blob: 2e3ce3a928b97dd8eeb54b42fb6f363d15a6d2ae [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...
7 *
8 * The routines in this file are used to kill a process when
Paul Jacksona49335c2005-09-06 15:18:09 -07009 * we're seriously out of memory. This gets called from __alloc_pages()
10 * in mm/page_alloc.c when we really run out of memory.
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 *
12 * Since we won't call these routines often (on a well-configured
13 * machine) this file will double as a 'coding guide' and a signpost
14 * for newbie kernel hackers. It features several pointers to major
15 * kernel subsystems and hints as to where to find out what things do.
16 */
17
Alexey Dobriyan8ac773b2006-10-19 23:28:32 -070018#include <linux/oom.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/mm.h>
20#include <linux/sched.h>
21#include <linux/swap.h>
22#include <linux/timex.h>
23#include <linux/jiffies.h>
Paul Jacksonef08e3b2005-09-06 15:18:13 -070024#include <linux/cpuset.h>
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -070025#include <linux/module.h>
26#include <linux/notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070027
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -070028int sysctl_panic_on_oom;
Linus Torvalds1da177e2005-04-16 15:20:36 -070029/* #define DEBUG */
30
31/**
Dave Peterson6937a252006-06-23 02:03:13 -070032 * badness - calculate a numeric value for how bad this task has been
Linus Torvalds1da177e2005-04-16 15:20:36 -070033 * @p: task struct of which task we should calculate
Paul Jacksona49335c2005-09-06 15:18:09 -070034 * @uptime: current uptime in seconds
Linus Torvalds1da177e2005-04-16 15:20:36 -070035 *
36 * The formula used is relatively simple and documented inline in the
37 * function. The main rationale is that we want to select a good task
38 * to kill when we run out of memory.
39 *
40 * Good in this context means that:
41 * 1) we lose the minimum amount of work done
42 * 2) we recover a large amount of memory
43 * 3) we don't kill anything innocent of eating tons of memory
44 * 4) we want to kill the minimum amount of processes (one)
45 * 5) we try to kill the process the user expects us to kill, this
46 * algorithm has been meticulously tuned to meet the principle
47 * of least surprise ... (be careful when you change it)
48 */
49
50unsigned long badness(struct task_struct *p, unsigned long uptime)
51{
52 unsigned long points, cpu_time, run_time, s;
Andrew Morton97c2c9b2006-04-18 22:20:38 -070053 struct mm_struct *mm;
54 struct task_struct *child;
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Andrew Morton97c2c9b2006-04-18 22:20:38 -070056 task_lock(p);
57 mm = p->mm;
58 if (!mm) {
59 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 return 0;
Andrew Morton97c2c9b2006-04-18 22:20:38 -070061 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63 /*
Nick Pigginaf5b9122006-09-25 23:31:31 -070064 * swapoff can easily use up all memory, so kill those first.
65 */
66 if (p->flags & PF_SWAPOFF)
67 return ULONG_MAX;
68
69 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 * The memory size of the process is the basis for the badness.
71 */
Andrew Morton97c2c9b2006-04-18 22:20:38 -070072 points = mm->total_vm;
73
74 /*
75 * After this unlock we can no longer dereference local variable `mm'
76 */
77 task_unlock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
79 /*
80 * Processes which fork a lot of child processes are likely
Kurt Garloff9827b782006-02-20 18:27:51 -080081 * a good choice. We add half the vmsize of the children if they
Linus Torvalds1da177e2005-04-16 15:20:36 -070082 * have an own mm. This prevents forking servers to flood the
Kurt Garloff9827b782006-02-20 18:27:51 -080083 * machine with an endless amount of children. In case a single
84 * child is eating the vast majority of memory, adding only half
85 * to the parents will make the child our kill candidate of choice.
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 */
Andrew Morton97c2c9b2006-04-18 22:20:38 -070087 list_for_each_entry(child, &p->children, sibling) {
88 task_lock(child);
89 if (child->mm != mm && child->mm)
90 points += child->mm->total_vm/2 + 1;
91 task_unlock(child);
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 }
93
94 /*
95 * CPU time is in tens of seconds and run time is in thousands
96 * of seconds. There is no particular reason for this other than
97 * that it turned out to work very well in practice.
98 */
99 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
100 >> (SHIFT_HZ + 3);
101
102 if (uptime >= p->start_time.tv_sec)
103 run_time = (uptime - p->start_time.tv_sec) >> 10;
104 else
105 run_time = 0;
106
107 s = int_sqrt(cpu_time);
108 if (s)
109 points /= s;
110 s = int_sqrt(int_sqrt(run_time));
111 if (s)
112 points /= s;
113
114 /*
115 * Niced processes are most likely less important, so double
116 * their badness points.
117 */
118 if (task_nice(p) > 0)
119 points *= 2;
120
121 /*
122 * Superuser processes are usually more important, so we make it
123 * less likely that we kill those.
124 */
125 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
126 p->uid == 0 || p->euid == 0)
127 points /= 4;
128
129 /*
130 * We don't want to kill a process with direct hardware access.
131 * Not only could that mess up the hardware, but usually users
132 * tend to only have this flag set on applications they think
133 * of as important.
134 */
135 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
136 points /= 4;
137
138 /*
Nick Piggin7887a3d2006-09-25 23:31:29 -0700139 * If p's nodes don't overlap ours, it may still help to kill p
140 * because p may have allocated or otherwise mapped memory on
141 * this node before. However it will be less likely.
142 */
143 if (!cpuset_excl_nodes_overlap(p))
144 points /= 8;
145
146 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 * Adjust the score by oomkilladj.
148 */
149 if (p->oomkilladj) {
150 if (p->oomkilladj > 0)
151 points <<= p->oomkilladj;
152 else
153 points >>= -(p->oomkilladj);
154 }
155
156#ifdef DEBUG
157 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
158 p->pid, p->comm, points);
159#endif
160 return points;
161}
162
163/*
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800164 * Types of limitations to the nodes from which allocations may occur
165 */
166#define CONSTRAINT_NONE 1
167#define CONSTRAINT_MEMORY_POLICY 2
168#define CONSTRAINT_CPUSET 3
169
170/*
171 * Determine the type of allocation constraint.
172 */
173static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
174{
175#ifdef CONFIG_NUMA
176 struct zone **z;
177 nodemask_t nodes = node_online_map;
178
179 for (z = zonelist->zones; *z; z++)
180 if (cpuset_zone_allowed(*z, gfp_mask))
Christoph Lameter89fa3022006-09-25 23:31:55 -0700181 node_clear(zone_to_nid(*z), nodes);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800182 else
183 return CONSTRAINT_CPUSET;
184
185 if (!nodes_empty(nodes))
186 return CONSTRAINT_MEMORY_POLICY;
187#endif
188
189 return CONSTRAINT_NONE;
190}
191
192/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 * Simple selection loop. We chose the process with the highest
194 * number of 'points'. We expect the caller will lock the tasklist.
195 *
196 * (not docbooked, we don't want this one cluttering up the manual)
197 */
Kurt Garloff9827b782006-02-20 18:27:51 -0800198static struct task_struct *select_bad_process(unsigned long *ppoints)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 struct task_struct *g, *p;
201 struct task_struct *chosen = NULL;
202 struct timespec uptime;
Kurt Garloff9827b782006-02-20 18:27:51 -0800203 *ppoints = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
205 do_posix_clock_monotonic_gettime(&uptime);
Paul Jacksona49335c2005-09-06 15:18:09 -0700206 do_each_thread(g, p) {
207 unsigned long points;
Paul Jacksona49335c2005-09-06 15:18:09 -0700208
Oleg Nesterov28324d12006-09-29 02:01:12 -0700209 /*
210 * skip kernel threads and tasks which have already released
211 * their mm.
212 */
Nick Piggin5081dde2006-09-25 23:31:32 -0700213 if (!p->mm)
214 continue;
Oleg Nesterov28324d12006-09-29 02:01:12 -0700215 /* skip the init task */
216 if (is_init(p))
Paul Jacksona49335c2005-09-06 15:18:09 -0700217 continue;
Paul Jacksonef08e3b2005-09-06 15:18:13 -0700218
Paul Jacksona49335c2005-09-06 15:18:09 -0700219 /*
Nick Pigginb78483a2006-09-29 02:01:14 -0700220 * This task already has access to memory reserves and is
221 * being killed. Don't allow any other task access to the
222 * memory reserve.
223 *
224 * Note: this may have a chance of deadlock if it gets
225 * blocked waiting for another task which itself is waiting
226 * for memory. Is there a better alternative?
227 */
228 if (test_tsk_thread_flag(p, TIF_MEMDIE))
229 return ERR_PTR(-1UL);
230
231 /*
Dave Peterson6937a252006-06-23 02:03:13 -0700232 * This is in the process of releasing memory so wait for it
Paul Jacksona49335c2005-09-06 15:18:09 -0700233 * to finish before killing some other task by mistake.
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700234 *
235 * However, if p is the current task, we allow the 'kill' to
236 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
237 * which will allow it to gain access to memory reserves in
238 * the process of exiting and releasing its resources.
Nick Pigginb78483a2006-09-29 02:01:14 -0700239 * Otherwise we could get an easy OOM deadlock.
Paul Jacksona49335c2005-09-06 15:18:09 -0700240 */
Nick Pigginb78483a2006-09-29 02:01:14 -0700241 if (p->flags & PF_EXITING) {
242 if (p != current)
243 return ERR_PTR(-1UL);
244
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700245 chosen = p;
246 *ppoints = ULONG_MAX;
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700247 }
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700248
Nick Piggin4a3ede12006-09-25 23:31:30 -0700249 if (p->oomkilladj == OOM_DISABLE)
250 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Paul Jacksona49335c2005-09-06 15:18:09 -0700252 points = badness(p, uptime.tv_sec);
Kurt Garloff9827b782006-02-20 18:27:51 -0800253 if (points > *ppoints || !chosen) {
Paul Jacksona49335c2005-09-06 15:18:09 -0700254 chosen = p;
Kurt Garloff9827b782006-02-20 18:27:51 -0800255 *ppoints = points;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 }
Paul Jacksona49335c2005-09-06 15:18:09 -0700257 } while_each_thread(g, p);
Oleg Nesterov972c4ea2006-09-29 02:01:12 -0700258
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 return chosen;
260}
261
262/**
Ram Gupta5a291b982006-09-25 23:31:54 -0700263 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
264 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
265 * set.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700267static void __oom_kill_task(struct task_struct *p, const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268{
Sukadev Bhattiproluf400e192006-09-29 02:00:07 -0700269 if (is_init(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 WARN_ON(1);
271 printk(KERN_WARNING "tried to kill init!\n");
272 return;
273 }
274
Oleg Nesterov01017a22006-09-29 02:01:13 -0700275 if (!p->mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 WARN_ON(1);
277 printk(KERN_WARNING "tried to kill an mm-less task!\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 return;
279 }
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700280
281 if (message) {
282 printk(KERN_ERR "%s: Killed process %d (%s).\n",
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800283 message, p->pid, p->comm);
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700284 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
286 /*
287 * We give our sacrificial lamb high priority and access to
288 * all the memory it needs. That way it should be able to
289 * exit() and clear out its resources quickly...
290 */
291 p->time_slice = HZ;
292 set_tsk_thread_flag(p, TIF_MEMDIE);
293
294 force_sig(SIGKILL, p);
295}
296
Ingo Molnar36c8b582006-07-03 00:25:41 -0700297static int oom_kill_task(struct task_struct *p, const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298{
Dave Peterson01315922006-04-18 22:20:44 -0700299 struct mm_struct *mm;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700300 struct task_struct *g, *q;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301
Dave Peterson01315922006-04-18 22:20:44 -0700302 mm = p->mm;
303
304 /* WARNING: mm may not be dereferenced since we did not obtain its
305 * value from get_task_mm(p). This is OK since all we need to do is
306 * compare mm to q->mm below.
307 *
308 * Furthermore, even if mm contains a non-NULL value, p->mm may
309 * change to NULL at any time since we do not hold task_lock(p).
310 * However, this is of no concern to us.
311 */
312
Oleg Nesterov01017a22006-09-29 02:01:13 -0700313 if (mm == NULL)
Dave Peterson01315922006-04-18 22:20:44 -0700314 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800316 __oom_kill_task(p, message);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317 /*
318 * kill all processes that share the ->mm (i.e. all threads),
319 * but are in a different thread group
320 */
321 do_each_thread(g, q)
322 if (q->mm == mm && q->tgid != p->tgid)
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800323 __oom_kill_task(q, message);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 while_each_thread(g, q);
325
Dave Peterson01315922006-04-18 22:20:44 -0700326 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327}
328
Dave Peterson01315922006-04-18 22:20:44 -0700329static int oom_kill_process(struct task_struct *p, unsigned long points,
330 const char *message)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 struct task_struct *c;
333 struct list_head *tsk;
334
Nick Piggin50ec3bb2006-09-25 23:31:29 -0700335 /*
336 * If the task is already exiting, don't alarm the sysadmin or kill
337 * its children or threads, just set TIF_MEMDIE so it can die quickly
338 */
339 if (p->flags & PF_EXITING) {
340 __oom_kill_task(p, NULL);
341 return 0;
342 }
343
344 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li"
345 " and children.\n", p->pid, p->comm, points);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 /* Try to kill a child first */
347 list_for_each(tsk, &p->children) {
348 c = list_entry(tsk, struct task_struct, sibling);
349 if (c->mm == p->mm)
350 continue;
Dave Peterson01315922006-04-18 22:20:44 -0700351 if (!oom_kill_task(c, message))
352 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 }
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800354 return oom_kill_task(p, message);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355}
356
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700357static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
358
359int register_oom_notifier(struct notifier_block *nb)
360{
361 return blocking_notifier_chain_register(&oom_notify_list, nb);
362}
363EXPORT_SYMBOL_GPL(register_oom_notifier);
364
365int unregister_oom_notifier(struct notifier_block *nb)
366{
367 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
368}
369EXPORT_SYMBOL_GPL(unregister_oom_notifier);
370
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371/**
Dave Peterson6937a252006-06-23 02:03:13 -0700372 * out_of_memory - kill the "best" process when we run out of memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 *
374 * If we run out of memory, we have the choice between either
375 * killing a random task (bad), letting the system crash (worse)
376 * OR try to be smart about which process to kill. Note that we
377 * don't have to be perfect here, we just have to be good.
378 */
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800379void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
Ingo Molnar36c8b582006-07-03 00:25:41 -0700381 struct task_struct *p;
Andrew Mortond6713e02006-02-28 16:59:19 -0800382 unsigned long points = 0;
Martin Schwidefsky8bc719d2006-09-25 23:31:20 -0700383 unsigned long freed = 0;
384
385 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
386 if (freed > 0)
387 /* Got some memory back in the last second. */
388 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Anton Blanchard42639262005-07-07 17:56:06 -0700390 if (printk_ratelimit()) {
Nick Pigginb72f1602006-09-25 23:31:32 -0700391 printk(KERN_WARNING "%s invoked oom-killer: "
392 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
393 current->comm, gfp_mask, order, current->oomkilladj);
Andrew Mortonb958f7d2006-02-01 03:05:51 -0800394 dump_stack();
Anton Blanchard42639262005-07-07 17:56:06 -0700395 show_mem();
396 }
Janet Morgan578c2fd2005-06-21 17:14:56 -0700397
Paul Jackson505970b2006-01-14 13:21:06 -0800398 cpuset_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 read_lock(&tasklist_lock);
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800400
401 /*
402 * Check if there were limitations on the allocation (only relevant for
403 * NUMA) that may require different handling.
404 */
405 switch (constrained_alloc(zonelist, gfp_mask)) {
406 case CONSTRAINT_MEMORY_POLICY:
Dave Peterson01315922006-04-18 22:20:44 -0700407 oom_kill_process(current, points,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800408 "No available memory (MPOL_BIND)");
409 break;
410
411 case CONSTRAINT_CPUSET:
Dave Peterson01315922006-04-18 22:20:44 -0700412 oom_kill_process(current, points,
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800413 "No available memory in cpuset");
414 break;
415
416 case CONSTRAINT_NONE:
KAMEZAWA Hiroyukifadd8fb2006-06-23 02:03:13 -0700417 if (sysctl_panic_on_oom)
418 panic("out of memory. panic_on_oom is selected\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419retry:
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800420 /*
421 * Rambo mode: Shoot down a process and hope it solves whatever
422 * issues we may have.
423 */
424 p = select_bad_process(&points);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800426 if (PTR_ERR(p) == -1UL)
427 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800429 /* Found nothing?!?! Either we hang forever, or we panic. */
430 if (!p) {
431 read_unlock(&tasklist_lock);
432 cpuset_unlock();
433 panic("Out of memory and no killable processes...\n");
434 }
435
Dave Peterson01315922006-04-18 22:20:44 -0700436 if (oom_kill_process(p, points, "Out of memory"))
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800437 goto retry;
438
439 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 }
441
Christoph Lameter9b0f8b02006-02-20 18:27:52 -0800442out:
Andrew Morton140ffce2006-03-02 02:54:28 -0800443 read_unlock(&tasklist_lock);
Paul Jackson505970b2006-01-14 13:21:06 -0800444 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445
446 /*
447 * Give "p" a good chance of killing itself before we
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800448 * retry to allocate memory unless "p" is current
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 */
Kirill Korotaev2f659f42006-01-08 01:01:05 -0800450 if (!test_thread_flag(TIF_MEMDIE))
Andrew Morton140ffce2006-03-02 02:54:28 -0800451 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}