blob: 7dcca55ede7ca1493df615af994f6ba23c1bb129 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -070035int sysctl_panic_on_oom;
David Rientjesfe071d72007-10-16 23:25:56 -070036int sysctl_oom_kill_allocating_task;
David Rientjesad915c42010-08-09 17:18:53 -070037int sysctl_oom_dump_tasks = 1;
David Rientjesc7d4cae2009-01-06 14:39:00 -080038static DEFINE_SPINLOCK(zone_scan_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
David Rientjes6f48d0eb2010-08-09 17:18:52 -070040#ifdef CONFIG_NUMA
41/**
42 * has_intersects_mems_allowed() - check task eligiblity for kill
43 * @tsk: task struct of which task to consider
44 * @mask: nodemask passed to page allocator for mempolicy ooms
45 *
46 * Task eligibility is determined by whether or not a candidate task, @tsk,
47 * shares the same mempolicy nodes as current if it is bound by such a policy
48 * and whether or not it has the same set of allowed cpuset nodes.
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070049 */
David Rientjes6f48d0eb2010-08-09 17:18:52 -070050static bool has_intersects_mems_allowed(struct task_struct *tsk,
51 const nodemask_t *mask)
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070052{
David Rientjes6f48d0eb2010-08-09 17:18:52 -070053 struct task_struct *start = tsk;
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070054
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070055 do {
David Rientjes6f48d0eb2010-08-09 17:18:52 -070056 if (mask) {
57 /*
58 * If this is a mempolicy constrained oom, tsk's
59 * cpuset is irrelevant. Only return true if its
60 * mempolicy intersects current, otherwise it may be
61 * needlessly killed.
62 */
63 if (mempolicy_nodemask_intersects(tsk, mask))
64 return true;
65 } else {
66 /*
67 * This is not a mempolicy constrained oom, so only
68 * check the mems of tsk's cpuset.
69 */
70 if (cpuset_mems_allowed_intersects(current, tsk))
71 return true;
72 }
KOSAKI Motohirodf1090a2010-08-09 17:19:39 -070073 } while_each_thread(start, tsk);
74
David Rientjes6f48d0eb2010-08-09 17:18:52 -070075 return false;
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070076}
David Rientjes6f48d0eb2010-08-09 17:18:52 -070077#else
78static bool has_intersects_mems_allowed(struct task_struct *tsk,
79 const nodemask_t *mask)
80{
81 return true;
82}
83#endif /* CONFIG_NUMA */
KOSAKI Motohiro495789a2009-09-21 17:03:14 -070084
David Rientjes6f48d0eb2010-08-09 17:18:52 -070085/*
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -070086 * If this is a system OOM (not a memcg OOM) and the task selected to be
87 * killed is not already running at high (RT) priorities, speed up the
88 * recovery by boosting the dying task to the lowest FIFO priority.
89 * That helps with the recovery and avoids interfering with RT tasks.
90 */
91static void boost_dying_task_prio(struct task_struct *p,
92 struct mem_cgroup *mem)
93{
94 struct sched_param param = { .sched_priority = 1 };
95
96 if (mem)
97 return;
98
99 if (!rt_task(p))
100 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
101}
102
103/*
David Rientjes6f48d0eb2010-08-09 17:18:52 -0700104 * The process p may have detached its own ->mm while exiting or through
105 * use_mm(), but one or more of its subthreads may still have a valid
106 * pointer. Return p, or any of its subthreads with a valid ->mm, with
107 * task_lock() held.
108 */
KAMEZAWA Hiroyuki158e0a22010-08-10 18:03:00 -0700109struct task_struct *find_lock_task_mm(struct task_struct *p)
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700110{
111 struct task_struct *t = p;
112
113 do {
114 task_lock(t);
115 if (likely(t->mm))
116 return t;
117 task_unlock(t);
118 } while_each_thread(p, t);
119
120 return NULL;
121}
122
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700123/* return true if the task is not adequate as candidate victim task. */
David Rientjese85bfd32010-09-22 13:05:10 -0700124static bool oom_unkillable_task(struct task_struct *p,
125 const struct mem_cgroup *mem, const nodemask_t *nodemask)
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700126{
127 if (is_global_init(p))
128 return true;
129 if (p->flags & PF_KTHREAD)
130 return true;
131
132 /* When mem_cgroup_out_of_memory() and p is not member of the group */
133 if (mem && !task_in_mem_cgroup(p, mem))
134 return true;
135
136 /* p may not have freeable memory in nodemask */
137 if (!has_intersects_mems_allowed(p, nodemask))
138 return true;
139
140 return false;
141}
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143/**
David Rientjesa63d83f2010-08-09 17:19:46 -0700144 * oom_badness - heuristic function to determine which candidate task to kill
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 * @p: task struct of which task we should calculate
David Rientjesa63d83f2010-08-09 17:19:46 -0700146 * @totalpages: total present RAM allowed for page allocation
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 *
David Rientjesa63d83f2010-08-09 17:19:46 -0700148 * The heuristic for determining which task to kill is made to be as simple and
149 * predictable as possible. The goal is to return the highest value for the
150 * task consuming the most memory to avoid subsequent oom failures.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700152unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
153 const nodemask_t *nodemask, unsigned long totalpages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154{
David Rientjesa63d83f2010-08-09 17:19:46 -0700155 int points;
KOSAKI Motohiro28b83c52009-09-21 17:03:13 -0700156
KOSAKI Motohiro26ebc982010-08-09 17:19:37 -0700157 if (oom_unkillable_task(p, mem, nodemask))
158 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700160 p = find_lock_task_mm(p);
161 if (!p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 return 0;
163
164 /*
David Rientjese18641e2010-10-26 14:21:23 -0700165 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
166 * so the entire heuristic doesn't need to be executed for something
167 * that cannot be killed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 */
David Rientjese18641e2010-10-26 14:21:23 -0700169 if (atomic_read(&p->mm->oom_disable_count)) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700170 task_unlock(p);
171 return 0;
172 }
173
174 /*
175 * When the PF_OOM_ORIGIN bit is set, it indicates the task should have
176 * priority for oom killing.
177 */
178 if (p->flags & PF_OOM_ORIGIN) {
179 task_unlock(p);
180 return 1000;
181 }
182
183 /*
184 * The memory controller may have a limit of 0 bytes, so avoid a divide
185 * by zero, if necessary.
186 */
187 if (!totalpages)
188 totalpages = 1;
189
190 /*
191 * The baseline for the badness score is the proportion of RAM that each
192 * task's rss and swap space use.
193 */
194 points = (get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS)) * 1000 /
195 totalpages;
Andrew Morton97c2c9b2006-04-18 22:20:38 -0700196 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197
198 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700199 * Root processes get 3% bonus, just like the __vm_enough_memory()
200 * implementation used by LSMs.
Hugh Dickins7ba34852007-01-05 16:37:03 -0800201 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700202 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
203 points -= 30;
Hugh Dickins7ba34852007-01-05 16:37:03 -0800204
205 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700206 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
207 * either completely disable oom killing or always prefer a certain
208 * task.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700210 points += p->signal->oom_score_adj;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
David Rientjesf19e8aa2010-09-22 13:04:52 -0700212 /*
213 * Never return 0 for an eligible task that may be killed since it's
214 * possible that no single user task uses more than 0.1% of memory and
215 * no single admin tasks uses more than 3.0%.
216 */
217 if (points <= 0)
218 return 1;
David Rientjesa63d83f2010-08-09 17:19:46 -0700219 return (points < 1000) ? points : 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220}
221
222/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800223 * Determine the type of allocation constraint.
224 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800225#ifdef CONFIG_NUMA
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800226static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
David Rientjesa63d83f2010-08-09 17:19:46 -0700227 gfp_t gfp_mask, nodemask_t *nodemask,
228 unsigned long *totalpages)
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800229{
Mel Gorman54a6eb52008-04-28 02:12:16 -0700230 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -0700231 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -0700232 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
David Rientjesa63d83f2010-08-09 17:19:46 -0700233 bool cpuset_limited = false;
234 int nid;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800235
David Rientjesa63d83f2010-08-09 17:19:46 -0700236 /* Default to all available memory */
237 *totalpages = totalram_pages + total_swap_pages;
238
239 if (!zonelist)
240 return CONSTRAINT_NONE;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800241 /*
242 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
243 * to kill current.We have to random task kill in this case.
244 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
245 */
246 if (gfp_mask & __GFP_THISNODE)
247 return CONSTRAINT_NONE;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800248
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800249 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700250 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
251 * the page allocator means a mempolicy is in effect. Cpuset policy
252 * is enforced in get_page_from_freelist().
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800253 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700254 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
255 *totalpages = total_swap_pages;
256 for_each_node_mask(nid, *nodemask)
257 *totalpages += node_spanned_pages(nid);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800258 return CONSTRAINT_MEMORY_POLICY;
David Rientjesa63d83f2010-08-09 17:19:46 -0700259 }
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800260
261 /* Check this allocation failure is caused by cpuset's wall function */
262 for_each_zone_zonelist_nodemask(zone, z, zonelist,
263 high_zoneidx, nodemask)
264 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
David Rientjesa63d83f2010-08-09 17:19:46 -0700265 cpuset_limited = true;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800266
David Rientjesa63d83f2010-08-09 17:19:46 -0700267 if (cpuset_limited) {
268 *totalpages = total_swap_pages;
269 for_each_node_mask(nid, cpuset_current_mems_allowed)
270 *totalpages += node_spanned_pages(nid);
271 return CONSTRAINT_CPUSET;
272 }
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800273 return CONSTRAINT_NONE;
274}
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800275#else
276static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
David Rientjesa63d83f2010-08-09 17:19:46 -0700277 gfp_t gfp_mask, nodemask_t *nodemask,
278 unsigned long *totalpages)
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800279{
David Rientjesa63d83f2010-08-09 17:19:46 -0700280 *totalpages = totalram_pages + total_swap_pages;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800281 return CONSTRAINT_NONE;
282}
283#endif
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800284
285/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 * Simple selection loop. We chose the process with the highest
287 * number of 'points'. We expect the caller will lock the tasklist.
288 *
289 * (not docbooked, we don't want this one cluttering up the manual)
290 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700291static struct task_struct *select_bad_process(unsigned int *ppoints,
292 unsigned long totalpages, struct mem_cgroup *mem,
293 const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700295 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 struct task_struct *chosen = NULL;
Kurt Garloff9827b782006-02-20 18:27:51 -0800297 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700299 for_each_process(p) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700300 unsigned int points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700301
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700302 if (oom_unkillable_task(p, mem, nodemask))
David Rientjes6cf86ac2010-08-09 17:18:50 -0700303 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700304
Paul Jacksona49335c2005-09-06 15:18:09 -0700305 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700306 * This task already has access to memory reserves and is
307 * being killed. Don't allow any other task access to the
308 * memory reserve.
309 *
310 * Note: this may have a chance of deadlock if it gets
311 * blocked waiting for another task which itself is waiting
312 * for memory. Is there a better alternative?
313 */
314 if (test_tsk_thread_flag(p, TIF_MEMDIE))
315 return ERR_PTR(-1UL);
316
317 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700318 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700319 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700320 *
321 * However, if p is the current task, we allow the 'kill' to
322 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
323 * which will allow it to gain access to memory reserves in
324 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700325 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700326 */
KOSAKI Motohirocef1d352010-08-09 17:19:42 -0700327 if (thread_group_empty(p) && (p->flags & PF_EXITING) && p->mm) {
Nick Pigginb78483a2006-09-29 02:01:14 -0700328 if (p != current)
329 return ERR_PTR(-1UL);
330
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700331 chosen = p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700332 *ppoints = 1000;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700333 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700334
David Rientjesa63d83f2010-08-09 17:19:46 -0700335 points = oom_badness(p, mem, nodemask, totalpages);
336 if (points > *ppoints) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700337 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800338 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 }
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700340 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700341
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 return chosen;
343}
344
345/**
Randy Dunlap1b578df2008-03-19 17:00:42 -0700346 * dump_tasks - dump current memory state of all system tasks
David Rientjes74ab7f12010-08-09 17:18:46 -0700347 * @mem: current's memory controller, if constrained
David Rientjese85bfd32010-09-22 13:05:10 -0700348 * @nodemask: nodemask passed to page allocator for mempolicy ooms
Randy Dunlap1b578df2008-03-19 17:00:42 -0700349 *
David Rientjese85bfd32010-09-22 13:05:10 -0700350 * Dumps the current memory state of all eligible tasks. Tasks not in the same
351 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
352 * are not shown.
David Rientjesfef1bdd2008-02-07 00:14:07 -0800353 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
David Rientjesa63d83f2010-08-09 17:19:46 -0700354 * value, oom_score_adj value, and name.
David Rientjesfef1bdd2008-02-07 00:14:07 -0800355 *
David Rientjesfef1bdd2008-02-07 00:14:07 -0800356 * Call with tasklist_lock read-locked.
357 */
David Rientjese85bfd32010-09-22 13:05:10 -0700358static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
David Rientjesfef1bdd2008-02-07 00:14:07 -0800359{
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700360 struct task_struct *p;
361 struct task_struct *task;
David Rientjesfef1bdd2008-02-07 00:14:07 -0800362
David Rientjesa63d83f2010-08-09 17:19:46 -0700363 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700364 for_each_process(p) {
David Rientjese85bfd32010-09-22 13:05:10 -0700365 if (oom_unkillable_task(p, mem, nodemask))
David Rientjesfef1bdd2008-02-07 00:14:07 -0800366 continue;
367
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700368 task = find_lock_task_mm(p);
369 if (!task) {
David Rientjes6d2661e2009-05-28 14:34:19 -0700370 /*
David Rientjes74ab7f12010-08-09 17:18:46 -0700371 * This is a kthread or all of p's threads have already
372 * detached their mm's. There's no need to report
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700373 * them; they can't be oom killed anyway.
David Rientjes6d2661e2009-05-28 14:34:19 -0700374 */
David Rientjes6d2661e2009-05-28 14:34:19 -0700375 continue;
376 }
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700377
David Rientjesa63d83f2010-08-09 17:19:46 -0700378 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
KOSAKI Motohiro8d6c83f2010-08-19 14:13:39 -0700379 task->pid, task_uid(task), task->tgid,
David Rientjesa63d83f2010-08-09 17:19:46 -0700380 task->mm->total_vm, get_mm_rss(task->mm),
381 task_cpu(task), task->signal->oom_adj,
382 task->signal->oom_score_adj, task->comm);
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700383 task_unlock(task);
384 }
David Rientjesfef1bdd2008-02-07 00:14:07 -0800385}
386
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800387static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjese85bfd32010-09-22 13:05:10 -0700388 struct mem_cgroup *mem, const nodemask_t *nodemask)
David Rientjes1b604d72009-12-14 17:57:47 -0800389{
David Rientjes5e9d8342010-08-09 17:18:51 -0700390 task_lock(current);
David Rientjes1b604d72009-12-14 17:57:47 -0800391 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
David Rientjesa63d83f2010-08-09 17:19:46 -0700392 "oom_adj=%d, oom_score_adj=%d\n",
393 current->comm, gfp_mask, order, current->signal->oom_adj,
394 current->signal->oom_score_adj);
David Rientjes1b604d72009-12-14 17:57:47 -0800395 cpuset_print_task_mems_allowed(current);
396 task_unlock(current);
397 dump_stack();
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800398 mem_cgroup_print_oom_info(mem, p);
David Rientjes1b604d72009-12-14 17:57:47 -0800399 show_mem();
400 if (sysctl_oom_dump_tasks)
David Rientjese85bfd32010-09-22 13:05:10 -0700401 dump_tasks(mem, nodemask);
David Rientjes1b604d72009-12-14 17:57:47 -0800402}
403
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800404#define K(x) ((x) << (PAGE_SHIFT-10))
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700405static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406{
David Rientjes1e99bad2010-10-26 14:21:24 -0700407 struct task_struct *q;
408 struct mm_struct *mm;
409
David Rientjesb940fd72010-08-09 17:18:57 -0700410 p = find_lock_task_mm(p);
KOSAKI Motohirobe71cf22010-08-19 14:13:38 -0700411 if (!p)
Dave Peterson01315922006-04-18 22:20:44 -0700412 return 1;
KOSAKI Motohirobe71cf22010-08-19 14:13:38 -0700413
David Rientjes1e99bad2010-10-26 14:21:24 -0700414 /* mm cannot be safely dereferenced after task_unlock(p) */
415 mm = p->mm;
416
David Rientjesb940fd72010-08-09 17:18:57 -0700417 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
418 task_pid_nr(p), p->comm, K(p->mm->total_vm),
419 K(get_mm_counter(p->mm, MM_ANONPAGES)),
420 K(get_mm_counter(p->mm, MM_FILEPAGES)));
421 task_unlock(p);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700422
David Rientjes1e99bad2010-10-26 14:21:24 -0700423 /*
424 * Kill all processes sharing p->mm in other thread groups, if any.
425 * They don't get access to memory reserves or a higher scheduler
426 * priority, though, to avoid depletion of all memory or task
427 * starvation. This prevents mm->mmap_sem livelock when an oom killed
428 * task cannot exit because it requires the semaphore and its contended
429 * by another thread trying to allocate memory itself. That thread will
430 * now get access to memory reserves since it has a pending fatal
431 * signal.
432 */
433 for_each_process(q)
434 if (q->mm == mm && !same_thread_group(q, p)) {
435 task_lock(q); /* Protect ->comm from prctl() */
436 pr_err("Kill process %d (%s) sharing same memory\n",
437 task_pid_nr(q), q->comm);
438 task_unlock(q);
439 force_sig(SIGKILL, q);
440 }
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700441
David Rientjesb940fd72010-08-09 17:18:57 -0700442 set_tsk_thread_flag(p, TIF_MEMDIE);
443 force_sig(SIGKILL, p);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700444
445 /*
446 * We give our sacrificial lamb high priority and access to
447 * all the memory it needs. That way it should be able to
448 * exit() and clear out its resources quickly...
449 */
450 boost_dying_task_prio(p, mem);
451
Dave Peterson01315922006-04-18 22:20:44 -0700452 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453}
David Rientjesb940fd72010-08-09 17:18:57 -0700454#undef K
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
David Rientjes7213f502007-10-16 23:25:57 -0700456static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesa63d83f2010-08-09 17:19:46 -0700457 unsigned int points, unsigned long totalpages,
458 struct mem_cgroup *mem, nodemask_t *nodemask,
459 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
David Rientjes5e9d8342010-08-09 17:18:51 -0700461 struct task_struct *victim = p;
462 struct task_struct *child;
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700463 struct task_struct *t = p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700464 unsigned int victim_points = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
David Rientjes1b604d72009-12-14 17:57:47 -0800466 if (printk_ratelimit())
David Rientjese85bfd32010-09-22 13:05:10 -0700467 dump_header(p, gfp_mask, order, mem, nodemask);
David Rientjes7213f502007-10-16 23:25:57 -0700468
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700469 /*
470 * If the task is already exiting, don't alarm the sysadmin or kill
471 * its children or threads, just set TIF_MEMDIE so it can die quickly
472 */
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700473 if (p->flags & PF_EXITING) {
David Rientjes43589972010-08-09 17:18:49 -0700474 set_tsk_thread_flag(p, TIF_MEMDIE);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700475 boost_dying_task_prio(p, mem);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700476 return 0;
477 }
478
David Rientjes5e9d8342010-08-09 17:18:51 -0700479 task_lock(p);
David Rientjesa63d83f2010-08-09 17:19:46 -0700480 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
David Rientjes5e9d8342010-08-09 17:18:51 -0700481 message, task_pid_nr(p), p->comm, points);
482 task_unlock(p);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800483
David Rientjes5e9d8342010-08-09 17:18:51 -0700484 /*
485 * If any of p's children has a different mm and is eligible for kill,
486 * the one with the highest badness() score is sacrificed for its
487 * parent. This attempts to lose the minimal amount of work done while
488 * still freeing memory.
489 */
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700490 do {
David Rientjes5e9d8342010-08-09 17:18:51 -0700491 list_for_each_entry(child, &t->children, sibling) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700492 unsigned int child_points;
David Rientjes5e9d8342010-08-09 17:18:51 -0700493
David Rientjesa63d83f2010-08-09 17:19:46 -0700494 /*
495 * oom_badness() returns 0 if the thread is unkillable
496 */
497 child_points = oom_badness(child, mem, nodemask,
498 totalpages);
David Rientjes5e9d8342010-08-09 17:18:51 -0700499 if (child_points > victim_points) {
500 victim = child;
501 victim_points = child_points;
502 }
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700503 }
504 } while_each_thread(p, t);
505
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700506 return oom_kill_task(victim, mem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507}
508
David Rientjes309ed882010-08-09 17:18:54 -0700509/*
510 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
511 */
512static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
David Rientjese85bfd32010-09-22 13:05:10 -0700513 int order, const nodemask_t *nodemask)
David Rientjes309ed882010-08-09 17:18:54 -0700514{
515 if (likely(!sysctl_panic_on_oom))
516 return;
517 if (sysctl_panic_on_oom != 2) {
518 /*
519 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
520 * does not panic for cpuset, mempolicy, or memcg allocation
521 * failures.
522 */
523 if (constraint != CONSTRAINT_NONE)
524 return;
525 }
526 read_lock(&tasklist_lock);
David Rientjese85bfd32010-09-22 13:05:10 -0700527 dump_header(NULL, gfp_mask, order, NULL, nodemask);
David Rientjes309ed882010-08-09 17:18:54 -0700528 read_unlock(&tasklist_lock);
529 panic("Out of memory: %s panic_on_oom is enabled\n",
530 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
531}
532
Balbir Singh00f0b822008-03-04 14:28:39 -0800533#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800534void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
535{
David Rientjesa63d83f2010-08-09 17:19:46 -0700536 unsigned long limit;
537 unsigned int points = 0;
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800538 struct task_struct *p;
539
David Rientjese85bfd32010-09-22 13:05:10 -0700540 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
David Rientjesa63d83f2010-08-09 17:19:46 -0700541 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
Li Zefane115f2d2008-04-15 14:34:37 -0700542 read_lock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800543retry:
David Rientjesa63d83f2010-08-09 17:19:46 -0700544 p = select_bad_process(&points, limit, mem, NULL);
David Rientjesdf64f812010-05-26 14:42:41 -0700545 if (!p || PTR_ERR(p) == -1UL)
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800546 goto out;
547
David Rientjesa63d83f2010-08-09 17:19:46 -0700548 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800549 "Memory cgroup out of memory"))
550 goto retry;
551out:
Li Zefane115f2d2008-04-15 14:34:37 -0700552 read_unlock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800553}
554#endif
555
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700556static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
557
558int register_oom_notifier(struct notifier_block *nb)
559{
560 return blocking_notifier_chain_register(&oom_notify_list, nb);
561}
562EXPORT_SYMBOL_GPL(register_oom_notifier);
563
564int unregister_oom_notifier(struct notifier_block *nb)
565{
566 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
567}
568EXPORT_SYMBOL_GPL(unregister_oom_notifier);
569
David Rientjes098d7f12007-10-16 23:25:55 -0700570/*
571 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
572 * if a parallel OOM killing is already taking place that includes a zone in
573 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
574 */
Minchan Kimff321fe2010-08-09 17:18:57 -0700575int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700576{
Mel Gormandd1a2392008-04-28 02:12:17 -0700577 struct zoneref *z;
578 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700579 int ret = 1;
580
David Rientjesc7d4cae2009-01-06 14:39:00 -0800581 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700582 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
583 if (zone_is_oom_locked(zone)) {
David Rientjes098d7f12007-10-16 23:25:55 -0700584 ret = 0;
585 goto out;
586 }
Mel Gormandd1a2392008-04-28 02:12:17 -0700587 }
David Rientjes098d7f12007-10-16 23:25:55 -0700588
Mel Gormandd1a2392008-04-28 02:12:17 -0700589 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
590 /*
David Rientjesc7d4cae2009-01-06 14:39:00 -0800591 * Lock each zone in the zonelist under zone_scan_lock so a
Minchan Kimff321fe2010-08-09 17:18:57 -0700592 * parallel invocation of try_set_zonelist_oom() doesn't succeed
Mel Gormandd1a2392008-04-28 02:12:17 -0700593 * when it shouldn't.
594 */
595 zone_set_flag(zone, ZONE_OOM_LOCKED);
596 }
597
David Rientjes098d7f12007-10-16 23:25:55 -0700598out:
David Rientjesc7d4cae2009-01-06 14:39:00 -0800599 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700600 return ret;
601}
602
603/*
604 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
605 * allocation attempts with zonelists containing them may now recall the OOM
606 * killer, if necessary.
607 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700608void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700609{
Mel Gormandd1a2392008-04-28 02:12:17 -0700610 struct zoneref *z;
611 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700612
David Rientjesc7d4cae2009-01-06 14:39:00 -0800613 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700614 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
615 zone_clear_flag(zone, ZONE_OOM_LOCKED);
616 }
David Rientjesc7d4cae2009-01-06 14:39:00 -0800617 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700618}
619
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800620/*
David Rientjese3658932010-08-09 17:18:55 -0700621 * Try to acquire the oom killer lock for all system zones. Returns zero if a
622 * parallel oom killing is taking place, otherwise locks all zones and returns
623 * non-zero.
624 */
625static int try_set_system_oom(void)
626{
627 struct zone *zone;
628 int ret = 1;
629
630 spin_lock(&zone_scan_lock);
631 for_each_populated_zone(zone)
632 if (zone_is_oom_locked(zone)) {
633 ret = 0;
634 goto out;
635 }
636 for_each_populated_zone(zone)
637 zone_set_flag(zone, ZONE_OOM_LOCKED);
638out:
639 spin_unlock(&zone_scan_lock);
640 return ret;
641}
642
643/*
644 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
645 * attempts or page faults may now recall the oom killer, if necessary.
646 */
647static void clear_system_oom(void)
648{
649 struct zone *zone;
650
651 spin_lock(&zone_scan_lock);
652 for_each_populated_zone(zone)
653 zone_clear_flag(zone, ZONE_OOM_LOCKED);
654 spin_unlock(&zone_scan_lock);
655}
656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657/**
Dave Peterson6937a252006-06-23 02:03:13 -0700658 * out_of_memory - kill the "best" process when we run out of memory
Randy Dunlap1b578df2008-03-19 17:00:42 -0700659 * @zonelist: zonelist pointer
660 * @gfp_mask: memory allocation flags
661 * @order: amount of memory being requested as a power of 2
David Rientjes6f48d0eb2010-08-09 17:18:52 -0700662 * @nodemask: nodemask passed to page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 *
664 * If we run out of memory, we have the choice between either
665 * killing a random task (bad), letting the system crash (worse)
666 * OR try to be smart about which process to kill. Note that we
667 * don't have to be perfect here, we just have to be good.
668 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800669void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
670 int order, nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671{
David Rientjese85bfd32010-09-22 13:05:10 -0700672 const nodemask_t *mpol_mask;
David Rientjes0aad4b32010-08-09 17:18:59 -0700673 struct task_struct *p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700674 unsigned long totalpages;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700675 unsigned long freed = 0;
David Rientjesa63d83f2010-08-09 17:19:46 -0700676 unsigned int points;
David Rientjese3658932010-08-09 17:18:55 -0700677 enum oom_constraint constraint = CONSTRAINT_NONE;
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700678 int killed = 0;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700679
680 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
681 if (freed > 0)
682 /* Got some memory back in the last second. */
683 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684
David Rientjes7b98c2e2010-08-09 17:18:48 -0700685 /*
686 * If current has a pending SIGKILL, then automatically select it. The
687 * goal is to allow it to allocate so that it may quickly exit and free
688 * its memory.
689 */
690 if (fatal_signal_pending(current)) {
691 set_thread_flag(TIF_MEMDIE);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700692 boost_dying_task_prio(current, NULL);
David Rientjes7b98c2e2010-08-09 17:18:48 -0700693 return;
694 }
695
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800696 /*
697 * Check if there were limitations on the allocation (only relevant for
698 * NUMA) that may require different handling.
699 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700700 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
701 &totalpages);
David Rientjese85bfd32010-09-22 13:05:10 -0700702 mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
703 check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
David Rientjes0aad4b32010-08-09 17:18:59 -0700704
David Rientjes2b45ab32007-05-06 14:50:00 -0700705 read_lock(&tasklist_lock);
KOSAKI Motohirof88ccad2010-08-09 17:19:36 -0700706 if (sysctl_oom_kill_allocating_task &&
KOSAKI Motohiroa96cfd62010-08-09 17:19:38 -0700707 !oom_unkillable_task(current, NULL, nodemask) &&
David Rientjese18641e2010-10-26 14:21:23 -0700708 current->mm && !atomic_read(&current->mm->oom_disable_count)) {
David Rientjes0aad4b32010-08-09 17:18:59 -0700709 /*
710 * oom_kill_process() needs tasklist_lock held. If it returns
711 * non-zero, current could not be killed so we must fallback to
712 * the tasklist scan.
713 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700714 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
715 NULL, nodemask,
David Rientjes0aad4b32010-08-09 17:18:59 -0700716 "Out of memory (oom_kill_allocating_task)"))
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700717 goto out;
David Rientjes0aad4b32010-08-09 17:18:59 -0700718 }
719
720retry:
David Rientjese85bfd32010-09-22 13:05:10 -0700721 p = select_bad_process(&points, totalpages, NULL, mpol_mask);
David Rientjes0aad4b32010-08-09 17:18:59 -0700722 if (PTR_ERR(p) == -1UL)
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700723 goto out;
David Rientjes0aad4b32010-08-09 17:18:59 -0700724
725 /* Found nothing?!?! Either we hang forever, or we panic. */
726 if (!p) {
David Rientjese85bfd32010-09-22 13:05:10 -0700727 dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
David Rientjes0aad4b32010-08-09 17:18:59 -0700728 read_unlock(&tasklist_lock);
729 panic("Out of memory and no killable processes...\n");
730 }
731
David Rientjesa63d83f2010-08-09 17:19:46 -0700732 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
733 nodemask, "Out of memory"))
David Rientjes0aad4b32010-08-09 17:18:59 -0700734 goto retry;
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700735 killed = 1;
736out:
Andrew Morton140ffce2006-03-02 02:54:28 -0800737 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738
739 /*
740 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800741 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 */
KOSAKI Motohirob52723c2010-08-19 14:13:39 -0700743 if (killed && !test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800744 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745}
David Rientjese3658932010-08-09 17:18:55 -0700746
747/*
748 * The pagefault handler calls here because it is out of memory, so kill a
749 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
750 * oom killing is already in progress so do nothing. If a task is found with
751 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
752 */
753void pagefault_out_of_memory(void)
754{
755 if (try_set_system_oom()) {
756 out_of_memory(NULL, 0, 0, NULL);
757 clear_system_oom();
758 }
759 if (!test_thread_flag(TIF_MEMDIE))
760 schedule_timeout_uninterruptible(1);
761}