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 } while_each_thread(start, tsk);
77 static bool has_intersects_mems_allowed(struct task_struct *tsk,
78 const nodemask_t *mask)
82 #endif /* CONFIG_NUMA */
85 * If this is a system OOM (not a memcg OOM) and the task selected to be
86 * killed is not already running at high (RT) priorities, speed up the
87 * recovery by boosting the dying task to the lowest FIFO priority.
88 * That helps with the recovery and avoids interfering with RT tasks.
90 static void boost_dying_task_prio(struct task_struct *p,
91 struct mem_cgroup *mem)
93 struct sched_param param = { .sched_priority = 1 };
99 sched_setscheduler_nocheck(p, SCHED_FIFO, ¶m);
103 * The process p may have detached its own ->mm while exiting or through
104 * use_mm(), but one or more of its subthreads may still have a valid
105 * pointer. Return p, or any of its subthreads with a valid ->mm, with
108 static struct task_struct *find_lock_task_mm(struct task_struct *p)
110 struct task_struct *t = p;
117 } while_each_thread(p, t);
122 /* return true if the task is not adequate as candidate victim task. */
123 static bool oom_unkillable_task(struct task_struct *p, struct mem_cgroup *mem,
124 const nodemask_t *nodemask)
126 if (is_global_init(p))
128 if (p->flags & PF_KTHREAD)
131 /* When mem_cgroup_out_of_memory() and p is not member of the group */
132 if (mem && !task_in_mem_cgroup(p, mem))
135 /* p may not have freeable memory in nodemask */
136 if (!has_intersects_mems_allowed(p, nodemask))
143 * badness - calculate a numeric value for how bad this task has been
144 * @p: task struct of which task we should calculate
145 * @uptime: current uptime in seconds
147 * The formula used is relatively simple and documented inline in the
148 * function. The main rationale is that we want to select a good task
149 * to kill when we run out of memory.
151 * Good in this context means that:
152 * 1) we lose the minimum amount of work done
153 * 2) we recover a large amount of memory
154 * 3) we don't kill anything innocent of eating tons of memory
155 * 4) we want to kill the minimum amount of processes (one)
156 * 5) we try to kill the process the user expects us to kill, this
157 * algorithm has been meticulously tuned to meet the principle
158 * of least surprise ... (be careful when you change it)
160 unsigned long badness(struct task_struct *p, struct mem_cgroup *mem,
161 const nodemask_t *nodemask, unsigned long uptime)
163 unsigned long points, cpu_time, run_time;
164 struct task_struct *child;
165 struct task_struct *c, *t;
166 int oom_adj = p->signal->oom_adj;
167 struct task_cputime task_time;
171 if (oom_unkillable_task(p, mem, nodemask))
173 if (oom_adj == OOM_DISABLE)
176 p = find_lock_task_mm(p);
181 * The memory size of the process is the basis for the badness.
183 points = p->mm->total_vm;
187 * swapoff can easily use up all memory, so kill those first.
189 if (p->flags & PF_OOM_ORIGIN)
193 * Processes which fork a lot of child processes are likely
194 * a good choice. We add half the vmsize of the children if they
195 * have an own mm. This prevents forking servers to flood the
196 * machine with an endless amount of children. In case a single
197 * child is eating the vast majority of memory, adding only half
198 * to the parents will make the child our kill candidate of choice.
202 list_for_each_entry(c, &t->children, sibling) {
203 child = find_lock_task_mm(c);
205 if (child->mm != p->mm)
206 points += child->mm->total_vm/2 + 1;
210 } while_each_thread(p, t);
213 * CPU time is in tens of seconds and run time is in thousands
214 * of seconds. There is no particular reason for this other than
215 * that it turned out to work very well in practice.
217 thread_group_cputime(p, &task_time);
218 utime = cputime_to_jiffies(task_time.utime);
219 stime = cputime_to_jiffies(task_time.stime);
220 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
223 if (uptime >= p->start_time.tv_sec)
224 run_time = (uptime - p->start_time.tv_sec) >> 10;
229 points /= int_sqrt(cpu_time);
231 points /= int_sqrt(int_sqrt(run_time));
234 * Niced processes are most likely less important, so double
235 * their badness points.
237 if (task_nice(p) > 0)
241 * Superuser processes are usually more important, so we make it
242 * less likely that we kill those.
244 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
245 has_capability_noaudit(p, CAP_SYS_RESOURCE))
249 * We don't want to kill a process with direct hardware access.
250 * Not only could that mess up the hardware, but usually users
251 * tend to only have this flag set on applications they think
254 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
258 * Adjust the score by oom_adj.
266 points >>= -(oom_adj);
270 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
271 p->pid, p->comm, points);
277 * Determine the type of allocation constraint.
280 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
281 gfp_t gfp_mask, nodemask_t *nodemask)
285 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
288 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
289 * to kill current.We have to random task kill in this case.
290 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
292 if (gfp_mask & __GFP_THISNODE)
293 return CONSTRAINT_NONE;
296 * The nodemask here is a nodemask passed to alloc_pages(). Now,
297 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
298 * feature. mempolicy is an only user of nodemask here.
299 * check mempolicy's nodemask contains all N_HIGH_MEMORY
301 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask))
302 return CONSTRAINT_MEMORY_POLICY;
304 /* Check this allocation failure is caused by cpuset's wall function */
305 for_each_zone_zonelist_nodemask(zone, z, zonelist,
306 high_zoneidx, nodemask)
307 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
308 return CONSTRAINT_CPUSET;
310 return CONSTRAINT_NONE;
313 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
314 gfp_t gfp_mask, nodemask_t *nodemask)
316 return CONSTRAINT_NONE;
321 * Simple selection loop. We chose the process with the highest
322 * number of 'points'. We expect the caller will lock the tasklist.
324 * (not docbooked, we don't want this one cluttering up the manual)
326 static struct task_struct *select_bad_process(unsigned long *ppoints,
327 struct mem_cgroup *mem, const nodemask_t *nodemask)
329 struct task_struct *p;
330 struct task_struct *chosen = NULL;
331 struct timespec uptime;
334 do_posix_clock_monotonic_gettime(&uptime);
335 for_each_process(p) {
336 unsigned long points;
338 if (oom_unkillable_task(p, mem, nodemask))
342 * This task already has access to memory reserves and is
343 * being killed. Don't allow any other task access to the
346 * Note: this may have a chance of deadlock if it gets
347 * blocked waiting for another task which itself is waiting
348 * for memory. Is there a better alternative?
350 if (test_tsk_thread_flag(p, TIF_MEMDIE))
351 return ERR_PTR(-1UL);
354 * This is in the process of releasing memory so wait for it
355 * to finish before killing some other task by mistake.
357 * However, if p is the current task, we allow the 'kill' to
358 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
359 * which will allow it to gain access to memory reserves in
360 * the process of exiting and releasing its resources.
361 * Otherwise we could get an easy OOM deadlock.
363 if ((p->flags & PF_EXITING) && p->mm) {
365 return ERR_PTR(-1UL);
368 *ppoints = ULONG_MAX;
371 points = badness(p, mem, nodemask, uptime.tv_sec);
372 if (points > *ppoints || !chosen) {
382 * dump_tasks - dump current memory state of all system tasks
383 * @mem: current's memory controller, if constrained
385 * Dumps the current memory state of all system tasks, excluding kernel threads.
386 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
389 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
392 * Call with tasklist_lock read-locked.
394 static void dump_tasks(const struct mem_cgroup *mem)
396 struct task_struct *p;
397 struct task_struct *task;
399 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
401 for_each_process(p) {
402 if (p->flags & PF_KTHREAD)
404 if (mem && !task_in_mem_cgroup(p, mem))
407 task = find_lock_task_mm(p);
410 * This is a kthread or all of p's threads have already
411 * detached their mm's. There's no need to report
412 * them; they can't be oom killed anyway.
417 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3u %3d %s\n",
418 task->pid, __task_cred(task)->uid, task->tgid,
419 task->mm->total_vm, get_mm_rss(task->mm),
420 task_cpu(task), task->signal->oom_adj, task->comm);
425 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
426 struct mem_cgroup *mem)
429 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
431 current->comm, gfp_mask, order, current->signal->oom_adj);
432 cpuset_print_task_mems_allowed(current);
433 task_unlock(current);
435 mem_cgroup_print_oom_info(mem, p);
437 if (sysctl_oom_dump_tasks)
441 #define K(x) ((x) << (PAGE_SHIFT-10))
442 static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
444 p = find_lock_task_mm(p);
449 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
450 task_pid_nr(p), p->comm, K(p->mm->total_vm),
451 K(get_mm_counter(p->mm, MM_ANONPAGES)),
452 K(get_mm_counter(p->mm, MM_FILEPAGES)));
456 set_tsk_thread_flag(p, TIF_MEMDIE);
457 force_sig(SIGKILL, p);
460 * We give our sacrificial lamb high priority and access to
461 * all the memory it needs. That way it should be able to
462 * exit() and clear out its resources quickly...
464 boost_dying_task_prio(p, mem);
470 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
471 unsigned long points, struct mem_cgroup *mem,
472 nodemask_t *nodemask, const char *message)
474 struct task_struct *victim = p;
475 struct task_struct *child;
476 struct task_struct *t = p;
477 unsigned long victim_points = 0;
478 struct timespec uptime;
480 if (printk_ratelimit())
481 dump_header(p, gfp_mask, order, mem);
484 * If the task is already exiting, don't alarm the sysadmin or kill
485 * its children or threads, just set TIF_MEMDIE so it can die quickly
487 if (p->flags & PF_EXITING) {
488 set_tsk_thread_flag(p, TIF_MEMDIE);
489 boost_dying_task_prio(p, mem);
494 pr_err("%s: Kill process %d (%s) score %lu or sacrifice child\n",
495 message, task_pid_nr(p), p->comm, points);
499 * If any of p's children has a different mm and is eligible for kill,
500 * the one with the highest badness() score is sacrificed for its
501 * parent. This attempts to lose the minimal amount of work done while
502 * still freeing memory.
504 do_posix_clock_monotonic_gettime(&uptime);
506 list_for_each_entry(child, &t->children, sibling) {
507 unsigned long child_points;
509 /* badness() returns 0 if the thread is unkillable */
510 child_points = badness(child, mem, nodemask,
512 if (child_points > victim_points) {
514 victim_points = child_points;
517 } while_each_thread(p, t);
519 return oom_kill_task(victim, mem);
523 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
525 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
528 if (likely(!sysctl_panic_on_oom))
530 if (sysctl_panic_on_oom != 2) {
532 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
533 * does not panic for cpuset, mempolicy, or memcg allocation
536 if (constraint != CONSTRAINT_NONE)
539 read_lock(&tasklist_lock);
540 dump_header(NULL, gfp_mask, order, NULL);
541 read_unlock(&tasklist_lock);
542 panic("Out of memory: %s panic_on_oom is enabled\n",
543 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
546 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
547 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
549 unsigned long points = 0;
550 struct task_struct *p;
552 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0);
553 read_lock(&tasklist_lock);
555 p = select_bad_process(&points, mem, NULL);
556 if (!p || PTR_ERR(p) == -1UL)
559 if (oom_kill_process(p, gfp_mask, 0, points, mem, NULL,
560 "Memory cgroup out of memory"))
563 read_unlock(&tasklist_lock);
567 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
569 int register_oom_notifier(struct notifier_block *nb)
571 return blocking_notifier_chain_register(&oom_notify_list, nb);
573 EXPORT_SYMBOL_GPL(register_oom_notifier);
575 int unregister_oom_notifier(struct notifier_block *nb)
577 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
579 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
582 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
583 * if a parallel OOM killing is already taking place that includes a zone in
584 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
586 int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
592 spin_lock(&zone_scan_lock);
593 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
594 if (zone_is_oom_locked(zone)) {
600 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
602 * Lock each zone in the zonelist under zone_scan_lock so a
603 * parallel invocation of try_set_zonelist_oom() doesn't succeed
606 zone_set_flag(zone, ZONE_OOM_LOCKED);
610 spin_unlock(&zone_scan_lock);
615 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
616 * allocation attempts with zonelists containing them may now recall the OOM
617 * killer, if necessary.
619 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
624 spin_lock(&zone_scan_lock);
625 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
626 zone_clear_flag(zone, ZONE_OOM_LOCKED);
628 spin_unlock(&zone_scan_lock);
632 * Try to acquire the oom killer lock for all system zones. Returns zero if a
633 * parallel oom killing is taking place, otherwise locks all zones and returns
636 static int try_set_system_oom(void)
641 spin_lock(&zone_scan_lock);
642 for_each_populated_zone(zone)
643 if (zone_is_oom_locked(zone)) {
647 for_each_populated_zone(zone)
648 zone_set_flag(zone, ZONE_OOM_LOCKED);
650 spin_unlock(&zone_scan_lock);
655 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
656 * attempts or page faults may now recall the oom killer, if necessary.
658 static void clear_system_oom(void)
662 spin_lock(&zone_scan_lock);
663 for_each_populated_zone(zone)
664 zone_clear_flag(zone, ZONE_OOM_LOCKED);
665 spin_unlock(&zone_scan_lock);
669 * out_of_memory - kill the "best" process when we run out of memory
670 * @zonelist: zonelist pointer
671 * @gfp_mask: memory allocation flags
672 * @order: amount of memory being requested as a power of 2
673 * @nodemask: nodemask passed to page allocator
675 * If we run out of memory, we have the choice between either
676 * killing a random task (bad), letting the system crash (worse)
677 * OR try to be smart about which process to kill. Note that we
678 * don't have to be perfect here, we just have to be good.
680 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
681 int order, nodemask_t *nodemask)
683 struct task_struct *p;
684 unsigned long freed = 0;
685 unsigned long points;
686 enum oom_constraint constraint = CONSTRAINT_NONE;
688 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
690 /* Got some memory back in the last second. */
694 * If current has a pending SIGKILL, then automatically select it. The
695 * goal is to allow it to allocate so that it may quickly exit and free
698 if (fatal_signal_pending(current)) {
699 set_thread_flag(TIF_MEMDIE);
700 boost_dying_task_prio(current, NULL);
705 * Check if there were limitations on the allocation (only relevant for
706 * NUMA) that may require different handling.
709 constraint = constrained_alloc(zonelist, gfp_mask, nodemask);
710 check_panic_on_oom(constraint, gfp_mask, order);
712 read_lock(&tasklist_lock);
713 if (sysctl_oom_kill_allocating_task &&
714 !oom_unkillable_task(current, NULL, nodemask) &&
715 (current->signal->oom_adj != OOM_DISABLE)) {
717 * oom_kill_process() needs tasklist_lock held. If it returns
718 * non-zero, current could not be killed so we must fallback to
721 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
723 "Out of memory (oom_kill_allocating_task)"))
728 p = select_bad_process(&points, NULL,
729 constraint == CONSTRAINT_MEMORY_POLICY ? nodemask :
731 if (PTR_ERR(p) == -1UL)
734 /* Found nothing?!?! Either we hang forever, or we panic. */
736 dump_header(NULL, gfp_mask, order, NULL);
737 read_unlock(&tasklist_lock);
738 panic("Out of memory and no killable processes...\n");
741 if (oom_kill_process(p, gfp_mask, order, points, NULL, nodemask,
744 read_unlock(&tasklist_lock);
747 * Give "p" a good chance of killing itself before we
748 * retry to allocate memory unless "p" is current
750 if (!test_thread_flag(TIF_MEMDIE))
751 schedule_timeout_uninterruptible(1);
755 * The pagefault handler calls here because it is out of memory, so kill a
756 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
757 * oom killing is already in progress so do nothing. If a task is found with
758 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
760 void pagefault_out_of_memory(void)
762 if (try_set_system_oom()) {
763 out_of_memory(NULL, 0, 0, NULL);
766 if (!test_thread_flag(TIF_MEMDIE))
767 schedule_timeout_uninterruptible(1);