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);
105 * badness - calculate a numeric value for how bad this task has been
106 * @p: task struct of which task we should calculate
107 * @uptime: current uptime in seconds
109 * The formula used is relatively simple and documented inline in the
110 * function. The main rationale is that we want to select a good task
111 * to kill when we run out of memory.
113 * Good in this context means that:
114 * 1) we lose the minimum amount of work done
115 * 2) we recover a large amount of memory
116 * 3) we don't kill anything innocent of eating tons of memory
117 * 4) we want to kill the minimum amount of processes (one)
118 * 5) we try to kill the process the user expects us to kill, this
119 * algorithm has been meticulously tuned to meet the principle
120 * of least surprise ... (be careful when you change it)
123 unsigned long badness(struct task_struct *p, unsigned long uptime)
125 unsigned long points, cpu_time, run_time;
126 struct task_struct *child;
127 struct task_struct *c, *t;
128 int oom_adj = p->signal->oom_adj;
129 struct task_cputime task_time;
133 if (oom_adj == OOM_DISABLE)
136 p = find_lock_task_mm(p);
141 * The memory size of the process is the basis for the badness.
143 points = p->mm->total_vm;
147 * swapoff can easily use up all memory, so kill those first.
149 if (p->flags & PF_OOM_ORIGIN)
153 * Processes which fork a lot of child processes are likely
154 * a good choice. We add half the vmsize of the children if they
155 * have an own mm. This prevents forking servers to flood the
156 * machine with an endless amount of children. In case a single
157 * child is eating the vast majority of memory, adding only half
158 * to the parents will make the child our kill candidate of choice.
162 list_for_each_entry(c, &t->children, sibling) {
163 child = find_lock_task_mm(c);
165 if (child->mm != p->mm)
166 points += child->mm->total_vm/2 + 1;
170 } while_each_thread(p, t);
173 * CPU time is in tens of seconds and run time is in thousands
174 * of seconds. There is no particular reason for this other than
175 * that it turned out to work very well in practice.
177 thread_group_cputime(p, &task_time);
178 utime = cputime_to_jiffies(task_time.utime);
179 stime = cputime_to_jiffies(task_time.stime);
180 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
183 if (uptime >= p->start_time.tv_sec)
184 run_time = (uptime - p->start_time.tv_sec) >> 10;
189 points /= int_sqrt(cpu_time);
191 points /= int_sqrt(int_sqrt(run_time));
194 * Niced processes are most likely less important, so double
195 * their badness points.
197 if (task_nice(p) > 0)
201 * Superuser processes are usually more important, so we make it
202 * less likely that we kill those.
204 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
205 has_capability_noaudit(p, CAP_SYS_RESOURCE))
209 * We don't want to kill a process with direct hardware access.
210 * Not only could that mess up the hardware, but usually users
211 * tend to only have this flag set on applications they think
214 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
218 * Adjust the score by oom_adj.
226 points >>= -(oom_adj);
230 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
231 p->pid, p->comm, points);
237 * Determine the type of allocation constraint.
240 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
241 gfp_t gfp_mask, nodemask_t *nodemask)
245 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
248 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
249 * to kill current.We have to random task kill in this case.
250 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
252 if (gfp_mask & __GFP_THISNODE)
253 return CONSTRAINT_NONE;
256 * The nodemask here is a nodemask passed to alloc_pages(). Now,
257 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
258 * feature. mempolicy is an only user of nodemask here.
259 * check mempolicy's nodemask contains all N_HIGH_MEMORY
261 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask))
262 return CONSTRAINT_MEMORY_POLICY;
264 /* Check this allocation failure is caused by cpuset's wall function */
265 for_each_zone_zonelist_nodemask(zone, z, zonelist,
266 high_zoneidx, nodemask)
267 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
268 return CONSTRAINT_CPUSET;
270 return CONSTRAINT_NONE;
273 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
274 gfp_t gfp_mask, nodemask_t *nodemask)
276 return CONSTRAINT_NONE;
281 * Simple selection loop. We chose the process with the highest
282 * number of 'points'. We expect the caller will lock the tasklist.
284 * (not docbooked, we don't want this one cluttering up the manual)
286 static struct task_struct *select_bad_process(unsigned long *ppoints,
287 struct mem_cgroup *mem, enum oom_constraint constraint,
288 const nodemask_t *mask)
290 struct task_struct *p;
291 struct task_struct *chosen = NULL;
292 struct timespec uptime;
295 do_posix_clock_monotonic_gettime(&uptime);
296 for_each_process(p) {
297 unsigned long points;
299 /* skip the init task and kthreads */
300 if (is_global_init(p) || (p->flags & PF_KTHREAD))
302 if (mem && !task_in_mem_cgroup(p, mem))
304 if (!has_intersects_mems_allowed(p,
305 constraint == CONSTRAINT_MEMORY_POLICY ? mask :
310 * This task already has access to memory reserves and is
311 * being killed. Don't allow any other task access to the
314 * Note: this may have a chance of deadlock if it gets
315 * blocked waiting for another task which itself is waiting
316 * for memory. Is there a better alternative?
318 if (test_tsk_thread_flag(p, TIF_MEMDIE))
319 return ERR_PTR(-1UL);
322 * This is in the process of releasing memory so wait for it
323 * to finish before killing some other task by mistake.
325 * However, if p is the current task, we allow the 'kill' to
326 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
327 * which will allow it to gain access to memory reserves in
328 * the process of exiting and releasing its resources.
329 * Otherwise we could get an easy OOM deadlock.
331 if ((p->flags & PF_EXITING) && p->mm) {
333 return ERR_PTR(-1UL);
336 *ppoints = ULONG_MAX;
339 if (p->signal->oom_adj == OOM_DISABLE)
342 points = badness(p, uptime.tv_sec);
343 if (points > *ppoints || !chosen) {
353 * dump_tasks - dump current memory state of all system tasks
354 * @mem: current's memory controller, if constrained
356 * Dumps the current memory state of all system tasks, excluding kernel threads.
357 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
360 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
363 * Call with tasklist_lock read-locked.
365 static void dump_tasks(const struct mem_cgroup *mem)
367 struct task_struct *p;
368 struct task_struct *task;
370 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
372 for_each_process(p) {
373 if (p->flags & PF_KTHREAD)
375 if (mem && !task_in_mem_cgroup(p, mem))
378 task = find_lock_task_mm(p);
381 * This is a kthread or all of p's threads have already
382 * detached their mm's. There's no need to report
383 * them; they can't be oom killed anyway.
388 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3u %3d %s\n",
389 task->pid, __task_cred(task)->uid, task->tgid,
390 task->mm->total_vm, get_mm_rss(task->mm),
391 task_cpu(task), task->signal->oom_adj, task->comm);
396 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
397 struct mem_cgroup *mem)
400 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
402 current->comm, gfp_mask, order, current->signal->oom_adj);
403 cpuset_print_task_mems_allowed(current);
404 task_unlock(current);
406 mem_cgroup_print_oom_info(mem, p);
408 if (sysctl_oom_dump_tasks)
412 #define K(x) ((x) << (PAGE_SHIFT-10))
415 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
416 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
419 static void __oom_kill_task(struct task_struct *p, int verbose)
421 if (is_global_init(p)) {
423 printk(KERN_WARNING "tried to kill init!\n");
427 p = find_lock_task_mm(p);
432 printk(KERN_ERR "Killed process %d (%s) "
433 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
434 task_pid_nr(p), p->comm,
436 K(get_mm_counter(p->mm, MM_ANONPAGES)),
437 K(get_mm_counter(p->mm, MM_FILEPAGES)));
441 * We give our sacrificial lamb high priority and access to
442 * all the memory it needs. That way it should be able to
443 * exit() and clear out its resources quickly...
445 p->rt.time_slice = HZ;
446 set_tsk_thread_flag(p, TIF_MEMDIE);
448 force_sig(SIGKILL, p);
451 static int oom_kill_task(struct task_struct *p)
453 /* WARNING: mm may not be dereferenced since we did not obtain its
454 * value from get_task_mm(p). This is OK since all we need to do is
455 * compare mm to q->mm below.
457 * Furthermore, even if mm contains a non-NULL value, p->mm may
458 * change to NULL at any time since we do not hold task_lock(p).
459 * However, this is of no concern to us.
461 if (!p->mm || p->signal->oom_adj == OOM_DISABLE)
464 __oom_kill_task(p, 1);
469 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
470 unsigned long points, struct mem_cgroup *mem,
473 struct task_struct *victim = p;
474 struct task_struct *child;
475 struct task_struct *t = p;
476 unsigned long victim_points = 0;
477 struct timespec uptime;
479 if (printk_ratelimit())
480 dump_header(p, gfp_mask, order, mem);
483 * If the task is already exiting, don't alarm the sysadmin or kill
484 * its children or threads, just set TIF_MEMDIE so it can die quickly
486 if (p->flags & PF_EXITING) {
487 set_tsk_thread_flag(p, TIF_MEMDIE);
492 pr_err("%s: Kill process %d (%s) score %lu or sacrifice child\n",
493 message, task_pid_nr(p), p->comm, points);
497 * If any of p's children has a different mm and is eligible for kill,
498 * the one with the highest badness() score is sacrificed for its
499 * parent. This attempts to lose the minimal amount of work done while
500 * still freeing memory.
502 do_posix_clock_monotonic_gettime(&uptime);
504 list_for_each_entry(child, &t->children, sibling) {
505 unsigned long child_points;
507 if (child->mm == p->mm)
509 if (mem && !task_in_mem_cgroup(child, mem))
512 /* badness() returns 0 if the thread is unkillable */
513 child_points = badness(child, uptime.tv_sec);
514 if (child_points > victim_points) {
516 victim_points = child_points;
519 } while_each_thread(p, t);
521 return oom_kill_task(victim);
524 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
525 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
527 unsigned long points = 0;
528 struct task_struct *p;
530 if (sysctl_panic_on_oom == 2)
531 panic("out of memory(memcg). panic_on_oom is selected.\n");
532 read_lock(&tasklist_lock);
534 p = select_bad_process(&points, mem, CONSTRAINT_NONE, NULL);
535 if (!p || PTR_ERR(p) == -1UL)
538 if (oom_kill_process(p, gfp_mask, 0, points, mem,
539 "Memory cgroup out of memory"))
542 read_unlock(&tasklist_lock);
546 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
548 int register_oom_notifier(struct notifier_block *nb)
550 return blocking_notifier_chain_register(&oom_notify_list, nb);
552 EXPORT_SYMBOL_GPL(register_oom_notifier);
554 int unregister_oom_notifier(struct notifier_block *nb)
556 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
558 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
561 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
562 * if a parallel OOM killing is already taking place that includes a zone in
563 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
565 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
571 spin_lock(&zone_scan_lock);
572 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
573 if (zone_is_oom_locked(zone)) {
579 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
581 * Lock each zone in the zonelist under zone_scan_lock so a
582 * parallel invocation of try_set_zone_oom() doesn't succeed
585 zone_set_flag(zone, ZONE_OOM_LOCKED);
589 spin_unlock(&zone_scan_lock);
594 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
595 * allocation attempts with zonelists containing them may now recall the OOM
596 * killer, if necessary.
598 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
603 spin_lock(&zone_scan_lock);
604 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
605 zone_clear_flag(zone, ZONE_OOM_LOCKED);
607 spin_unlock(&zone_scan_lock);
611 * Must be called with tasklist_lock held for read.
613 static void __out_of_memory(gfp_t gfp_mask, int order,
614 enum oom_constraint constraint, const nodemask_t *mask)
616 struct task_struct *p;
617 unsigned long points;
619 if (sysctl_oom_kill_allocating_task)
620 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
621 "Out of memory (oom_kill_allocating_task)"))
625 * Rambo mode: Shoot down a process and hope it solves whatever
626 * issues we may have.
628 p = select_bad_process(&points, NULL, constraint, mask);
630 if (PTR_ERR(p) == -1UL)
633 /* Found nothing?!?! Either we hang forever, or we panic. */
635 read_unlock(&tasklist_lock);
636 dump_header(NULL, gfp_mask, order, NULL);
637 panic("Out of memory and no killable processes...\n");
640 if (oom_kill_process(p, gfp_mask, order, points, NULL,
646 * pagefault handler calls into here because it is out of memory but
647 * doesn't know exactly how or why.
649 void pagefault_out_of_memory(void)
651 unsigned long freed = 0;
653 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
655 /* Got some memory back in the last second. */
658 if (sysctl_panic_on_oom)
659 panic("out of memory from page fault. panic_on_oom is selected.\n");
661 read_lock(&tasklist_lock);
662 /* unknown gfp_mask and order */
663 __out_of_memory(0, 0, CONSTRAINT_NONE, NULL);
664 read_unlock(&tasklist_lock);
667 * Give "p" a good chance of killing itself before we
668 * retry to allocate memory.
670 if (!test_thread_flag(TIF_MEMDIE))
671 schedule_timeout_uninterruptible(1);
675 * out_of_memory - kill the "best" process when we run out of memory
676 * @zonelist: zonelist pointer
677 * @gfp_mask: memory allocation flags
678 * @order: amount of memory being requested as a power of 2
679 * @nodemask: nodemask passed to page allocator
681 * If we run out of memory, we have the choice between either
682 * killing a random task (bad), letting the system crash (worse)
683 * OR try to be smart about which process to kill. Note that we
684 * don't have to be perfect here, we just have to be good.
686 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
687 int order, nodemask_t *nodemask)
689 unsigned long freed = 0;
690 enum oom_constraint constraint;
692 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
694 /* Got some memory back in the last second. */
698 * If current has a pending SIGKILL, then automatically select it. The
699 * goal is to allow it to allocate so that it may quickly exit and free
702 if (fatal_signal_pending(current)) {
703 set_thread_flag(TIF_MEMDIE);
707 if (sysctl_panic_on_oom == 2) {
708 dump_header(NULL, gfp_mask, order, NULL);
709 panic("out of memory. Compulsory panic_on_oom is selected.\n");
713 * Check if there were limitations on the allocation (only relevant for
714 * NUMA) that may require different handling.
716 constraint = constrained_alloc(zonelist, gfp_mask, nodemask);
717 read_lock(&tasklist_lock);
718 if (unlikely(sysctl_panic_on_oom)) {
720 * panic_on_oom only affects CONSTRAINT_NONE, the kernel
721 * should not panic for cpuset or mempolicy induced memory
724 if (constraint == CONSTRAINT_NONE) {
725 dump_header(NULL, gfp_mask, order, NULL);
726 read_unlock(&tasklist_lock);
727 panic("Out of memory: panic_on_oom is enabled\n");
730 __out_of_memory(gfp_mask, order, constraint, nodemask);
731 read_unlock(&tasklist_lock);
734 * Give "p" a good chance of killing itself before we
735 * retry to allocate memory unless "p" is current
737 if (!test_thread_flag(TIF_MEMDIE))
738 schedule_timeout_uninterruptible(1);