Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/oom_kill.c
4 *
5 * Copyright (C) 1998,2000 Rik van Riel
6 * Thanks go out to Claus Fischer for some serious inspiration and
7 * for goading me into coding this file...
8 * Copyright (C) 2010 Google, Inc.
9 * Rewritten by David Rientjes
10 *
11 * The routines in this file are used to kill a process when
12 * we're seriously out of memory. This gets called from __alloc_pages()
13 * in mm/page_alloc.c when we really run out of memory.
14 *
15 * Since we won't call these routines often (on a well-configured
16 * machine) this file will double as a 'coding guide' and a signpost
17 * for newbie kernel hackers. It features several pointers to major
18 * kernel subsystems and hints as to where to find out what things do.
19 */
20
21#include <linux/oom.h>
22#include <linux/mm.h>
23#include <linux/err.h>
24#include <linux/gfp.h>
25#include <linux/sched.h>
26#include <linux/sched/mm.h>
27#include <linux/sched/coredump.h>
28#include <linux/sched/task.h>
29#include <linux/swap.h>
30#include <linux/timex.h>
31#include <linux/jiffies.h>
32#include <linux/cpuset.h>
33#include <linux/export.h>
34#include <linux/notifier.h>
35#include <linux/memcontrol.h>
36#include <linux/mempolicy.h>
37#include <linux/security.h>
38#include <linux/ptrace.h>
39#include <linux/freezer.h>
40#include <linux/ftrace.h>
41#include <linux/ratelimit.h>
42#include <linux/kthread.h>
43#include <linux/init.h>
44#include <linux/mmu_notifier.h>
45
46#include <asm/tlb.h>
47#include "internal.h"
48#include "slab.h"
49
50#define CREATE_TRACE_POINTS
51#include <trace/events/oom.h>
52
53int sysctl_panic_on_oom;
54int sysctl_oom_kill_allocating_task;
55int sysctl_oom_dump_tasks = 1;
56
57/*
58 * Serializes oom killer invocations (out_of_memory()) from all contexts to
59 * prevent from over eager oom killing (e.g. when the oom killer is invoked
60 * from different domains).
61 *
62 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
63 * and mark_oom_victim
64 */
65DEFINE_MUTEX(oom_lock);
66
67static inline bool is_memcg_oom(struct oom_control *oc)
68{
69 return oc->memcg != NULL;
70}
71
72#ifdef CONFIG_NUMA
73/**
74 * oom_cpuset_eligible() - check task eligiblity for kill
75 * @start: task struct of which task to consider
76 * @oc: pointer to struct oom_control
77 *
78 * Task eligibility is determined by whether or not a candidate task, @tsk,
79 * shares the same mempolicy nodes as current if it is bound by such a policy
80 * and whether or not it has the same set of allowed cpuset nodes.
81 *
82 * This function is assuming oom-killer context and 'current' has triggered
83 * the oom-killer.
84 */
85static bool oom_cpuset_eligible(struct task_struct *start,
86 struct oom_control *oc)
87{
88 struct task_struct *tsk;
89 bool ret = false;
90 const nodemask_t *mask = oc->nodemask;
91
92 if (is_memcg_oom(oc))
93 return true;
94
95 rcu_read_lock();
96 for_each_thread(start, tsk) {
97 if (mask) {
98 /*
99 * If this is a mempolicy constrained oom, tsk's
100 * cpuset is irrelevant. Only return true if its
101 * mempolicy intersects current, otherwise it may be
102 * needlessly killed.
103 */
104 ret = mempolicy_nodemask_intersects(tsk, mask);
105 } else {
106 /*
107 * This is not a mempolicy constrained oom, so only
108 * check the mems of tsk's cpuset.
109 */
110 ret = cpuset_mems_allowed_intersects(current, tsk);
111 }
112 if (ret)
113 break;
114 }
115 rcu_read_unlock();
116
117 return ret;
118}
119#else
120static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
121{
122 return true;
123}
124#endif /* CONFIG_NUMA */
125
126/*
127 * The process p may have detached its own ->mm while exiting or through
128 * use_mm(), but one or more of its subthreads may still have a valid
129 * pointer. Return p, or any of its subthreads with a valid ->mm, with
130 * task_lock() held.
131 */
132struct task_struct *find_lock_task_mm(struct task_struct *p)
133{
134 struct task_struct *t;
135
136 rcu_read_lock();
137
138 for_each_thread(p, t) {
139 task_lock(t);
140 if (likely(t->mm))
141 goto found;
142 task_unlock(t);
143 }
144 t = NULL;
145found:
146 rcu_read_unlock();
147
148 return t;
149}
150
151/*
152 * order == -1 means the oom kill is required by sysrq, otherwise only
153 * for display purposes.
154 */
155static inline bool is_sysrq_oom(struct oom_control *oc)
156{
157 return oc->order == -1;
158}
159
160/* return true if the task is not adequate as candidate victim task. */
161static bool oom_unkillable_task(struct task_struct *p)
162{
163 if (is_global_init(p))
164 return true;
165 if (p->flags & PF_KTHREAD)
166 return true;
167 return false;
168}
169
170/*
171 * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
172 * than all user memory (LRU pages)
173 */
174static bool is_dump_unreclaim_slabs(void)
175{
176 unsigned long nr_lru;
177
178 nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
179 global_node_page_state(NR_INACTIVE_ANON) +
180 global_node_page_state(NR_ACTIVE_FILE) +
181 global_node_page_state(NR_INACTIVE_FILE) +
182 global_node_page_state(NR_ISOLATED_ANON) +
183 global_node_page_state(NR_ISOLATED_FILE) +
184 global_node_page_state(NR_UNEVICTABLE);
185
186 return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
187}
188
189/**
190 * oom_badness - heuristic function to determine which candidate task to kill
191 * @p: task struct of which task we should calculate
192 * @totalpages: total present RAM allowed for page allocation
193 *
194 * The heuristic for determining which task to kill is made to be as simple and
195 * predictable as possible. The goal is to return the highest value for the
196 * task consuming the most memory to avoid subsequent oom failures.
197 */
198unsigned long oom_badness(struct task_struct *p, unsigned long totalpages)
199{
200 long points;
201 long adj;
202
203 if (oom_unkillable_task(p))
204 return 0;
205
206 p = find_lock_task_mm(p);
207 if (!p)
208 return 0;
209
210 /*
211 * Do not even consider tasks which are explicitly marked oom
212 * unkillable or have been already oom reaped or the are in
213 * the middle of vfork
214 */
215 adj = (long)p->signal->oom_score_adj;
216 if (adj == OOM_SCORE_ADJ_MIN ||
217 test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
218 in_vfork(p)) {
219 task_unlock(p);
220 return 0;
221 }
222
223 /*
224 * The baseline for the badness score is the proportion of RAM that each
225 * task's rss, pagetable and swap space use.
226 */
227 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
228 mm_pgtables_bytes(p->mm) / PAGE_SIZE;
229 task_unlock(p);
230
231 /* Normalize to oom_score_adj units */
232 adj *= totalpages / 1000;
233 points += adj;
234
235 /*
236 * Never return 0 for an eligible task regardless of the root bonus and
237 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
238 */
239 return points > 0 ? points : 1;
240}
241
242static const char * const oom_constraint_text[] = {
243 [CONSTRAINT_NONE] = "CONSTRAINT_NONE",
244 [CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
245 [CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
246 [CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
247};
248
249/*
250 * Determine the type of allocation constraint.
251 */
252static enum oom_constraint constrained_alloc(struct oom_control *oc)
253{
254 struct zone *zone;
255 struct zoneref *z;
256 enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
257 bool cpuset_limited = false;
258 int nid;
259
260 if (is_memcg_oom(oc)) {
261 oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
262 return CONSTRAINT_MEMCG;
263 }
264
265 /* Default to all available memory */
266 oc->totalpages = totalram_pages() + total_swap_pages;
267
268 if (!IS_ENABLED(CONFIG_NUMA))
269 return CONSTRAINT_NONE;
270
271 if (!oc->zonelist)
272 return CONSTRAINT_NONE;
273 /*
274 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
275 * to kill current.We have to random task kill in this case.
276 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
277 */
278 if (oc->gfp_mask & __GFP_THISNODE)
279 return CONSTRAINT_NONE;
280
281 /*
282 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
283 * the page allocator means a mempolicy is in effect. Cpuset policy
284 * is enforced in get_page_from_freelist().
285 */
286 if (oc->nodemask &&
287 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
288 oc->totalpages = total_swap_pages;
289 for_each_node_mask(nid, *oc->nodemask)
290 oc->totalpages += node_present_pages(nid);
291 return CONSTRAINT_MEMORY_POLICY;
292 }
293
294 /* Check this allocation failure is caused by cpuset's wall function */
295 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
296 high_zoneidx, oc->nodemask)
297 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
298 cpuset_limited = true;
299
300 if (cpuset_limited) {
301 oc->totalpages = total_swap_pages;
302 for_each_node_mask(nid, cpuset_current_mems_allowed)
303 oc->totalpages += node_present_pages(nid);
304 return CONSTRAINT_CPUSET;
305 }
306 return CONSTRAINT_NONE;
307}
308
309static int oom_evaluate_task(struct task_struct *task, void *arg)
310{
311 struct oom_control *oc = arg;
312 unsigned long points;
313
314 if (oom_unkillable_task(task))
315 goto next;
316
317 /* p may not have freeable memory in nodemask */
318 if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
319 goto next;
320
321 /*
322 * This task already has access to memory reserves and is being killed.
323 * Don't allow any other task to have access to the reserves unless
324 * the task has MMF_OOM_SKIP because chances that it would release
325 * any memory is quite low.
326 */
327 if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
328 if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
329 goto next;
330 goto abort;
331 }
332
333 /*
334 * If task is allocating a lot of memory and has been marked to be
335 * killed first if it triggers an oom, then select it.
336 */
337 if (oom_task_origin(task)) {
338 points = ULONG_MAX;
339 goto select;
340 }
341
342 points = oom_badness(task, oc->totalpages);
343 if (!points || points < oc->chosen_points)
344 goto next;
345
346select:
347 if (oc->chosen)
348 put_task_struct(oc->chosen);
349 get_task_struct(task);
350 oc->chosen = task;
351 oc->chosen_points = points;
352next:
353 return 0;
354abort:
355 if (oc->chosen)
356 put_task_struct(oc->chosen);
357 oc->chosen = (void *)-1UL;
358 return 1;
359}
360
361/*
362 * Simple selection loop. We choose the process with the highest number of
363 * 'points'. In case scan was aborted, oc->chosen is set to -1.
364 */
365static void select_bad_process(struct oom_control *oc)
366{
367 if (is_memcg_oom(oc))
368 mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
369 else {
370 struct task_struct *p;
371
372 rcu_read_lock();
373 for_each_process(p)
374 if (oom_evaluate_task(p, oc))
375 break;
376 rcu_read_unlock();
377 }
378}
379
380static int dump_task(struct task_struct *p, void *arg)
381{
382 struct oom_control *oc = arg;
383 struct task_struct *task;
384
385 if (oom_unkillable_task(p))
386 return 0;
387
388 /* p may not have freeable memory in nodemask */
389 if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
390 return 0;
391
392 task = find_lock_task_mm(p);
393 if (!task) {
394 /*
395 * This is a kthread or all of p's threads have already
396 * detached their mm's. There's no need to report
397 * them; they can't be oom killed anyway.
398 */
399 return 0;
400 }
401
402 pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu %5hd %s\n",
403 task->pid, from_kuid(&init_user_ns, task_uid(task)),
404 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
405 mm_pgtables_bytes(task->mm),
406 get_mm_counter(task->mm, MM_SWAPENTS),
407 task->signal->oom_score_adj, task->comm);
408 task_unlock(task);
409
410 return 0;
411}
412
413/**
414 * dump_tasks - dump current memory state of all system tasks
415 * @oc: pointer to struct oom_control
416 *
417 * Dumps the current memory state of all eligible tasks. Tasks not in the same
418 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
419 * are not shown.
420 * State information includes task's pid, uid, tgid, vm size, rss,
421 * pgtables_bytes, swapents, oom_score_adj value, and name.
422 */
423static void dump_tasks(struct oom_control *oc)
424{
425 pr_info("Tasks state (memory values in pages):\n");
426 pr_info("[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name\n");
427
428 if (is_memcg_oom(oc))
429 mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
430 else {
431 struct task_struct *p;
432
433 rcu_read_lock();
434 for_each_process(p)
435 dump_task(p, oc);
436 rcu_read_unlock();
437 }
438}
439
440static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
441{
442 /* one line summary of the oom killer context. */
443 pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
444 oom_constraint_text[oc->constraint],
445 nodemask_pr_args(oc->nodemask));
446 cpuset_print_current_mems_allowed();
447 mem_cgroup_print_oom_context(oc->memcg, victim);
448 pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
449 from_kuid(&init_user_ns, task_uid(victim)));
450}
451
452static void dump_header(struct oom_control *oc, struct task_struct *p)
453{
454 pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
455 current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
456 current->signal->oom_score_adj);
457 if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
458 pr_warn("COMPACTION is disabled!!!\n");
459
460 dump_stack();
461 if (is_memcg_oom(oc))
462 mem_cgroup_print_oom_meminfo(oc->memcg);
463 else {
464 show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
465 if (is_dump_unreclaim_slabs())
466 dump_unreclaimable_slab();
467 }
468 if (sysctl_oom_dump_tasks)
469 dump_tasks(oc);
470 if (p)
471 dump_oom_summary(oc, p);
472}
473
474/*
475 * Number of OOM victims in flight
476 */
477static atomic_t oom_victims = ATOMIC_INIT(0);
478static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
479
480static bool oom_killer_disabled __read_mostly;
481
482#define K(x) ((x) << (PAGE_SHIFT-10))
483
484/*
485 * task->mm can be NULL if the task is the exited group leader. So to
486 * determine whether the task is using a particular mm, we examine all the
487 * task's threads: if one of those is using this mm then this task was also
488 * using it.
489 */
490bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
491{
492 struct task_struct *t;
493
494 for_each_thread(p, t) {
495 struct mm_struct *t_mm = READ_ONCE(t->mm);
496 if (t_mm)
497 return t_mm == mm;
498 }
499 return false;
500}
501
502#ifdef CONFIG_MMU
503/*
504 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
505 * victim (if that is possible) to help the OOM killer to move on.
506 */
507static struct task_struct *oom_reaper_th;
508static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
509static struct task_struct *oom_reaper_list;
510static DEFINE_SPINLOCK(oom_reaper_lock);
511
512bool __oom_reap_task_mm(struct mm_struct *mm)
513{
514 struct vm_area_struct *vma;
515 bool ret = true;
516
517 /*
518 * Tell all users of get_user/copy_from_user etc... that the content
519 * is no longer stable. No barriers really needed because unmapping
520 * should imply barriers already and the reader would hit a page fault
521 * if it stumbled over a reaped memory.
522 */
523 set_bit(MMF_UNSTABLE, &mm->flags);
524
525 for (vma = mm->mmap ; vma; vma = vma->vm_next) {
526 if (!can_madv_lru_vma(vma))
527 continue;
528
529 /*
530 * Only anonymous pages have a good chance to be dropped
531 * without additional steps which we cannot afford as we
532 * are OOM already.
533 *
534 * We do not even care about fs backed pages because all
535 * which are reclaimable have already been reclaimed and
536 * we do not want to block exit_mmap by keeping mm ref
537 * count elevated without a good reason.
538 */
539 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
540 struct mmu_notifier_range range;
541 struct mmu_gather tlb;
542
543 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
544 vma, mm, vma->vm_start,
545 vma->vm_end);
546 tlb_gather_mmu(&tlb, mm, range.start, range.end);
547 if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
548 tlb_finish_mmu(&tlb, range.start, range.end);
549 ret = false;
550 continue;
551 }
552 unmap_page_range(&tlb, vma, range.start, range.end, NULL);
553 mmu_notifier_invalidate_range_end(&range);
554 tlb_finish_mmu(&tlb, range.start, range.end);
555 }
556 }
557
558 return ret;
559}
560
561/*
562 * Reaps the address space of the give task.
563 *
564 * Returns true on success and false if none or part of the address space
565 * has been reclaimed and the caller should retry later.
566 */
567static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
568{
569 bool ret = true;
570
571 if (!down_read_trylock(&mm->mmap_sem)) {
572 trace_skip_task_reaping(tsk->pid);
573 return false;
574 }
575
576 /*
577 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
578 * work on the mm anymore. The check for MMF_OOM_SKIP must run
579 * under mmap_sem for reading because it serializes against the
580 * down_write();up_write() cycle in exit_mmap().
581 */
582 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
583 trace_skip_task_reaping(tsk->pid);
584 goto out_unlock;
585 }
586
587 trace_start_task_reaping(tsk->pid);
588
589 /* failed to reap part of the address space. Try again later */
590 ret = __oom_reap_task_mm(mm);
591 if (!ret)
592 goto out_finish;
593
594 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
595 task_pid_nr(tsk), tsk->comm,
596 K(get_mm_counter(mm, MM_ANONPAGES)),
597 K(get_mm_counter(mm, MM_FILEPAGES)),
598 K(get_mm_counter(mm, MM_SHMEMPAGES)));
599out_finish:
600 trace_finish_task_reaping(tsk->pid);
601out_unlock:
602 up_read(&mm->mmap_sem);
603
604 return ret;
605}
606
607#define MAX_OOM_REAP_RETRIES 10
608static void oom_reap_task(struct task_struct *tsk)
609{
610 int attempts = 0;
611 struct mm_struct *mm = tsk->signal->oom_mm;
612
613 /* Retry the down_read_trylock(mmap_sem) a few times */
614 while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
615 schedule_timeout_idle(HZ/10);
616
617 if (attempts <= MAX_OOM_REAP_RETRIES ||
618 test_bit(MMF_OOM_SKIP, &mm->flags))
619 goto done;
620
621 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
622 task_pid_nr(tsk), tsk->comm);
623 debug_show_all_locks();
624
625done:
626 tsk->oom_reaper_list = NULL;
627
628 /*
629 * Hide this mm from OOM killer because it has been either reaped or
630 * somebody can't call up_write(mmap_sem).
631 */
632 set_bit(MMF_OOM_SKIP, &mm->flags);
633
634 /* Drop a reference taken by wake_oom_reaper */
635 put_task_struct(tsk);
636}
637
638static int oom_reaper(void *unused)
639{
640 while (true) {
641 struct task_struct *tsk = NULL;
642
643 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
644 spin_lock(&oom_reaper_lock);
645 if (oom_reaper_list != NULL) {
646 tsk = oom_reaper_list;
647 oom_reaper_list = tsk->oom_reaper_list;
648 }
649 spin_unlock(&oom_reaper_lock);
650
651 if (tsk)
652 oom_reap_task(tsk);
653 }
654
655 return 0;
656}
657
658static void wake_oom_reaper(struct task_struct *tsk)
659{
660 /* mm is already queued? */
661 if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
662 return;
663
664 get_task_struct(tsk);
665
666 spin_lock(&oom_reaper_lock);
667 tsk->oom_reaper_list = oom_reaper_list;
668 oom_reaper_list = tsk;
669 spin_unlock(&oom_reaper_lock);
670 trace_wake_reaper(tsk->pid);
671 wake_up(&oom_reaper_wait);
672}
673
674static int __init oom_init(void)
675{
676 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
677 return 0;
678}
679subsys_initcall(oom_init)
680#else
681static inline void wake_oom_reaper(struct task_struct *tsk)
682{
683}
684#endif /* CONFIG_MMU */
685
686/**
687 * mark_oom_victim - mark the given task as OOM victim
688 * @tsk: task to mark
689 *
690 * Has to be called with oom_lock held and never after
691 * oom has been disabled already.
692 *
693 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
694 * under task_lock or operate on the current).
695 */
696static void mark_oom_victim(struct task_struct *tsk)
697{
698 struct mm_struct *mm = tsk->mm;
699
700 WARN_ON(oom_killer_disabled);
701 /* OOM killer might race with memcg OOM */
702 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
703 return;
704
705 /* oom_mm is bound to the signal struct life time. */
706 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
707 mmgrab(tsk->signal->oom_mm);
708 set_bit(MMF_OOM_VICTIM, &mm->flags);
709 }
710
711 /*
712 * Make sure that the task is woken up from uninterruptible sleep
713 * if it is frozen because OOM killer wouldn't be able to free
714 * any memory and livelock. freezing_slow_path will tell the freezer
715 * that TIF_MEMDIE tasks should be ignored.
716 */
717 __thaw_task(tsk);
718 atomic_inc(&oom_victims);
719 trace_mark_victim(tsk->pid);
720}
721
722/**
723 * exit_oom_victim - note the exit of an OOM victim
724 */
725void exit_oom_victim(void)
726{
727 clear_thread_flag(TIF_MEMDIE);
728
729 if (!atomic_dec_return(&oom_victims))
730 wake_up_all(&oom_victims_wait);
731}
732
733/**
734 * oom_killer_enable - enable OOM killer
735 */
736void oom_killer_enable(void)
737{
738 oom_killer_disabled = false;
739 pr_info("OOM killer enabled.\n");
740}
741
742/**
743 * oom_killer_disable - disable OOM killer
744 * @timeout: maximum timeout to wait for oom victims in jiffies
745 *
746 * Forces all page allocations to fail rather than trigger OOM killer.
747 * Will block and wait until all OOM victims are killed or the given
748 * timeout expires.
749 *
750 * The function cannot be called when there are runnable user tasks because
751 * the userspace would see unexpected allocation failures as a result. Any
752 * new usage of this function should be consulted with MM people.
753 *
754 * Returns true if successful and false if the OOM killer cannot be
755 * disabled.
756 */
757bool oom_killer_disable(signed long timeout)
758{
759 signed long ret;
760
761 /*
762 * Make sure to not race with an ongoing OOM killer. Check that the
763 * current is not killed (possibly due to sharing the victim's memory).
764 */
765 if (mutex_lock_killable(&oom_lock))
766 return false;
767 oom_killer_disabled = true;
768 mutex_unlock(&oom_lock);
769
770 ret = wait_event_interruptible_timeout(oom_victims_wait,
771 !atomic_read(&oom_victims), timeout);
772 if (ret <= 0) {
773 oom_killer_enable();
774 return false;
775 }
776 pr_info("OOM killer disabled.\n");
777
778 return true;
779}
780
781static inline bool __task_will_free_mem(struct task_struct *task)
782{
783 struct signal_struct *sig = task->signal;
784
785 /*
786 * A coredumping process may sleep for an extended period in exit_mm(),
787 * so the oom killer cannot assume that the process will promptly exit
788 * and release memory.
789 */
790 if (sig->flags & SIGNAL_GROUP_COREDUMP)
791 return false;
792
793 if (sig->flags & SIGNAL_GROUP_EXIT)
794 return true;
795
796 if (thread_group_empty(task) && (task->flags & PF_EXITING))
797 return true;
798
799 return false;
800}
801
802/*
803 * Checks whether the given task is dying or exiting and likely to
804 * release its address space. This means that all threads and processes
805 * sharing the same mm have to be killed or exiting.
806 * Caller has to make sure that task->mm is stable (hold task_lock or
807 * it operates on the current).
808 */
809static bool task_will_free_mem(struct task_struct *task)
810{
811 struct mm_struct *mm = task->mm;
812 struct task_struct *p;
813 bool ret = true;
814
815 /*
816 * Skip tasks without mm because it might have passed its exit_mm and
817 * exit_oom_victim. oom_reaper could have rescued that but do not rely
818 * on that for now. We can consider find_lock_task_mm in future.
819 */
820 if (!mm)
821 return false;
822
823 if (!__task_will_free_mem(task))
824 return false;
825
826 /*
827 * This task has already been drained by the oom reaper so there are
828 * only small chances it will free some more
829 */
830 if (test_bit(MMF_OOM_SKIP, &mm->flags))
831 return false;
832
833 if (atomic_read(&mm->mm_users) <= 1)
834 return true;
835
836 /*
837 * Make sure that all tasks which share the mm with the given tasks
838 * are dying as well to make sure that a) nobody pins its mm and
839 * b) the task is also reapable by the oom reaper.
840 */
841 rcu_read_lock();
842 for_each_process(p) {
843 if (!process_shares_mm(p, mm))
844 continue;
845 if (same_thread_group(task, p))
846 continue;
847 ret = __task_will_free_mem(p);
848 if (!ret)
849 break;
850 }
851 rcu_read_unlock();
852
853 return ret;
854}
855
856static void __oom_kill_process(struct task_struct *victim, const char *message)
857{
858 struct task_struct *p;
859 struct mm_struct *mm;
860 bool can_oom_reap = true;
861
862 p = find_lock_task_mm(victim);
863 if (!p) {
864 put_task_struct(victim);
865 return;
866 } else if (victim != p) {
867 get_task_struct(p);
868 put_task_struct(victim);
869 victim = p;
870 }
871
872 /* Get a reference to safely compare mm after task_unlock(victim) */
873 mm = victim->mm;
874 mmgrab(mm);
875
876 /* Raise event before sending signal: task reaper must see this */
877 count_vm_event(OOM_KILL);
878 memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
879
880 /*
881 * We should send SIGKILL before granting access to memory reserves
882 * in order to prevent the OOM victim from depleting the memory
883 * reserves from the user space under its control.
884 */
885 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
886 mark_oom_victim(victim);
887 pr_err("%s: Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB, UID:%u pgtables:%lukB oom_score_adj:%hd\n",
888 message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
889 K(get_mm_counter(mm, MM_ANONPAGES)),
890 K(get_mm_counter(mm, MM_FILEPAGES)),
891 K(get_mm_counter(mm, MM_SHMEMPAGES)),
892 from_kuid(&init_user_ns, task_uid(victim)),
893 mm_pgtables_bytes(mm), victim->signal->oom_score_adj);
894 task_unlock(victim);
895
896 /*
897 * Kill all user processes sharing victim->mm in other thread groups, if
898 * any. They don't get access to memory reserves, though, to avoid
899 * depletion of all memory. This prevents mm->mmap_sem livelock when an
900 * oom killed thread cannot exit because it requires the semaphore and
901 * its contended by another thread trying to allocate memory itself.
902 * That thread will now get access to memory reserves since it has a
903 * pending fatal signal.
904 */
905 rcu_read_lock();
906 for_each_process(p) {
907 if (!process_shares_mm(p, mm))
908 continue;
909 if (same_thread_group(p, victim))
910 continue;
911 if (is_global_init(p)) {
912 can_oom_reap = false;
913 set_bit(MMF_OOM_SKIP, &mm->flags);
914 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
915 task_pid_nr(victim), victim->comm,
916 task_pid_nr(p), p->comm);
917 continue;
918 }
919 /*
920 * No use_mm() user needs to read from the userspace so we are
921 * ok to reap it.
922 */
923 if (unlikely(p->flags & PF_KTHREAD))
924 continue;
925 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
926 }
927 rcu_read_unlock();
928
929 if (can_oom_reap)
930 wake_oom_reaper(victim);
931
932 mmdrop(mm);
933 put_task_struct(victim);
934}
935#undef K
936
937/*
938 * Kill provided task unless it's secured by setting
939 * oom_score_adj to OOM_SCORE_ADJ_MIN.
940 */
941static int oom_kill_memcg_member(struct task_struct *task, void *message)
942{
943 if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
944 !is_global_init(task)) {
945 get_task_struct(task);
946 __oom_kill_process(task, message);
947 }
948 return 0;
949}
950
951static void oom_kill_process(struct oom_control *oc, const char *message)
952{
953 struct task_struct *victim = oc->chosen;
954 struct mem_cgroup *oom_group;
955 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
956 DEFAULT_RATELIMIT_BURST);
957
958 /*
959 * If the task is already exiting, don't alarm the sysadmin or kill
960 * its children or threads, just give it access to memory reserves
961 * so it can die quickly
962 */
963 task_lock(victim);
964 if (task_will_free_mem(victim)) {
965 mark_oom_victim(victim);
966 wake_oom_reaper(victim);
967 task_unlock(victim);
968 put_task_struct(victim);
969 return;
970 }
971 task_unlock(victim);
972
973 if (__ratelimit(&oom_rs))
974 dump_header(oc, victim);
975
976 /*
977 * Do we need to kill the entire memory cgroup?
978 * Or even one of the ancestor memory cgroups?
979 * Check this out before killing the victim task.
980 */
981 oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
982
983 __oom_kill_process(victim, message);
984
985 /*
986 * If necessary, kill all tasks in the selected memory cgroup.
987 */
988 if (oom_group) {
989 mem_cgroup_print_oom_group(oom_group);
990 mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
991 (void*)message);
992 mem_cgroup_put(oom_group);
993 }
994}
995
996/*
997 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
998 */
999static void check_panic_on_oom(struct oom_control *oc)
1000{
1001 if (likely(!sysctl_panic_on_oom))
1002 return;
1003 if (sysctl_panic_on_oom != 2) {
1004 /*
1005 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1006 * does not panic for cpuset, mempolicy, or memcg allocation
1007 * failures.
1008 */
1009 if (oc->constraint != CONSTRAINT_NONE)
1010 return;
1011 }
1012 /* Do not panic for oom kills triggered by sysrq */
1013 if (is_sysrq_oom(oc))
1014 return;
1015 dump_header(oc, NULL);
1016 panic("Out of memory: %s panic_on_oom is enabled\n",
1017 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1018}
1019
1020static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1021
1022int register_oom_notifier(struct notifier_block *nb)
1023{
1024 return blocking_notifier_chain_register(&oom_notify_list, nb);
1025}
1026EXPORT_SYMBOL_GPL(register_oom_notifier);
1027
1028int unregister_oom_notifier(struct notifier_block *nb)
1029{
1030 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1031}
1032EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1033
1034/**
1035 * out_of_memory - kill the "best" process when we run out of memory
1036 * @oc: pointer to struct oom_control
1037 *
1038 * If we run out of memory, we have the choice between either
1039 * killing a random task (bad), letting the system crash (worse)
1040 * OR try to be smart about which process to kill. Note that we
1041 * don't have to be perfect here, we just have to be good.
1042 */
1043bool out_of_memory(struct oom_control *oc)
1044{
1045 unsigned long freed = 0;
1046
1047 if (oom_killer_disabled)
1048 return false;
1049
1050 if (!is_memcg_oom(oc)) {
1051 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1052 if (freed > 0)
1053 /* Got some memory back in the last second. */
1054 return true;
1055 }
1056
1057 /*
1058 * If current has a pending SIGKILL or is exiting, then automatically
1059 * select it. The goal is to allow it to allocate so that it may
1060 * quickly exit and free its memory.
1061 */
1062 if (task_will_free_mem(current)) {
1063 mark_oom_victim(current);
1064 wake_oom_reaper(current);
1065 return true;
1066 }
1067
1068 /*
1069 * The OOM killer does not compensate for IO-less reclaim.
1070 * pagefault_out_of_memory lost its gfp context so we have to
1071 * make sure exclude 0 mask - all other users should have at least
1072 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1073 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1074 */
1075 if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1076 return true;
1077
1078 /*
1079 * Check if there were limitations on the allocation (only relevant for
1080 * NUMA and memcg) that may require different handling.
1081 */
1082 oc->constraint = constrained_alloc(oc);
1083 if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1084 oc->nodemask = NULL;
1085 check_panic_on_oom(oc);
1086
1087 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1088 current->mm && !oom_unkillable_task(current) &&
1089 oom_cpuset_eligible(current, oc) &&
1090 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1091 get_task_struct(current);
1092 oc->chosen = current;
1093 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1094 return true;
1095 }
1096
1097 select_bad_process(oc);
1098 /* Found nothing?!?! */
1099 if (!oc->chosen) {
1100 dump_header(oc, NULL);
1101 pr_warn("Out of memory and no killable processes...\n");
1102 /*
1103 * If we got here due to an actual allocation at the
1104 * system level, we cannot survive this and will enter
1105 * an endless loop in the allocator. Bail out now.
1106 */
1107 if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1108 panic("System is deadlocked on memory\n");
1109 }
1110 if (oc->chosen && oc->chosen != (void *)-1UL)
1111 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1112 "Memory cgroup out of memory");
1113 return !!oc->chosen;
1114}
1115
1116/*
1117 * The pagefault handler calls here because it is out of memory, so kill a
1118 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1119 * killing is already in progress so do nothing.
1120 */
1121void pagefault_out_of_memory(void)
1122{
1123 struct oom_control oc = {
1124 .zonelist = NULL,
1125 .nodemask = NULL,
1126 .memcg = NULL,
1127 .gfp_mask = 0,
1128 .order = 0,
1129 };
1130
1131 if (mem_cgroup_oom_synchronize(true))
1132 return;
1133
1134 if (!mutex_trylock(&oom_lock))
1135 return;
1136 out_of_memory(&oc);
1137 mutex_unlock(&oom_lock);
1138}
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}