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/security.h>
32 int sysctl_panic_on_oom;
33 int sysctl_oom_kill_allocating_task;
34 int sysctl_oom_dump_tasks;
35 static DEFINE_SPINLOCK(zone_scan_lock);
39 * Is all threads of the target process nodes overlap ours?
41 static int has_intersects_mems_allowed(struct task_struct *tsk)
43 struct task_struct *t;
47 if (cpuset_mems_allowed_intersects(current, t))
56 * badness - calculate a numeric value for how bad this task has been
57 * @p: task struct of which task we should calculate
58 * @uptime: current uptime in seconds
60 * The formula used is relatively simple and documented inline in the
61 * function. The main rationale is that we want to select a good task
62 * to kill when we run out of memory.
64 * Good in this context means that:
65 * 1) we lose the minimum amount of work done
66 * 2) we recover a large amount of memory
67 * 3) we don't kill anything innocent of eating tons of memory
68 * 4) we want to kill the minimum amount of processes (one)
69 * 5) we try to kill the process the user expects us to kill, this
70 * algorithm has been meticulously tuned to meet the principle
71 * of least surprise ... (be careful when you change it)
74 unsigned long badness(struct task_struct *p, unsigned long uptime)
76 unsigned long points, cpu_time, run_time;
78 struct task_struct *child;
79 int oom_adj = p->signal->oom_adj;
80 struct task_cputime task_time;
84 if (oom_adj == OOM_DISABLE)
95 * The memory size of the process is the basis for the badness.
97 points = mm->total_vm;
100 * After this unlock we can no longer dereference local variable `mm'
105 * swapoff can easily use up all memory, so kill those first.
107 if (p->flags & PF_OOM_ORIGIN)
111 * Processes which fork a lot of child processes are likely
112 * a good choice. We add half the vmsize of the children if they
113 * have an own mm. This prevents forking servers to flood the
114 * machine with an endless amount of children. In case a single
115 * child is eating the vast majority of memory, adding only half
116 * to the parents will make the child our kill candidate of choice.
118 list_for_each_entry(child, &p->children, sibling) {
120 if (child->mm != mm && child->mm)
121 points += child->mm->total_vm/2 + 1;
126 * CPU time is in tens of seconds and run time is in thousands
127 * of seconds. There is no particular reason for this other than
128 * that it turned out to work very well in practice.
130 thread_group_cputime(p, &task_time);
131 utime = cputime_to_jiffies(task_time.utime);
132 stime = cputime_to_jiffies(task_time.stime);
133 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
136 if (uptime >= p->start_time.tv_sec)
137 run_time = (uptime - p->start_time.tv_sec) >> 10;
142 points /= int_sqrt(cpu_time);
144 points /= int_sqrt(int_sqrt(run_time));
147 * Niced processes are most likely less important, so double
148 * their badness points.
150 if (task_nice(p) > 0)
154 * Superuser processes are usually more important, so we make it
155 * less likely that we kill those.
157 if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
158 has_capability_noaudit(p, CAP_SYS_RESOURCE))
162 * We don't want to kill a process with direct hardware access.
163 * Not only could that mess up the hardware, but usually users
164 * tend to only have this flag set on applications they think
167 if (has_capability_noaudit(p, CAP_SYS_RAWIO))
171 * If p's nodes don't overlap ours, it may still help to kill p
172 * because p may have allocated or otherwise mapped memory on
173 * this node before. However it will be less likely.
175 if (!has_intersects_mems_allowed(p))
179 * Adjust the score by oom_adj.
187 points >>= -(oom_adj);
191 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
192 p->pid, p->comm, points);
198 * Determine the type of allocation constraint.
201 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
202 gfp_t gfp_mask, nodemask_t *nodemask)
206 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
209 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
210 * to kill current.We have to random task kill in this case.
211 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
213 if (gfp_mask & __GFP_THISNODE)
214 return CONSTRAINT_NONE;
217 * The nodemask here is a nodemask passed to alloc_pages(). Now,
218 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy
219 * feature. mempolicy is an only user of nodemask here.
220 * check mempolicy's nodemask contains all N_HIGH_MEMORY
222 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask))
223 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 return CONSTRAINT_CPUSET;
231 return CONSTRAINT_NONE;
234 static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
235 gfp_t gfp_mask, nodemask_t *nodemask)
237 return CONSTRAINT_NONE;
242 * Simple selection loop. We chose the process with the highest
243 * number of 'points'. We expect the caller will lock the tasklist.
245 * (not docbooked, we don't want this one cluttering up the manual)
247 static struct task_struct *select_bad_process(unsigned long *ppoints,
248 struct mem_cgroup *mem)
250 struct task_struct *p;
251 struct task_struct *chosen = NULL;
252 struct timespec uptime;
255 do_posix_clock_monotonic_gettime(&uptime);
256 for_each_process(p) {
257 unsigned long points;
260 * skip kernel threads and tasks which have already released
265 /* skip the init task */
266 if (is_global_init(p))
268 if (mem && !task_in_mem_cgroup(p, mem))
272 * This task already has access to memory reserves and is
273 * being killed. Don't allow any other task access to the
276 * Note: this may have a chance of deadlock if it gets
277 * blocked waiting for another task which itself is waiting
278 * for memory. Is there a better alternative?
280 if (test_tsk_thread_flag(p, TIF_MEMDIE))
281 return ERR_PTR(-1UL);
284 * This is in the process of releasing memory so wait for it
285 * to finish before killing some other task by mistake.
287 * However, if p is the current task, we allow the 'kill' to
288 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
289 * which will allow it to gain access to memory reserves in
290 * the process of exiting and releasing its resources.
291 * Otherwise we could get an easy OOM deadlock.
293 if (p->flags & PF_EXITING) {
295 return ERR_PTR(-1UL);
298 *ppoints = ULONG_MAX;
301 if (p->signal->oom_adj == OOM_DISABLE)
304 points = badness(p, uptime.tv_sec);
305 if (points > *ppoints || !chosen) {
315 * dump_tasks - dump current memory state of all system tasks
316 * @mem: target memory controller
318 * Dumps the current memory state of all system tasks, excluding kernel threads.
319 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
322 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
325 * Call with tasklist_lock read-locked.
327 static void dump_tasks(const struct mem_cgroup *mem)
329 struct task_struct *g, *p;
331 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
333 do_each_thread(g, p) {
334 struct mm_struct *mm;
336 if (mem && !task_in_mem_cgroup(p, mem))
338 if (!thread_group_leader(p))
345 * total_vm and rss sizes do not exist for tasks with no
346 * mm so there's no need to report them; they can't be
352 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
353 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
354 get_mm_rss(mm), (int)task_cpu(p), p->signal->oom_adj,
357 } while_each_thread(g, p);
360 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
361 struct mem_cgroup *mem)
363 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
365 current->comm, gfp_mask, order, current->signal->oom_adj);
367 cpuset_print_task_mems_allowed(current);
368 task_unlock(current);
370 mem_cgroup_print_oom_info(mem, p);
372 if (sysctl_oom_dump_tasks)
376 #define K(x) ((x) << (PAGE_SHIFT-10))
379 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
380 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
383 static void __oom_kill_task(struct task_struct *p, int verbose)
385 if (is_global_init(p)) {
387 printk(KERN_WARNING "tried to kill init!\n");
394 printk(KERN_WARNING "tried to kill an mm-less task %d (%s)!\n",
395 task_pid_nr(p), p->comm);
401 printk(KERN_ERR "Killed process %d (%s) "
402 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
403 task_pid_nr(p), p->comm,
405 K(get_mm_counter(p->mm, MM_ANONPAGES)),
406 K(get_mm_counter(p->mm, MM_FILEPAGES)));
410 * We give our sacrificial lamb high priority and access to
411 * all the memory it needs. That way it should be able to
412 * exit() and clear out its resources quickly...
414 p->rt.time_slice = HZ;
415 set_tsk_thread_flag(p, TIF_MEMDIE);
417 force_sig(SIGKILL, p);
420 static int oom_kill_task(struct task_struct *p)
422 /* WARNING: mm may not be dereferenced since we did not obtain its
423 * value from get_task_mm(p). This is OK since all we need to do is
424 * compare mm to q->mm below.
426 * Furthermore, even if mm contains a non-NULL value, p->mm may
427 * change to NULL at any time since we do not hold task_lock(p).
428 * However, this is of no concern to us.
430 if (!p->mm || p->signal->oom_adj == OOM_DISABLE)
433 __oom_kill_task(p, 1);
438 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
439 unsigned long points, struct mem_cgroup *mem,
442 struct task_struct *c;
444 if (printk_ratelimit())
445 dump_header(p, gfp_mask, order, mem);
448 * If the task is already exiting, don't alarm the sysadmin or kill
449 * its children or threads, just set TIF_MEMDIE so it can die quickly
451 if (p->flags & PF_EXITING) {
452 __oom_kill_task(p, 0);
456 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
457 message, task_pid_nr(p), p->comm, points);
459 /* Try to kill a child first */
460 list_for_each_entry(c, &p->children, sibling) {
463 if (mem && !task_in_mem_cgroup(c, mem))
465 if (!oom_kill_task(c))
468 return oom_kill_task(p);
471 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
472 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
474 unsigned long points = 0;
475 struct task_struct *p;
477 if (sysctl_panic_on_oom == 2)
478 panic("out of memory(memcg). panic_on_oom is selected.\n");
479 read_lock(&tasklist_lock);
481 p = select_bad_process(&points, mem);
482 if (PTR_ERR(p) == -1UL)
488 if (oom_kill_process(p, gfp_mask, 0, points, mem,
489 "Memory cgroup out of memory"))
492 read_unlock(&tasklist_lock);
496 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
498 int register_oom_notifier(struct notifier_block *nb)
500 return blocking_notifier_chain_register(&oom_notify_list, nb);
502 EXPORT_SYMBOL_GPL(register_oom_notifier);
504 int unregister_oom_notifier(struct notifier_block *nb)
506 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
508 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
511 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
512 * if a parallel OOM killing is already taking place that includes a zone in
513 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
515 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
521 spin_lock(&zone_scan_lock);
522 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
523 if (zone_is_oom_locked(zone)) {
529 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
531 * Lock each zone in the zonelist under zone_scan_lock so a
532 * parallel invocation of try_set_zone_oom() doesn't succeed
535 zone_set_flag(zone, ZONE_OOM_LOCKED);
539 spin_unlock(&zone_scan_lock);
544 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
545 * allocation attempts with zonelists containing them may now recall the OOM
546 * killer, if necessary.
548 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
553 spin_lock(&zone_scan_lock);
554 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
555 zone_clear_flag(zone, ZONE_OOM_LOCKED);
557 spin_unlock(&zone_scan_lock);
561 * Must be called with tasklist_lock held for read.
563 static void __out_of_memory(gfp_t gfp_mask, int order)
565 struct task_struct *p;
566 unsigned long points;
568 if (sysctl_oom_kill_allocating_task)
569 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
570 "Out of memory (oom_kill_allocating_task)"))
574 * Rambo mode: Shoot down a process and hope it solves whatever
575 * issues we may have.
577 p = select_bad_process(&points, NULL);
579 if (PTR_ERR(p) == -1UL)
582 /* Found nothing?!?! Either we hang forever, or we panic. */
584 read_unlock(&tasklist_lock);
585 dump_header(NULL, gfp_mask, order, NULL);
586 panic("Out of memory and no killable processes...\n");
589 if (oom_kill_process(p, gfp_mask, order, points, NULL,
595 * pagefault handler calls into here because it is out of memory but
596 * doesn't know exactly how or why.
598 void pagefault_out_of_memory(void)
600 unsigned long freed = 0;
602 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
604 /* Got some memory back in the last second. */
607 if (sysctl_panic_on_oom)
608 panic("out of memory from page fault. panic_on_oom is selected.\n");
610 read_lock(&tasklist_lock);
611 __out_of_memory(0, 0); /* unknown gfp_mask and order */
612 read_unlock(&tasklist_lock);
615 * Give "p" a good chance of killing itself before we
616 * retry to allocate memory.
618 if (!test_thread_flag(TIF_MEMDIE))
619 schedule_timeout_uninterruptible(1);
623 * out_of_memory - kill the "best" process when we run out of memory
624 * @zonelist: zonelist pointer
625 * @gfp_mask: memory allocation flags
626 * @order: amount of memory being requested as a power of 2
628 * If we run out of memory, we have the choice between either
629 * killing a random task (bad), letting the system crash (worse)
630 * OR try to be smart about which process to kill. Note that we
631 * don't have to be perfect here, we just have to be good.
633 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
634 int order, nodemask_t *nodemask)
636 unsigned long freed = 0;
637 enum oom_constraint constraint;
639 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
641 /* Got some memory back in the last second. */
644 if (sysctl_panic_on_oom == 2) {
645 dump_header(NULL, gfp_mask, order, NULL);
646 panic("out of memory. Compulsory panic_on_oom is selected.\n");
650 * Check if there were limitations on the allocation (only relevant for
651 * NUMA) that may require different handling.
653 constraint = constrained_alloc(zonelist, gfp_mask, nodemask);
654 read_lock(&tasklist_lock);
656 switch (constraint) {
657 case CONSTRAINT_MEMORY_POLICY:
658 oom_kill_process(current, gfp_mask, order, 0, NULL,
659 "No available memory (MPOL_BIND)");
662 case CONSTRAINT_NONE:
663 if (sysctl_panic_on_oom) {
664 dump_header(NULL, gfp_mask, order, NULL);
665 panic("out of memory. panic_on_oom is selected\n");
668 case CONSTRAINT_CPUSET:
669 __out_of_memory(gfp_mask, order);
673 read_unlock(&tasklist_lock);
676 * Give "p" a good chance of killing itself before we
677 * retry to allocate memory unless "p" is current
679 if (!test_thread_flag(TIF_MEMDIE))
680 schedule_timeout_uninterruptible(1);