blob: f52e85c80e8d554fcae1a7ad40e0617c0bb1318f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/oom_kill.c
3 *
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...
David Rientjesa63d83f2010-08-09 17:19:46 -07007 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 *
10 * The routines in this file are used to kill a process when
Paul Jacksona49335c2005-09-06 15:18:09 -070011 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 *
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.
18 */
19
Alexey Dobriyan8ac773b2006-10-19 23:28:32 -070020#include <linux/oom.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/mm.h>
Alexey Dobriyan4e950f62007-07-30 02:36:13 +040022#include <linux/err.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090023#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/sched.h>
25#include <linux/swap.h>
26#include <linux/timex.h>
27#include <linux/jiffies.h>
Paul Jacksonef08e3b2005-09-06 15:18:13 -070028#include <linux/cpuset.h>
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -070029#include <linux/module.h>
30#include <linux/notifier.h>
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -080031#include <linux/memcontrol.h>
David Rientjes6f48d0eb2010-08-09 17:18:52 -070032#include <linux/mempolicy.h>
David Howells5cd9c582008-08-14 11:37:28 +010033#include <linux/security.h>
David Rientjesedd45542011-03-22 16:30:12 -070034#include <linux/ptrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -070036int sysctl_panic_on_oom;
David Rientjesfe071d72007-10-16 23:25:56 -070037int sysctl_oom_kill_allocating_task;
David Rientjesad915c42010-08-09 17:18:53 -070038int sysctl_oom_dump_tasks = 1;
David Rientjesc7d4cae2009-01-06 14:39:00 -080039static DEFINE_SPINLOCK(zone_scan_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
David Rientjes6f48d0eb2010-08-09 17:18:52 -070041#ifdef CONFIG_NUMA
42/**
43 * has_intersects_mems_allowed() - check task eligiblity for kill
44 * @tsk: task struct of which task to consider
45 * @mask: nodemask passed to page allocator for mempolicy ooms
46 *
47 * Task eligibility is determined by whether or not a candidate task, @tsk,
48 * shares the same mempolicy nodes as current if it is bound by such a policy
49 * and whether or not it has the same set of allowed cpuset nodes.
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070050 */
David Rientjes6f48d0eb2010-08-09 17:18:52 -070051static bool has_intersects_mems_allowed(struct task_struct *tsk,
52 const nodemask_t *mask)
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070053{
David Rientjes6f48d0eb2010-08-09 17:18:52 -070054 struct task_struct *start = tsk;
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070055
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070056 do {
David Rientjes6f48d0eb2010-08-09 17:18:52 -070057 if (mask) {
58 /*
59 * If this is a mempolicy constrained oom, tsk's
60 * cpuset is irrelevant. Only return true if its
61 * mempolicy intersects current, otherwise it may be
62 * needlessly killed.
63 */
64 if (mempolicy_nodemask_intersects(tsk, mask))
65 return true;
66 } else {
67 /*
68 * This is not a mempolicy constrained oom, so only
69 * check the mems of tsk's cpuset.
70 */
71 if (cpuset_mems_allowed_intersects(current, tsk))
72 return true;
73 }
KOSAKI Motohirodf1090a2010-08-09 17:19:39 -070074 } while_each_thread(start, tsk);
75
David Rientjes6f48d0eb2010-08-09 17:18:52 -070076 return false;
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070077}
David Rientjes6f48d0eb2010-08-09 17:18:52 -070078#else
79static bool has_intersects_mems_allowed(struct task_struct *tsk,
80 const nodemask_t *mask)
81{
82 return true;
83}
84#endif /* CONFIG_NUMA */
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070085
David Rientjes6f48d0eb2010-08-09 17:18:52 -070086/*
87 * The process p may have detached its own ->mm while exiting or through
88 * use_mm(), but one or more of its subthreads may still have a valid
89 * pointer. Return p, or any of its subthreads with a valid ->mm, with
90 * task_lock() held.
91 */
KAMEZAWA Hiroyuki158e0a22010-08-10 18:03:00 -070092struct task_struct *find_lock_task_mm(struct task_struct *p)
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -070093{
94 struct task_struct *t = p;
95
96 do {
97 task_lock(t);
98 if (likely(t->mm))
99 return t;
100 task_unlock(t);
101 } while_each_thread(p, t);
102
103 return NULL;
104}
105
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700106/* return true if the task is not adequate as candidate victim task. */
David Rientjese85bfd32010-09-22 13:05:10 -0700107static bool oom_unkillable_task(struct task_struct *p,
108 const struct mem_cgroup *mem, const nodemask_t *nodemask)
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700109{
110 if (is_global_init(p))
111 return true;
112 if (p->flags & PF_KTHREAD)
113 return true;
114
115 /* When mem_cgroup_out_of_memory() and p is not member of the group */
116 if (mem && !task_in_mem_cgroup(p, mem))
117 return true;
118
119 /* p may not have freeable memory in nodemask */
120 if (!has_intersects_mems_allowed(p, nodemask))
121 return true;
122
123 return false;
124}
125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126/**
David Rientjesa63d83f2010-08-09 17:19:46 -0700127 * oom_badness - heuristic function to determine which candidate task to kill
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 * @p: task struct of which task we should calculate
David Rientjesa63d83f2010-08-09 17:19:46 -0700129 * @totalpages: total present RAM allowed for page allocation
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 *
David Rientjesa63d83f2010-08-09 17:19:46 -0700131 * The heuristic for determining which task to kill is made to be as simple and
132 * predictable as possible. The goal is to return the highest value for the
133 * task consuming the most memory to avoid subsequent oom failures.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700135unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
136 const nodemask_t *nodemask, unsigned long totalpages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137{
David Rientjesa63d83f2010-08-09 17:19:46 -0700138 int points;
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700139
KOSAKI Motohiro26ebc982010-08-09 17:19:37 -0700140 if (oom_unkillable_task(p, mem, nodemask))
141 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700143 p = find_lock_task_mm(p);
144 if (!p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 return 0;
146
147 /*
David Rientjese18641e2010-10-26 14:21:23 -0700148 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
149 * so the entire heuristic doesn't need to be executed for something
150 * that cannot be killed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 */
David Rientjese18641e2010-10-26 14:21:23 -0700152 if (atomic_read(&p->mm->oom_disable_count)) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700153 task_unlock(p);
154 return 0;
155 }
156
157 /*
158 * When the PF_OOM_ORIGIN bit is set, it indicates the task should have
159 * priority for oom killing.
160 */
161 if (p->flags & PF_OOM_ORIGIN) {
162 task_unlock(p);
163 return 1000;
164 }
165
166 /*
167 * The memory controller may have a limit of 0 bytes, so avoid a divide
168 * by zero, if necessary.
169 */
170 if (!totalpages)
171 totalpages = 1;
172
173 /*
174 * The baseline for the badness score is the proportion of RAM that each
KOSAKI Motohirof755a042011-04-27 15:26:50 -0700175 * task's rss, pagetable and swap space use.
David Rientjesa63d83f2010-08-09 17:19:46 -0700176 */
KOSAKI Motohirof755a042011-04-27 15:26:50 -0700177 points = get_mm_rss(p->mm) + p->mm->nr_ptes;
178 points += get_mm_counter(p->mm, MM_SWAPENTS);
179
180 points *= 1000;
181 points /= totalpages;
Andrew Morton97c2c9b82006-04-18 22:20:38 -0700182 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183
184 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700185 * Root processes get 3% bonus, just like the __vm_enough_memory()
186 * implementation used by LSMs.
Hugh Dickins7ba34852007-01-05 16:37:03 -0800187 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700188 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
189 points -= 30;
Hugh Dickins7ba34852007-01-05 16:37:03 -0800190
191 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700192 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
193 * either completely disable oom killing or always prefer a certain
194 * task.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700196 points += p->signal->oom_score_adj;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197
David Rientjesf19e8aa2010-09-22 13:04:52 -0700198 /*
199 * Never return 0 for an eligible task that may be killed since it's
200 * possible that no single user task uses more than 0.1% of memory and
201 * no single admin tasks uses more than 3.0%.
202 */
203 if (points <= 0)
204 return 1;
David Rientjesa63d83f2010-08-09 17:19:46 -0700205 return (points < 1000) ? points : 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206}
207
208/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800209 * Determine the type of allocation constraint.
210 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800211#ifdef CONFIG_NUMA
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800212static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
David Rientjesa63d83f2010-08-09 17:19:46 -0700213 gfp_t gfp_mask, nodemask_t *nodemask,
214 unsigned long *totalpages)
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800215{
Mel Gorman54a6eb52008-04-28 02:12:16 -0700216 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -0700217 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -0700218 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
David Rientjesa63d83f2010-08-09 17:19:46 -0700219 bool cpuset_limited = false;
220 int nid;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800221
David Rientjesa63d83f2010-08-09 17:19:46 -0700222 /* Default to all available memory */
223 *totalpages = totalram_pages + total_swap_pages;
224
225 if (!zonelist)
226 return CONSTRAINT_NONE;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800227 /*
228 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
229 * to kill current.We have to random task kill in this case.
230 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
231 */
232 if (gfp_mask & __GFP_THISNODE)
233 return CONSTRAINT_NONE;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800234
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800235 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700236 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
237 * the page allocator means a mempolicy is in effect. Cpuset policy
238 * is enforced in get_page_from_freelist().
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800239 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700240 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
241 *totalpages = total_swap_pages;
242 for_each_node_mask(nid, *nodemask)
243 *totalpages += node_spanned_pages(nid);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800244 return CONSTRAINT_MEMORY_POLICY;
David Rientjesa63d83f2010-08-09 17:19:46 -0700245 }
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800246
247 /* Check this allocation failure is caused by cpuset's wall function */
248 for_each_zone_zonelist_nodemask(zone, z, zonelist,
249 high_zoneidx, nodemask)
250 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
David Rientjesa63d83f2010-08-09 17:19:46 -0700251 cpuset_limited = true;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800252
David Rientjesa63d83f2010-08-09 17:19:46 -0700253 if (cpuset_limited) {
254 *totalpages = total_swap_pages;
255 for_each_node_mask(nid, cpuset_current_mems_allowed)
256 *totalpages += node_spanned_pages(nid);
257 return CONSTRAINT_CPUSET;
258 }
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800259 return CONSTRAINT_NONE;
260}
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800261#else
262static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
David Rientjesa63d83f2010-08-09 17:19:46 -0700263 gfp_t gfp_mask, nodemask_t *nodemask,
264 unsigned long *totalpages)
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800265{
David Rientjesa63d83f2010-08-09 17:19:46 -0700266 *totalpages = totalram_pages + total_swap_pages;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800267 return CONSTRAINT_NONE;
268}
269#endif
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800270
271/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 * Simple selection loop. We chose the process with the highest
273 * number of 'points'. We expect the caller will lock the tasklist.
274 *
275 * (not docbooked, we don't want this one cluttering up the manual)
276 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700277static struct task_struct *select_bad_process(unsigned int *ppoints,
278 unsigned long totalpages, struct mem_cgroup *mem,
279 const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280{
David Rientjes3a5dda72011-03-22 16:30:09 -0700281 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 struct task_struct *chosen = NULL;
Kurt Garloff9827b782006-02-20 18:27:51 -0800283 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284
David Rientjes3a5dda72011-03-22 16:30:09 -0700285 do_each_thread(g, p) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700286 unsigned int points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700287
Andrey Vagin30e2b412011-03-22 16:30:11 -0700288 if (!p->mm)
289 continue;
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700290 if (oom_unkillable_task(p, mem, nodemask))
David Rientjes6cf86ac2010-08-09 17:18:50 -0700291 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700292
Paul Jacksona49335c2005-09-06 15:18:09 -0700293 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700294 * This task already has access to memory reserves and is
295 * being killed. Don't allow any other task access to the
296 * memory reserve.
297 *
298 * Note: this may have a chance of deadlock if it gets
299 * blocked waiting for another task which itself is waiting
300 * for memory. Is there a better alternative?
301 */
302 if (test_tsk_thread_flag(p, TIF_MEMDIE))
303 return ERR_PTR(-1UL);
304
Andrey Vagin30e2b412011-03-22 16:30:11 -0700305 if (p->flags & PF_EXITING) {
David Rientjesedd45542011-03-22 16:30:12 -0700306 /*
307 * If p is the current task and is in the process of
308 * releasing memory, we allow the "kill" to set
309 * TIF_MEMDIE, which will allow it to gain access to
310 * memory reserves. Otherwise, it may stall forever.
311 *
312 * The loop isn't broken here, however, in case other
313 * threads are found to have already been oom killed.
314 */
315 if (p == current) {
316 chosen = p;
317 *ppoints = 1000;
318 } else {
319 /*
320 * If this task is not being ptraced on exit,
321 * then wait for it to finish before killing
322 * some other task unnecessarily.
323 */
324 if (!(task_ptrace(p->group_leader) &
325 PT_TRACE_EXIT))
326 return ERR_PTR(-1UL);
327 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700328 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700329
David Rientjesa63d83f2010-08-09 17:19:46 -0700330 points = oom_badness(p, mem, nodemask, totalpages);
331 if (points > *ppoints) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700332 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800333 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 }
David Rientjes3a5dda72011-03-22 16:30:09 -0700335 } while_each_thread(g, p);
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700336
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 return chosen;
338}
339
340/**
Randy Dunlap1b578df2008-03-19 17:00:42 -0700341 * dump_tasks - dump current memory state of all system tasks
David Rientjes74ab7f12010-08-09 17:18:46 -0700342 * @mem: current's memory controller, if constrained
David Rientjese85bfd32010-09-22 13:05:10 -0700343 * @nodemask: nodemask passed to page allocator for mempolicy ooms
Randy Dunlap1b578df2008-03-19 17:00:42 -0700344 *
David Rientjese85bfd32010-09-22 13:05:10 -0700345 * Dumps the current memory state of all eligible tasks. Tasks not in the same
346 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
347 * are not shown.
David Rientjesfef1bdd2008-02-07 00:14:07 -0800348 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
David Rientjesa63d83f2010-08-09 17:19:46 -0700349 * value, oom_score_adj value, and name.
David Rientjesfef1bdd2008-02-07 00:14:07 -0800350 *
David Rientjesfef1bdd2008-02-07 00:14:07 -0800351 * Call with tasklist_lock read-locked.
352 */
David Rientjese85bfd32010-09-22 13:05:10 -0700353static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
David Rientjesfef1bdd2008-02-07 00:14:07 -0800354{
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700355 struct task_struct *p;
356 struct task_struct *task;
David Rientjesfef1bdd2008-02-07 00:14:07 -0800357
David Rientjesa63d83f2010-08-09 17:19:46 -0700358 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700359 for_each_process(p) {
David Rientjese85bfd32010-09-22 13:05:10 -0700360 if (oom_unkillable_task(p, mem, nodemask))
David Rientjesfef1bdd2008-02-07 00:14:07 -0800361 continue;
362
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700363 task = find_lock_task_mm(p);
364 if (!task) {
David Rientjes6d2661e2009-05-28 14:34:19 -0700365 /*
David Rientjes74ab7f12010-08-09 17:18:46 -0700366 * This is a kthread or all of p's threads have already
367 * detached their mm's. There's no need to report
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700368 * them; they can't be oom killed anyway.
David Rientjes6d2661e2009-05-28 14:34:19 -0700369 */
David Rientjes6d2661e2009-05-28 14:34:19 -0700370 continue;
371 }
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700372
David Rientjesa63d83f2010-08-09 17:19:46 -0700373 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
KOSAKI Motohiro8d6c83f2010-08-19 14:13:39 -0700374 task->pid, task_uid(task), task->tgid,
David Rientjesa63d83f2010-08-09 17:19:46 -0700375 task->mm->total_vm, get_mm_rss(task->mm),
376 task_cpu(task), task->signal->oom_adj,
377 task->signal->oom_score_adj, task->comm);
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700378 task_unlock(task);
379 }
David Rientjesfef1bdd2008-02-07 00:14:07 -0800380}
381
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800382static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjese85bfd32010-09-22 13:05:10 -0700383 struct mem_cgroup *mem, const nodemask_t *nodemask)
David Rientjes1b604d72009-12-14 17:57:47 -0800384{
David Rientjes5e9d8342010-08-09 17:18:51 -0700385 task_lock(current);
David Rientjes1b604d72009-12-14 17:57:47 -0800386 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
David Rientjesa63d83f2010-08-09 17:19:46 -0700387 "oom_adj=%d, oom_score_adj=%d\n",
388 current->comm, gfp_mask, order, current->signal->oom_adj,
389 current->signal->oom_score_adj);
David Rientjes1b604d72009-12-14 17:57:47 -0800390 cpuset_print_task_mems_allowed(current);
391 task_unlock(current);
392 dump_stack();
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800393 mem_cgroup_print_oom_info(mem, p);
David Rientjesb2b755b2011-03-24 15:18:15 -0700394 show_mem(SHOW_MEM_FILTER_NODES);
David Rientjes1b604d72009-12-14 17:57:47 -0800395 if (sysctl_oom_dump_tasks)
David Rientjese85bfd32010-09-22 13:05:10 -0700396 dump_tasks(mem, nodemask);
David Rientjes1b604d72009-12-14 17:57:47 -0800397}
398
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800399#define K(x) ((x) << (PAGE_SHIFT-10))
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700400static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401{
David Rientjes1e99bad2010-10-26 14:21:24 -0700402 struct task_struct *q;
403 struct mm_struct *mm;
404
David Rientjesb940fd72010-08-09 17:18:57 -0700405 p = find_lock_task_mm(p);
KOSAKI Motohirobe71cf22010-08-19 14:13:38 -0700406 if (!p)
Dave Peterson01315922006-04-18 22:20:44 -0700407 return 1;
KOSAKI Motohirobe71cf22010-08-19 14:13:38 -0700408
David Rientjes1e99bad2010-10-26 14:21:24 -0700409 /* mm cannot be safely dereferenced after task_unlock(p) */
410 mm = p->mm;
411
David Rientjesb940fd72010-08-09 17:18:57 -0700412 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
413 task_pid_nr(p), p->comm, K(p->mm->total_vm),
414 K(get_mm_counter(p->mm, MM_ANONPAGES)),
415 K(get_mm_counter(p->mm, MM_FILEPAGES)));
416 task_unlock(p);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700417
David Rientjes1e99bad2010-10-26 14:21:24 -0700418 /*
419 * Kill all processes sharing p->mm in other thread groups, if any.
420 * They don't get access to memory reserves or a higher scheduler
421 * priority, though, to avoid depletion of all memory or task
422 * starvation. This prevents mm->mmap_sem livelock when an oom killed
423 * task cannot exit because it requires the semaphore and its contended
424 * by another thread trying to allocate memory itself. That thread will
425 * now get access to memory reserves since it has a pending fatal
426 * signal.
427 */
428 for_each_process(q)
429 if (q->mm == mm && !same_thread_group(q, p)) {
430 task_lock(q); /* Protect ->comm from prctl() */
431 pr_err("Kill process %d (%s) sharing same memory\n",
432 task_pid_nr(q), q->comm);
433 task_unlock(q);
434 force_sig(SIGKILL, q);
435 }
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700436
David Rientjesb940fd72010-08-09 17:18:57 -0700437 set_tsk_thread_flag(p, TIF_MEMDIE);
438 force_sig(SIGKILL, p);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700439
Dave Peterson01315922006-04-18 22:20:44 -0700440 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441}
David Rientjesb940fd72010-08-09 17:18:57 -0700442#undef K
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
David Rientjes7213f502007-10-16 23:25:57 -0700444static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesa63d83f2010-08-09 17:19:46 -0700445 unsigned int points, unsigned long totalpages,
446 struct mem_cgroup *mem, nodemask_t *nodemask,
447 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448{
Linus Torvalds52d3c032011-03-14 15:17:07 -0700449 struct task_struct *victim = p;
David Rientjes5e9d8342010-08-09 17:18:51 -0700450 struct task_struct *child;
Linus Torvalds52d3c032011-03-14 15:17:07 -0700451 struct task_struct *t = p;
452 unsigned int victim_points = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
David Rientjes1b604d72009-12-14 17:57:47 -0800454 if (printk_ratelimit())
David Rientjese85bfd32010-09-22 13:05:10 -0700455 dump_header(p, gfp_mask, order, mem, nodemask);
David Rientjes7213f502007-10-16 23:25:57 -0700456
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700457 /*
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
460 */
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700461 if (p->flags & PF_EXITING) {
David Rientjes43589972010-08-09 17:18:49 -0700462 set_tsk_thread_flag(p, TIF_MEMDIE);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700463 return 0;
464 }
465
David Rientjes5e9d8342010-08-09 17:18:51 -0700466 task_lock(p);
David Rientjesa63d83f2010-08-09 17:19:46 -0700467 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
David Rientjes5e9d8342010-08-09 17:18:51 -0700468 message, task_pid_nr(p), p->comm, points);
469 task_unlock(p);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800470
David Rientjes5e9d8342010-08-09 17:18:51 -0700471 /*
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.
476 */
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700477 do {
David Rientjes5e9d8342010-08-09 17:18:51 -0700478 list_for_each_entry(child, &t->children, sibling) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700479 unsigned int child_points;
David Rientjes5e9d8342010-08-09 17:18:51 -0700480
David Rientjesedd45542011-03-22 16:30:12 -0700481 if (child->mm == p->mm)
482 continue;
David Rientjesa63d83f2010-08-09 17:19:46 -0700483 /*
484 * oom_badness() returns 0 if the thread is unkillable
485 */
486 child_points = oom_badness(child, mem, nodemask,
487 totalpages);
David Rientjes5e9d8342010-08-09 17:18:51 -0700488 if (child_points > victim_points) {
489 victim = child;
490 victim_points = child_points;
491 }
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700492 }
493 } while_each_thread(p, t);
494
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700495 return oom_kill_task(victim, mem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496}
497
David Rientjes309ed882010-08-09 17:18:54 -0700498/*
499 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
500 */
501static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
David Rientjese85bfd32010-09-22 13:05:10 -0700502 int order, const nodemask_t *nodemask)
David Rientjes309ed882010-08-09 17:18:54 -0700503{
504 if (likely(!sysctl_panic_on_oom))
505 return;
506 if (sysctl_panic_on_oom != 2) {
507 /*
508 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
509 * does not panic for cpuset, mempolicy, or memcg allocation
510 * failures.
511 */
512 if (constraint != CONSTRAINT_NONE)
513 return;
514 }
515 read_lock(&tasklist_lock);
David Rientjese85bfd32010-09-22 13:05:10 -0700516 dump_header(NULL, gfp_mask, order, NULL, nodemask);
David Rientjes309ed882010-08-09 17:18:54 -0700517 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");
520}
521
Balbir Singh00f0b822008-03-04 14:28:39 -0800522#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800523void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
524{
David Rientjesa63d83f2010-08-09 17:19:46 -0700525 unsigned long limit;
526 unsigned int points = 0;
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800527 struct task_struct *p;
528
David Rientjesf9434ad2011-03-23 16:42:44 -0700529 /*
530 * If current has a pending SIGKILL, then automatically select it. The
531 * goal is to allow it to allocate so that it may quickly exit and free
532 * its memory.
533 */
534 if (fatal_signal_pending(current)) {
535 set_thread_flag(TIF_MEMDIE);
David Rientjesf9434ad2011-03-23 16:42:44 -0700536 return;
537 }
538
David Rientjese85bfd32010-09-22 13:05:10 -0700539 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
David Rientjesa63d83f2010-08-09 17:19:46 -0700540 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
Li Zefane115f2d2008-04-15 14:34:37 -0700541 read_lock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800542retry:
David Rientjesa63d83f2010-08-09 17:19:46 -0700543 p = select_bad_process(&points, limit, mem, NULL);
David Rientjesdf64f812010-05-26 14:42:41 -0700544 if (!p || PTR_ERR(p) == -1UL)
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800545 goto out;
546
David Rientjesa63d83f2010-08-09 17:19:46 -0700547 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800548 "Memory cgroup out of memory"))
549 goto retry;
550out:
Li Zefane115f2d2008-04-15 14:34:37 -0700551 read_unlock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800552}
553#endif
554
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700555static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
556
557int register_oom_notifier(struct notifier_block *nb)
558{
559 return blocking_notifier_chain_register(&oom_notify_list, nb);
560}
561EXPORT_SYMBOL_GPL(register_oom_notifier);
562
563int unregister_oom_notifier(struct notifier_block *nb)
564{
565 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
566}
567EXPORT_SYMBOL_GPL(unregister_oom_notifier);
568
David Rientjes098d7f12007-10-16 23:25:55 -0700569/*
570 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
571 * if a parallel OOM killing is already taking place that includes a zone in
572 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
573 */
Minchan Kimff321fe2010-08-09 17:18:57 -0700574int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700575{
Mel Gormandd1a2392008-04-28 02:12:17 -0700576 struct zoneref *z;
577 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700578 int ret = 1;
579
David Rientjesc7d4cae2009-01-06 14:39:00 -0800580 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700581 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
582 if (zone_is_oom_locked(zone)) {
David Rientjes098d7f12007-10-16 23:25:55 -0700583 ret = 0;
584 goto out;
585 }
Mel Gormandd1a2392008-04-28 02:12:17 -0700586 }
David Rientjes098d7f12007-10-16 23:25:55 -0700587
Mel Gormandd1a2392008-04-28 02:12:17 -0700588 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
589 /*
David Rientjesc7d4cae2009-01-06 14:39:00 -0800590 * Lock each zone in the zonelist under zone_scan_lock so a
Minchan Kimff321fe2010-08-09 17:18:57 -0700591 * parallel invocation of try_set_zonelist_oom() doesn't succeed
Mel Gormandd1a2392008-04-28 02:12:17 -0700592 * when it shouldn't.
593 */
594 zone_set_flag(zone, ZONE_OOM_LOCKED);
595 }
596
David Rientjes098d7f12007-10-16 23:25:55 -0700597out:
David Rientjesc7d4cae2009-01-06 14:39:00 -0800598 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700599 return ret;
600}
601
602/*
603 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
604 * allocation attempts with zonelists containing them may now recall the OOM
605 * killer, if necessary.
606 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700607void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700608{
Mel Gormandd1a2392008-04-28 02:12:17 -0700609 struct zoneref *z;
610 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700611
David Rientjesc7d4cae2009-01-06 14:39:00 -0800612 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700613 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
614 zone_clear_flag(zone, ZONE_OOM_LOCKED);
615 }
David Rientjesc7d4cae2009-01-06 14:39:00 -0800616 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700617}
618
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800619/*
David Rientjese3658932010-08-09 17:18:55 -0700620 * Try to acquire the oom killer lock for all system zones. Returns zero if a
621 * parallel oom killing is taking place, otherwise locks all zones and returns
622 * non-zero.
623 */
624static int try_set_system_oom(void)
625{
626 struct zone *zone;
627 int ret = 1;
628
629 spin_lock(&zone_scan_lock);
630 for_each_populated_zone(zone)
631 if (zone_is_oom_locked(zone)) {
632 ret = 0;
633 goto out;
634 }
635 for_each_populated_zone(zone)
636 zone_set_flag(zone, ZONE_OOM_LOCKED);
637out:
638 spin_unlock(&zone_scan_lock);
639 return ret;
640}
641
642/*
643 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
644 * attempts or page faults may now recall the oom killer, if necessary.
645 */
646static void clear_system_oom(void)
647{
648 struct zone *zone;
649
650 spin_lock(&zone_scan_lock);
651 for_each_populated_zone(zone)
652 zone_clear_flag(zone, ZONE_OOM_LOCKED);
653 spin_unlock(&zone_scan_lock);
654}
655
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656/**
Dave Peterson6937a252006-06-23 02:03:13 -0700657 * out_of_memory - kill the "best" process when we run out of memory
Randy Dunlap1b578df2008-03-19 17:00:42 -0700658 * @zonelist: zonelist pointer
659 * @gfp_mask: memory allocation flags
660 * @order: amount of memory being requested as a power of 2
David Rientjes6f48d0eb2010-08-09 17:18:52 -0700661 * @nodemask: nodemask passed to page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 *
663 * If we run out of memory, we have the choice between either
664 * killing a random task (bad), letting the system crash (worse)
665 * OR try to be smart about which process to kill. Note that we
666 * don't have to be perfect here, we just have to be good.
667 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800668void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
669 int order, nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670{
David Rientjese85bfd32010-09-22 13:05:10 -0700671 const nodemask_t *mpol_mask;
David Rientjes0aad4b32010-08-09 17:18:59 -0700672 struct task_struct *p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700673 unsigned long totalpages;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700674 unsigned long freed = 0;
David Rientjesa63d83f2010-08-09 17:19:46 -0700675 unsigned int points;
David Rientjese3658932010-08-09 17:18:55 -0700676 enum oom_constraint constraint = CONSTRAINT_NONE;
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700677 int killed = 0;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700678
679 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
680 if (freed > 0)
681 /* Got some memory back in the last second. */
682 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
David Rientjes7b98c2e2010-08-09 17:18:48 -0700684 /*
685 * If current has a pending SIGKILL, then automatically select it. The
686 * goal is to allow it to allocate so that it may quickly exit and free
687 * its memory.
688 */
689 if (fatal_signal_pending(current)) {
690 set_thread_flag(TIF_MEMDIE);
691 return;
692 }
693
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800694 /*
695 * Check if there were limitations on the allocation (only relevant for
696 * NUMA) that may require different handling.
697 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700698 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
699 &totalpages);
David Rientjese85bfd32010-09-22 13:05:10 -0700700 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
701 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
David Rientjes0aad4b32010-08-09 17:18:59 -0700702
David Rientjes2b45ab32007-05-06 14:50:00 -0700703 read_lock(&tasklist_lock);
KOSAKI Motohirof88ccad2010-08-09 17:19:36 -0700704 if (sysctl_oom_kill_allocating_task &&
KOSAKI Motohiroa96cfd62010-08-09 17:19:38 -0700705 !oom_unkillable_task(current, NULL, nodemask) &&
David Rientjese18641e2010-10-26 14:21:23 -0700706 current->mm && !atomic_read(&current->mm->oom_disable_count)) {
David Rientjes0aad4b32010-08-09 17:18:59 -0700707 /*
708 * oom_kill_process() needs tasklist_lock held. If it returns
709 * non-zero, current could not be killed so we must fallback to
710 * the tasklist scan.
711 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700712 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
713 NULL, nodemask,
David Rientjes0aad4b32010-08-09 17:18:59 -0700714 "Out of memory (oom_kill_allocating_task)"))
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700715 goto out;
David Rientjes0aad4b32010-08-09 17:18:59 -0700716 }
717
718retry:
David Rientjese85bfd32010-09-22 13:05:10 -0700719 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
David Rientjes0aad4b32010-08-09 17:18:59 -0700720 if (PTR_ERR(p) == -1UL)
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700721 goto out;
David Rientjes0aad4b32010-08-09 17:18:59 -0700722
723 /* Found nothing?!?! Either we hang forever, or we panic. */
724 if (!p) {
David Rientjese85bfd32010-09-22 13:05:10 -0700725 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
David Rientjes0aad4b32010-08-09 17:18:59 -0700726 read_unlock(&tasklist_lock);
727 panic("Out of memory and no killable processes...\n");
728 }
729
David Rientjesa63d83f2010-08-09 17:19:46 -0700730 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
731 nodemask, "Out of memory"))
David Rientjes0aad4b32010-08-09 17:18:59 -0700732 goto retry;
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700733 killed = 1;
734out:
Andrew Morton140ffce2006-03-02 02:54:28 -0800735 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736
737 /*
738 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800739 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 */
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700741 if (killed && !test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800742 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743}
David Rientjese3658932010-08-09 17:18:55 -0700744
745/*
746 * The pagefault handler calls here because it is out of memory, so kill a
747 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
748 * oom killing is already in progress so do nothing. If a task is found with
749 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
750 */
751void pagefault_out_of_memory(void)
752{
753 if (try_set_system_oom()) {
754 out_of_memory(NULL, 0, 0, NULL);
755 clear_system_oom();
756 }
757 if (!test_thread_flag(TIF_MEMDIE))
758 schedule_timeout_uninterruptible(1);
759}