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
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.
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.
20 #include <linux/oom.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>
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/oom.h>
42 int sysctl_panic_on_oom;
43 int sysctl_oom_kill_allocating_task;
44 int sysctl_oom_dump_tasks = 1;
45 static DEFINE_SPINLOCK(zone_scan_lock);
49 * has_intersects_mems_allowed() - check task eligiblity for kill
50 * @tsk: task struct of which task to consider
51 * @mask: nodemask passed to page allocator for mempolicy ooms
53 * Task eligibility is determined by whether or not a candidate task, @tsk,
54 * shares the same mempolicy nodes as current if it is bound by such a policy
55 * and whether or not it has the same set of allowed cpuset nodes.
57 static bool has_intersects_mems_allowed(struct task_struct *tsk,
58 const nodemask_t *mask)
60 struct task_struct *start = tsk;
65 * If this is a mempolicy constrained oom, tsk's
66 * cpuset is irrelevant. Only return true if its
67 * mempolicy intersects current, otherwise it may be
70 if (mempolicy_nodemask_intersects(tsk, mask))
74 * This is not a mempolicy constrained oom, so only
75 * check the mems of tsk's cpuset.
77 if (cpuset_mems_allowed_intersects(current, tsk))
80 } while_each_thread(start, tsk);
85 static bool has_intersects_mems_allowed(struct task_struct *tsk,
86 const nodemask_t *mask)
90 #endif /* CONFIG_NUMA */
93 * The process p may have detached its own ->mm while exiting or through
94 * use_mm(), but one or more of its subthreads may still have a valid
95 * pointer. Return p, or any of its subthreads with a valid ->mm, with
98 struct task_struct *find_lock_task_mm(struct task_struct *p)
100 struct task_struct *t = p;
107 } while_each_thread(p, t);
112 /* return true if the task is not adequate as candidate victim task. */
113 static bool oom_unkillable_task(struct task_struct *p,
114 const struct mem_cgroup *memcg, const nodemask_t *nodemask)
116 if (is_global_init(p))
118 if (p->flags & PF_KTHREAD)
121 /* When mem_cgroup_out_of_memory() and p is not member of the group */
122 if (memcg && !task_in_mem_cgroup(p, memcg))
125 /* p may not have freeable memory in nodemask */
126 if (!has_intersects_mems_allowed(p, nodemask))
133 * oom_badness - heuristic function to determine which candidate task to kill
134 * @p: task struct of which task we should calculate
135 * @totalpages: total present RAM allowed for page allocation
137 * The heuristic for determining which task to kill is made to be as simple and
138 * predictable as possible. The goal is to return the highest value for the
139 * task consuming the most memory to avoid subsequent oom failures.
141 unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
142 const nodemask_t *nodemask, unsigned long totalpages)
147 if (oom_unkillable_task(p, memcg, nodemask))
150 p = find_lock_task_mm(p);
154 adj = (long)p->signal->oom_score_adj;
155 if (adj == OOM_SCORE_ADJ_MIN) {
161 * The baseline for the badness score is the proportion of RAM that each
162 * task's rss, pagetable and swap space use.
164 points = get_mm_rss(p->mm) + p->mm->nr_ptes +
165 get_mm_counter(p->mm, MM_SWAPENTS);
169 * Root processes get 3% bonus, just like the __vm_enough_memory()
170 * implementation used by LSMs.
172 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
175 /* Normalize to oom_score_adj units */
176 adj *= totalpages / 1000;
180 * Never return 0 for an eligible task regardless of the root bonus and
181 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
183 return points > 0 ? points : 1;
187 * Determine the type of allocation constraint.
190 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
191 gfp_t gfp_mask, nodemask_t *nodemask,
192 unsigned long *totalpages)
196 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
197 bool cpuset_limited = false;
200 /* Default to all available memory */
201 *totalpages = totalram_pages + total_swap_pages;
204 return CONSTRAINT_NONE;
206 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
207 * to kill current.We have to random task kill in this case.
208 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
210 if (gfp_mask & __GFP_THISNODE)
211 return CONSTRAINT_NONE;
214 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
215 * the page allocator means a mempolicy is in effect. Cpuset policy
216 * is enforced in get_page_from_freelist().
218 if (nodemask && !nodes_subset(node_states[N_MEMORY], *nodemask)) {
219 *totalpages = total_swap_pages;
220 for_each_node_mask(nid, *nodemask)
221 *totalpages += node_spanned_pages(nid);
222 return CONSTRAINT_MEMORY_POLICY;
225 /* Check this allocation failure is caused by cpuset's wall function */
226 for_each_zone_zonelist_nodemask(zone, z, zonelist,
227 high_zoneidx, nodemask)
228 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
229 cpuset_limited = true;
231 if (cpuset_limited) {
232 *totalpages = total_swap_pages;
233 for_each_node_mask(nid, cpuset_current_mems_allowed)
234 *totalpages += node_spanned_pages(nid);
235 return CONSTRAINT_CPUSET;
237 return CONSTRAINT_NONE;
240 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
241 gfp_t gfp_mask, nodemask_t *nodemask,
242 unsigned long *totalpages)
244 *totalpages = totalram_pages + total_swap_pages;
245 return CONSTRAINT_NONE;
249 enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
250 unsigned long totalpages, const nodemask_t *nodemask,
253 if (task->exit_state)
254 return OOM_SCAN_CONTINUE;
255 if (oom_unkillable_task(task, NULL, nodemask))
256 return OOM_SCAN_CONTINUE;
259 * This task already has access to memory reserves and is being killed.
260 * Don't allow any other task to have access to the reserves.
262 if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
263 if (unlikely(frozen(task)))
266 return OOM_SCAN_ABORT;
269 return OOM_SCAN_CONTINUE;
272 * If task is allocating a lot of memory and has been marked to be
273 * killed first if it triggers an oom, then select it.
275 if (oom_task_origin(task))
276 return OOM_SCAN_SELECT;
278 if (task->flags & PF_EXITING && !force_kill) {
280 * If this task is not being ptraced on exit, then wait for it
281 * to finish before killing some other task unnecessarily.
283 if (!(task->group_leader->ptrace & PT_TRACE_EXIT))
284 return OOM_SCAN_ABORT;
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. Returns -1 on scan abort.
293 * (not docbooked, we don't want this one cluttering up the manual)
295 static struct task_struct *select_bad_process(unsigned int *ppoints,
296 unsigned long totalpages, const nodemask_t *nodemask,
299 struct task_struct *g, *p;
300 struct task_struct *chosen = NULL;
301 unsigned long chosen_points = 0;
304 do_each_thread(g, p) {
307 switch (oom_scan_process_thread(p, totalpages, nodemask,
309 case OOM_SCAN_SELECT:
311 chosen_points = ULONG_MAX;
313 case OOM_SCAN_CONTINUE:
317 return (struct task_struct *)(-1UL);
321 points = oom_badness(p, NULL, nodemask, totalpages);
322 if (points > chosen_points) {
324 chosen_points = points;
326 } while_each_thread(g, p);
328 get_task_struct(chosen);
331 *ppoints = chosen_points * 1000 / totalpages;
336 * dump_tasks - dump current memory state of all system tasks
337 * @memcg: current's memory controller, if constrained
338 * @nodemask: nodemask passed to page allocator for mempolicy ooms
340 * Dumps the current memory state of all eligible tasks. Tasks not in the same
341 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
343 * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes,
344 * swapents, oom_score_adj value, and name.
346 static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
348 struct task_struct *p;
349 struct task_struct *task;
351 pr_info("[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name\n");
353 for_each_process(p) {
354 if (oom_unkillable_task(p, memcg, nodemask))
357 task = find_lock_task_mm(p);
360 * This is a kthread or all of p's threads have already
361 * detached their mm's. There's no need to report
362 * them; they can't be oom killed anyway.
367 pr_info("[%5d] %5d %5d %8lu %8lu %7lu %8lu %5hd %s\n",
368 task->pid, from_kuid(&init_user_ns, task_uid(task)),
369 task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
371 get_mm_counter(task->mm, MM_SWAPENTS),
372 task->signal->oom_score_adj, task->comm);
378 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
379 struct mem_cgroup *memcg, const nodemask_t *nodemask)
382 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
383 "oom_score_adj=%hd\n",
384 current->comm, gfp_mask, order,
385 current->signal->oom_score_adj);
386 cpuset_print_task_mems_allowed(current);
387 task_unlock(current);
390 mem_cgroup_print_oom_info(memcg, p);
392 show_mem(SHOW_MEM_FILTER_NODES);
393 if (sysctl_oom_dump_tasks)
394 dump_tasks(memcg, nodemask);
397 #define K(x) ((x) << (PAGE_SHIFT-10))
399 * Must be called while holding a reference to p, which will be released upon
402 void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
403 unsigned int points, unsigned long totalpages,
404 struct mem_cgroup *memcg, nodemask_t *nodemask,
407 struct task_struct *victim = p;
408 struct task_struct *child;
409 struct task_struct *t = p;
410 struct mm_struct *mm;
411 unsigned int victim_points = 0;
412 static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
413 DEFAULT_RATELIMIT_BURST);
416 * If the task is already exiting, don't alarm the sysadmin or kill
417 * its children or threads, just set TIF_MEMDIE so it can die quickly
419 if (p->flags & PF_EXITING) {
420 set_tsk_thread_flag(p, TIF_MEMDIE);
425 if (__ratelimit(&oom_rs))
426 dump_header(p, gfp_mask, order, memcg, nodemask);
429 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
430 message, task_pid_nr(p), p->comm, points);
434 * If any of p's children has a different mm and is eligible for kill,
435 * the one with the highest oom_badness() score is sacrificed for its
436 * parent. This attempts to lose the minimal amount of work done while
437 * still freeing memory.
439 read_lock(&tasklist_lock);
441 list_for_each_entry(child, &t->children, sibling) {
442 unsigned int child_points;
444 if (child->mm == p->mm)
447 * oom_badness() returns 0 if the thread is unkillable
449 child_points = oom_badness(child, memcg, nodemask,
451 if (child_points > victim_points) {
452 put_task_struct(victim);
454 victim_points = child_points;
455 get_task_struct(victim);
458 } while_each_thread(p, t);
459 read_unlock(&tasklist_lock);
462 p = find_lock_task_mm(victim);
465 put_task_struct(victim);
467 } else if (victim != p) {
469 put_task_struct(victim);
473 /* mm cannot safely be dereferenced after task_unlock(victim) */
475 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
476 task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
477 K(get_mm_counter(victim->mm, MM_ANONPAGES)),
478 K(get_mm_counter(victim->mm, MM_FILEPAGES)));
482 * Kill all user processes sharing victim->mm in other thread groups, if
483 * any. They don't get access to memory reserves, though, to avoid
484 * depletion of all memory. This prevents mm->mmap_sem livelock when an
485 * oom killed thread cannot exit because it requires the semaphore and
486 * its contended by another thread trying to allocate memory itself.
487 * That thread will now get access to memory reserves since it has a
488 * pending fatal signal.
491 if (p->mm == mm && !same_thread_group(p, victim) &&
492 !(p->flags & PF_KTHREAD)) {
493 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
496 task_lock(p); /* Protect ->comm from prctl() */
497 pr_err("Kill process %d (%s) sharing same memory\n",
498 task_pid_nr(p), p->comm);
500 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
504 set_tsk_thread_flag(victim, TIF_MEMDIE);
505 do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
506 put_task_struct(victim);
511 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
513 void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
514 int order, const nodemask_t *nodemask)
516 if (likely(!sysctl_panic_on_oom))
518 if (sysctl_panic_on_oom != 2) {
520 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
521 * does not panic for cpuset, mempolicy, or memcg allocation
524 if (constraint != CONSTRAINT_NONE)
527 dump_header(NULL, gfp_mask, order, NULL, nodemask);
528 panic("Out of memory: %s panic_on_oom is enabled\n",
529 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
532 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
534 int register_oom_notifier(struct notifier_block *nb)
536 return blocking_notifier_chain_register(&oom_notify_list, nb);
538 EXPORT_SYMBOL_GPL(register_oom_notifier);
540 int unregister_oom_notifier(struct notifier_block *nb)
542 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
544 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
547 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
548 * if a parallel OOM killing is already taking place that includes a zone in
549 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
551 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
557 spin_lock(&zone_scan_lock);
558 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
559 if (zone_is_oom_locked(zone)) {
565 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
567 * Lock each zone in the zonelist under zone_scan_lock so a
568 * parallel invocation of try_set_zonelist_oom() doesn't succeed
571 zone_set_flag(zone, ZONE_OOM_LOCKED);
575 spin_unlock(&zone_scan_lock);
580 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
581 * allocation attempts with zonelists containing them may now recall the OOM
582 * killer, if necessary.
584 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
589 spin_lock(&zone_scan_lock);
590 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
591 zone_clear_flag(zone, ZONE_OOM_LOCKED);
593 spin_unlock(&zone_scan_lock);
597 * out_of_memory - kill the "best" process when we run out of memory
598 * @zonelist: zonelist pointer
599 * @gfp_mask: memory allocation flags
600 * @order: amount of memory being requested as a power of 2
601 * @nodemask: nodemask passed to page allocator
602 * @force_kill: true if a task must be killed, even if others are exiting
604 * If we run out of memory, we have the choice between either
605 * killing a random task (bad), letting the system crash (worse)
606 * OR try to be smart about which process to kill. Note that we
607 * don't have to be perfect here, we just have to be good.
609 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
610 int order, nodemask_t *nodemask, bool force_kill)
612 const nodemask_t *mpol_mask;
613 struct task_struct *p;
614 unsigned long totalpages;
615 unsigned long freed = 0;
616 unsigned int uninitialized_var(points);
617 enum oom_constraint constraint = CONSTRAINT_NONE;
620 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
622 /* Got some memory back in the last second. */
626 * If current has a pending SIGKILL or is exiting, then automatically
627 * select it. The goal is to allow it to allocate so that it may
628 * quickly exit and free its memory.
630 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
631 set_thread_flag(TIF_MEMDIE);
636 * Check if there were limitations on the allocation (only relevant for
637 * NUMA) that may require different handling.
639 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
641 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
642 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
644 if (sysctl_oom_kill_allocating_task && current->mm &&
645 !oom_unkillable_task(current, NULL, nodemask) &&
646 current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
647 get_task_struct(current);
648 oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
650 "Out of memory (oom_kill_allocating_task)");
654 p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
655 /* Found nothing?!?! Either we hang forever, or we panic. */
657 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
658 panic("Out of memory and no killable processes...\n");
660 if (p != (void *)-1UL) {
661 oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
662 nodemask, "Out of memory");
667 * Give the killed threads a good chance of exiting before trying to
668 * allocate memory again.
671 schedule_timeout_killable(1);
675 * The pagefault handler calls here because it is out of memory, so kill a
676 * memory-hogging task. If any populated zone has ZONE_OOM_LOCKED set, a
677 * parallel oom killing is already in progress so do nothing.
679 void pagefault_out_of_memory(void)
681 struct zonelist *zonelist = node_zonelist(first_online_node,
684 if (try_set_zonelist_oom(zonelist, GFP_KERNEL)) {
685 out_of_memory(NULL, 0, 0, NULL, false);
686 clear_zonelist_oom(zonelist, GFP_KERNEL);