mempolicy: use MPOL_PREFERRED for system-wide default policy
[linux-2.6/mini2440.git] / mm / oom_kill.c
blobe41504aa5da9efa8d5155e37097f425bdf0b193c
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...
8 * The routines in this file are used to kill a process when
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.
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.
18 #include <linux/oom.h>
19 #include <linux/mm.h>
20 #include <linux/err.h>
21 #include <linux/sched.h>
22 #include <linux/swap.h>
23 #include <linux/timex.h>
24 #include <linux/jiffies.h>
25 #include <linux/cpuset.h>
26 #include <linux/module.h>
27 #include <linux/notifier.h>
28 #include <linux/memcontrol.h>
30 int sysctl_panic_on_oom;
31 int sysctl_oom_kill_allocating_task;
32 int sysctl_oom_dump_tasks;
33 static DEFINE_SPINLOCK(zone_scan_mutex);
34 /* #define DEBUG */
36 /**
37 * badness - calculate a numeric value for how bad this task has been
38 * @p: task struct of which task we should calculate
39 * @uptime: current uptime in seconds
40 * @mem: target memory controller
42 * The formula used is relatively simple and documented inline in the
43 * function. The main rationale is that we want to select a good task
44 * to kill when we run out of memory.
46 * Good in this context means that:
47 * 1) we lose the minimum amount of work done
48 * 2) we recover a large amount of memory
49 * 3) we don't kill anything innocent of eating tons of memory
50 * 4) we want to kill the minimum amount of processes (one)
51 * 5) we try to kill the process the user expects us to kill, this
52 * algorithm has been meticulously tuned to meet the principle
53 * of least surprise ... (be careful when you change it)
56 unsigned long badness(struct task_struct *p, unsigned long uptime,
57 struct mem_cgroup *mem)
59 unsigned long points, cpu_time, run_time, s;
60 struct mm_struct *mm;
61 struct task_struct *child;
63 task_lock(p);
64 mm = p->mm;
65 if (!mm) {
66 task_unlock(p);
67 return 0;
71 * The memory size of the process is the basis for the badness.
73 points = mm->total_vm;
76 * After this unlock we can no longer dereference local variable `mm'
78 task_unlock(p);
81 * swapoff can easily use up all memory, so kill those first.
83 if (p->flags & PF_SWAPOFF)
84 return ULONG_MAX;
87 * Processes which fork a lot of child processes are likely
88 * a good choice. We add half the vmsize of the children if they
89 * have an own mm. This prevents forking servers to flood the
90 * machine with an endless amount of children. In case a single
91 * child is eating the vast majority of memory, adding only half
92 * to the parents will make the child our kill candidate of choice.
94 list_for_each_entry(child, &p->children, sibling) {
95 task_lock(child);
96 if (child->mm != mm && child->mm)
97 points += child->mm->total_vm/2 + 1;
98 task_unlock(child);
102 * CPU time is in tens of seconds and run time is in thousands
103 * of seconds. There is no particular reason for this other than
104 * that it turned out to work very well in practice.
106 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
107 >> (SHIFT_HZ + 3);
109 if (uptime >= p->start_time.tv_sec)
110 run_time = (uptime - p->start_time.tv_sec) >> 10;
111 else
112 run_time = 0;
114 s = int_sqrt(cpu_time);
115 if (s)
116 points /= s;
117 s = int_sqrt(int_sqrt(run_time));
118 if (s)
119 points /= s;
122 * Niced processes are most likely less important, so double
123 * their badness points.
125 if (task_nice(p) > 0)
126 points *= 2;
129 * Superuser processes are usually more important, so we make it
130 * less likely that we kill those.
132 if (__capable(p, CAP_SYS_ADMIN) || __capable(p, CAP_SYS_RESOURCE))
133 points /= 4;
136 * We don't want to kill a process with direct hardware access.
137 * Not only could that mess up the hardware, but usually users
138 * tend to only have this flag set on applications they think
139 * of as important.
141 if (__capable(p, CAP_SYS_RAWIO))
142 points /= 4;
145 * If p's nodes don't overlap ours, it may still help to kill p
146 * because p may have allocated or otherwise mapped memory on
147 * this node before. However it will be less likely.
149 if (!cpuset_mems_allowed_intersects(current, p))
150 points /= 8;
153 * Adjust the score by oomkilladj.
155 if (p->oomkilladj) {
156 if (p->oomkilladj > 0) {
157 if (!points)
158 points = 1;
159 points <<= p->oomkilladj;
160 } else
161 points >>= -(p->oomkilladj);
164 #ifdef DEBUG
165 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
166 p->pid, p->comm, points);
167 #endif
168 return points;
172 * Determine the type of allocation constraint.
174 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
175 gfp_t gfp_mask)
177 #ifdef CONFIG_NUMA
178 struct zone *zone;
179 struct zoneref *z;
180 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
181 nodemask_t nodes = node_states[N_HIGH_MEMORY];
183 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
184 if (cpuset_zone_allowed_softwall(zone, gfp_mask))
185 node_clear(zone_to_nid(zone), nodes);
186 else
187 return CONSTRAINT_CPUSET;
189 if (!nodes_empty(nodes))
190 return CONSTRAINT_MEMORY_POLICY;
191 #endif
193 return CONSTRAINT_NONE;
197 * Simple selection loop. We chose the process with the highest
198 * number of 'points'. We expect the caller will lock the tasklist.
200 * (not docbooked, we don't want this one cluttering up the manual)
202 static struct task_struct *select_bad_process(unsigned long *ppoints,
203 struct mem_cgroup *mem)
205 struct task_struct *g, *p;
206 struct task_struct *chosen = NULL;
207 struct timespec uptime;
208 *ppoints = 0;
210 do_posix_clock_monotonic_gettime(&uptime);
211 do_each_thread(g, p) {
212 unsigned long points;
215 * skip kernel threads and tasks which have already released
216 * their mm.
218 if (!p->mm)
219 continue;
220 /* skip the init task */
221 if (is_global_init(p))
222 continue;
223 if (mem && !task_in_mem_cgroup(p, mem))
224 continue;
227 * This task already has access to memory reserves and is
228 * being killed. Don't allow any other task access to the
229 * memory reserve.
231 * Note: this may have a chance of deadlock if it gets
232 * blocked waiting for another task which itself is waiting
233 * for memory. Is there a better alternative?
235 if (test_tsk_thread_flag(p, TIF_MEMDIE))
236 return ERR_PTR(-1UL);
239 * This is in the process of releasing memory so wait for it
240 * to finish before killing some other task by mistake.
242 * However, if p is the current task, we allow the 'kill' to
243 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
244 * which will allow it to gain access to memory reserves in
245 * the process of exiting and releasing its resources.
246 * Otherwise we could get an easy OOM deadlock.
248 if (p->flags & PF_EXITING) {
249 if (p != current)
250 return ERR_PTR(-1UL);
252 chosen = p;
253 *ppoints = ULONG_MAX;
256 if (p->oomkilladj == OOM_DISABLE)
257 continue;
259 points = badness(p, uptime.tv_sec, mem);
260 if (points > *ppoints || !chosen) {
261 chosen = p;
262 *ppoints = points;
264 } while_each_thread(g, p);
266 return chosen;
270 * dump_tasks - dump current memory state of all system tasks
271 * @mem: target memory controller
273 * Dumps the current memory state of all system tasks, excluding kernel threads.
274 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
275 * score, and name.
277 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
278 * shown.
280 * Call with tasklist_lock read-locked.
282 static void dump_tasks(const struct mem_cgroup *mem)
284 struct task_struct *g, *p;
286 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
287 "name\n");
288 do_each_thread(g, p) {
290 * total_vm and rss sizes do not exist for tasks with a
291 * detached mm so there's no need to report them.
293 if (!p->mm)
294 continue;
295 if (mem && !task_in_mem_cgroup(p, mem))
296 continue;
298 task_lock(p);
299 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n",
300 p->pid, p->uid, p->tgid, p->mm->total_vm,
301 get_mm_rss(p->mm), (int)task_cpu(p), p->oomkilladj,
302 p->comm);
303 task_unlock(p);
304 } while_each_thread(g, p);
308 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
309 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
310 * set.
312 static void __oom_kill_task(struct task_struct *p, int verbose)
314 if (is_global_init(p)) {
315 WARN_ON(1);
316 printk(KERN_WARNING "tried to kill init!\n");
317 return;
320 if (!p->mm) {
321 WARN_ON(1);
322 printk(KERN_WARNING "tried to kill an mm-less task!\n");
323 return;
326 if (verbose)
327 printk(KERN_ERR "Killed process %d (%s)\n",
328 task_pid_nr(p), p->comm);
331 * We give our sacrificial lamb high priority and access to
332 * all the memory it needs. That way it should be able to
333 * exit() and clear out its resources quickly...
335 p->rt.time_slice = HZ;
336 set_tsk_thread_flag(p, TIF_MEMDIE);
338 force_sig(SIGKILL, p);
341 static int oom_kill_task(struct task_struct *p)
343 struct mm_struct *mm;
344 struct task_struct *g, *q;
346 mm = p->mm;
348 /* WARNING: mm may not be dereferenced since we did not obtain its
349 * value from get_task_mm(p). This is OK since all we need to do is
350 * compare mm to q->mm below.
352 * Furthermore, even if mm contains a non-NULL value, p->mm may
353 * change to NULL at any time since we do not hold task_lock(p).
354 * However, this is of no concern to us.
357 if (mm == NULL)
358 return 1;
361 * Don't kill the process if any threads are set to OOM_DISABLE
363 do_each_thread(g, q) {
364 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
365 return 1;
366 } while_each_thread(g, q);
368 __oom_kill_task(p, 1);
371 * kill all processes that share the ->mm (i.e. all threads),
372 * but are in a different thread group. Don't let them have access
373 * to memory reserves though, otherwise we might deplete all memory.
375 do_each_thread(g, q) {
376 if (q->mm == mm && !same_thread_group(q, p))
377 force_sig(SIGKILL, q);
378 } while_each_thread(g, q);
380 return 0;
383 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
384 unsigned long points, struct mem_cgroup *mem,
385 const char *message)
387 struct task_struct *c;
389 if (printk_ratelimit()) {
390 printk(KERN_WARNING "%s invoked oom-killer: "
391 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
392 current->comm, gfp_mask, order, current->oomkilladj);
393 dump_stack();
394 show_mem();
395 if (sysctl_oom_dump_tasks)
396 dump_tasks(mem);
400 * If the task is already exiting, don't alarm the sysadmin or kill
401 * its children or threads, just set TIF_MEMDIE so it can die quickly
403 if (p->flags & PF_EXITING) {
404 __oom_kill_task(p, 0);
405 return 0;
408 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
409 message, task_pid_nr(p), p->comm, points);
411 /* Try to kill a child first */
412 list_for_each_entry(c, &p->children, sibling) {
413 if (c->mm == p->mm)
414 continue;
415 if (!oom_kill_task(c))
416 return 0;
418 return oom_kill_task(p);
421 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
422 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
424 unsigned long points = 0;
425 struct task_struct *p;
427 cgroup_lock();
428 read_lock(&tasklist_lock);
429 retry:
430 p = select_bad_process(&points, mem);
431 if (PTR_ERR(p) == -1UL)
432 goto out;
434 if (!p)
435 p = current;
437 if (oom_kill_process(p, gfp_mask, 0, points, mem,
438 "Memory cgroup out of memory"))
439 goto retry;
440 out:
441 read_unlock(&tasklist_lock);
442 cgroup_unlock();
444 #endif
446 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
448 int register_oom_notifier(struct notifier_block *nb)
450 return blocking_notifier_chain_register(&oom_notify_list, nb);
452 EXPORT_SYMBOL_GPL(register_oom_notifier);
454 int unregister_oom_notifier(struct notifier_block *nb)
456 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
458 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
461 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero
462 * if a parallel OOM killing is already taking place that includes a zone in
463 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
465 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
467 struct zoneref *z;
468 struct zone *zone;
469 int ret = 1;
471 spin_lock(&zone_scan_mutex);
472 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
473 if (zone_is_oom_locked(zone)) {
474 ret = 0;
475 goto out;
479 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
481 * Lock each zone in the zonelist under zone_scan_mutex so a
482 * parallel invocation of try_set_zone_oom() doesn't succeed
483 * when it shouldn't.
485 zone_set_flag(zone, ZONE_OOM_LOCKED);
488 out:
489 spin_unlock(&zone_scan_mutex);
490 return ret;
494 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
495 * allocation attempts with zonelists containing them may now recall the OOM
496 * killer, if necessary.
498 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
500 struct zoneref *z;
501 struct zone *zone;
503 spin_lock(&zone_scan_mutex);
504 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
505 zone_clear_flag(zone, ZONE_OOM_LOCKED);
507 spin_unlock(&zone_scan_mutex);
511 * out_of_memory - kill the "best" process when we run out of memory
512 * @zonelist: zonelist pointer
513 * @gfp_mask: memory allocation flags
514 * @order: amount of memory being requested as a power of 2
516 * If we run out of memory, we have the choice between either
517 * killing a random task (bad), letting the system crash (worse)
518 * OR try to be smart about which process to kill. Note that we
519 * don't have to be perfect here, we just have to be good.
521 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
523 struct task_struct *p;
524 unsigned long points = 0;
525 unsigned long freed = 0;
526 enum oom_constraint constraint;
528 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
529 if (freed > 0)
530 /* Got some memory back in the last second. */
531 return;
533 if (sysctl_panic_on_oom == 2)
534 panic("out of memory. Compulsory panic_on_oom is selected.\n");
537 * Check if there were limitations on the allocation (only relevant for
538 * NUMA) that may require different handling.
540 constraint = constrained_alloc(zonelist, gfp_mask);
541 read_lock(&tasklist_lock);
543 switch (constraint) {
544 case CONSTRAINT_MEMORY_POLICY:
545 oom_kill_process(current, gfp_mask, order, points, NULL,
546 "No available memory (MPOL_BIND)");
547 break;
549 case CONSTRAINT_NONE:
550 if (sysctl_panic_on_oom)
551 panic("out of memory. panic_on_oom is selected\n");
552 /* Fall-through */
553 case CONSTRAINT_CPUSET:
554 if (sysctl_oom_kill_allocating_task) {
555 oom_kill_process(current, gfp_mask, order, points, NULL,
556 "Out of memory (oom_kill_allocating_task)");
557 break;
559 retry:
561 * Rambo mode: Shoot down a process and hope it solves whatever
562 * issues we may have.
564 p = select_bad_process(&points, NULL);
566 if (PTR_ERR(p) == -1UL)
567 goto out;
569 /* Found nothing?!?! Either we hang forever, or we panic. */
570 if (!p) {
571 read_unlock(&tasklist_lock);
572 panic("Out of memory and no killable processes...\n");
575 if (oom_kill_process(p, gfp_mask, order, points, NULL,
576 "Out of memory"))
577 goto retry;
579 break;
582 out:
583 read_unlock(&tasklist_lock);
586 * Give "p" a good chance of killing itself before we
587 * retry to allocate memory unless "p" is current
589 if (!test_thread_flag(TIF_MEMDIE))
590 schedule_timeout_uninterruptible(1);