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