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>
20 #include <linux/sched.h>
21 #include <linux/swap.h>
22 #include <linux/timex.h>
23 #include <linux/jiffies.h>
24 #include <linux/cpuset.h>
25 #include <linux/module.h>
26 #include <linux/notifier.h>
28 int sysctl_panic_on_oom;
32 * badness - calculate a numeric value for how bad this task has been
33 * @p: task struct of which task we should calculate
34 * @uptime: current uptime in seconds
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.
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)
50 unsigned long badness(struct task_struct *p, unsigned long uptime)
52 unsigned long points, cpu_time, run_time, s;
54 struct task_struct *child;
64 * The memory size of the process is the basis for the badness.
66 points = mm->total_vm;
69 * After this unlock we can no longer dereference local variable `mm'
74 * swapoff can easily use up all memory, so kill those first.
76 if (p->flags & PF_SWAPOFF)
80 * Processes which fork a lot of child processes are likely
81 * a good choice. We add half the vmsize of the children if they
82 * have an own mm. This prevents forking servers to flood the
83 * 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.
87 list_for_each_entry(child, &p->children, sibling) {
89 if (child->mm != mm && child->mm)
90 points += child->mm->total_vm/2 + 1;
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.
99 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
102 if (uptime >= p->start_time.tv_sec)
103 run_time = (uptime - p->start_time.tv_sec) >> 10;
107 s = int_sqrt(cpu_time);
110 s = int_sqrt(int_sqrt(run_time));
115 * Niced processes are most likely less important, so double
116 * their badness points.
118 if (task_nice(p) > 0)
122 * Superuser processes are usually more important, so we make it
123 * less likely that we kill those.
125 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
126 p->uid == 0 || p->euid == 0)
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
135 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
139 * 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.
143 if (!cpuset_excl_nodes_overlap(p))
147 * Adjust the score by oomkilladj.
150 if (p->oomkilladj > 0)
151 points <<= p->oomkilladj;
153 points >>= -(p->oomkilladj);
157 printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
158 p->pid, p->comm, points);
164 * Types of limitations to the nodes from which allocations may occur
166 #define CONSTRAINT_NONE 1
167 #define CONSTRAINT_MEMORY_POLICY 2
168 #define CONSTRAINT_CPUSET 3
171 * Determine the type of allocation constraint.
173 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
179 /* node has memory ? */
180 for_each_online_node(node)
181 if (NODE_DATA(node)->node_present_pages)
182 node_set(node, nodes);
184 for (z = zonelist->zones; *z; z++)
185 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
186 node_clear(zone_to_nid(*z), nodes);
188 return CONSTRAINT_CPUSET;
190 if (!nodes_empty(nodes))
191 return CONSTRAINT_MEMORY_POLICY;
194 return CONSTRAINT_NONE;
198 * Simple selection loop. We chose the process with the highest
199 * number of 'points'. We expect the caller will lock the tasklist.
201 * (not docbooked, we don't want this one cluttering up the manual)
203 static struct task_struct *select_bad_process(unsigned long *ppoints)
205 struct task_struct *g, *p;
206 struct task_struct *chosen = NULL;
207 struct timespec uptime;
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
220 /* skip the init task */
225 * This task already has access to memory reserves and is
226 * being killed. Don't allow any other task access to the
229 * Note: this may have a chance of deadlock if it gets
230 * blocked waiting for another task which itself is waiting
231 * for memory. Is there a better alternative?
233 if (test_tsk_thread_flag(p, TIF_MEMDIE))
234 return ERR_PTR(-1UL);
237 * This is in the process of releasing memory so wait for it
238 * to finish before killing some other task by mistake.
240 * However, if p is the current task, we allow the 'kill' to
241 * go ahead if it is exiting: this will simply set TIF_MEMDIE,
242 * which will allow it to gain access to memory reserves in
243 * the process of exiting and releasing its resources.
244 * Otherwise we could get an easy OOM deadlock.
246 if (p->flags & PF_EXITING) {
248 return ERR_PTR(-1UL);
251 *ppoints = ULONG_MAX;
254 if (p->oomkilladj == OOM_DISABLE)
257 points = badness(p, uptime.tv_sec);
258 if (points > *ppoints || !chosen) {
262 } while_each_thread(g, p);
268 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
269 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
272 static void __oom_kill_task(struct task_struct *p, int verbose)
276 printk(KERN_WARNING "tried to kill init!\n");
282 printk(KERN_WARNING "tried to kill an mm-less task!\n");
287 printk(KERN_ERR "Killed process %d (%s)\n", p->pid, p->comm);
290 * We give our sacrificial lamb high priority and access to
291 * all the memory it needs. That way it should be able to
292 * exit() and clear out its resources quickly...
295 set_tsk_thread_flag(p, TIF_MEMDIE);
297 force_sig(SIGKILL, p);
300 static int oom_kill_task(struct task_struct *p)
302 struct mm_struct *mm;
303 struct task_struct *g, *q;
307 /* WARNING: mm may not be dereferenced since we did not obtain its
308 * value from get_task_mm(p). This is OK since all we need to do is
309 * compare mm to q->mm below.
311 * Furthermore, even if mm contains a non-NULL value, p->mm may
312 * change to NULL at any time since we do not hold task_lock(p).
313 * However, this is of no concern to us.
320 * Don't kill the process if any threads are set to OOM_DISABLE
322 do_each_thread(g, q) {
323 if (q->mm == mm && p->oomkilladj == OOM_DISABLE)
325 } while_each_thread(g, q);
327 __oom_kill_task(p, 1);
330 * kill all processes that share the ->mm (i.e. all threads),
331 * but are in a different thread group. Don't let them have access
332 * to memory reserves though, otherwise we might deplete all memory.
334 do_each_thread(g, q) {
335 if (q->mm == mm && q->tgid != p->tgid)
336 force_sig(SIGKILL, p);
337 } while_each_thread(g, q);
342 static int oom_kill_process(struct task_struct *p, unsigned long points,
345 struct task_struct *c;
346 struct list_head *tsk;
349 * If the task is already exiting, don't alarm the sysadmin or kill
350 * its children or threads, just set TIF_MEMDIE so it can die quickly
352 if (p->flags & PF_EXITING) {
353 __oom_kill_task(p, 0);
357 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
358 message, p->pid, p->comm, points);
360 /* Try to kill a child first */
361 list_for_each(tsk, &p->children) {
362 c = list_entry(tsk, struct task_struct, sibling);
365 if (!oom_kill_task(c))
368 return oom_kill_task(p);
371 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
373 int register_oom_notifier(struct notifier_block *nb)
375 return blocking_notifier_chain_register(&oom_notify_list, nb);
377 EXPORT_SYMBOL_GPL(register_oom_notifier);
379 int unregister_oom_notifier(struct notifier_block *nb)
381 return blocking_notifier_chain_unregister(&oom_notify_list, nb);
383 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
386 * out_of_memory - kill the "best" process when we run out of memory
388 * If we run out of memory, we have the choice between either
389 * killing a random task (bad), letting the system crash (worse)
390 * OR try to be smart about which process to kill. Note that we
391 * don't have to be perfect here, we just have to be good.
393 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
395 struct task_struct *p;
396 unsigned long points = 0;
397 unsigned long freed = 0;
399 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
401 /* Got some memory back in the last second. */
404 if (printk_ratelimit()) {
405 printk(KERN_WARNING "%s invoked oom-killer: "
406 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
407 current->comm, gfp_mask, order, current->oomkilladj);
413 read_lock(&tasklist_lock);
416 * Check if there were limitations on the allocation (only relevant for
417 * NUMA) that may require different handling.
419 switch (constrained_alloc(zonelist, gfp_mask)) {
420 case CONSTRAINT_MEMORY_POLICY:
421 oom_kill_process(current, points,
422 "No available memory (MPOL_BIND)");
425 case CONSTRAINT_CPUSET:
426 oom_kill_process(current, points,
427 "No available memory in cpuset");
430 case CONSTRAINT_NONE:
431 if (sysctl_panic_on_oom)
432 panic("out of memory. panic_on_oom is selected\n");
435 * Rambo mode: Shoot down a process and hope it solves whatever
436 * issues we may have.
438 p = select_bad_process(&points);
440 if (PTR_ERR(p) == -1UL)
443 /* Found nothing?!?! Either we hang forever, or we panic. */
445 read_unlock(&tasklist_lock);
447 panic("Out of memory and no killable processes...\n");
450 if (oom_kill_process(p, points, "Out of memory"))
457 read_unlock(&tasklist_lock);
461 * Give "p" a good chance of killing itself before we
462 * retry to allocate memory unless "p" is current
464 if (!test_thread_flag(TIF_MEMDIE))
465 schedule_timeout_uninterruptible(1);