_thread(p, t) {
119 task_lock(t);
120 if (likely(t->mm))
121 goto found;
122 task_unlock(t);
123 }
124 t = NULL;
125found:
126 rcu_read_unlock();
127
128 return t;
129}
130
131/*
132 * order == -1 means the oom kill is required by sysrq, otherwise only
133 * for display purposes.
134 */
135static inline bool is_sysrq_oom(struct oom_control *oc)
136{
137 return oc->order == -1;
138}
139
140static inline bool is_memcg_oom(struct oom_control *oc)
141{
142 return oc->memcg != NULL;
143}
144
145/* return true if the task is not adequate as candidate victim task. */
146static bool oom_unkillable_task(struct task_struct *p,
147 struct mem_cgroup *memcg, const nodemask_t *nodemask)
148{
149 if (is_global_init(p))
150 return true;
151 if (p->flags & PF_KTHREAD)
152 return true;
153
154 /* When mem_cgroup_out_of_memory() and p is not member of the group */
155 if (memcg && !task_in_mem_cgroup(p, memcg))
156 return true;
157
158 /* p may not have freeable memory in nodemask */
159 if (!has_intersects_mems_allowed(p, nodemask))
160 return true;
161
162 return false;
163}
164
165/*
166 * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
167 * than all user memory (LRU pages)
168 */
169static bool is_dump_unreclaim_slabs(void)
170{
171 unsigned long nr_lru;
172
173 nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
174 global_node_page_state(NR_INACTIVE_ANON) +
175 global_node_page_state(NR_ACTIVE_FILE) +
176 global_node_page_state(NR_INACTIVE_FILE) +
177 global_node_page_state(NR_ISOLATED_ANON) +
178 global_node_page_state(NR_ISOLATED_FILE) +
179 global_node_page_state(NR_UNEVICTABLE);
180
181 return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
182}
183
184/**
185 * oom_badness - heuristic function to determine which candidate task to kill
186 * @p: task struct of which task we should calculate
187 * @totalpages: total present RAM allowed for page allocation
188 * @memcg: task's memory controller, if constrained
189 * @nodemask: nodemask passed to page allocator for mempolicy ooms
190 *
191 * The heuristic for determining which task to kill is made to be as simple and
192 * predictable as possible. The goal is to return the highest value for the
193 * task consuming the most memory to avoid subsequent oom failures.
194 */
195unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
196 const nodemask_t *nodemask, unsigned long totalpages)
197{
198 long points;
199 long adj;
200
201 if (oom_unkillable_task(p, memcg, nodemask))
202 return 0;
203
204 p = find_lock_task_mm(p);
205 if (!p)
206 return 0;
207
208 /*
209 * Do not even consider tasks which are explicitly marked oom
210 * unkillable or have been already oom reaped or the are in
211 * the middle of vfork
212 */
213 adj = (long)p->signal->oom_score_adj;
214 if (adj == OOM_SCORE_ADJ_MIN ||
215 test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
216 in_vfork(p)) {
217 task_unlock(p);
218 return 0;
219 }
220
221 /*
222 * The baseline for the badness score is the proportion of RAM that each
223 * task's rss, pagetable and swap space use.
224 */
225 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
226 mm_pgtables_bytes(p->mm) / PAGE_SIZE;
227 task_unlock(p);
228
229 /* Normalize to oom_score_adj units */
230 adj *= totalpages / 1000;
231 points += adj;
232
233 /*
234 * Never return 0 for an eligible task regardless of the root bonus and
235 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
236 */
237 return points > 0 ? points : 1;
238}
239
240enum oom_constraint {
241 CONSTRAINT_NONE,
242 CONSTRAINT_CPUSET,
243 CONSTRAINT_MEMORY_POLICY,
244 CONSTRAINT_MEMCG,
245};
246
247/*
248 * Determine the type of allocation constraint.
249 */
250static enum oom_constraint constrained_alloc(struct oom_control *oc)
251{
252 struct zone *zone;
253 struct zoneref *z;
254 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
255 bool cpuset_limited = false;
256 int nid;
257
258 if (is_memcg_oom(oc)) {
259 oc->totalpages = mem_cgroup_get_limit(oc->memcg) ?: 1;
260 return CONSTRAINT_MEMCG;
261 }
262
263 /* Default to all available memory */
264 oc->totalpages = totalram_pages + total_swap_pages;
265
266 if (!IS_ENABLED(CONFIG_NUMA))
267 return CONSTRAINT_NONE;
268
269 if (!oc->zonelist)
270 return CONSTRAINT_NONE;
271 /*
272 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
273 * to kill current.We have to random task kill in this case.
274 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
275 */
276 if (oc->gfp_mask & __GFP_THISNODE)
277 return CONSTRAINT_NONE;
278
279 /*
280 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
281 * the page allocator means a mempolicy is in effect. Cpuset policy
282 * is enforced in get_page_from_freelist().
283 */
284 if (oc->nodemask &&
285 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
286 oc->totalpages = total_swap_pages;
287 for_each_node_mask(nid, *oc->nodemask)
288 oc->totalpages += node_spanned_pages(nid);
289 return CONSTRAINT_MEMORY_POLICY;
290 }
291
292 /* Check this allocation failure is caused by cpuset's wall function */
293 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
294 high_zoneidx, oc->nodemask)
295 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
296 cpuset_limited = true;
297
298 if (cpuset_limited) {
299 oc->totalpages = total_swap_pages;
300 for_each_node_mask(nid, cpuset_current_mems_allowed)
301 oc->totalpages += node_spanned_pages(nid);
302 return CONSTRAINT_CPUSET;
303 }
304 return CONSTRAINT_NONE;
305}
306
307static int oom_evaluate_task(struct task_struct *task, void *arg)
308{
309 struct oom_control *oc = arg;
310 unsigned long points;
311
312 if (oom_unkillable_task(task, NULL, oc->nodemask))
313 goto next;
314
315 /*
316 * This task already has access to memory reserves and is being killed.
317 * Don't allow any other task to have access to the reserves unless
318 * the task has MMF_OOM_SKIP because chances that it would release
319 * any memory is quite low.
320 */
321 if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
322 if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
323 goto next;
324 goto abort;
325 }
326
327 /*
328 * If task is allocating a lot of memory and has been marked to be
329 * killed first if it triggers an oom, then select it.
330 */
331 if (oom_task_origin(task)) {
332 points = ULONG_MAX;
333 goto select;
334 }
335
336 points = oom_badness(task, NULL, oc->nodemask, oc->totalpages);
337 if (!points || points < oc->chosen_points)
338 goto next;
339
340 /* Prefer thread group leaders for display purposes */
341 if (points == oc->chosen_points && thread_group_leader(oc->chosen))
342 goto next;
343select:
344 if (oc->chosen)
345 put_task_struct(oc->chosen);
346 get_task_struct(task);
347 oc->chosen = task;
348 oc->chosen_points = points;
349next:
350 return 0;
351abort:
352 if (oc->chosen)
353 put_task_struct(oc->chosen);
354 oc->chosen = (void *)-1UL;
355 return 1;
356}
357
358/*
359 * Simple selection loop. We choose the process with the highest number of
360 * 'points'. In case scan was aborted, oc->chosen is set to -1.
361 */
362static void select_bad_process(struct oom_control *oc)
363{
364 if (is_memcg_oom(oc))
365 mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
366 else {
367 struct task_struct *p;
368
369 rcu_read_lock();
370 for_each_process(p)
371 if (oom_evaluate_task(p, oc))
372 break;
373 rcu_read_unlock();
374 }
375
376 oc->chosen_points = oc->chosen_points * 1000 / oc->totalpages;
377}
378
379/**
380 * dump_tasks - dump current memory state of all system tasks
381 * @memcg: current's memory controller, if constrained
382 * @nodemask: nodemask passed to page allocator for mempolicy ooms
383 *
384 * Dumps the current memory state of all eligible tasks. Tasks not in the same
385 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
386 * are not shown.
387 * State information includes task's pid, uid, tgid, vm size, rss,
388 * pgtables_bytes, swapents, oom_score_adj value, and name.
389 */
390static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
391{
392 struct task_struct *p;
393 struct task_struct *task;
394
395 pr_info("[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name\n");
396 rcu_read_lock();
397 for_each_process(p) {
398 if (oom_unkillable_task(p, memcg, nodemask))
399 continue;
400
401 task = find_lock_task_mm(p);
402 if (!task) {
403 /*
404 * This is a kthread or all of p's threads have already
405 * detached their mm's. There's no need to report
406 * them; they can't be oom killed anyway.
407 */
408 continue;
409 }
410
411 pr_info("[%5d] %5d %5d %8lu %8lu %8ld %8lu %5hd %s\n",
412 task->pid, from_kuid(&init_user_ns, task_uid(task)),
413 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
414 mm_pgtables_bytes(task->mm),
415 get_mm_counter(task->mm, MM_SWAPENTS),
416 task->signal->oom_score_adj, task->comm);
417 task_unlock(task);
418 }
419 rcu_read_unlock();
420}
421
422static void dump_header(struct oom_control *oc, struct task_struct *p)
423{
424 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=%*pbl, order=%d, oom_score_adj=%hd\n",
425 current->comm, oc->gfp_mask, &oc->gfp_mask,
426 nodemask_pr_args(oc->nodemask), oc->order,
427 current->signal->oom_score_adj);
428 if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
429 pr_warn("COMPACTION is disabled!!!\n");
430
431 cpuset_print_current_mems_allowed();
432 dump_stack();
433 if (is_memcg_oom(oc))
434 mem_cgroup_print_oom_info(oc->memcg, p);
435 else {
436 show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
437 if (is_dump_unreclaim_slabs())
438 dump_unreclaimable_slab();
439 }
440 if (sysctl_oom_dump_tasks)
441 dump_tasks(oc->memcg, oc->nodemask);
442}
443
444/*
445 * Number of OOM victims in flight
446 */
447static atomic_t oom_victims = ATOMIC_INIT(0);
448static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
449
450static bool oom_killer_disabled __read_mostly;
451
452#define K(x) ((x) << (PAGE_SHIFT-10))
453
454/*
455 * task->mm can be NULL if the task is the exited group leader. So to
456 * determine whether the task is using a particular mm, we examine all the
457 * task's threads: if one of those is using this mm then this task was also
458 * using it.
459 */
460bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
461{
462 struct task_struct *t;
463
464 for_each_thread(p, t) {
465 struct mm_struct *t_mm = READ_ONCE(t->mm);
466 if (t_mm)
467 return t_mm == mm;
468 }
469 return false;
470}
471
472#ifdef CONFIG_MMU
473/*
474 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
475 * victim (if that is possible) to help the OOM killer to move on.
476 */
477static struct task_struct *oom_reaper_th;
478static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
479static struct task_struct *oom_reaper_list;
480static DEFINE_SPINLOCK(oom_reaper_lock);
481
482void __oom_reap_task_mm(struct mm_struct *mm)
483{
484 struct vm_area_struct *vma;
485
486 /*
487 * Tell all users of get_user/copy_from_user etc... that the content
488 * is no longer stable. No barriers really needed because unmapping
489 * should imply barriers already and the reader would hit a page fault
490 * if it stumbled over a reaped memory.
491 */
492 set_bit(MMF_UNSTABLE, &mm->flags);
493
494 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
495 if (!can_madv_dontneed_vma(vma))
496 continue;
497
498 /*
499 * Only anonymous pages have a good chance to be dropped
500 * without additional steps which we cannot afford as we
501 * are OOM already.
502 *
503 * We do not even care about fs backed pages because all
504 * which are reclaimable have already been reclaimed and
505 * we do not want to block exit_mmap by keeping mm ref
506 * count elevated without a good reason.
507 */
508 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
509 const unsigned long start = vma->vm_start;
510 const unsigned long end = vma->vm_end;
511 struct mmu_gather tlb;
512
513 tlb_gather_mmu(&tlb, mm, start, end);
514 mmu_notifier_invalidate_range_start(mm, start, end);
515 unmap_page_range(&tlb, vma, start, end, NULL);
516 mmu_notifier_invalidate_range_end(mm, start, end);
517 tlb_finish_mmu(&tlb, start, end);
518 }
519 }
520}
521
522static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
523{
524 bool ret = true;
525
526 /*
527 * We have to make sure to not race with the victim exit path
528 * and cause premature new oom victim selection:
529 * oom_reap_task_mm exit_mm
530 * mmget_not_zero
531 * mmput
532 * atomic_dec_and_test
533 * exit_oom_victim
534 * [...]
535 * out_of_memory
536 * select_bad_process
537 * # no TIF_MEMDIE task selects new victim
538 * unmap_page_range # frees some memory
539 */
540 mutex_lock(&oom_lock);
541
542 if (!down_read_trylock(&mm->mmap_sem)) {
543 ret = false;
544 trace_skip_task_reaping(tsk->pid);
545 goto unlock_oom;
546 }
547
548 /*
549 * If the mm has invalidate_{start,end}() notifiers that could block,
550 * sleep to give the oom victim some more time.
551 * TODO: we really want to get rid of this ugly hack and make sure that
552 * notifiers cannot block for unbounded amount of time
553 */
554 if (mm_has_blockable_invalidate_notifiers(mm)) {
555 up_read(&mm->mmap_sem);
556 schedule_timeout_idle(HZ);
557 goto unlock_oom;
558 }
559
560 /*
561 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
562 * work on the mm anymore. The check for MMF_OOM_SKIP must run
563 * under mmap_sem for reading because it serializes against the
564 * down_write();up_write() cycle in exit_mmap().
565 */
566 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
567 up_read(&mm->mmap_sem);
568 trace_skip_task_reaping(tsk->pid);
569 goto unlock_oom;
570 }
571
572 trace_start_task_reaping(tsk->pid);
573
574 __oom_reap_task_mm(mm);
575
576 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
577 task_pid_nr(tsk), tsk->comm,
578 K(get_mm_counter(mm, MM_ANONPAGES)),
579 K(get_mm_counter(mm, MM_FILEPAGES)),
580 K(get_mm_counter(mm, MM_SHMEMPAGES)));
581 up_read(&mm->mmap_sem);
582
583 trace_finish_task_reaping(tsk->pid);
584unlock_oom:
585 mutex_unlock(&oom_lock);
586 return ret;
587}
588
589#define MAX_OOM_REAP_RETRIES 10
590static void oom_reap_task(struct task_struct *tsk)
591{
592 int attempts = 0;
593 struct mm_struct *mm = tsk->signal->oom_mm;
594
595 /* Retry the down_read_trylock(mmap_sem) a few times */
596 while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
597 schedule_timeout_idle(HZ/10);
598
599 if (attempts <= MAX_OOM_REAP_RETRIES ||
600 test_bit(MMF_OOM_SKIP, &mm->flags))
601 goto done;
602
603 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
604 task_pid_nr(tsk), tsk->comm);
605 debug_show_all_locks();
606
607done:
608 tsk->oom_reaper_list = NULL;
609
610 /*
611 * Hide this mm from OOM killer because it has been either reaped or
612 * somebody can't call up_write(mmap_sem).
613 */
614 set_bit(MMF_OOM_SKIP, &mm->flags);
615
616 /* Drop a reference taken by wake_oom_reaper */
617 put_task_struct(tsk);
618}
619
620static int oom_reaper(void *unused)
621{
622 while (true) {
623 struct task_struct *tsk = NULL;
624
625 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
626 spin_lock(&oom_reaper_lock);
627 if (oom_reaper_list != NULL) {
628 tsk = oom_reaper_list;
629 oom_reaper_list = tsk->oom_reaper_list;
630 }
631 spin_unlock(&oom_reaper_lock);
632
633 if (tsk)
634 oom_reap_task(tsk);
635 }
636
637 return 0;
638}
639
640static void wake_oom_reaper(struct task_struct *tsk)
641{
642 /* tsk is already queued? */
643 if (tsk == oom_reaper_list || tsk->oom_reaper_list)
644 return;
645
646 get_task_struct(tsk);
647
648 spin_lock(&oom_reaper_lock);
649 tsk->oom_reaper_list = oom_reaper_list;
650 oom_reaper_list = tsk;
651 spin_unlock(&oom_reaper_lock);
652 trace_wake_reaper(tsk->pid);
653 wake_up(&oom_reaper_wait);
654}
655
656static int __init oom_init(void)
657{
658 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
659 return 0;
660}
661subsys_initcall(oom_init)
662#else
663static inline void wake_oom_reaper(struct task_struct *tsk)
664{
665}
666#endif /* CONFIG_MMU */
667
668/**
669 * mark_oom_victim - mark the given task as OOM victim
670 * @tsk: task to mark
671 *
672 * Has to be called with oom_lock held and never after
673 * oom has been disabled already.
674 *
675 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
676 * under task_lock or operate on the current).
677 */
678static void mark_oom_victim(struct task_struct *tsk)
679{
680 struct mm_struct *mm = tsk->mm;
681
682 WARN_ON(oom_killer_disabled);
683 /* OOM killer might race with memcg OOM */
684 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
685 return;
686
687 /* oom_mm is bound to the signal struct life time. */
688 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
689 mmgrab(tsk->signal->oom_mm);
690 set_bit(MMF_OOM_VICTIM, &mm->flags);
691 }
692
693 /*
694 * Make sure that the task is woken up from uninterruptible sleep
695 * if it is frozen because OOM killer wouldn't be able to free
696 * any memory and livelock. freezing_slow_path will tell the freezer
697 * that TIF_MEMDIE tasks should be ignored.
698 */
699 __thaw_task(tsk);
700 atomic_inc(&oom_victims);
701 trace_mark_victim(tsk->pid);
702}
703
704/**
705 * exit_oom_victim - note the exit of an OOM victim
706 */
707void exit_oom_victim(void)
708{
709 clear_thread_flag(TIF_MEMDIE);
710
711 if (!atomic_dec_return(&oom_victims))
712 wake_up_all(&oom_victims_wait);
713}
714
715/**
716 * oom_killer_enable - enable OOM killer
717 */
718void oom_killer_enable(void)
719{
720 oom_killer_disabled = false;
721 pr_info("OOM killer enabled.\n");
722}
723
724/**
725 * oom_killer_disable - disable OOM killer
726 * @timeout: maximum timeout to wait for oom victims in jiffies
727 *
728 * Forces all page allocations to fail rather than trigger OOM killer.
729 * Will block and wait until all OOM victims are killed or the given
730 * timeout expires.
731 *
732 * The function cannot be called when there are runnable user tasks because
733 * the userspace would see unexpected allocation failures as a result. Any
734 * new usage of this function should be consulted with MM people.
735 *
736 * Returns true if successful and false if the OOM killer cannot be
737 * disabled.
738 */
739bool oom_killer_disable(signed long timeout)
740{
741 signed long ret;
742
743 /*
744 * Make sure to not race with an ongoing OOM killer. Check that the
745 * current is not killed (possibly due to sharing the victim's memory).
746 */
747 if (mutex_lock_killable(&oom_lock))
748 return false;
749 oom_killer_disabled = true;
750 mutex_unlock(&oom_lock);
751
752 ret = wait_event_interruptible_timeout(oom_victims_wait,
753 !atomic_read(&oom_victims), timeout);
754 if (ret <= 0) {
755 oom_killer_enable();
756 return false;
757 }
758 pr_info("OOM killer disabled.\n");
759
760 return true;
761}
762
763static inline bool __task_will_free_mem(struct task_struct *task)
764{
765 struct signal_struct *sig = task->signal;
766
767 /*
768 * A coredumping process may sleep for an extended period in exit_mm(),
769 * so the oom killer cannot assume that the process will promptly exit
770 * and release memory.
771 */
772 if (sig->flags & SIGNAL_GROUP_COREDUMP)
773 return false;
774
775 if (sig->flags & SIGNAL_GROUP_EXIT)
776 return true;
777
778 if (thread_group_empty(task) && (task->flags & PF_EXITING))
779 return true;
780
781 return false;
782}
783
784/*
785 * Checks whether the given task is dying or exiting and likely to
786 * release its address space. This means that all threads and processes
787 * sharing the same mm have to be killed or exiting.
788 * Caller has to make sure that task->mm is stable (hold task_lock or
789 * it operates on the current).
790 */
791static bool task_will_free_mem(struct task_struct *task)
792{
793 struct mm_struct *mm = task->mm;
794 struct task_struct *p;
795 bool ret = true;
796
797 /*
798 * Skip tasks without mm because it might have passed its exit_mm and
799 * exit_oom_victim. oom_reaper could have rescued that but do not rely
800 * on that for now. We can consider find_lock_task_mm in future.
801 */
802 if (!mm)
803 return false;
804
805 if (!__task_will_free_mem(task))
806 return false;
807
808 /*
809 * This task has already been drained by the oom reaper so there are
810 * only small chances it will free some more
811 */
812 if (test_bit(MMF_OOM_SKIP, &mm->flags))
813 return false;
814
815 if (atomic_read(&mm->mm_users) <= 1)
816 return true;
817
818 /*
819 * Make sure that all tasks which share the mm with the given tasks
820 * are dying as well to make sure that a) nobody pins its mm and
821 * b) the task is also reapable by the oom reaper.
822 */
823 rcu_read_lock();
824 for_each_process(p) {
825 if (!process_shares_mm(p, mm))
826 continue;
827 if (same_thread_group(task, p))
828 continue;
829 ret = __task_will_free_mem(p);
830 if (!ret)
831 break;
832 }
833 rcu_read_unlock();
834
835 return ret;
836}
837
838static void oom_kill_process(struct oom_control *oc, const char *message)
839{
840 struct task_struct *p = oc->chosen;
841 unsigned int points = oc->chosen_points;
842 struct task_struct *victim = p;
843 struct task_struct *child;
844 struct task_struct *t;
845 struct mm_struct *mm;
846 unsigned int victim_points = 0;
847 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
848 DEFAULT_RATELIMIT_BURST);
849 bool can_oom_reap = true;
850
851 /*
852 * If the task is already exiting, don't alarm the sysadmin or kill
853 * its children or threads, just give it access to memory reserves
854 * so it can die quickly
855 */
856 task_lock(p);
857 if (task_will_free_mem(p)) {
858 mark_oom_victim(p);
859 wake_oom_reaper(p);
860 task_unlock(p);
861 put_task_struct(p);
862 return;
863 }
864 task_unlock(p);
865
866 if (__ratelimit(&oom_rs))
867 dump_header(oc, p);
868
869 pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
870 message, task_pid_nr(p), p->comm, points);
871
872 /*
873 * If any of p's children has a different mm and is eligible for kill,
874 * the one with the highest oom_badness() score is sacrificed for its
875 * parent. This attempts to lose the minimal amount of work done while
876 * still freeing memory.
877 */
878 read_lock(&tasklist_lock);
879 for_each_thread(p, t) {
880 list_for_each_entry(child, &t->children, sibling) {
881 unsigned int child_points;
882
883 if (process_shares_mm(child, p->mm))
884 continue;
885 /*
886 * oom_badness() returns 0 if the thread is unkillable
887 */
888 child_points = oom_badness(child,
889 oc->memcg, oc->nodemask, oc->totalpages);
890 if (child_points > victim_points) {
891 put_task_struct(victim);
892 victim = child;
893 victim_points = child_points;
894 get_task_struct(victim);
895 }
896 }
897 }
898 read_unlock(&tasklist_lock);
899
900 p = find_lock_task_mm(victim);
901 if (!p) {
902 put_task_struct(victim);
903 return;
904 } else if (victim != p) {
905 get_task_struct(p);
906 put_task_struct(victim);
907 victim = p;
908 }
909
910 /* Get a reference to safely compare mm after task_unlock(victim) */
911 mm = victim->mm;
912 mmgrab(mm);
913
914 /* Raise event before sending signal: task reaper must see this */
915 count_vm_event(OOM_KILL);
916 count_memcg_event_mm(mm, OOM_KILL);
917
918 /*
919 * We should send SIGKILL before granting access to memory reserves
920 * in order to prevent the OOM victim from depleting the memory
921 * reserves from the user space under its control.
922 */
923 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
924 mark_oom_victim(victim);
925 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
926 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
927 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
928 K(get_mm_counter(victim->mm, MM_FILEPAGES)),
929 K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
930 task_unlock(victim);
931
932 /*
933 * Kill all user processes sharing victim->mm in other thread groups, if
934 * any. They don't get access to memory reserves, though, to avoid
935 * depletion of all memory. This prevents mm->mmap_sem livelock when an
936 * oom killed thread cannot exit because it requires the semaphore and
937 * its contended by another thread trying to allocate memory itself.
938 * That thread will now get access to memory reserves since it has a
939 * pending fatal signal.
940 */
941 rcu_read_lock();
942 for_each_process(p) {
943 if (!process_shares_mm(p, mm))
944 continue;
945 if (same_thread_group(p, victim))
946 continue;
947 if (is_global_init(p)) {
948 can_oom_reap = false;
949 set_bit(MMF_OOM_SKIP, &mm->flags);
950 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
951 task_pid_nr(victim), victim->comm,
952 task_pid_nr(p), p->comm);
953 continue;
954 }
955 /*
956 * No use_mm() user needs to read from the userspace so we are
957 * ok to reap it.
958 */
959 if (unlikely(p->flags & PF_KTHREAD))
960 continue;
961 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
962 }
963 rcu_read_unlock();
964
965 if (can_oom_reap)
966 wake_oom_reaper(victim);
967
968 mmdrop(mm);
969 put_task_struct(victim);
970}
971#undef K
972
973/*
974 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
975 */
976static void check_panic_on_oom(struct oom_control *oc,
977 enum oom_constraint constraint)
978{
979 if (likely(!sysctl_panic_on_oom))
980 return;
981 if (sysctl_panic_on_oom != 2) {
982 /*
983 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
984 * does not panic for cpuset, mempolicy, or memcg allocation
985 * failures.
986 */
987 if (constraint != CONSTRAINT_NONE)
988 return;
989 }
990 /* Do not panic for oom kills triggered by sysrq */
991 if (is_sysrq_oom(oc))
992 return;
993 dump_header(oc, NULL);
994 panic("Out of memory: %s panic_on_oom is enabled\n",
995 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
996}
997
998static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
999
1000int register_oom_notifier(struct notifier_block *nb)
1001{
1002 return blocking_notifier_chain_register(&oom_notify_list, nb);
1003}
1004EXPORT_SYMBOL_GPL(register_oom_notifier);
1005
1006int unregister_oom_notifier(struct notifier_block *nb)
1007{
1008 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1009}
1010EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1011
1012/**
1013 * out_of_memory - kill the "best" process when we run out of memory
1014 * @oc: pointer to struct oom_control
1015 *
1016 * If we run out of memory, we have the choice between either
1017 * killing a random task (bad), letting the system crash (worse)
1018 * OR try to be smart about which process to kill. Note that we
1019 * don't have to be perfect here, we just have to be good.
1020 */
1021bool out_of_memory(struct oom_control *oc)
1022{
1023 unsigned long freed = 0;
1024 enum oom_constraint constraint = CONSTRAINT_NONE;
1025
1026 if (oom_killer_disabled)
1027 return false;
1028
1029 if (!is_memcg_oom(oc)) {
1030 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1031 if (freed > 0)
1032 /* Got some memory back in the last second. */
1033 return true;
1034 }
1035
1036 /*
1037 * If current has a pending SIGKILL or is exiting, then automatically
1038 * select it. The goal is to allow it to allocate so that it may
1039 * quickly exit and free its memory.
1040 */
1041 if (task_will_free_mem(current)) {
1042 mark_oom_victim(current);
1043 wake_oom_reaper(current);
1044 return true;
1045 }
1046
1047 /*
1048 * The OOM killer does not compensate for IO-less reclaim.
1049 * pagefault_out_of_memory lost its gfp context so we have to
1050 * make sure exclude 0 mask - all other users should have at least
1051 * ___GFP_DIRECT_RECLAIM to get here.
1052 */
1053 if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS))
1054 return true;
1055
1056 /*
1057 * Check if there were limitations on the allocation (only relevant for
1058 * NUMA and memcg) that may require different handling.
1059 */
1060 constraint = constrained_alloc(oc);
1061 if (constraint != CONSTRAINT_MEMORY_POLICY)
1062 oc->nodemask = NULL;
1063 check_panic_on_oom(oc, constraint);
1064
1065 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1066 current->mm && !oom_unkillable_task(current, NULL, oc->nodemask) &&
1067 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1068 get_task_struct(current);
1069 oc->chosen = current;
1070 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1071 return true;
1072 }
1073
1074 select_bad_process(oc);
1075 /* Found nothing?!?! Either we hang forever, or we panic. */
1076 if (!oc->chosen && !is_sysrq_oom(oc) && !is_memcg_oom(oc)) {
1077 dump_header(oc, NULL);
1078 panic("Out of memory and no killable processes...\n");
1079 }
1080 if (oc->chosen && oc->chosen != (void *)-1UL) {
1081 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1082 "Memory cgroup out of memory");
1083 /*
1084 * Give the killed process a good chance to exit before trying
1085 * to allocate memory again.
1086 */
1087 schedule_timeout_killable(1);
1088 }
1089 return !!oc->chosen;
1090}
1091
1092/*
1093 * The pagefault handler calls here because it is out of memory, so kill a
1094 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1095 * killing is already in progress so do nothing.
1096 */
1097void pagefault_out_of_memory(void)
1098{
1099 struct oom_control oc = {
1100 .zonelist = NULL,
1101 .nodemask = NULL,
1102 .memcg = NULL,
1103 .gfp_mask = 0,
1104 .order = 0,
1105 };
1106
1107 if (mem_cgroup_oom_synchronize(true))
1108 return;
1109
1110 if (!mutex_trylock(&oom_lock))
1111 return;
1112 out_of_memory(&oc);
1113 mutex_unlock(&oom_lock);
1114}
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}