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...
8 * The routines in this file are used to kill a process when
9 * we're seriously out of memory. This gets called from __alloc_pages()
10 * in mm/page_alloc.c when we really run out of memory.
12 * Since we won't call these routines often (on a well-configured
13 * machine) this file will double as a 'coding guide' and a signpost
14 * for newbie kernel hackers. It features several pointers to major
15 * kernel subsystems and hints as to where to find out what things do.
18 #include <linux/oom.h>
20 #include <linux/err.h>
21 #include <linux/gfp.h>
22 #include <linux/sched.h>
23 #include <linux/swap.h>
24 #include <linux/timex.h>
25 #include <linux/jiffies.h>
26 #include <linux/cpuset.h>
27 #include <linux/module.h>
28 #include <linux/notifier.h>
29 #include <linux/memcontrol.h>
30 #include <linux/mempolicy.h>
31 #include <linux/security.h>
33 int sysctl_panic_on_oom;
34 int sysctl_oom_kill_allocating_task;
35 int sysctl_oom_dump_tasks = 1;
36 static DEFINE_SPINLOCK(zone_scan_lock);
41 * has_intersects_mems_allowed() - check task eligiblity for kill
42 * @tsk: task struct of which task to consider
43 * @mask: nodemask passed to page allocator for mempolicy ooms
45 * Task eligibility is determined by whether or not a candidate task, @tsk,
46 * shares the same mempolicy nodes as current if it is bound by such a policy
47 * and whether or not it has the same set of allowed cpuset nodes.
49 static bool has_intersects_mems_allowed(struct task_struct *tsk,
50 const nodemask_t *mask)
52 struct task_struct *start = tsk;
57 * If this is a mempolicy constrained oom, tsk's
58 * cpuset is irrelevant. Only return true if its
59 * mempolicy intersects current, otherwise it may be
62 if (mempolicy_nodemask_intersects(tsk, mask))
66 * This is not a mempolicy constrained oom, so only
67 * check the mems of tsk's cpuset.
69 if (cpuset_mems_allowed_intersects(current, tsk))
72 tsk = next_thread(tsk);
73 } while (tsk != start);
77 static bool has_intersects_mems_allowed(struct task_struct *tsk,
78 const nodemask_t *mask)
82 #endif /* CONFIG_NUMA */
85 * The process p may have detached its own ->mm while exiting or through
86 * use_mm(), but one or more of its subthreads may still have a valid
87 * pointer. Return p, or any of its subthreads with a valid ->mm, with
90 static struct task_struct *find_lock_task_mm(struct task_struct *p)
92 struct task_struct *t = p;
99 } while_each_thread(p, t);
104 /* return true if the task is not adequate as candidate victim task. */
105 static bool oom_unkillable_task(struct task_struct *p, struct mem_cgroup *mem,
106 const nodemask_t *nodemask)
108 if (is_global_init(p))
110 if (p->flags & PF_KTHREAD)
113 /* When mem_cgroup_out_of_memory() and p is not member of the group */
114 if (mem && !task_in_mem_cgroup(p, mem))
117 /* p may not have freeable memory in nodemask */
118 if (!has_intersects_mems_allowed(p, nodemask))
125 * badness - calculate a numeric value for how bad this task has been
126 * @p: task struct of which task we should calculate
127 * @uptime: current uptime in seconds
129 * The formula used is relatively simple and documented inline in the
130 * function. The main rationale is that we want to select a good task
131 * to kill when we run out of memory.
133 * Good in this context means that:
134 * 1) we lose the minimum amount of work done
135 * 2) we recover a large amount of memory
136 * 3) we don't kill anything innocent of eating tons of memory
137 * 4) we want to kill the minimum amount of processes (one)
138 * 5) we try to kill the process the user expects us to kill, this
139 * algorithm has been meticulously tuned to meet the principle
140 * of least surprise ... (be careful when you change it)
142 unsigned long badness(struct task_struct *p, struct mem_cgroup *mem,
143 const nodemask_t *nodemask, unsigned long uptime)
145 unsigned long points, cpu_time, run_time;
146 struct task_struct *child;
147 struct task_struct *c, *t;
148 int oom_adj = p->signal->oom_adj;
149 struct task_cputime task_time;
153 if (oom_unkillable_task(p, mem, nodemask))
155 if (oom_adj == OOM_DISABLE)
158 p = find_lock_task_mm(p);
163 * The memory size of the process is the basis for the badness.
165 points = p->mm->total_vm;
169 * swapoff can easily use up all memory, so kill those first.
171 if (p->flags & PF_OOM_ORIGIN)
175 * Processes which fork a lot of child processes are likely
176 * a good choice. We add half the vmsize of the children if they
177 * have an own mm. This prevents forking servers to flood the
178 * machine with an endless amount of children. In case a single
179 * child is eating the vast majority of memory, adding only half
180 * to the parents will make the child our kill candidate of choice.
184 list_for_each_entry(c, &t->children, sibling) {
185 child = find_lock_task_mm(c);
187 if (child->mm != p->mm)
188 points += child->mm->total_vm/2 + 1;
192 } while_each_thread(p, t);
195 * CPU time is in tens of seconds and run time is in thousands
196 * of seconds. There is no particular reason for this other than
197 * that it turned out to work very well in practice.
199 thread_group_cputime(p, &task_time);
200 utime = cputime_to_jiffies(task_time.utime);
201 stime = cputime_to_jiffies(task_time.stime);
202 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
205 if (uptime >= p->start_time.tv_sec)
206 run_time = (uptime - p->start_time.tv_sec) >> 10;
211 points /= int_sqrt(cpu_time);
213 points /= int_sqrt(int_sqrt(run_time));
216 * Niced processes are most likely less important, so double
217 * their badness points.
219 if (task_nice(p) > 0)
223 * Superuser processes are usually more important, so we make it
224 * less likely that we kill those.
226 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
227 has_capability_noaudit(p, CAP_SYS_RESOURCE))
231 * We don't want to kill a process with direct hardware access.
232 * Not only could that mess up the hardware, but usually users
233 * tend to only have this flag set on applications they think
236 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
240 * Adjust the score by oom_adj.
248 points >>= -(oom_adj);
252 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
253 p->pid, p->comm, points);
259 * Determine the type of allocation constraint.
262 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
263 gfp_t gfp_mask, nodemask_t *nodemask)
267 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
270 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
271 * to kill current.We have to random task kill in this case.
272 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
274 if (gfp_mask & __GFP_THISNODE)
275 return CONSTRAINT_NONE;
278 * The nodemask here is a nodemask passed to alloc_pages(). Now,
279 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
280 * feature. mempolicy is an only user of nodemask here.
281 * check mempolicy's nodemask contains all N_HIGH_MEMORY
283 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask))
284 return CONSTRAINT_MEMORY_POLICY;
286 /* Check this allocation failure is caused by cpuset's wall function */
287 for_each_zone_zonelist_nodemask(zone, z, zonelist,
288 high_zoneidx, nodemask)
289 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
290 return CONSTRAINT_CPUSET;
292 return CONSTRAINT_NONE;
295 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
296 gfp_t gfp_mask, nodemask_t *nodemask)
298 return CONSTRAINT_NONE;
303 * Simple selection loop. We chose the process with the highest
304 * number of 'points'. We expect the caller will lock the tasklist.
306 * (not docbooked, we don't want this one cluttering up the manual)
308 static struct task_struct *select_bad_process(unsigned long *ppoints,
309 struct mem_cgroup *mem, const nodemask_t *nodemask)
311 struct task_struct *p;
312 struct task_struct *chosen = NULL;
313 struct timespec uptime;
316 do_posix_clock_monotonic_gettime(&uptime);
317 for_each_process(p) {
318 unsigned long points;
320 if (oom_unkillable_task(p, mem, nodemask))
324 * This task already has access to memory reserves and is
325 * being killed. Don't allow any other task access to the
328 * Note: this may have a chance of deadlock if it gets
329 * blocked waiting for another task which itself is waiting
330 * for memory. Is there a better alternative?
332 if (test_tsk_thread_flag(p, TIF_MEMDIE))
333 return ERR_PTR(-1UL);
336 * This is in the process of releasing memory so wait for it
337 * to finish before killing some other task by mistake.
339 * However, if p is the current task, we allow the 'kill' to
340 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
341 * which will allow it to gain access to memory reserves in
342 * the process of exiting and releasing its resources.
343 * Otherwise we could get an easy OOM deadlock.
345 if ((p->flags & PF_EXITING) && p->mm) {
347 return ERR_PTR(-1UL);
350 *ppoints = ULONG_MAX;
353 points = badness(p, mem, nodemask, uptime.tv_sec);
354 if (points > *ppoints || !chosen) {
364 * dump_tasks - dump current memory state of all system tasks
365 * @mem: current's memory controller, if constrained
367 * Dumps the current memory state of all system tasks, excluding kernel threads.
368 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
371 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
374 * Call with tasklist_lock read-locked.
376 static void dump_tasks(const struct mem_cgroup *mem)
378 struct task_struct *p;
379 struct task_struct *task;
381 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
383 for_each_process(p) {
384 if (p->flags & PF_KTHREAD)
386 if (mem && !task_in_mem_cgroup(p, mem))
389 task = find_lock_task_mm(p);
392 * This is a kthread or all of p's threads have already
393 * detached their mm's. There's no need to report
394 * them; they can't be oom killed anyway.
399 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3u %3d %s\n",
400 task->pid, __task_cred(task)->uid, task->tgid,
401 task->mm->total_vm, get_mm_rss(task->mm),
402 task_cpu(task), task->signal->oom_adj, task->comm);
407 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
408 struct mem_cgroup *mem)
411 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
413 current->comm, gfp_mask, order, current->signal->oom_adj);
414 cpuset_print_task_mems_allowed(current);
415 task_unlock(current);
417 mem_cgroup_print_oom_info(mem, p);
419 if (sysctl_oom_dump_tasks)
423 #define K(x) ((x) << (PAGE_SHIFT-10))
424 static int oom_kill_task(struct task_struct *p)
426 p = find_lock_task_mm(p);
427 if (!p || p->signal->oom_adj == OOM_DISABLE) {
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)));
437 p->rt.time_slice = HZ;
438 set_tsk_thread_flag(p, TIF_MEMDIE);
439 force_sig(SIGKILL, p);
444 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
445 unsigned long points, struct mem_cgroup *mem,
446 nodemask_t *nodemask, const char *message)
448 struct task_struct *victim = p;
449 struct task_struct *child;
450 struct task_struct *t = p;
451 unsigned long victim_points = 0;
452 struct timespec uptime;
454 if (printk_ratelimit())
455 dump_header(p, gfp_mask, order, mem);
458 * If the task is already exiting, don't alarm the sysadmin or kill
459 * its children or threads, just set TIF_MEMDIE so it can die quickly
461 if (p->flags & PF_EXITING) {
462 set_tsk_thread_flag(p, TIF_MEMDIE);
467 pr_err("%s: Kill process %d (%s) score %lu or sacrifice child\n",
468 message, task_pid_nr(p), p->comm, points);
472 * If any of p's children has a different mm and is eligible for kill,
473 * the one with the highest badness() score is sacrificed for its
474 * parent. This attempts to lose the minimal amount of work done while
475 * still freeing memory.
477 do_posix_clock_monotonic_gettime(&uptime);
479 list_for_each_entry(child, &t->children, sibling) {
480 unsigned long child_points;
482 if (child->mm == p->mm)
485 /* badness() returns 0 if the thread is unkillable */
486 child_points = badness(child, mem, nodemask,
488 if (child_points > victim_points) {
490 victim_points = child_points;
493 } while_each_thread(p, t);
495 return oom_kill_task(victim);
499 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
501 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
504 if (likely(!sysctl_panic_on_oom))
506 if (sysctl_panic_on_oom != 2) {
508 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
509 * does not panic for cpuset, mempolicy, or memcg allocation
512 if (constraint != CONSTRAINT_NONE)
515 read_lock(&tasklist_lock);
516 dump_header(NULL, gfp_mask, order, NULL);
517 read_unlock(&tasklist_lock);
518 panic("Out of memory: %s panic_on_oom is enabled\n",
519 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
522 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
523 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
525 unsigned long points = 0;
526 struct task_struct *p;
528 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0);
529 read_lock(&tasklist_lock);
531 p = select_bad_process(&points, mem, NULL);
532 if (!p || PTR_ERR(p) == -1UL)
535 if (oom_kill_process(p, gfp_mask, 0, points, mem, NULL,
536 "Memory cgroup out of memory"))
539 read_unlock(&tasklist_lock);
543 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
545 int register_oom_notifier(struct notifier_block *nb)
547 return blocking_notifier_chain_register(&oom_notify_list, nb);
549 EXPORT_SYMBOL_GPL(register_oom_notifier);
551 int unregister_oom_notifier(struct notifier_block *nb)
553 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
555 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
558 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
559 * if a parallel OOM killing is already taking place that includes a zone in
560 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
562 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
568 spin_lock(&zone_scan_lock);
569 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
570 if (zone_is_oom_locked(zone)) {
576 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
578 * Lock each zone in the zonelist under zone_scan_lock so a
579 * parallel invocation of try_set_zonelist_oom() doesn't succeed
582 zone_set_flag(zone, ZONE_OOM_LOCKED);
586 spin_unlock(&zone_scan_lock);
591 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
592 * allocation attempts with zonelists containing them may now recall the OOM
593 * killer, if necessary.
595 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
600 spin_lock(&zone_scan_lock);
601 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
602 zone_clear_flag(zone, ZONE_OOM_LOCKED);
604 spin_unlock(&zone_scan_lock);
608 * Try to acquire the oom killer lock for all system zones. Returns zero if a
609 * parallel oom killing is taking place, otherwise locks all zones and returns
612 static int try_set_system_oom(void)
617 spin_lock(&zone_scan_lock);
618 for_each_populated_zone(zone)
619 if (zone_is_oom_locked(zone)) {
623 for_each_populated_zone(zone)
624 zone_set_flag(zone, ZONE_OOM_LOCKED);
626 spin_unlock(&zone_scan_lock);
631 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
632 * attempts or page faults may now recall the oom killer, if necessary.
634 static void clear_system_oom(void)
638 spin_lock(&zone_scan_lock);
639 for_each_populated_zone(zone)
640 zone_clear_flag(zone, ZONE_OOM_LOCKED);
641 spin_unlock(&zone_scan_lock);
645 * out_of_memory - kill the "best" process when we run out of memory
646 * @zonelist: zonelist pointer
647 * @gfp_mask: memory allocation flags
648 * @order: amount of memory being requested as a power of 2
649 * @nodemask: nodemask passed to page allocator
651 * If we run out of memory, we have the choice between either
652 * killing a random task (bad), letting the system crash (worse)
653 * OR try to be smart about which process to kill. Note that we
654 * don't have to be perfect here, we just have to be good.
656 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
657 int order, nodemask_t *nodemask)
659 struct task_struct *p;
660 unsigned long freed = 0;
661 unsigned long points;
662 enum oom_constraint constraint = CONSTRAINT_NONE;
664 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
666 /* Got some memory back in the last second. */
670 * If current has a pending SIGKILL, then automatically select it. The
671 * goal is to allow it to allocate so that it may quickly exit and free
674 if (fatal_signal_pending(current)) {
675 set_thread_flag(TIF_MEMDIE);
680 * Check if there were limitations on the allocation (only relevant for
681 * NUMA) that may require different handling.
684 constraint = constrained_alloc(zonelist, gfp_mask, nodemask);
685 check_panic_on_oom(constraint, gfp_mask, order);
687 read_lock(&tasklist_lock);
688 if (sysctl_oom_kill_allocating_task &&
689 !oom_unkillable_task(current, NULL, nodemask)) {
691 * oom_kill_process() needs tasklist_lock held. If it returns
692 * non-zero, current could not be killed so we must fallback to
695 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
697 "Out of memory (oom_kill_allocating_task)"))
702 p = select_bad_process(&points, NULL,
703 constraint == CONSTRAINT_MEMORY_POLICY ? nodemask :
705 if (PTR_ERR(p) == -1UL)
708 /* Found nothing?!?! Either we hang forever, or we panic. */
710 dump_header(NULL, gfp_mask, order, NULL);
711 read_unlock(&tasklist_lock);
712 panic("Out of memory and no killable processes...\n");
715 if (oom_kill_process(p, gfp_mask, order, points, NULL, nodemask,
718 read_unlock(&tasklist_lock);
721 * Give "p" a good chance of killing itself before we
722 * retry to allocate memory unless "p" is current
724 if (!test_thread_flag(TIF_MEMDIE))
725 schedule_timeout_uninterruptible(1);
729 * The pagefault handler calls here because it is out of memory, so kill a
730 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
731 * oom killing is already in progress so do nothing. If a task is found with
732 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
734 void pagefault_out_of_memory(void)
736 if (try_set_system_oom()) {
737 out_of_memory(NULL, 0, 0, NULL);
740 if (!test_thread_flag(TIF_MEMDIE))
741 schedule_timeout_uninterruptible(1);