blob: d3def05a33d99f6699e6dd6ea9472e6e7da54443 [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 */
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700109static struct task_struct *find_lock_task_mm(struct task_struct *p)
110{
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. */
124static bool oom_unkillable_task(struct task_struct *p, struct mem_cgroup *mem,
125 const nodemask_t *nodemask)
126{
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 Rientjesa63d83f2010-08-09 17:19:46 -0700165 * Shortcut check for OOM_SCORE_ADJ_MIN so the entire heuristic doesn't
166 * need to be executed for something that cannot be killed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700168 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
169 task_unlock(p);
170 return 0;
171 }
172
173 /*
174 * When the PF_OOM_ORIGIN bit is set, it indicates the task should have
175 * priority for oom killing.
176 */
177 if (p->flags & PF_OOM_ORIGIN) {
178 task_unlock(p);
179 return 1000;
180 }
181
182 /*
183 * The memory controller may have a limit of 0 bytes, so avoid a divide
184 * by zero, if necessary.
185 */
186 if (!totalpages)
187 totalpages = 1;
188
189 /*
190 * The baseline for the badness score is the proportion of RAM that each
191 * task's rss and swap space use.
192 */
193 points = (get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS)) * 1000 /
194 totalpages;
Andrew Morton97c2c9b82006-04-18 22:20:38 -0700195 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196
197 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700198 * Root processes get 3% bonus, just like the __vm_enough_memory()
199 * implementation used by LSMs.
Hugh Dickins7ba34852007-01-05 16:37:03 -0800200 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700201 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
202 points -= 30;
Hugh Dickins7ba34852007-01-05 16:37:03 -0800203
204 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700205 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
206 * either completely disable oom killing or always prefer a certain
207 * task.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700209 points += p->signal->oom_score_adj;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210
David Rientjesa63d83f2010-08-09 17:19:46 -0700211 if (points < 0)
212 return 0;
213 return (points < 1000) ? points : 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
215
216/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800217 * Determine the type of allocation constraint.
218 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800219#ifdef CONFIG_NUMA
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800220static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
David Rientjesa63d83f2010-08-09 17:19:46 -0700221 gfp_t gfp_mask, nodemask_t *nodemask,
222 unsigned long *totalpages)
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800223{
Mel Gorman54a6eb52008-04-28 02:12:16 -0700224 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -0700225 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -0700226 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
David Rientjesa63d83f2010-08-09 17:19:46 -0700227 bool cpuset_limited = false;
228 int nid;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800229
David Rientjesa63d83f2010-08-09 17:19:46 -0700230 /* Default to all available memory */
231 *totalpages = totalram_pages + total_swap_pages;
232
233 if (!zonelist)
234 return CONSTRAINT_NONE;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800235 /*
236 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
237 * to kill current.We have to random task kill in this case.
238 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
239 */
240 if (gfp_mask & __GFP_THISNODE)
241 return CONSTRAINT_NONE;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800242
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800243 /*
David Rientjesa63d83f2010-08-09 17:19:46 -0700244 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
245 * the page allocator means a mempolicy is in effect. Cpuset policy
246 * is enforced in get_page_from_freelist().
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800247 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700248 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
249 *totalpages = total_swap_pages;
250 for_each_node_mask(nid, *nodemask)
251 *totalpages += node_spanned_pages(nid);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800252 return CONSTRAINT_MEMORY_POLICY;
David Rientjesa63d83f2010-08-09 17:19:46 -0700253 }
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800254
255 /* Check this allocation failure is caused by cpuset's wall function */
256 for_each_zone_zonelist_nodemask(zone, z, zonelist,
257 high_zoneidx, nodemask)
258 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
David Rientjesa63d83f2010-08-09 17:19:46 -0700259 cpuset_limited = true;
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800260
David Rientjesa63d83f2010-08-09 17:19:46 -0700261 if (cpuset_limited) {
262 *totalpages = total_swap_pages;
263 for_each_node_mask(nid, cpuset_current_mems_allowed)
264 *totalpages += node_spanned_pages(nid);
265 return CONSTRAINT_CPUSET;
266 }
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800267 return CONSTRAINT_NONE;
268}
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800269#else
270static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
David Rientjesa63d83f2010-08-09 17:19:46 -0700271 gfp_t gfp_mask, nodemask_t *nodemask,
272 unsigned long *totalpages)
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800273{
David Rientjesa63d83f2010-08-09 17:19:46 -0700274 *totalpages = totalram_pages + total_swap_pages;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800275 return CONSTRAINT_NONE;
276}
277#endif
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800278
279/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 * Simple selection loop. We chose the process with the highest
281 * number of 'points'. We expect the caller will lock the tasklist.
282 *
283 * (not docbooked, we don't want this one cluttering up the manual)
284 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700285static struct task_struct *select_bad_process(unsigned int *ppoints,
286 unsigned long totalpages, struct mem_cgroup *mem,
287 const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288{
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700289 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 struct task_struct *chosen = NULL;
Kurt Garloff9827b782006-02-20 18:27:51 -0800291 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700293 for_each_process(p) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700294 unsigned int points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700295
KOSAKI Motohiroab290ad2010-08-09 17:19:35 -0700296 if (oom_unkillable_task(p, mem, nodemask))
David Rientjes6cf86ac2010-08-09 17:18:50 -0700297 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700298
Paul Jacksona49335c2005-09-06 15:18:09 -0700299 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700300 * This task already has access to memory reserves and is
301 * being killed. Don't allow any other task access to the
302 * memory reserve.
303 *
304 * Note: this may have a chance of deadlock if it gets
305 * blocked waiting for another task which itself is waiting
306 * for memory. Is there a better alternative?
307 */
308 if (test_tsk_thread_flag(p, TIF_MEMDIE))
309 return ERR_PTR(-1UL);
310
311 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700312 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700313 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700314 *
315 * However, if p is the current task, we allow the 'kill' to
316 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
317 * which will allow it to gain access to memory reserves in
318 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700319 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700320 */
KOSAKI Motohirocef1d352010-08-09 17:19:42 -0700321 if (thread_group_empty(p) && (p->flags & PF_EXITING) && p->mm) {
Nick Pigginb78483a2006-09-29 02:01:14 -0700322 if (p != current)
323 return ERR_PTR(-1UL);
324
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700325 chosen = p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700326 *ppoints = 1000;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700327 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700328
David Rientjesa63d83f2010-08-09 17:19:46 -0700329 points = oom_badness(p, mem, nodemask, totalpages);
330 if (points > *ppoints) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700331 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800332 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 }
KOSAKI Motohiro495789a2009-09-21 17:03:14 -0700334 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700335
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 return chosen;
337}
338
339/**
Randy Dunlap1b578df2008-03-19 17:00:42 -0700340 * dump_tasks - dump current memory state of all system tasks
David Rientjes74ab7f12010-08-09 17:18:46 -0700341 * @mem: current's memory controller, if constrained
Randy Dunlap1b578df2008-03-19 17:00:42 -0700342 *
David Rientjesfef1bdd2008-02-07 00:14:07 -0800343 * Dumps the current memory state of all system tasks, excluding kernel threads.
344 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
David Rientjesa63d83f2010-08-09 17:19:46 -0700345 * value, oom_score_adj value, and name.
David Rientjesfef1bdd2008-02-07 00:14:07 -0800346 *
347 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
348 * shown.
349 *
350 * Call with tasklist_lock read-locked.
351 */
352static void dump_tasks(const struct mem_cgroup *mem)
353{
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700354 struct task_struct *p;
355 struct task_struct *task;
David Rientjesfef1bdd2008-02-07 00:14:07 -0800356
David Rientjesa63d83f2010-08-09 17:19:46 -0700357 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700358 for_each_process(p) {
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700359 if (p->flags & PF_KTHREAD)
360 continue;
David Rientjesfef1bdd2008-02-07 00:14:07 -0800361 if (mem && !task_in_mem_cgroup(p, mem))
362 continue;
363
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700364 task = find_lock_task_mm(p);
365 if (!task) {
David Rientjes6d2661e2009-05-28 14:34:19 -0700366 /*
David Rientjes74ab7f12010-08-09 17:18:46 -0700367 * This is a kthread or all of p's threads have already
368 * detached their mm's. There's no need to report
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700369 * them; they can't be oom killed anyway.
David Rientjes6d2661e2009-05-28 14:34:19 -0700370 */
David Rientjes6d2661e2009-05-28 14:34:19 -0700371 continue;
372 }
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700373
David Rientjesa63d83f2010-08-09 17:19:46 -0700374 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
375 task->pid, __task_cred(task)->uid, task->tgid,
376 task->mm->total_vm, get_mm_rss(task->mm),
377 task_cpu(task), task->signal->oom_adj,
378 task->signal->oom_score_adj, task->comm);
KOSAKI Motohiroc55db952010-08-09 17:18:46 -0700379 task_unlock(task);
380 }
David Rientjesfef1bdd2008-02-07 00:14:07 -0800381}
382
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800383static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
384 struct mem_cgroup *mem)
David Rientjes1b604d72009-12-14 17:57:47 -0800385{
David Rientjes5e9d8342010-08-09 17:18:51 -0700386 task_lock(current);
David Rientjes1b604d72009-12-14 17:57:47 -0800387 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
David Rientjesa63d83f2010-08-09 17:19:46 -0700388 "oom_adj=%d, oom_score_adj=%d\n",
389 current->comm, gfp_mask, order, current->signal->oom_adj,
390 current->signal->oom_score_adj);
David Rientjes1b604d72009-12-14 17:57:47 -0800391 cpuset_print_task_mems_allowed(current);
392 task_unlock(current);
393 dump_stack();
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800394 mem_cgroup_print_oom_info(mem, p);
David Rientjes1b604d72009-12-14 17:57:47 -0800395 show_mem();
396 if (sysctl_oom_dump_tasks)
397 dump_tasks(mem);
398}
399
KOSAKI Motohiro3b4798c2009-12-15 16:45:32 -0800400#define K(x) ((x) << (PAGE_SHIFT-10))
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700401static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402{
David Rientjesb940fd72010-08-09 17:18:57 -0700403 p = find_lock_task_mm(p);
KOSAKI Motohiroa96cfd62010-08-09 17:19:38 -0700404 if (!p) {
David Rientjesb940fd72010-08-09 17:18:57 -0700405 task_unlock(p);
Dave Peterson01315922006-04-18 22:20:44 -0700406 return 1;
David Rientjesb940fd72010-08-09 17:18:57 -0700407 }
408 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
409 task_pid_nr(p), p->comm, K(p->mm->total_vm),
410 K(get_mm_counter(p->mm, MM_ANONPAGES)),
411 K(get_mm_counter(p->mm, MM_FILEPAGES)));
412 task_unlock(p);
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700413
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700414
David Rientjesb940fd72010-08-09 17:18:57 -0700415 set_tsk_thread_flag(p, TIF_MEMDIE);
416 force_sig(SIGKILL, p);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700417
418 /*
419 * We give our sacrificial lamb high priority and access to
420 * all the memory it needs. That way it should be able to
421 * exit() and clear out its resources quickly...
422 */
423 boost_dying_task_prio(p, mem);
424
Dave Peterson01315922006-04-18 22:20:44 -0700425 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426}
David Rientjesb940fd72010-08-09 17:18:57 -0700427#undef K
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
David Rientjes7213f502007-10-16 23:25:57 -0700429static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
David Rientjesa63d83f2010-08-09 17:19:46 -0700430 unsigned int points, unsigned long totalpages,
431 struct mem_cgroup *mem, nodemask_t *nodemask,
432 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
David Rientjes5e9d8342010-08-09 17:18:51 -0700434 struct task_struct *victim = p;
435 struct task_struct *child;
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700436 struct task_struct *t = p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700437 unsigned int victim_points = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
David Rientjes1b604d72009-12-14 17:57:47 -0800439 if (printk_ratelimit())
Daisuke Nishimurad31f56d2009-12-15 16:47:12 -0800440 dump_header(p, gfp_mask, order, mem);
David Rientjes7213f502007-10-16 23:25:57 -0700441
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700442 /*
443 * If the task is already exiting, don't alarm the sysadmin or kill
444 * its children or threads, just set TIF_MEMDIE so it can die quickly
445 */
KOSAKI Motohiro0753ba02009-08-18 14:11:10 -0700446 if (p->flags & PF_EXITING) {
David Rientjes43589972010-08-09 17:18:49 -0700447 set_tsk_thread_flag(p, TIF_MEMDIE);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700448 boost_dying_task_prio(p, mem);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700449 return 0;
450 }
451
David Rientjes5e9d8342010-08-09 17:18:51 -0700452 task_lock(p);
David Rientjesa63d83f2010-08-09 17:19:46 -0700453 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
David Rientjes5e9d8342010-08-09 17:18:51 -0700454 message, task_pid_nr(p), p->comm, points);
455 task_unlock(p);
Nick Pigginf3af38d2006-12-06 20:31:51 -0800456
David Rientjes5e9d8342010-08-09 17:18:51 -0700457 /*
458 * If any of p's children has a different mm and is eligible for kill,
459 * the one with the highest badness() score is sacrificed for its
460 * parent. This attempts to lose the minimal amount of work done while
461 * still freeing memory.
462 */
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700463 do {
David Rientjes5e9d8342010-08-09 17:18:51 -0700464 list_for_each_entry(child, &t->children, sibling) {
David Rientjesa63d83f2010-08-09 17:19:46 -0700465 unsigned int child_points;
David Rientjes5e9d8342010-08-09 17:18:51 -0700466
David Rientjesa63d83f2010-08-09 17:19:46 -0700467 /*
468 * oom_badness() returns 0 if the thread is unkillable
469 */
470 child_points = oom_badness(child, mem, nodemask,
471 totalpages);
David Rientjes5e9d8342010-08-09 17:18:51 -0700472 if (child_points > victim_points) {
473 victim = child;
474 victim_points = child_points;
475 }
Oleg Nesterovdd8e8f42010-08-09 17:18:45 -0700476 }
477 } while_each_thread(p, t);
478
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700479 return oom_kill_task(victim, mem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480}
481
David Rientjes309ed882010-08-09 17:18:54 -0700482/*
483 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
484 */
485static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
486 int order)
487{
488 if (likely(!sysctl_panic_on_oom))
489 return;
490 if (sysctl_panic_on_oom != 2) {
491 /*
492 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
493 * does not panic for cpuset, mempolicy, or memcg allocation
494 * failures.
495 */
496 if (constraint != CONSTRAINT_NONE)
497 return;
498 }
499 read_lock(&tasklist_lock);
500 dump_header(NULL, gfp_mask, order, NULL);
501 read_unlock(&tasklist_lock);
502 panic("Out of memory: %s panic_on_oom is enabled\n",
503 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
504}
505
Balbir Singh00f0b822008-03-04 14:28:39 -0800506#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800507void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
508{
David Rientjesa63d83f2010-08-09 17:19:46 -0700509 unsigned long limit;
510 unsigned int points = 0;
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800511 struct task_struct *p;
512
David Rientjes309ed882010-08-09 17:18:54 -0700513 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0);
David Rientjesa63d83f2010-08-09 17:19:46 -0700514 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
Li Zefane115f2d2008-04-15 14:34:37 -0700515 read_lock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800516retry:
David Rientjesa63d83f2010-08-09 17:19:46 -0700517 p = select_bad_process(&points, limit, mem, NULL);
David Rientjesdf64f812010-05-26 14:42:41 -0700518 if (!p || PTR_ERR(p) == -1UL)
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800519 goto out;
520
David Rientjesa63d83f2010-08-09 17:19:46 -0700521 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800522 "Memory cgroup out of memory"))
523 goto retry;
524out:
Li Zefane115f2d2008-04-15 14:34:37 -0700525 read_unlock(&tasklist_lock);
Pavel Emelianovc7ba5c92008-02-07 00:13:58 -0800526}
527#endif
528
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700529static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
530
531int register_oom_notifier(struct notifier_block *nb)
532{
533 return blocking_notifier_chain_register(&oom_notify_list, nb);
534}
535EXPORT_SYMBOL_GPL(register_oom_notifier);
536
537int unregister_oom_notifier(struct notifier_block *nb)
538{
539 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
540}
541EXPORT_SYMBOL_GPL(unregister_oom_notifier);
542
David Rientjes098d7f12007-10-16 23:25:55 -0700543/*
544 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
545 * if a parallel OOM killing is already taking place that includes a zone in
546 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
547 */
Minchan Kimff321fe2010-08-09 17:18:57 -0700548int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700549{
Mel Gormandd1a2392008-04-28 02:12:17 -0700550 struct zoneref *z;
551 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700552 int ret = 1;
553
David Rientjesc7d4cae2009-01-06 14:39:00 -0800554 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700555 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
556 if (zone_is_oom_locked(zone)) {
David Rientjes098d7f12007-10-16 23:25:55 -0700557 ret = 0;
558 goto out;
559 }
Mel Gormandd1a2392008-04-28 02:12:17 -0700560 }
David Rientjes098d7f12007-10-16 23:25:55 -0700561
Mel Gormandd1a2392008-04-28 02:12:17 -0700562 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
563 /*
David Rientjesc7d4cae2009-01-06 14:39:00 -0800564 * Lock each zone in the zonelist under zone_scan_lock so a
Minchan Kimff321fe2010-08-09 17:18:57 -0700565 * parallel invocation of try_set_zonelist_oom() doesn't succeed
Mel Gormandd1a2392008-04-28 02:12:17 -0700566 * when it shouldn't.
567 */
568 zone_set_flag(zone, ZONE_OOM_LOCKED);
569 }
570
David Rientjes098d7f12007-10-16 23:25:55 -0700571out:
David Rientjesc7d4cae2009-01-06 14:39:00 -0800572 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700573 return ret;
574}
575
576/*
577 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
578 * allocation attempts with zonelists containing them may now recall the OOM
579 * killer, if necessary.
580 */
Mel Gormandd1a2392008-04-28 02:12:17 -0700581void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
David Rientjes098d7f12007-10-16 23:25:55 -0700582{
Mel Gormandd1a2392008-04-28 02:12:17 -0700583 struct zoneref *z;
584 struct zone *zone;
David Rientjes098d7f12007-10-16 23:25:55 -0700585
David Rientjesc7d4cae2009-01-06 14:39:00 -0800586 spin_lock(&zone_scan_lock);
Mel Gormandd1a2392008-04-28 02:12:17 -0700587 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
588 zone_clear_flag(zone, ZONE_OOM_LOCKED);
589 }
David Rientjesc7d4cae2009-01-06 14:39:00 -0800590 spin_unlock(&zone_scan_lock);
David Rientjes098d7f12007-10-16 23:25:55 -0700591}
592
Nick Piggin1c0fe6e2009-01-06 14:38:59 -0800593/*
David Rientjese3658932010-08-09 17:18:55 -0700594 * Try to acquire the oom killer lock for all system zones. Returns zero if a
595 * parallel oom killing is taking place, otherwise locks all zones and returns
596 * non-zero.
597 */
598static int try_set_system_oom(void)
599{
600 struct zone *zone;
601 int ret = 1;
602
603 spin_lock(&zone_scan_lock);
604 for_each_populated_zone(zone)
605 if (zone_is_oom_locked(zone)) {
606 ret = 0;
607 goto out;
608 }
609 for_each_populated_zone(zone)
610 zone_set_flag(zone, ZONE_OOM_LOCKED);
611out:
612 spin_unlock(&zone_scan_lock);
613 return ret;
614}
615
616/*
617 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
618 * attempts or page faults may now recall the oom killer, if necessary.
619 */
620static void clear_system_oom(void)
621{
622 struct zone *zone;
623
624 spin_lock(&zone_scan_lock);
625 for_each_populated_zone(zone)
626 zone_clear_flag(zone, ZONE_OOM_LOCKED);
627 spin_unlock(&zone_scan_lock);
628}
629
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630/**
Dave Peterson6937a252006-06-23 02:03:13 -0700631 * out_of_memory - kill the "best" process when we run out of memory
Randy Dunlap1b578df2008-03-19 17:00:42 -0700632 * @zonelist: zonelist pointer
633 * @gfp_mask: memory allocation flags
634 * @order: amount of memory being requested as a power of 2
David Rientjes6f48d0eb2010-08-09 17:18:52 -0700635 * @nodemask: nodemask passed to page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 *
637 * If we run out of memory, we have the choice between either
638 * killing a random task (bad), letting the system crash (worse)
639 * OR try to be smart about which process to kill. Note that we
640 * don't have to be perfect here, we just have to be good.
641 */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -0800642void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
643 int order, nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644{
David Rientjes0aad4b32010-08-09 17:18:59 -0700645 struct task_struct *p;
David Rientjesa63d83f2010-08-09 17:19:46 -0700646 unsigned long totalpages;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700647 unsigned long freed = 0;
David Rientjesa63d83f2010-08-09 17:19:46 -0700648 unsigned int points;
David Rientjese3658932010-08-09 17:18:55 -0700649 enum oom_constraint constraint = CONSTRAINT_NONE;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700650
651 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
652 if (freed > 0)
653 /* Got some memory back in the last second. */
654 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655
David Rientjes7b98c2e2010-08-09 17:18:48 -0700656 /*
657 * If current has a pending SIGKILL, then automatically select it. The
658 * goal is to allow it to allocate so that it may quickly exit and free
659 * its memory.
660 */
661 if (fatal_signal_pending(current)) {
662 set_thread_flag(TIF_MEMDIE);
Luis Claudio R. Goncalves93b43fa2010-08-09 17:19:41 -0700663 boost_dying_task_prio(current, NULL);
David Rientjes7b98c2e2010-08-09 17:18:48 -0700664 return;
665 }
666
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800667 /*
668 * Check if there were limitations on the allocation (only relevant for
669 * NUMA) that may require different handling.
670 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700671 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
672 &totalpages);
David Rientjes309ed882010-08-09 17:18:54 -0700673 check_panic_on_oom(constraint, gfp_mask, order);
David Rientjes0aad4b32010-08-09 17:18:59 -0700674
David Rientjes2b45ab32007-05-06 14:50:00 -0700675 read_lock(&tasklist_lock);
KOSAKI Motohirof88ccad2010-08-09 17:19:36 -0700676 if (sysctl_oom_kill_allocating_task &&
KOSAKI Motohiroa96cfd62010-08-09 17:19:38 -0700677 !oom_unkillable_task(current, NULL, nodemask) &&
678 (current->signal->oom_adj != OOM_DISABLE)) {
David Rientjes0aad4b32010-08-09 17:18:59 -0700679 /*
680 * oom_kill_process() needs tasklist_lock held. If it returns
681 * non-zero, current could not be killed so we must fallback to
682 * the tasklist scan.
683 */
David Rientjesa63d83f2010-08-09 17:19:46 -0700684 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
685 NULL, nodemask,
David Rientjes0aad4b32010-08-09 17:18:59 -0700686 "Out of memory (oom_kill_allocating_task)"))
687 return;
688 }
689
690retry:
David Rientjesa63d83f2010-08-09 17:19:46 -0700691 p = select_bad_process(&points, totalpages, NULL,
David Rientjesf4420032010-08-09 17:18:58 -0700692 constraint == CONSTRAINT_MEMORY_POLICY ? nodemask :
693 NULL);
David Rientjes0aad4b32010-08-09 17:18:59 -0700694 if (PTR_ERR(p) == -1UL)
695 return;
696
697 /* Found nothing?!?! Either we hang forever, or we panic. */
698 if (!p) {
699 dump_header(NULL, gfp_mask, order, NULL);
700 read_unlock(&tasklist_lock);
701 panic("Out of memory and no killable processes...\n");
702 }
703
David Rientjesa63d83f2010-08-09 17:19:46 -0700704 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
705 nodemask, "Out of memory"))
David Rientjes0aad4b32010-08-09 17:18:59 -0700706 goto retry;
Andrew Morton140ffce2006-03-02 02:54:28 -0800707 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
709 /*
710 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800711 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800713 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800714 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715}
David Rientjese3658932010-08-09 17:18:55 -0700716
717/*
718 * The pagefault handler calls here because it is out of memory, so kill a
719 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
720 * oom killing is already in progress so do nothing. If a task is found with
721 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
722 */
723void pagefault_out_of_memory(void)
724{
725 if (try_set_system_oom()) {
726 out_of_memory(NULL, 0, 0, NULL);
727 clear_system_oom();
728 }
729 if (!test_thread_flag(TIF_MEMDIE))
730 schedule_timeout_uninterruptible(1);
731}