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/sched/debug.h>
30#include <linux/swap.h>
31#include <linux/syscalls.h>
32#include <linux/timex.h>
33#include <linux/jiffies.h>
34#include <linux/cpuset.h>
35#include <linux/export.h>
36#include <linux/notifier.h>
37#include <linux/memcontrol.h>
38#include <linux/mempolicy.h>
39#include <linux/security.h>
40#include <linux/ptrace.h>
41#include <linux/freezer.h>
42#include <linux/ftrace.h>
43#include <linux/ratelimit.h>
44#include <linux/kthread.h>
45#include <linux/init.h>
46#include <linux/mmu_notifier.h>
47
48#include <asm/tlb.h>
49#include "internal.h"
50#include "slab.h"
51
52#define CREATE_TRACE_POINTS
53#include <trace/events/oom.h>
54
55static int sysctl_panic_on_oom;
56static int sysctl_oom_kill_allocating_task;
57static int sysctl_oom_dump_tasks = 1;
58
59/*
60 * Serializes oom killer invocations (out_of_memory()) from all contexts to
61 * prevent from over eager oom killing (e.g. when the oom killer is invoked
62 * from different domains).
63 *
64 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
65 * and mark_oom_victim
66 */
67DEFINE_MUTEX(oom_lock);
68/* Serializes oom_score_adj and oom_score_adj_min updates */
69DEFINE_MUTEX(oom_adj_mutex);
70
71static inline bool is_memcg_oom(struct oom_control *oc)
72{
73 return oc->memcg != NULL;
74}
75
76#ifdef CONFIG_NUMA
77/**
78 * oom_cpuset_eligible() - check task eligibility for kill
79 * @start: task struct of which task to consider
80 * @oc: pointer to struct oom_control
81 *
82 * Task eligibility is determined by whether or not a candidate task, @tsk,
83 * shares the same mempolicy nodes as current if it is bound by such a policy
84 * and whether or not it has the same set of allowed cpuset nodes.
85 *
86 * This function is assuming oom-killer context and 'current' has triggered
87 * the oom-killer.
88 */
89static bool oom_cpuset_eligible(struct task_struct *start,
90 struct oom_control *oc)
91{
92 struct task_struct *tsk;
93 bool ret = false;
94 const nodemask_t *mask = oc->nodemask;
95
96 rcu_read_lock();
97 for_each_thread(start, tsk) {
98 if (mask) {
99 /*
100 * If this is a mempolicy constrained oom, tsk's
101 * cpuset is irrelevant. Only return true if its
102 * mempolicy intersects current, otherwise it may be
103 * needlessly killed.
104 */
105 ret = mempolicy_in_oom_domain(tsk, mask);
106 } else {
107 /*
108 * This is not a mempolicy constrained oom, so only
109 * check the mems of tsk's cpuset.
110 */
111 ret = cpuset_mems_allowed_intersects(current, tsk);
112 }
113 if (ret)
114 break;
115 }
116 rcu_read_unlock();
117
118 return ret;
119}
120#else
121static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
122{
123 return true;
124}
125#endif /* CONFIG_NUMA */
126
127/*
128 * The process p may have detached its own ->mm while exiting or through
129 * kthread_use_mm(), but one or more of its subthreads may still have a valid
130 * pointer. Return p, or any of its subthreads with a valid ->mm, with
131 * task_lock() held.
132 */
133struct task_struct *find_lock_task_mm(struct task_struct *p)
134{
135 struct task_struct *t;
136
137 rcu_read_lock();
138
139 for_each_thread(p, t) {
140 task_lock(t);
141 if (likely(t->mm))
142 goto found;
143 task_unlock(t);
144 }
145 t = NULL;
146found:
147 rcu_read_unlock();
148
149 return t;
150}
151
152/*
153 * order == -1 means the oom kill is required by sysrq, otherwise only
154 * for display purposes.
155 */
156static inline bool is_sysrq_oom(struct oom_control *oc)
157{
158 return oc->order == -1;
159}
160
161/* return true if the task is not adequate as candidate victim task. */
162static bool oom_unkillable_task(struct task_struct *p)
163{
164 if (is_global_init(p))
165 return true;
166 if (p->flags & PF_KTHREAD)
167 return true;
168 return false;
169}
170
171/*
172 * Check whether unreclaimable slab amount is greater than
173 * all user memory(LRU pages).
174 * dump_unreclaimable_slab() could help in the case that
175 * oom due to too much unreclaimable slab used by kernel.
176*/
177static bool should_dump_unreclaim_slab(void)
178{
179 unsigned long nr_lru;
180
181 nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
182 global_node_page_state(NR_INACTIVE_ANON) +
183 global_node_page_state(NR_ACTIVE_FILE) +
184 global_node_page_state(NR_INACTIVE_FILE) +
185 global_node_page_state(NR_ISOLATED_ANON) +
186 global_node_page_state(NR_ISOLATED_FILE) +
187 global_node_page_state(NR_UNEVICTABLE);
188
189 return (global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B) > nr_lru);
190}
191
192/**
193 * oom_badness - heuristic function to determine which candidate task to kill
194 * @p: task struct of which task we should calculate
195 * @totalpages: total present RAM allowed for page allocation
196 *
197 * The heuristic for determining which task to kill is made to be as simple and
198 * predictable as possible. The goal is to return the highest value for the
199 * task consuming the most memory to avoid subsequent oom failures.
200 */
201long oom_badness(struct task_struct *p, unsigned long totalpages)
202{
203 long points;
204 long adj;
205
206 if (oom_unkillable_task(p))
207 return LONG_MIN;
208
209 p = find_lock_task_mm(p);
210 if (!p)
211 return LONG_MIN;
212
213 /*
214 * Do not even consider tasks which are explicitly marked oom
215 * unkillable or have been already oom reaped or the are in
216 * the middle of vfork
217 */
218 adj = (long)p->signal->oom_score_adj;
219 if (adj == OOM_SCORE_ADJ_MIN ||
220 test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
221 in_vfork(p)) {
222 task_unlock(p);
223 return LONG_MIN;
224 }
225
226 /*
227 * The baseline for the badness score is the proportion of RAM that each
228 * task's rss, pagetable and swap space use.
229 */
230 points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
231 mm_pgtables_bytes(p->mm) / PAGE_SIZE;
232 task_unlock(p);
233
234 /* Normalize to oom_score_adj units */
235 adj *= totalpages / 1000;
236 points += adj;
237
238 return points;
239}
240
241static const char * const oom_constraint_text[] = {
242 [CONSTRAINT_NONE] = "CONSTRAINT_NONE",
243 [CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
244 [CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
245 [CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
246};
247
248/*
249 * Determine the type of allocation constraint.
250 */
251static enum oom_constraint constrained_alloc(struct oom_control *oc)
252{
253 struct zone *zone;
254 struct zoneref *z;
255 enum zone_type highest_zoneidx = gfp_zone(oc->gfp_mask);
256 bool cpuset_limited = false;
257 int nid;
258
259 if (is_memcg_oom(oc)) {
260 oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
261 return CONSTRAINT_MEMCG;
262 }
263
264 /* Default to all available memory */
265 oc->totalpages = totalram_pages() + total_swap_pages;
266
267 if (!IS_ENABLED(CONFIG_NUMA))
268 return CONSTRAINT_NONE;
269
270 if (!oc->zonelist)
271 return CONSTRAINT_NONE;
272 /*
273 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
274 * to kill current.We have to random task kill in this case.
275 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
276 */
277 if (oc->gfp_mask & __GFP_THISNODE)
278 return CONSTRAINT_NONE;
279
280 /*
281 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
282 * the page allocator means a mempolicy is in effect. Cpuset policy
283 * is enforced in get_page_from_freelist().
284 */
285 if (oc->nodemask &&
286 !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
287 oc->totalpages = total_swap_pages;
288 for_each_node_mask(nid, *oc->nodemask)
289 oc->totalpages += node_present_pages(nid);
290 return CONSTRAINT_MEMORY_POLICY;
291 }
292
293 /* Check this allocation failure is caused by cpuset's wall function */
294 for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
295 highest_zoneidx, oc->nodemask)
296 if (!cpuset_zone_allowed(zone, oc->gfp_mask))
297 cpuset_limited = true;
298
299 if (cpuset_limited) {
300 oc->totalpages = total_swap_pages;
301 for_each_node_mask(nid, cpuset_current_mems_allowed)
302 oc->totalpages += node_present_pages(nid);
303 return CONSTRAINT_CPUSET;
304 }
305 return CONSTRAINT_NONE;
306}
307
308static int oom_evaluate_task(struct task_struct *task, void *arg)
309{
310 struct oom_control *oc = arg;
311 long points;
312
313 if (oom_unkillable_task(task))
314 goto next;
315
316 /* p may not have freeable memory in nodemask */
317 if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
318 goto next;
319
320 /*
321 * This task already has access to memory reserves and is being killed.
322 * Don't allow any other task to have access to the reserves unless
323 * the task has MMF_OOM_SKIP because chances that it would release
324 * any memory is quite low.
325 */
326 if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
327 if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
328 goto next;
329 goto abort;
330 }
331
332 /*
333 * If task is allocating a lot of memory and has been marked to be
334 * killed first if it triggers an oom, then select it.
335 */
336 if (oom_task_origin(task)) {
337 points = LONG_MAX;
338 goto select;
339 }
340
341 points = oom_badness(task, oc->totalpages);
342 if (points == LONG_MIN || points < oc->chosen_points)
343 goto next;
344
345select:
346 if (oc->chosen)
347 put_task_struct(oc->chosen);
348 get_task_struct(task);
349 oc->chosen = task;
350 oc->chosen_points = points;
351next:
352 return 0;
353abort:
354 if (oc->chosen)
355 put_task_struct(oc->chosen);
356 oc->chosen = (void *)-1UL;
357 return 1;
358}
359
360/*
361 * Simple selection loop. We choose the process with the highest number of
362 * 'points'. In case scan was aborted, oc->chosen is set to -1.
363 */
364static void select_bad_process(struct oom_control *oc)
365{
366 oc->chosen_points = LONG_MIN;
367
368 if (is_memcg_oom(oc))
369 mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
370 else {
371 struct task_struct *p;
372
373 rcu_read_lock();
374 for_each_process(p)
375 if (oom_evaluate_task(p, oc))
376 break;
377 rcu_read_unlock();
378 }
379}
380
381static int dump_task(struct task_struct *p, void *arg)
382{
383 struct oom_control *oc = arg;
384 struct task_struct *task;
385
386 if (oom_unkillable_task(p))
387 return 0;
388
389 /* p may not have freeable memory in nodemask */
390 if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
391 return 0;
392
393 task = find_lock_task_mm(p);
394 if (!task) {
395 /*
396 * All of p's threads have already detached their mm's. There's
397 * no need to report 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, gfp_zone(oc->gfp_mask));
465 if (should_dump_unreclaim_slab())
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
512static bool __oom_reap_task_mm(struct mm_struct *mm)
513{
514 struct vm_area_struct *vma;
515 bool ret = true;
516 VMA_ITERATOR(vmi, mm, 0);
517
518 /*
519 * Tell all users of get_user/copy_from_user etc... that the content
520 * is no longer stable. No barriers really needed because unmapping
521 * should imply barriers already and the reader would hit a page fault
522 * if it stumbled over a reaped memory.
523 */
524 set_bit(MMF_UNSTABLE, &mm->flags);
525
526 for_each_vma(vmi, vma) {
527 if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
528 continue;
529
530 /*
531 * Only anonymous pages have a good chance to be dropped
532 * without additional steps which we cannot afford as we
533 * are OOM already.
534 *
535 * We do not even care about fs backed pages because all
536 * which are reclaimable have already been reclaimed and
537 * we do not want to block exit_mmap by keeping mm ref
538 * count elevated without a good reason.
539 */
540 if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
541 struct mmu_notifier_range range;
542 struct mmu_gather tlb;
543
544 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
545 vma, mm, vma->vm_start,
546 vma->vm_end);
547 tlb_gather_mmu(&tlb, mm);
548 if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
549 tlb_finish_mmu(&tlb);
550 ret = false;
551 continue;
552 }
553 unmap_page_range(&tlb, vma, range.start, range.end, NULL);
554 mmu_notifier_invalidate_range_end(&range);
555 tlb_finish_mmu(&tlb);
556 }
557 }
558
559 return ret;
560}
561
562/*
563 * Reaps the address space of the give task.
564 *
565 * Returns true on success and false if none or part of the address space
566 * has been reclaimed and the caller should retry later.
567 */
568static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
569{
570 bool ret = true;
571
572 if (!mmap_read_trylock(mm)) {
573 trace_skip_task_reaping(tsk->pid);
574 return false;
575 }
576
577 /*
578 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
579 * work on the mm anymore. The check for MMF_OOM_SKIP must run
580 * under mmap_lock for reading because it serializes against the
581 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
582 */
583 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
584 trace_skip_task_reaping(tsk->pid);
585 goto out_unlock;
586 }
587
588 trace_start_task_reaping(tsk->pid);
589
590 /* failed to reap part of the address space. Try again later */
591 ret = __oom_reap_task_mm(mm);
592 if (!ret)
593 goto out_finish;
594
595 pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
596 task_pid_nr(tsk), tsk->comm,
597 K(get_mm_counter(mm, MM_ANONPAGES)),
598 K(get_mm_counter(mm, MM_FILEPAGES)),
599 K(get_mm_counter(mm, MM_SHMEMPAGES)));
600out_finish:
601 trace_finish_task_reaping(tsk->pid);
602out_unlock:
603 mmap_read_unlock(mm);
604
605 return ret;
606}
607
608#define MAX_OOM_REAP_RETRIES 10
609static void oom_reap_task(struct task_struct *tsk)
610{
611 int attempts = 0;
612 struct mm_struct *mm = tsk->signal->oom_mm;
613
614 /* Retry the mmap_read_trylock(mm) a few times */
615 while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
616 schedule_timeout_idle(HZ/10);
617
618 if (attempts <= MAX_OOM_REAP_RETRIES ||
619 test_bit(MMF_OOM_SKIP, &mm->flags))
620 goto done;
621
622 pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
623 task_pid_nr(tsk), tsk->comm);
624 sched_show_task(tsk);
625 debug_show_all_locks();
626
627done:
628 tsk->oom_reaper_list = NULL;
629
630 /*
631 * Hide this mm from OOM killer because it has been either reaped or
632 * somebody can't call mmap_write_unlock(mm).
633 */
634 set_bit(MMF_OOM_SKIP, &mm->flags);
635
636 /* Drop a reference taken by queue_oom_reaper */
637 put_task_struct(tsk);
638}
639
640static int oom_reaper(void *unused)
641{
642 set_freezable();
643
644 while (true) {
645 struct task_struct *tsk = NULL;
646
647 wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
648 spin_lock_irq(&oom_reaper_lock);
649 if (oom_reaper_list != NULL) {
650 tsk = oom_reaper_list;
651 oom_reaper_list = tsk->oom_reaper_list;
652 }
653 spin_unlock_irq(&oom_reaper_lock);
654
655 if (tsk)
656 oom_reap_task(tsk);
657 }
658
659 return 0;
660}
661
662static void wake_oom_reaper(struct timer_list *timer)
663{
664 struct task_struct *tsk = container_of(timer, struct task_struct,
665 oom_reaper_timer);
666 struct mm_struct *mm = tsk->signal->oom_mm;
667 unsigned long flags;
668
669 /* The victim managed to terminate on its own - see exit_mmap */
670 if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
671 put_task_struct(tsk);
672 return;
673 }
674
675 spin_lock_irqsave(&oom_reaper_lock, flags);
676 tsk->oom_reaper_list = oom_reaper_list;
677 oom_reaper_list = tsk;
678 spin_unlock_irqrestore(&oom_reaper_lock, flags);
679 trace_wake_reaper(tsk->pid);
680 wake_up(&oom_reaper_wait);
681}
682
683/*
684 * Give the OOM victim time to exit naturally before invoking the oom_reaping.
685 * The timers timeout is arbitrary... the longer it is, the longer the worst
686 * case scenario for the OOM can take. If it is too small, the oom_reaper can
687 * get in the way and release resources needed by the process exit path.
688 * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
689 * before the exit path is able to wake the futex waiters.
690 */
691#define OOM_REAPER_DELAY (2*HZ)
692static void queue_oom_reaper(struct task_struct *tsk)
693{
694 /* mm is already queued? */
695 if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
696 return;
697
698 get_task_struct(tsk);
699 timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
700 tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
701 add_timer(&tsk->oom_reaper_timer);
702}
703
704#ifdef CONFIG_SYSCTL
705static struct ctl_table vm_oom_kill_table[] = {
706 {
707 .procname = "panic_on_oom",
708 .data = &sysctl_panic_on_oom,
709 .maxlen = sizeof(sysctl_panic_on_oom),
710 .mode = 0644,
711 .proc_handler = proc_dointvec_minmax,
712 .extra1 = SYSCTL_ZERO,
713 .extra2 = SYSCTL_TWO,
714 },
715 {
716 .procname = "oom_kill_allocating_task",
717 .data = &sysctl_oom_kill_allocating_task,
718 .maxlen = sizeof(sysctl_oom_kill_allocating_task),
719 .mode = 0644,
720 .proc_handler = proc_dointvec,
721 },
722 {
723 .procname = "oom_dump_tasks",
724 .data = &sysctl_oom_dump_tasks,
725 .maxlen = sizeof(sysctl_oom_dump_tasks),
726 .mode = 0644,
727 .proc_handler = proc_dointvec,
728 },
729 {}
730};
731#endif
732
733static int __init oom_init(void)
734{
735 oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
736#ifdef CONFIG_SYSCTL
737 register_sysctl_init("vm", vm_oom_kill_table);
738#endif
739 return 0;
740}
741subsys_initcall(oom_init)
742#else
743static inline void queue_oom_reaper(struct task_struct *tsk)
744{
745}
746#endif /* CONFIG_MMU */
747
748/**
749 * mark_oom_victim - mark the given task as OOM victim
750 * @tsk: task to mark
751 *
752 * Has to be called with oom_lock held and never after
753 * oom has been disabled already.
754 *
755 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
756 * under task_lock or operate on the current).
757 */
758static void mark_oom_victim(struct task_struct *tsk)
759{
760 struct mm_struct *mm = tsk->mm;
761
762 WARN_ON(oom_killer_disabled);
763 /* OOM killer might race with memcg OOM */
764 if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
765 return;
766
767 /* oom_mm is bound to the signal struct life time. */
768 if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
769 mmgrab(tsk->signal->oom_mm);
770
771 /*
772 * Make sure that the task is woken up from uninterruptible sleep
773 * if it is frozen because OOM killer wouldn't be able to free
774 * any memory and livelock. freezing_slow_path will tell the freezer
775 * that TIF_MEMDIE tasks should be ignored.
776 */
777 __thaw_task(tsk);
778 atomic_inc(&oom_victims);
779 trace_mark_victim(tsk->pid);
780}
781
782/**
783 * exit_oom_victim - note the exit of an OOM victim
784 */
785void exit_oom_victim(void)
786{
787 clear_thread_flag(TIF_MEMDIE);
788
789 if (!atomic_dec_return(&oom_victims))
790 wake_up_all(&oom_victims_wait);
791}
792
793/**
794 * oom_killer_enable - enable OOM killer
795 */
796void oom_killer_enable(void)
797{
798 oom_killer_disabled = false;
799 pr_info("OOM killer enabled.\n");
800}
801
802/**
803 * oom_killer_disable - disable OOM killer
804 * @timeout: maximum timeout to wait for oom victims in jiffies
805 *
806 * Forces all page allocations to fail rather than trigger OOM killer.
807 * Will block and wait until all OOM victims are killed or the given
808 * timeout expires.
809 *
810 * The function cannot be called when there are runnable user tasks because
811 * the userspace would see unexpected allocation failures as a result. Any
812 * new usage of this function should be consulted with MM people.
813 *
814 * Returns true if successful and false if the OOM killer cannot be
815 * disabled.
816 */
817bool oom_killer_disable(signed long timeout)
818{
819 signed long ret;
820
821 /*
822 * Make sure to not race with an ongoing OOM killer. Check that the
823 * current is not killed (possibly due to sharing the victim's memory).
824 */
825 if (mutex_lock_killable(&oom_lock))
826 return false;
827 oom_killer_disabled = true;
828 mutex_unlock(&oom_lock);
829
830 ret = wait_event_interruptible_timeout(oom_victims_wait,
831 !atomic_read(&oom_victims), timeout);
832 if (ret <= 0) {
833 oom_killer_enable();
834 return false;
835 }
836 pr_info("OOM killer disabled.\n");
837
838 return true;
839}
840
841static inline bool __task_will_free_mem(struct task_struct *task)
842{
843 struct signal_struct *sig = task->signal;
844
845 /*
846 * A coredumping process may sleep for an extended period in
847 * coredump_task_exit(), so the oom killer cannot assume that
848 * the process will promptly exit and release memory.
849 */
850 if (sig->core_state)
851 return false;
852
853 if (sig->flags & SIGNAL_GROUP_EXIT)
854 return true;
855
856 if (thread_group_empty(task) && (task->flags & PF_EXITING))
857 return true;
858
859 return false;
860}
861
862/*
863 * Checks whether the given task is dying or exiting and likely to
864 * release its address space. This means that all threads and processes
865 * sharing the same mm have to be killed or exiting.
866 * Caller has to make sure that task->mm is stable (hold task_lock or
867 * it operates on the current).
868 */
869static bool task_will_free_mem(struct task_struct *task)
870{
871 struct mm_struct *mm = task->mm;
872 struct task_struct *p;
873 bool ret = true;
874
875 /*
876 * Skip tasks without mm because it might have passed its exit_mm and
877 * exit_oom_victim. oom_reaper could have rescued that but do not rely
878 * on that for now. We can consider find_lock_task_mm in future.
879 */
880 if (!mm)
881 return false;
882
883 if (!__task_will_free_mem(task))
884 return false;
885
886 /*
887 * This task has already been drained by the oom reaper so there are
888 * only small chances it will free some more
889 */
890 if (test_bit(MMF_OOM_SKIP, &mm->flags))
891 return false;
892
893 if (atomic_read(&mm->mm_users) <= 1)
894 return true;
895
896 /*
897 * Make sure that all tasks which share the mm with the given tasks
898 * are dying as well to make sure that a) nobody pins its mm and
899 * b) the task is also reapable by the oom reaper.
900 */
901 rcu_read_lock();
902 for_each_process(p) {
903 if (!process_shares_mm(p, mm))
904 continue;
905 if (same_thread_group(task, p))
906 continue;
907 ret = __task_will_free_mem(p);
908 if (!ret)
909 break;
910 }
911 rcu_read_unlock();
912
913 return ret;
914}
915
916static void __oom_kill_process(struct task_struct *victim, const char *message)
917{
918 struct task_struct *p;
919 struct mm_struct *mm;
920 bool can_oom_reap = true;
921
922 p = find_lock_task_mm(victim);
923 if (!p) {
924 pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
925 message, task_pid_nr(victim), victim->comm);
926 put_task_struct(victim);
927 return;
928 } else if (victim != p) {
929 get_task_struct(p);
930 put_task_struct(victim);
931 victim = p;
932 }
933
934 /* Get a reference to safely compare mm after task_unlock(victim) */
935 mm = victim->mm;
936 mmgrab(mm);
937
938 /* Raise event before sending signal: task reaper must see this */
939 count_vm_event(OOM_KILL);
940 memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
941
942 /*
943 * We should send SIGKILL before granting access to memory reserves
944 * in order to prevent the OOM victim from depleting the memory
945 * reserves from the user space under its control.
946 */
947 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
948 mark_oom_victim(victim);
949 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",
950 message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
951 K(get_mm_counter(mm, MM_ANONPAGES)),
952 K(get_mm_counter(mm, MM_FILEPAGES)),
953 K(get_mm_counter(mm, MM_SHMEMPAGES)),
954 from_kuid(&init_user_ns, task_uid(victim)),
955 mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
956 task_unlock(victim);
957
958 /*
959 * Kill all user processes sharing victim->mm in other thread groups, if
960 * any. They don't get access to memory reserves, though, to avoid
961 * depletion of all memory. This prevents mm->mmap_lock livelock when an
962 * oom killed thread cannot exit because it requires the semaphore and
963 * its contended by another thread trying to allocate memory itself.
964 * That thread will now get access to memory reserves since it has a
965 * pending fatal signal.
966 */
967 rcu_read_lock();
968 for_each_process(p) {
969 if (!process_shares_mm(p, mm))
970 continue;
971 if (same_thread_group(p, victim))
972 continue;
973 if (is_global_init(p)) {
974 can_oom_reap = false;
975 set_bit(MMF_OOM_SKIP, &mm->flags);
976 pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
977 task_pid_nr(victim), victim->comm,
978 task_pid_nr(p), p->comm);
979 continue;
980 }
981 /*
982 * No kthread_use_mm() user needs to read from the userspace so
983 * we are ok to reap it.
984 */
985 if (unlikely(p->flags & PF_KTHREAD))
986 continue;
987 do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
988 }
989 rcu_read_unlock();
990
991 if (can_oom_reap)
992 queue_oom_reaper(victim);
993
994 mmdrop(mm);
995 put_task_struct(victim);
996}
997#undef K
998
999/*
1000 * Kill provided task unless it's secured by setting
1001 * oom_score_adj to OOM_SCORE_ADJ_MIN.
1002 */
1003static int oom_kill_memcg_member(struct task_struct *task, void *message)
1004{
1005 if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1006 !is_global_init(task)) {
1007 get_task_struct(task);
1008 __oom_kill_process(task, message);
1009 }
1010 return 0;
1011}
1012
1013static void oom_kill_process(struct oom_control *oc, const char *message)
1014{
1015 struct task_struct *victim = oc->chosen;
1016 struct mem_cgroup *oom_group;
1017 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1018 DEFAULT_RATELIMIT_BURST);
1019
1020 /*
1021 * If the task is already exiting, don't alarm the sysadmin or kill
1022 * its children or threads, just give it access to memory reserves
1023 * so it can die quickly
1024 */
1025 task_lock(victim);
1026 if (task_will_free_mem(victim)) {
1027 mark_oom_victim(victim);
1028 queue_oom_reaper(victim);
1029 task_unlock(victim);
1030 put_task_struct(victim);
1031 return;
1032 }
1033 task_unlock(victim);
1034
1035 if (__ratelimit(&oom_rs))
1036 dump_header(oc, victim);
1037
1038 /*
1039 * Do we need to kill the entire memory cgroup?
1040 * Or even one of the ancestor memory cgroups?
1041 * Check this out before killing the victim task.
1042 */
1043 oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1044
1045 __oom_kill_process(victim, message);
1046
1047 /*
1048 * If necessary, kill all tasks in the selected memory cgroup.
1049 */
1050 if (oom_group) {
1051 memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1052 mem_cgroup_print_oom_group(oom_group);
1053 mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1054 (void *)message);
1055 mem_cgroup_put(oom_group);
1056 }
1057}
1058
1059/*
1060 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1061 */
1062static void check_panic_on_oom(struct oom_control *oc)
1063{
1064 if (likely(!sysctl_panic_on_oom))
1065 return;
1066 if (sysctl_panic_on_oom != 2) {
1067 /*
1068 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1069 * does not panic for cpuset, mempolicy, or memcg allocation
1070 * failures.
1071 */
1072 if (oc->constraint != CONSTRAINT_NONE)
1073 return;
1074 }
1075 /* Do not panic for oom kills triggered by sysrq */
1076 if (is_sysrq_oom(oc))
1077 return;
1078 dump_header(oc, NULL);
1079 panic("Out of memory: %s panic_on_oom is enabled\n",
1080 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1081}
1082
1083static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1084
1085int register_oom_notifier(struct notifier_block *nb)
1086{
1087 return blocking_notifier_chain_register(&oom_notify_list, nb);
1088}
1089EXPORT_SYMBOL_GPL(register_oom_notifier);
1090
1091int unregister_oom_notifier(struct notifier_block *nb)
1092{
1093 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1094}
1095EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1096
1097/**
1098 * out_of_memory - kill the "best" process when we run out of memory
1099 * @oc: pointer to struct oom_control
1100 *
1101 * If we run out of memory, we have the choice between either
1102 * killing a random task (bad), letting the system crash (worse)
1103 * OR try to be smart about which process to kill. Note that we
1104 * don't have to be perfect here, we just have to be good.
1105 */
1106bool out_of_memory(struct oom_control *oc)
1107{
1108 unsigned long freed = 0;
1109
1110 if (oom_killer_disabled)
1111 return false;
1112
1113 if (!is_memcg_oom(oc)) {
1114 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1115 if (freed > 0 && !is_sysrq_oom(oc))
1116 /* Got some memory back in the last second. */
1117 return true;
1118 }
1119
1120 /*
1121 * If current has a pending SIGKILL or is exiting, then automatically
1122 * select it. The goal is to allow it to allocate so that it may
1123 * quickly exit and free its memory.
1124 */
1125 if (task_will_free_mem(current)) {
1126 mark_oom_victim(current);
1127 queue_oom_reaper(current);
1128 return true;
1129 }
1130
1131 /*
1132 * The OOM killer does not compensate for IO-less reclaim.
1133 * pagefault_out_of_memory lost its gfp context so we have to
1134 * make sure exclude 0 mask - all other users should have at least
1135 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1136 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1137 */
1138 if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1139 return true;
1140
1141 /*
1142 * Check if there were limitations on the allocation (only relevant for
1143 * NUMA and memcg) that may require different handling.
1144 */
1145 oc->constraint = constrained_alloc(oc);
1146 if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1147 oc->nodemask = NULL;
1148 check_panic_on_oom(oc);
1149
1150 if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1151 current->mm && !oom_unkillable_task(current) &&
1152 oom_cpuset_eligible(current, oc) &&
1153 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1154 get_task_struct(current);
1155 oc->chosen = current;
1156 oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1157 return true;
1158 }
1159
1160 select_bad_process(oc);
1161 /* Found nothing?!?! */
1162 if (!oc->chosen) {
1163 dump_header(oc, NULL);
1164 pr_warn("Out of memory and no killable processes...\n");
1165 /*
1166 * If we got here due to an actual allocation at the
1167 * system level, we cannot survive this and will enter
1168 * an endless loop in the allocator. Bail out now.
1169 */
1170 if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1171 panic("System is deadlocked on memory\n");
1172 }
1173 if (oc->chosen && oc->chosen != (void *)-1UL)
1174 oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1175 "Memory cgroup out of memory");
1176 return !!oc->chosen;
1177}
1178
1179/*
1180 * The pagefault handler calls here because some allocation has failed. We have
1181 * to take care of the memcg OOM here because this is the only safe context without
1182 * any locks held but let the oom killer triggered from the allocation context care
1183 * about the global OOM.
1184 */
1185void pagefault_out_of_memory(void)
1186{
1187 static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1188 DEFAULT_RATELIMIT_BURST);
1189
1190 if (mem_cgroup_oom_synchronize(true))
1191 return;
1192
1193 if (fatal_signal_pending(current))
1194 return;
1195
1196 if (__ratelimit(&pfoom_rs))
1197 pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1198}
1199
1200SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1201{
1202#ifdef CONFIG_MMU
1203 struct mm_struct *mm = NULL;
1204 struct task_struct *task;
1205 struct task_struct *p;
1206 unsigned int f_flags;
1207 bool reap = false;
1208 long ret = 0;
1209
1210 if (flags)
1211 return -EINVAL;
1212
1213 task = pidfd_get_task(pidfd, &f_flags);
1214 if (IS_ERR(task))
1215 return PTR_ERR(task);
1216
1217 /*
1218 * Make sure to choose a thread which still has a reference to mm
1219 * during the group exit
1220 */
1221 p = find_lock_task_mm(task);
1222 if (!p) {
1223 ret = -ESRCH;
1224 goto put_task;
1225 }
1226
1227 mm = p->mm;
1228 mmgrab(mm);
1229
1230 if (task_will_free_mem(p))
1231 reap = true;
1232 else {
1233 /* Error only if the work has not been done already */
1234 if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1235 ret = -EINVAL;
1236 }
1237 task_unlock(p);
1238
1239 if (!reap)
1240 goto drop_mm;
1241
1242 if (mmap_read_lock_killable(mm)) {
1243 ret = -EINTR;
1244 goto drop_mm;
1245 }
1246 /*
1247 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1248 * possible change in exit_mmap is seen
1249 */
1250 if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1251 ret = -EAGAIN;
1252 mmap_read_unlock(mm);
1253
1254drop_mm:
1255 mmdrop(mm);
1256put_task:
1257 put_task_struct(task);
1258 return ret;
1259#else
1260 return -ENOSYS;
1261#endif /* CONFIG_MMU */
1262}
1/*
2 * linux/mm/oom_kill.c
3 *
4 * Copyright (C) 1998,2000 Rik van Riel
5 * Thanks go out to Claus Fischer for some serious inspiration and
6 * for goading me into coding this file...
7 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
9 *
10 * The routines in this file are used to kill a process when
11 * we're seriously out of memory. This gets called from __alloc_pages()
12 * in mm/page_alloc.c when we really run out of memory.
13 *
14 * Since we won't call these routines often (on a well-configured
15 * machine) this file will double as a 'coding guide' and a signpost
16 * for newbie kernel hackers. It features several pointers to major
17 * kernel subsystems and hints as to where to find out what things do.
18 */
19
20#include <linux/oom.h>
21#include <linux/mm.h>
22#include <linux/err.h>
23#include <linux/gfp.h>
24#include <linux/sched.h>
25#include <linux/swap.h>
26#include <linux/timex.h>
27#include <linux/jiffies.h>
28#include <linux/cpuset.h>
29#include <linux/module.h>
30#include <linux/notifier.h>
31#include <linux/memcontrol.h>
32#include <linux/mempolicy.h>
33#include <linux/security.h>
34#include <linux/ptrace.h>
35
36int sysctl_panic_on_oom;
37int sysctl_oom_kill_allocating_task;
38int sysctl_oom_dump_tasks = 1;
39static DEFINE_SPINLOCK(zone_scan_lock);
40
41/**
42 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
43 * @new_val: new oom_score_adj value
44 *
45 * Sets the oom_score_adj value for current to @new_val with proper
46 * synchronization and returns the old value. Usually used to temporarily
47 * set a value, save the old value in the caller, and then reinstate it later.
48 */
49int test_set_oom_score_adj(int new_val)
50{
51 struct sighand_struct *sighand = current->sighand;
52 int old_val;
53
54 spin_lock_irq(&sighand->siglock);
55 old_val = current->signal->oom_score_adj;
56 if (new_val != old_val) {
57 if (new_val == OOM_SCORE_ADJ_MIN)
58 atomic_inc(¤t->mm->oom_disable_count);
59 else if (old_val == OOM_SCORE_ADJ_MIN)
60 atomic_dec(¤t->mm->oom_disable_count);
61 current->signal->oom_score_adj = new_val;
62 }
63 spin_unlock_irq(&sighand->siglock);
64
65 return old_val;
66}
67
68#ifdef CONFIG_NUMA
69/**
70 * has_intersects_mems_allowed() - check task eligiblity for kill
71 * @tsk: task struct of which task to consider
72 * @mask: nodemask passed to page allocator for mempolicy ooms
73 *
74 * Task eligibility is determined by whether or not a candidate task, @tsk,
75 * shares the same mempolicy nodes as current if it is bound by such a policy
76 * and whether or not it has the same set of allowed cpuset nodes.
77 */
78static bool has_intersects_mems_allowed(struct task_struct *tsk,
79 const nodemask_t *mask)
80{
81 struct task_struct *start = tsk;
82
83 do {
84 if (mask) {
85 /*
86 * If this is a mempolicy constrained oom, tsk's
87 * cpuset is irrelevant. Only return true if its
88 * mempolicy intersects current, otherwise it may be
89 * needlessly killed.
90 */
91 if (mempolicy_nodemask_intersects(tsk, mask))
92 return true;
93 } else {
94 /*
95 * This is not a mempolicy constrained oom, so only
96 * check the mems of tsk's cpuset.
97 */
98 if (cpuset_mems_allowed_intersects(current, tsk))
99 return true;
100 }
101 } while_each_thread(start, tsk);
102
103 return false;
104}
105#else
106static bool has_intersects_mems_allowed(struct task_struct *tsk,
107 const nodemask_t *mask)
108{
109 return true;
110}
111#endif /* CONFIG_NUMA */
112
113/*
114 * The process p may have detached its own ->mm while exiting or through
115 * use_mm(), but one or more of its subthreads may still have a valid
116 * pointer. Return p, or any of its subthreads with a valid ->mm, with
117 * task_lock() held.
118 */
119struct task_struct *find_lock_task_mm(struct task_struct *p)
120{
121 struct task_struct *t = p;
122
123 do {
124 task_lock(t);
125 if (likely(t->mm))
126 return t;
127 task_unlock(t);
128 } while_each_thread(p, t);
129
130 return NULL;
131}
132
133/* return true if the task is not adequate as candidate victim task. */
134static bool oom_unkillable_task(struct task_struct *p,
135 const struct mem_cgroup *mem, const nodemask_t *nodemask)
136{
137 if (is_global_init(p))
138 return true;
139 if (p->flags & PF_KTHREAD)
140 return true;
141
142 /* When mem_cgroup_out_of_memory() and p is not member of the group */
143 if (mem && !task_in_mem_cgroup(p, mem))
144 return true;
145
146 /* p may not have freeable memory in nodemask */
147 if (!has_intersects_mems_allowed(p, nodemask))
148 return true;
149
150 return false;
151}
152
153/**
154 * oom_badness - heuristic function to determine which candidate task to kill
155 * @p: task struct of which task we should calculate
156 * @totalpages: total present RAM allowed for page allocation
157 *
158 * The heuristic for determining which task to kill is made to be as simple and
159 * predictable as possible. The goal is to return the highest value for the
160 * task consuming the most memory to avoid subsequent oom failures.
161 */
162unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
163 const nodemask_t *nodemask, unsigned long totalpages)
164{
165 int points;
166
167 if (oom_unkillable_task(p, mem, nodemask))
168 return 0;
169
170 p = find_lock_task_mm(p);
171 if (!p)
172 return 0;
173
174 /*
175 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
176 * so the entire heuristic doesn't need to be executed for something
177 * that cannot be killed.
178 */
179 if (atomic_read(&p->mm->oom_disable_count)) {
180 task_unlock(p);
181 return 0;
182 }
183
184 /*
185 * The memory controller may have a limit of 0 bytes, so avoid a divide
186 * by zero, if necessary.
187 */
188 if (!totalpages)
189 totalpages = 1;
190
191 /*
192 * The baseline for the badness score is the proportion of RAM that each
193 * task's rss, pagetable and swap space use.
194 */
195 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
196 points += get_mm_counter(p->mm, MM_SWAPENTS);
197
198 points *= 1000;
199 points /= totalpages;
200 task_unlock(p);
201
202 /*
203 * Root processes get 3% bonus, just like the __vm_enough_memory()
204 * implementation used by LSMs.
205 */
206 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
207 points -= 30;
208
209 /*
210 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
211 * either completely disable oom killing or always prefer a certain
212 * task.
213 */
214 points += p->signal->oom_score_adj;
215
216 /*
217 * Never return 0 for an eligible task that may be killed since it's
218 * possible that no single user task uses more than 0.1% of memory and
219 * no single admin tasks uses more than 3.0%.
220 */
221 if (points <= 0)
222 return 1;
223 return (points < 1000) ? points : 1000;
224}
225
226/*
227 * Determine the type of allocation constraint.
228 */
229#ifdef CONFIG_NUMA
230static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
231 gfp_t gfp_mask, nodemask_t *nodemask,
232 unsigned long *totalpages)
233{
234 struct zone *zone;
235 struct zoneref *z;
236 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
237 bool cpuset_limited = false;
238 int nid;
239
240 /* Default to all available memory */
241 *totalpages = totalram_pages + total_swap_pages;
242
243 if (!zonelist)
244 return CONSTRAINT_NONE;
245 /*
246 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
247 * to kill current.We have to random task kill in this case.
248 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
249 */
250 if (gfp_mask & __GFP_THISNODE)
251 return CONSTRAINT_NONE;
252
253 /*
254 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
255 * the page allocator means a mempolicy is in effect. Cpuset policy
256 * is enforced in get_page_from_freelist().
257 */
258 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
259 *totalpages = total_swap_pages;
260 for_each_node_mask(nid, *nodemask)
261 *totalpages += node_spanned_pages(nid);
262 return CONSTRAINT_MEMORY_POLICY;
263 }
264
265 /* Check this allocation failure is caused by cpuset's wall function */
266 for_each_zone_zonelist_nodemask(zone, z, zonelist,
267 high_zoneidx, nodemask)
268 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
269 cpuset_limited = true;
270
271 if (cpuset_limited) {
272 *totalpages = total_swap_pages;
273 for_each_node_mask(nid, cpuset_current_mems_allowed)
274 *totalpages += node_spanned_pages(nid);
275 return CONSTRAINT_CPUSET;
276 }
277 return CONSTRAINT_NONE;
278}
279#else
280static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
281 gfp_t gfp_mask, nodemask_t *nodemask,
282 unsigned long *totalpages)
283{
284 *totalpages = totalram_pages + total_swap_pages;
285 return CONSTRAINT_NONE;
286}
287#endif
288
289/*
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. We expect the caller will lock the tasklist.
292 *
293 * (not docbooked, we don't want this one cluttering up the manual)
294 */
295static struct task_struct *select_bad_process(unsigned int *ppoints,
296 unsigned long totalpages, struct mem_cgroup *mem,
297 const nodemask_t *nodemask)
298{
299 struct task_struct *g, *p;
300 struct task_struct *chosen = NULL;
301 *ppoints = 0;
302
303 do_each_thread(g, p) {
304 unsigned int points;
305
306 if (p->exit_state)
307 continue;
308 if (oom_unkillable_task(p, mem, nodemask))
309 continue;
310
311 /*
312 * This task already has access to memory reserves and is
313 * being killed. Don't allow any other task access to the
314 * memory reserve.
315 *
316 * Note: this may have a chance of deadlock if it gets
317 * blocked waiting for another task which itself is waiting
318 * for memory. Is there a better alternative?
319 */
320 if (test_tsk_thread_flag(p, TIF_MEMDIE))
321 return ERR_PTR(-1UL);
322 if (!p->mm)
323 continue;
324
325 if (p->flags & PF_EXITING) {
326 /*
327 * If p is the current task and is in the process of
328 * releasing memory, we allow the "kill" to set
329 * TIF_MEMDIE, which will allow it to gain access to
330 * memory reserves. Otherwise, it may stall forever.
331 *
332 * The loop isn't broken here, however, in case other
333 * threads are found to have already been oom killed.
334 */
335 if (p == current) {
336 chosen = p;
337 *ppoints = 1000;
338 } else {
339 /*
340 * If this task is not being ptraced on exit,
341 * then wait for it to finish before killing
342 * some other task unnecessarily.
343 */
344 if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
345 return ERR_PTR(-1UL);
346 }
347 }
348
349 points = oom_badness(p, mem, nodemask, totalpages);
350 if (points > *ppoints) {
351 chosen = p;
352 *ppoints = points;
353 }
354 } while_each_thread(g, p);
355
356 return chosen;
357}
358
359/**
360 * dump_tasks - dump current memory state of all system tasks
361 * @mem: current's memory controller, if constrained
362 * @nodemask: nodemask passed to page allocator for mempolicy ooms
363 *
364 * Dumps the current memory state of all eligible tasks. Tasks not in the same
365 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
366 * are not shown.
367 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
368 * value, oom_score_adj value, and name.
369 *
370 * Call with tasklist_lock read-locked.
371 */
372static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
373{
374 struct task_struct *p;
375 struct task_struct *task;
376
377 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
378 for_each_process(p) {
379 if (oom_unkillable_task(p, mem, nodemask))
380 continue;
381
382 task = find_lock_task_mm(p);
383 if (!task) {
384 /*
385 * This is a kthread or all of p's threads have already
386 * detached their mm's. There's no need to report
387 * them; they can't be oom killed anyway.
388 */
389 continue;
390 }
391
392 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
393 task->pid, task_uid(task), task->tgid,
394 task->mm->total_vm, get_mm_rss(task->mm),
395 task_cpu(task), task->signal->oom_adj,
396 task->signal->oom_score_adj, task->comm);
397 task_unlock(task);
398 }
399}
400
401static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
402 struct mem_cgroup *mem, const nodemask_t *nodemask)
403{
404 task_lock(current);
405 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
406 "oom_adj=%d, oom_score_adj=%d\n",
407 current->comm, gfp_mask, order, current->signal->oom_adj,
408 current->signal->oom_score_adj);
409 cpuset_print_task_mems_allowed(current);
410 task_unlock(current);
411 dump_stack();
412 mem_cgroup_print_oom_info(mem, p);
413 show_mem(SHOW_MEM_FILTER_NODES);
414 if (sysctl_oom_dump_tasks)
415 dump_tasks(mem, nodemask);
416}
417
418#define K(x) ((x) << (PAGE_SHIFT-10))
419static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
420{
421 struct task_struct *q;
422 struct mm_struct *mm;
423
424 p = find_lock_task_mm(p);
425 if (!p)
426 return 1;
427
428 /* mm cannot be safely dereferenced after task_unlock(p) */
429 mm = p->mm;
430
431 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
432 task_pid_nr(p), p->comm, K(p->mm->total_vm),
433 K(get_mm_counter(p->mm, MM_ANONPAGES)),
434 K(get_mm_counter(p->mm, MM_FILEPAGES)));
435 task_unlock(p);
436
437 /*
438 * Kill all processes sharing p->mm in other thread groups, if any.
439 * They don't get access to memory reserves or a higher scheduler
440 * priority, though, to avoid depletion of all memory or task
441 * starvation. This prevents mm->mmap_sem livelock when an oom killed
442 * task cannot exit because it requires the semaphore and its contended
443 * by another thread trying to allocate memory itself. That thread will
444 * now get access to memory reserves since it has a pending fatal
445 * signal.
446 */
447 for_each_process(q)
448 if (q->mm == mm && !same_thread_group(q, p)) {
449 task_lock(q); /* Protect ->comm from prctl() */
450 pr_err("Kill process %d (%s) sharing same memory\n",
451 task_pid_nr(q), q->comm);
452 task_unlock(q);
453 force_sig(SIGKILL, q);
454 }
455
456 set_tsk_thread_flag(p, TIF_MEMDIE);
457 force_sig(SIGKILL, p);
458
459 return 0;
460}
461#undef K
462
463static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
464 unsigned int points, unsigned long totalpages,
465 struct mem_cgroup *mem, nodemask_t *nodemask,
466 const char *message)
467{
468 struct task_struct *victim = p;
469 struct task_struct *child;
470 struct task_struct *t = p;
471 unsigned int victim_points = 0;
472
473 if (printk_ratelimit())
474 dump_header(p, gfp_mask, order, mem, nodemask);
475
476 /*
477 * If the task is already exiting, don't alarm the sysadmin or kill
478 * its children or threads, just set TIF_MEMDIE so it can die quickly
479 */
480 if (p->flags & PF_EXITING) {
481 set_tsk_thread_flag(p, TIF_MEMDIE);
482 return 0;
483 }
484
485 task_lock(p);
486 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
487 message, task_pid_nr(p), p->comm, points);
488 task_unlock(p);
489
490 /*
491 * If any of p's children has a different mm and is eligible for kill,
492 * the one with the highest oom_badness() score is sacrificed for its
493 * parent. This attempts to lose the minimal amount of work done while
494 * still freeing memory.
495 */
496 do {
497 list_for_each_entry(child, &t->children, sibling) {
498 unsigned int child_points;
499
500 if (child->mm == p->mm)
501 continue;
502 /*
503 * oom_badness() returns 0 if the thread is unkillable
504 */
505 child_points = oom_badness(child, mem, nodemask,
506 totalpages);
507 if (child_points > victim_points) {
508 victim = child;
509 victim_points = child_points;
510 }
511 }
512 } while_each_thread(p, t);
513
514 return oom_kill_task(victim, mem);
515}
516
517/*
518 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
519 */
520static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
521 int order, const nodemask_t *nodemask)
522{
523 if (likely(!sysctl_panic_on_oom))
524 return;
525 if (sysctl_panic_on_oom != 2) {
526 /*
527 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
528 * does not panic for cpuset, mempolicy, or memcg allocation
529 * failures.
530 */
531 if (constraint != CONSTRAINT_NONE)
532 return;
533 }
534 read_lock(&tasklist_lock);
535 dump_header(NULL, gfp_mask, order, NULL, nodemask);
536 read_unlock(&tasklist_lock);
537 panic("Out of memory: %s panic_on_oom is enabled\n",
538 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
539}
540
541#ifdef CONFIG_CGROUP_MEM_RES_CTLR
542void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
543{
544 unsigned long limit;
545 unsigned int points = 0;
546 struct task_struct *p;
547
548 /*
549 * If current has a pending SIGKILL, then automatically select it. The
550 * goal is to allow it to allocate so that it may quickly exit and free
551 * its memory.
552 */
553 if (fatal_signal_pending(current)) {
554 set_thread_flag(TIF_MEMDIE);
555 return;
556 }
557
558 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
559 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
560 read_lock(&tasklist_lock);
561retry:
562 p = select_bad_process(&points, limit, mem, NULL);
563 if (!p || PTR_ERR(p) == -1UL)
564 goto out;
565
566 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
567 "Memory cgroup out of memory"))
568 goto retry;
569out:
570 read_unlock(&tasklist_lock);
571}
572#endif
573
574static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
575
576int register_oom_notifier(struct notifier_block *nb)
577{
578 return blocking_notifier_chain_register(&oom_notify_list, nb);
579}
580EXPORT_SYMBOL_GPL(register_oom_notifier);
581
582int unregister_oom_notifier(struct notifier_block *nb)
583{
584 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
585}
586EXPORT_SYMBOL_GPL(unregister_oom_notifier);
587
588/*
589 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
590 * if a parallel OOM killing is already taking place that includes a zone in
591 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
592 */
593int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
594{
595 struct zoneref *z;
596 struct zone *zone;
597 int ret = 1;
598
599 spin_lock(&zone_scan_lock);
600 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
601 if (zone_is_oom_locked(zone)) {
602 ret = 0;
603 goto out;
604 }
605 }
606
607 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
608 /*
609 * Lock each zone in the zonelist under zone_scan_lock so a
610 * parallel invocation of try_set_zonelist_oom() doesn't succeed
611 * when it shouldn't.
612 */
613 zone_set_flag(zone, ZONE_OOM_LOCKED);
614 }
615
616out:
617 spin_unlock(&zone_scan_lock);
618 return ret;
619}
620
621/*
622 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
623 * allocation attempts with zonelists containing them may now recall the OOM
624 * killer, if necessary.
625 */
626void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
627{
628 struct zoneref *z;
629 struct zone *zone;
630
631 spin_lock(&zone_scan_lock);
632 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
633 zone_clear_flag(zone, ZONE_OOM_LOCKED);
634 }
635 spin_unlock(&zone_scan_lock);
636}
637
638/*
639 * Try to acquire the oom killer lock for all system zones. Returns zero if a
640 * parallel oom killing is taking place, otherwise locks all zones and returns
641 * non-zero.
642 */
643static int try_set_system_oom(void)
644{
645 struct zone *zone;
646 int ret = 1;
647
648 spin_lock(&zone_scan_lock);
649 for_each_populated_zone(zone)
650 if (zone_is_oom_locked(zone)) {
651 ret = 0;
652 goto out;
653 }
654 for_each_populated_zone(zone)
655 zone_set_flag(zone, ZONE_OOM_LOCKED);
656out:
657 spin_unlock(&zone_scan_lock);
658 return ret;
659}
660
661/*
662 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
663 * attempts or page faults may now recall the oom killer, if necessary.
664 */
665static void clear_system_oom(void)
666{
667 struct zone *zone;
668
669 spin_lock(&zone_scan_lock);
670 for_each_populated_zone(zone)
671 zone_clear_flag(zone, ZONE_OOM_LOCKED);
672 spin_unlock(&zone_scan_lock);
673}
674
675/**
676 * out_of_memory - kill the "best" process when we run out of memory
677 * @zonelist: zonelist pointer
678 * @gfp_mask: memory allocation flags
679 * @order: amount of memory being requested as a power of 2
680 * @nodemask: nodemask passed to page allocator
681 *
682 * If we run out of memory, we have the choice between either
683 * killing a random task (bad), letting the system crash (worse)
684 * OR try to be smart about which process to kill. Note that we
685 * don't have to be perfect here, we just have to be good.
686 */
687void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
688 int order, nodemask_t *nodemask)
689{
690 const nodemask_t *mpol_mask;
691 struct task_struct *p;
692 unsigned long totalpages;
693 unsigned long freed = 0;
694 unsigned int points;
695 enum oom_constraint constraint = CONSTRAINT_NONE;
696 int killed = 0;
697
698 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
699 if (freed > 0)
700 /* Got some memory back in the last second. */
701 return;
702
703 /*
704 * If current has a pending SIGKILL, then automatically select it. The
705 * goal is to allow it to allocate so that it may quickly exit and free
706 * its memory.
707 */
708 if (fatal_signal_pending(current)) {
709 set_thread_flag(TIF_MEMDIE);
710 return;
711 }
712
713 /*
714 * Check if there were limitations on the allocation (only relevant for
715 * NUMA) that may require different handling.
716 */
717 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
718 &totalpages);
719 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
720 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
721
722 read_lock(&tasklist_lock);
723 if (sysctl_oom_kill_allocating_task &&
724 !oom_unkillable_task(current, NULL, nodemask) &&
725 current->mm && !atomic_read(¤t->mm->oom_disable_count)) {
726 /*
727 * oom_kill_process() needs tasklist_lock held. If it returns
728 * non-zero, current could not be killed so we must fallback to
729 * the tasklist scan.
730 */
731 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
732 NULL, nodemask,
733 "Out of memory (oom_kill_allocating_task)"))
734 goto out;
735 }
736
737retry:
738 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
739 if (PTR_ERR(p) == -1UL)
740 goto out;
741
742 /* Found nothing?!?! Either we hang forever, or we panic. */
743 if (!p) {
744 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
745 read_unlock(&tasklist_lock);
746 panic("Out of memory and no killable processes...\n");
747 }
748
749 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
750 nodemask, "Out of memory"))
751 goto retry;
752 killed = 1;
753out:
754 read_unlock(&tasklist_lock);
755
756 /*
757 * Give "p" a good chance of killing itself before we
758 * retry to allocate memory unless "p" is current
759 */
760 if (killed && !test_thread_flag(TIF_MEMDIE))
761 schedule_timeout_uninterruptible(1);
762}
763
764/*
765 * The pagefault handler calls here because it is out of memory, so kill a
766 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
767 * oom killing is already in progress so do nothing. If a task is found with
768 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
769 */
770void pagefault_out_of_memory(void)
771{
772 if (try_set_system_oom()) {
773 out_of_memory(NULL, 0, 0, NULL);
774 clear_system_oom();
775 }
776 if (!test_thread_flag(TIF_MEMDIE))
777 schedule_timeout_uninterruptible(1);
778}