Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
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 | |
a49335cc PJ |
9 | * we're seriously out of memory. This gets called from __alloc_pages() |
10 | * in mm/page_alloc.c when we really run out of memory. | |
1da177e4 LT |
11 | * |
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 | ||
18 | #include <linux/mm.h> | |
19 | #include <linux/sched.h> | |
20 | #include <linux/swap.h> | |
21 | #include <linux/timex.h> | |
22 | #include <linux/jiffies.h> | |
ef08e3b4 | 23 | #include <linux/cpuset.h> |
8bc719d3 MS |
24 | #include <linux/module.h> |
25 | #include <linux/notifier.h> | |
1da177e4 | 26 | |
fadd8fbd | 27 | int sysctl_panic_on_oom; |
1da177e4 LT |
28 | /* #define DEBUG */ |
29 | ||
30 | /** | |
6937a25c | 31 | * badness - calculate a numeric value for how bad this task has been |
1da177e4 | 32 | * @p: task struct of which task we should calculate |
a49335cc | 33 | * @uptime: current uptime in seconds |
1da177e4 LT |
34 | * |
35 | * The formula used is relatively simple and documented inline in the | |
36 | * function. The main rationale is that we want to select a good task | |
37 | * to kill when we run out of memory. | |
38 | * | |
39 | * Good in this context means that: | |
40 | * 1) we lose the minimum amount of work done | |
41 | * 2) we recover a large amount of memory | |
42 | * 3) we don't kill anything innocent of eating tons of memory | |
43 | * 4) we want to kill the minimum amount of processes (one) | |
44 | * 5) we try to kill the process the user expects us to kill, this | |
45 | * algorithm has been meticulously tuned to meet the principle | |
46 | * of least surprise ... (be careful when you change it) | |
47 | */ | |
48 | ||
49 | unsigned long badness(struct task_struct *p, unsigned long uptime) | |
50 | { | |
51 | unsigned long points, cpu_time, run_time, s; | |
97c2c9b8 AM |
52 | struct mm_struct *mm; |
53 | struct task_struct *child; | |
1da177e4 | 54 | |
97c2c9b8 AM |
55 | task_lock(p); |
56 | mm = p->mm; | |
57 | if (!mm) { | |
58 | task_unlock(p); | |
1da177e4 | 59 | return 0; |
97c2c9b8 | 60 | } |
1da177e4 | 61 | |
af5b9124 NP |
62 | /* |
63 | * swapoff can easily use up all memory, so kill those first. | |
64 | */ | |
65 | if (p->flags & PF_SWAPOFF) | |
66 | return ULONG_MAX; | |
67 | ||
1da177e4 LT |
68 | /* |
69 | * The memory size of the process is the basis for the badness. | |
70 | */ | |
97c2c9b8 AM |
71 | points = mm->total_vm; |
72 | ||
73 | /* | |
74 | * After this unlock we can no longer dereference local variable `mm' | |
75 | */ | |
76 | task_unlock(p); | |
1da177e4 LT |
77 | |
78 | /* | |
79 | * Processes which fork a lot of child processes are likely | |
9827b781 | 80 | * a good choice. We add half the vmsize of the children if they |
1da177e4 | 81 | * have an own mm. This prevents forking servers to flood the |
9827b781 KG |
82 | * machine with an endless amount of children. In case a single |
83 | * child is eating the vast majority of memory, adding only half | |
84 | * to the parents will make the child our kill candidate of choice. | |
1da177e4 | 85 | */ |
97c2c9b8 AM |
86 | list_for_each_entry(child, &p->children, sibling) { |
87 | task_lock(child); | |
88 | if (child->mm != mm && child->mm) | |
89 | points += child->mm->total_vm/2 + 1; | |
90 | task_unlock(child); | |
1da177e4 LT |
91 | } |
92 | ||
93 | /* | |
94 | * CPU time is in tens of seconds and run time is in thousands | |
95 | * of seconds. There is no particular reason for this other than | |
96 | * that it turned out to work very well in practice. | |
97 | */ | |
98 | cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime)) | |
99 | >> (SHIFT_HZ + 3); | |
100 | ||
101 | if (uptime >= p->start_time.tv_sec) | |
102 | run_time = (uptime - p->start_time.tv_sec) >> 10; | |
103 | else | |
104 | run_time = 0; | |
105 | ||
106 | s = int_sqrt(cpu_time); | |
107 | if (s) | |
108 | points /= s; | |
109 | s = int_sqrt(int_sqrt(run_time)); | |
110 | if (s) | |
111 | points /= s; | |
112 | ||
113 | /* | |
114 | * Niced processes are most likely less important, so double | |
115 | * their badness points. | |
116 | */ | |
117 | if (task_nice(p) > 0) | |
118 | points *= 2; | |
119 | ||
120 | /* | |
121 | * Superuser processes are usually more important, so we make it | |
122 | * less likely that we kill those. | |
123 | */ | |
124 | if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) || | |
125 | p->uid == 0 || p->euid == 0) | |
126 | points /= 4; | |
127 | ||
128 | /* | |
129 | * We don't want to kill a process with direct hardware access. | |
130 | * Not only could that mess up the hardware, but usually users | |
131 | * tend to only have this flag set on applications they think | |
132 | * of as important. | |
133 | */ | |
134 | if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO)) | |
135 | points /= 4; | |
136 | ||
7887a3da NP |
137 | /* |
138 | * If p's nodes don't overlap ours, it may still help to kill p | |
139 | * because p may have allocated or otherwise mapped memory on | |
140 | * this node before. However it will be less likely. | |
141 | */ | |
142 | if (!cpuset_excl_nodes_overlap(p)) | |
143 | points /= 8; | |
144 | ||
1da177e4 LT |
145 | /* |
146 | * Adjust the score by oomkilladj. | |
147 | */ | |
148 | if (p->oomkilladj) { | |
149 | if (p->oomkilladj > 0) | |
150 | points <<= p->oomkilladj; | |
151 | else | |
152 | points >>= -(p->oomkilladj); | |
153 | } | |
154 | ||
155 | #ifdef DEBUG | |
156 | printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n", | |
157 | p->pid, p->comm, points); | |
158 | #endif | |
159 | return points; | |
160 | } | |
161 | ||
9b0f8b04 CL |
162 | /* |
163 | * Types of limitations to the nodes from which allocations may occur | |
164 | */ | |
165 | #define CONSTRAINT_NONE 1 | |
166 | #define CONSTRAINT_MEMORY_POLICY 2 | |
167 | #define CONSTRAINT_CPUSET 3 | |
168 | ||
169 | /* | |
170 | * Determine the type of allocation constraint. | |
171 | */ | |
172 | static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask) | |
173 | { | |
174 | #ifdef CONFIG_NUMA | |
175 | struct zone **z; | |
176 | nodemask_t nodes = node_online_map; | |
177 | ||
178 | for (z = zonelist->zones; *z; z++) | |
179 | if (cpuset_zone_allowed(*z, gfp_mask)) | |
89fa3024 | 180 | node_clear(zone_to_nid(*z), nodes); |
9b0f8b04 CL |
181 | else |
182 | return CONSTRAINT_CPUSET; | |
183 | ||
184 | if (!nodes_empty(nodes)) | |
185 | return CONSTRAINT_MEMORY_POLICY; | |
186 | #endif | |
187 | ||
188 | return CONSTRAINT_NONE; | |
189 | } | |
190 | ||
1da177e4 LT |
191 | /* |
192 | * Simple selection loop. We chose the process with the highest | |
193 | * number of 'points'. We expect the caller will lock the tasklist. | |
194 | * | |
195 | * (not docbooked, we don't want this one cluttering up the manual) | |
196 | */ | |
9827b781 | 197 | static struct task_struct *select_bad_process(unsigned long *ppoints) |
1da177e4 | 198 | { |
1da177e4 LT |
199 | struct task_struct *g, *p; |
200 | struct task_struct *chosen = NULL; | |
201 | struct timespec uptime; | |
9827b781 | 202 | *ppoints = 0; |
1da177e4 LT |
203 | |
204 | do_posix_clock_monotonic_gettime(&uptime); | |
a49335cc PJ |
205 | do_each_thread(g, p) { |
206 | unsigned long points; | |
207 | int releasing; | |
208 | ||
5081dde3 NP |
209 | /* skip kernel threads */ |
210 | if (!p->mm) | |
211 | continue; | |
1da177e4 | 212 | /* skip the init task with pid == 1 */ |
a49335cc PJ |
213 | if (p->pid == 1) |
214 | continue; | |
ef08e3b4 | 215 | |
a49335cc | 216 | /* |
6937a25c | 217 | * This is in the process of releasing memory so wait for it |
a49335cc | 218 | * to finish before killing some other task by mistake. |
50ec3bbf NP |
219 | * |
220 | * However, if p is the current task, we allow the 'kill' to | |
221 | * go ahead if it is exiting: this will simply set TIF_MEMDIE, | |
222 | * which will allow it to gain access to memory reserves in | |
223 | * the process of exiting and releasing its resources. | |
224 | * Otherwise we could get an OOM deadlock. | |
a49335cc PJ |
225 | */ |
226 | releasing = test_tsk_thread_flag(p, TIF_MEMDIE) || | |
227 | p->flags & PF_EXITING; | |
50ec3bbf NP |
228 | if (releasing) { |
229 | /* PF_DEAD tasks have already released their mm */ | |
230 | if (p->flags & PF_DEAD) | |
231 | continue; | |
232 | if (p->flags & PF_EXITING && p == current) { | |
233 | chosen = p; | |
234 | *ppoints = ULONG_MAX; | |
235 | break; | |
236 | } | |
a49335cc | 237 | return ERR_PTR(-1UL); |
50ec3bbf | 238 | } |
4a3ede10 NP |
239 | if (p->oomkilladj == OOM_DISABLE) |
240 | continue; | |
a49335cc PJ |
241 | |
242 | points = badness(p, uptime.tv_sec); | |
9827b781 | 243 | if (points > *ppoints || !chosen) { |
a49335cc | 244 | chosen = p; |
9827b781 | 245 | *ppoints = points; |
1da177e4 | 246 | } |
a49335cc | 247 | } while_each_thread(g, p); |
1da177e4 LT |
248 | return chosen; |
249 | } | |
250 | ||
251 | /** | |
5a291b98 RG |
252 | * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO |
253 | * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO | |
254 | * set. | |
1da177e4 | 255 | */ |
36c8b586 | 256 | static void __oom_kill_task(struct task_struct *p, const char *message) |
1da177e4 LT |
257 | { |
258 | if (p->pid == 1) { | |
259 | WARN_ON(1); | |
260 | printk(KERN_WARNING "tried to kill init!\n"); | |
261 | return; | |
262 | } | |
263 | ||
264 | task_lock(p); | |
265 | if (!p->mm || p->mm == &init_mm) { | |
266 | WARN_ON(1); | |
267 | printk(KERN_WARNING "tried to kill an mm-less task!\n"); | |
268 | task_unlock(p); | |
269 | return; | |
270 | } | |
271 | task_unlock(p); | |
50ec3bbf NP |
272 | |
273 | if (message) { | |
274 | printk(KERN_ERR "%s: Killed process %d (%s).\n", | |
9b0f8b04 | 275 | message, p->pid, p->comm); |
50ec3bbf | 276 | } |
1da177e4 LT |
277 | |
278 | /* | |
279 | * We give our sacrificial lamb high priority and access to | |
280 | * all the memory it needs. That way it should be able to | |
281 | * exit() and clear out its resources quickly... | |
282 | */ | |
283 | p->time_slice = HZ; | |
284 | set_tsk_thread_flag(p, TIF_MEMDIE); | |
285 | ||
286 | force_sig(SIGKILL, p); | |
287 | } | |
288 | ||
36c8b586 | 289 | static int oom_kill_task(struct task_struct *p, const char *message) |
1da177e4 | 290 | { |
01315922 | 291 | struct mm_struct *mm; |
36c8b586 | 292 | struct task_struct *g, *q; |
1da177e4 | 293 | |
01315922 DP |
294 | mm = p->mm; |
295 | ||
296 | /* WARNING: mm may not be dereferenced since we did not obtain its | |
297 | * value from get_task_mm(p). This is OK since all we need to do is | |
298 | * compare mm to q->mm below. | |
299 | * | |
300 | * Furthermore, even if mm contains a non-NULL value, p->mm may | |
301 | * change to NULL at any time since we do not hold task_lock(p). | |
302 | * However, this is of no concern to us. | |
303 | */ | |
304 | ||
305 | if (mm == NULL || mm == &init_mm) | |
306 | return 1; | |
1da177e4 | 307 | |
9b0f8b04 | 308 | __oom_kill_task(p, message); |
1da177e4 LT |
309 | /* |
310 | * kill all processes that share the ->mm (i.e. all threads), | |
311 | * but are in a different thread group | |
312 | */ | |
313 | do_each_thread(g, q) | |
314 | if (q->mm == mm && q->tgid != p->tgid) | |
9b0f8b04 | 315 | __oom_kill_task(q, message); |
1da177e4 LT |
316 | while_each_thread(g, q); |
317 | ||
01315922 | 318 | return 0; |
1da177e4 LT |
319 | } |
320 | ||
01315922 DP |
321 | static int oom_kill_process(struct task_struct *p, unsigned long points, |
322 | const char *message) | |
1da177e4 | 323 | { |
1da177e4 LT |
324 | struct task_struct *c; |
325 | struct list_head *tsk; | |
326 | ||
50ec3bbf NP |
327 | /* |
328 | * If the task is already exiting, don't alarm the sysadmin or kill | |
329 | * its children or threads, just set TIF_MEMDIE so it can die quickly | |
330 | */ | |
331 | if (p->flags & PF_EXITING) { | |
332 | __oom_kill_task(p, NULL); | |
333 | return 0; | |
334 | } | |
335 | ||
336 | printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li" | |
337 | " and children.\n", p->pid, p->comm, points); | |
1da177e4 LT |
338 | /* Try to kill a child first */ |
339 | list_for_each(tsk, &p->children) { | |
340 | c = list_entry(tsk, struct task_struct, sibling); | |
341 | if (c->mm == p->mm) | |
342 | continue; | |
01315922 DP |
343 | if (!oom_kill_task(c, message)) |
344 | return 0; | |
1da177e4 | 345 | } |
9b0f8b04 | 346 | return oom_kill_task(p, message); |
1da177e4 LT |
347 | } |
348 | ||
8bc719d3 MS |
349 | static BLOCKING_NOTIFIER_HEAD(oom_notify_list); |
350 | ||
351 | int register_oom_notifier(struct notifier_block *nb) | |
352 | { | |
353 | return blocking_notifier_chain_register(&oom_notify_list, nb); | |
354 | } | |
355 | EXPORT_SYMBOL_GPL(register_oom_notifier); | |
356 | ||
357 | int unregister_oom_notifier(struct notifier_block *nb) | |
358 | { | |
359 | return blocking_notifier_chain_unregister(&oom_notify_list, nb); | |
360 | } | |
361 | EXPORT_SYMBOL_GPL(unregister_oom_notifier); | |
362 | ||
1da177e4 | 363 | /** |
6937a25c | 364 | * out_of_memory - kill the "best" process when we run out of memory |
1da177e4 LT |
365 | * |
366 | * If we run out of memory, we have the choice between either | |
367 | * killing a random task (bad), letting the system crash (worse) | |
368 | * OR try to be smart about which process to kill. Note that we | |
369 | * don't have to be perfect here, we just have to be good. | |
370 | */ | |
9b0f8b04 | 371 | void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order) |
1da177e4 | 372 | { |
36c8b586 | 373 | struct task_struct *p; |
d6713e04 | 374 | unsigned long points = 0; |
8bc719d3 MS |
375 | unsigned long freed = 0; |
376 | ||
377 | blocking_notifier_call_chain(&oom_notify_list, 0, &freed); | |
378 | if (freed > 0) | |
379 | /* Got some memory back in the last second. */ | |
380 | return; | |
1da177e4 | 381 | |
42639269 | 382 | if (printk_ratelimit()) { |
b72f1604 NP |
383 | printk(KERN_WARNING "%s invoked oom-killer: " |
384 | "gfp_mask=0x%x, order=%d, oomkilladj=%d\n", | |
385 | current->comm, gfp_mask, order, current->oomkilladj); | |
b958f7d9 | 386 | dump_stack(); |
42639269 AB |
387 | show_mem(); |
388 | } | |
578c2fd6 | 389 | |
505970b9 | 390 | cpuset_lock(); |
1da177e4 | 391 | read_lock(&tasklist_lock); |
9b0f8b04 CL |
392 | |
393 | /* | |
394 | * Check if there were limitations on the allocation (only relevant for | |
395 | * NUMA) that may require different handling. | |
396 | */ | |
397 | switch (constrained_alloc(zonelist, gfp_mask)) { | |
398 | case CONSTRAINT_MEMORY_POLICY: | |
01315922 | 399 | oom_kill_process(current, points, |
9b0f8b04 CL |
400 | "No available memory (MPOL_BIND)"); |
401 | break; | |
402 | ||
403 | case CONSTRAINT_CPUSET: | |
01315922 | 404 | oom_kill_process(current, points, |
9b0f8b04 CL |
405 | "No available memory in cpuset"); |
406 | break; | |
407 | ||
408 | case CONSTRAINT_NONE: | |
fadd8fbd KH |
409 | if (sysctl_panic_on_oom) |
410 | panic("out of memory. panic_on_oom is selected\n"); | |
1da177e4 | 411 | retry: |
9b0f8b04 CL |
412 | /* |
413 | * Rambo mode: Shoot down a process and hope it solves whatever | |
414 | * issues we may have. | |
415 | */ | |
416 | p = select_bad_process(&points); | |
1da177e4 | 417 | |
9b0f8b04 CL |
418 | if (PTR_ERR(p) == -1UL) |
419 | goto out; | |
1da177e4 | 420 | |
9b0f8b04 CL |
421 | /* Found nothing?!?! Either we hang forever, or we panic. */ |
422 | if (!p) { | |
423 | read_unlock(&tasklist_lock); | |
424 | cpuset_unlock(); | |
425 | panic("Out of memory and no killable processes...\n"); | |
426 | } | |
1da177e4 | 427 | |
01315922 | 428 | if (oom_kill_process(p, points, "Out of memory")) |
9b0f8b04 CL |
429 | goto retry; |
430 | ||
431 | break; | |
432 | } | |
1da177e4 | 433 | |
9b0f8b04 | 434 | out: |
140ffcec | 435 | read_unlock(&tasklist_lock); |
505970b9 | 436 | cpuset_unlock(); |
1da177e4 LT |
437 | |
438 | /* | |
439 | * Give "p" a good chance of killing itself before we | |
2f659f46 | 440 | * retry to allocate memory unless "p" is current |
1da177e4 | 441 | */ |
2f659f46 | 442 | if (!test_thread_flag(TIF_MEMDIE)) |
140ffcec | 443 | schedule_timeout_uninterruptible(1); |
1da177e4 | 444 | } |