Linux Audio

Check our new training course

Loading...
v3.1
 
  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 *  Copyright (C)  2010  Google, Inc.
  8 *	Rewritten by David Rientjes
  9 *
 10 *  The routines in this file are used to kill a process when
 11 *  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.
 13 *
 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
 20#include <linux/oom.h>
 21#include <linux/mm.h>
 22#include <linux/err.h>
 23#include <linux/gfp.h>
 24#include <linux/sched.h>
 
 
 
 
 25#include <linux/swap.h>
 26#include <linux/timex.h>
 27#include <linux/jiffies.h>
 28#include <linux/cpuset.h>
 29#include <linux/module.h>
 30#include <linux/notifier.h>
 31#include <linux/memcontrol.h>
 32#include <linux/mempolicy.h>
 33#include <linux/security.h>
 34#include <linux/ptrace.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 35
 36int sysctl_panic_on_oom;
 37int sysctl_oom_kill_allocating_task;
 38int sysctl_oom_dump_tasks = 1;
 39static DEFINE_SPINLOCK(zone_scan_lock);
 40
 41/**
 42 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
 43 * @new_val: new oom_score_adj value
 
 44 *
 45 * Sets the oom_score_adj value for current to @new_val with proper
 46 * synchronization and returns the old value.  Usually used to temporarily
 47 * set a value, save the old value in the caller, and then reinstate it later.
 48 */
 49int test_set_oom_score_adj(int new_val)
 50{
 51	struct sighand_struct *sighand = current->sighand;
 52	int old_val;
 53
 54	spin_lock_irq(&sighand->siglock);
 55	old_val = current->signal->oom_score_adj;
 56	if (new_val != old_val) {
 57		if (new_val == OOM_SCORE_ADJ_MIN)
 58			atomic_inc(&current->mm->oom_disable_count);
 59		else if (old_val == OOM_SCORE_ADJ_MIN)
 60			atomic_dec(&current->mm->oom_disable_count);
 61		current->signal->oom_score_adj = new_val;
 62	}
 63	spin_unlock_irq(&sighand->siglock);
 64
 65	return old_val;
 
 
 66}
 67
 68#ifdef CONFIG_NUMA
 69/**
 70 * has_intersects_mems_allowed() - check task eligiblity for kill
 71 * @tsk: task struct of which task to consider
 72 * @mask: nodemask passed to page allocator for mempolicy ooms
 73 *
 74 * Task eligibility is determined by whether or not a candidate task, @tsk,
 75 * shares the same mempolicy nodes as current if it is bound by such a policy
 76 * and whether or not it has the same set of allowed cpuset nodes.
 
 
 
 77 */
 78static bool has_intersects_mems_allowed(struct task_struct *tsk,
 79					const nodemask_t *mask)
 80{
 81	struct task_struct *start = tsk;
 
 
 
 
 
 82
 83	do {
 
 84		if (mask) {
 85			/*
 86			 * If this is a mempolicy constrained oom, tsk's
 87			 * cpuset is irrelevant.  Only return true if its
 88			 * mempolicy intersects current, otherwise it may be
 89			 * needlessly killed.
 90			 */
 91			if (mempolicy_nodemask_intersects(tsk, mask))
 92				return true;
 93		} else {
 94			/*
 95			 * This is not a mempolicy constrained oom, so only
 96			 * check the mems of tsk's cpuset.
 97			 */
 98			if (cpuset_mems_allowed_intersects(current, tsk))
 99				return true;
100		}
101	} while_each_thread(start, tsk);
 
 
 
102
103	return false;
104}
105#else
106static bool has_intersects_mems_allowed(struct task_struct *tsk,
107					const nodemask_t *mask)
108{
109	return true;
110}
111#endif /* CONFIG_NUMA */
112
113/*
114 * The process p may have detached its own ->mm while exiting or through
115 * use_mm(), but one or more of its subthreads may still have a valid
116 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
117 * task_lock() held.
118 */
119struct task_struct *find_lock_task_mm(struct task_struct *p)
120{
121	struct task_struct *t = p;
 
 
122
123	do {
124		task_lock(t);
125		if (likely(t->mm))
126			return t;
127		task_unlock(t);
128	} while_each_thread(p, t);
 
 
 
129
130	return NULL;
 
 
 
 
 
 
 
 
 
131}
132
133/* return true if the task is not adequate as candidate victim task. */
134static bool oom_unkillable_task(struct task_struct *p,
135		const struct mem_cgroup *mem, const nodemask_t *nodemask)
136{
137	if (is_global_init(p))
138		return true;
139	if (p->flags & PF_KTHREAD)
140		return true;
 
 
141
142	/* When mem_cgroup_out_of_memory() and p is not member of the group */
143	if (mem && !task_in_mem_cgroup(p, mem))
144		return true;
145
146	/* p may not have freeable memory in nodemask */
147	if (!has_intersects_mems_allowed(p, nodemask))
148		return true;
 
 
 
 
 
 
 
 
 
 
149
150	return false;
151}
152
153/**
154 * oom_badness - heuristic function to determine which candidate task to kill
155 * @p: task struct of which task we should calculate
156 * @totalpages: total present RAM allowed for page allocation
157 *
158 * The heuristic for determining which task to kill is made to be as simple and
159 * predictable as possible.  The goal is to return the highest value for the
160 * task consuming the most memory to avoid subsequent oom failures.
161 */
162unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
163		      const nodemask_t *nodemask, unsigned long totalpages)
164{
165	int points;
 
166
167	if (oom_unkillable_task(p, mem, nodemask))
168		return 0;
169
170	p = find_lock_task_mm(p);
171	if (!p)
172		return 0;
173
174	/*
175	 * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
176	 * so the entire heuristic doesn't need to be executed for something
177	 * that cannot be killed.
178	 */
179	if (atomic_read(&p->mm->oom_disable_count)) {
 
 
 
180		task_unlock(p);
181		return 0;
182	}
183
184	/*
185	 * The memory controller may have a limit of 0 bytes, so avoid a divide
186	 * by zero, if necessary.
187	 */
188	if (!totalpages)
189		totalpages = 1;
190
191	/*
192	 * The baseline for the badness score is the proportion of RAM that each
193	 * task's rss, pagetable and swap space use.
194	 */
195	points = get_mm_rss(p->mm) + p->mm->nr_ptes;
196	points += get_mm_counter(p->mm, MM_SWAPENTS);
197
198	points *= 1000;
199	points /= totalpages;
200	task_unlock(p);
201
202	/*
203	 * Root processes get 3% bonus, just like the __vm_enough_memory()
204	 * implementation used by LSMs.
205	 */
206	if (has_capability_noaudit(p, CAP_SYS_ADMIN))
207		points -= 30;
208
209	/*
210	 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
211	 * either completely disable oom killing or always prefer a certain
212	 * task.
213	 */
214	points += p->signal->oom_score_adj;
215
216	/*
217	 * Never return 0 for an eligible task that may be killed since it's
218	 * possible that no single user task uses more than 0.1% of memory and
219	 * no single admin tasks uses more than 3.0%.
220	 */
221	if (points <= 0)
222		return 1;
223	return (points < 1000) ? points : 1000;
224}
225
 
 
 
 
 
 
 
226/*
227 * Determine the type of allocation constraint.
228 */
229#ifdef CONFIG_NUMA
230static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
231				gfp_t gfp_mask, nodemask_t *nodemask,
232				unsigned long *totalpages)
233{
234	struct zone *zone;
235	struct zoneref *z;
236	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
237	bool cpuset_limited = false;
238	int nid;
239
 
 
 
 
 
240	/* Default to all available memory */
241	*totalpages = totalram_pages + total_swap_pages;
 
 
 
242
243	if (!zonelist)
244		return CONSTRAINT_NONE;
245	/*
246	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
247	 * to kill current.We have to random task kill in this case.
248	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
249	 */
250	if (gfp_mask & __GFP_THISNODE)
251		return CONSTRAINT_NONE;
252
253	/*
254	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
255	 * the page allocator means a mempolicy is in effect.  Cpuset policy
256	 * is enforced in get_page_from_freelist().
257	 */
258	if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
259		*totalpages = total_swap_pages;
260		for_each_node_mask(nid, *nodemask)
261			*totalpages += node_spanned_pages(nid);
 
262		return CONSTRAINT_MEMORY_POLICY;
263	}
264
265	/* Check this allocation failure is caused by cpuset's wall function */
266	for_each_zone_zonelist_nodemask(zone, z, zonelist,
267			high_zoneidx, nodemask)
268		if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
269			cpuset_limited = true;
270
271	if (cpuset_limited) {
272		*totalpages = total_swap_pages;
273		for_each_node_mask(nid, cpuset_current_mems_allowed)
274			*totalpages += node_spanned_pages(nid);
275		return CONSTRAINT_CPUSET;
276	}
277	return CONSTRAINT_NONE;
278}
279#else
280static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
281				gfp_t gfp_mask, nodemask_t *nodemask,
282				unsigned long *totalpages)
283{
284	*totalpages = totalram_pages + total_swap_pages;
285	return CONSTRAINT_NONE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
286}
287#endif
288
289/*
290 * Simple selection loop. We chose the process with the highest
291 * number of 'points'. We expect the caller will lock the tasklist.
292 *
293 * (not docbooked, we don't want this one cluttering up the manual)
294 */
295static struct task_struct *select_bad_process(unsigned int *ppoints,
296		unsigned long totalpages, struct mem_cgroup *mem,
297		const nodemask_t *nodemask)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
298{
299	struct task_struct *g, *p;
300	struct task_struct *chosen = NULL;
301	*ppoints = 0;
302
303	do_each_thread(g, p) {
304		unsigned int points;
305
306		if (p->exit_state)
307			continue;
308		if (oom_unkillable_task(p, mem, nodemask))
309			continue;
310
 
 
311		/*
312		 * This task already has access to memory reserves and is
313		 * being killed. Don't allow any other task access to the
314		 * memory reserve.
315		 *
316		 * Note: this may have a chance of deadlock if it gets
317		 * blocked waiting for another task which itself is waiting
318		 * for memory. Is there a better alternative?
319		 */
320		if (test_tsk_thread_flag(p, TIF_MEMDIE))
321			return ERR_PTR(-1UL);
322		if (!p->mm)
323			continue;
324
325		if (p->flags & PF_EXITING) {
326			/*
327			 * If p is the current task and is in the process of
328			 * releasing memory, we allow the "kill" to set
329			 * TIF_MEMDIE, which will allow it to gain access to
330			 * memory reserves.  Otherwise, it may stall forever.
331			 *
332			 * The loop isn't broken here, however, in case other
333			 * threads are found to have already been oom killed.
334			 */
335			if (p == current) {
336				chosen = p;
337				*ppoints = 1000;
338			} else {
339				/*
340				 * If this task is not being ptraced on exit,
341				 * then wait for it to finish before killing
342				 * some other task unnecessarily.
343				 */
344				if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
345					return ERR_PTR(-1UL);
346			}
347		}
348
349		points = oom_badness(p, mem, nodemask, totalpages);
350		if (points > *ppoints) {
351			chosen = p;
352			*ppoints = points;
353		}
354	} while_each_thread(g, p);
 
355
356	return chosen;
357}
358
359/**
360 * dump_tasks - dump current memory state of all system tasks
361 * @mem: current's memory controller, if constrained
362 * @nodemask: nodemask passed to page allocator for mempolicy ooms
363 *
364 * Dumps the current memory state of all eligible tasks.  Tasks not in the same
365 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
366 * are not shown.
367 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
368 * value, oom_score_adj value, and name.
369 *
370 * Call with tasklist_lock read-locked.
371 */
372static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
373{
374	struct task_struct *p;
375	struct task_struct *task;
376
377	pr_info("[ pid ]   uid  tgid total_vm      rss cpu oom_adj oom_score_adj name\n");
378	for_each_process(p) {
379		if (oom_unkillable_task(p, mem, nodemask))
380			continue;
381
382		task = find_lock_task_mm(p);
383		if (!task) {
384			/*
385			 * This is a kthread or all of p's threads have already
386			 * detached their mm's.  There's no need to report
387			 * them; they can't be oom killed anyway.
388			 */
389			continue;
390		}
391
392		pr_info("[%5d] %5d %5d %8lu %8lu %3u     %3d         %5d %s\n",
393			task->pid, task_uid(task), task->tgid,
394			task->mm->total_vm, get_mm_rss(task->mm),
395			task_cpu(task), task->signal->oom_adj,
396			task->signal->oom_score_adj, task->comm);
397		task_unlock(task);
398	}
399}
400
401static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
402			struct mem_cgroup *mem, const nodemask_t *nodemask)
403{
404	task_lock(current);
405	pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
406		"oom_adj=%d, oom_score_adj=%d\n",
407		current->comm, gfp_mask, order, current->signal->oom_adj,
408		current->signal->oom_score_adj);
409	cpuset_print_task_mems_allowed(current);
410	task_unlock(current);
 
 
 
 
 
 
 
 
 
 
411	dump_stack();
412	mem_cgroup_print_oom_info(mem, p);
413	show_mem(SHOW_MEM_FILTER_NODES);
 
 
 
 
 
414	if (sysctl_oom_dump_tasks)
415		dump_tasks(mem, nodemask);
 
 
416}
417
 
 
 
 
 
 
 
 
418#define K(x) ((x) << (PAGE_SHIFT-10))
419static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
 
 
 
 
 
 
 
420{
421	struct task_struct *q;
422	struct mm_struct *mm;
423
424	p = find_lock_task_mm(p);
425	if (!p)
426		return 1;
 
 
 
 
427
428	/* mm cannot be safely dereferenced after task_unlock(p) */
429	mm = p->mm;
 
 
 
 
 
 
 
430
431	pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
432		task_pid_nr(p), p->comm, K(p->mm->total_vm),
433		K(get_mm_counter(p->mm, MM_ANONPAGES)),
434		K(get_mm_counter(p->mm, MM_FILEPAGES)));
435	task_unlock(p);
436
437	/*
438	 * Kill all processes sharing p->mm in other thread groups, if any.
439	 * They don't get access to memory reserves or a higher scheduler
440	 * priority, though, to avoid depletion of all memory or task
441	 * starvation.  This prevents mm->mmap_sem livelock when an oom killed
442	 * task cannot exit because it requires the semaphore and its contended
443	 * by another thread trying to allocate memory itself.  That thread will
444	 * now get access to memory reserves since it has a pending fatal
445	 * signal.
446	 */
447	for_each_process(q)
448		if (q->mm == mm && !same_thread_group(q, p)) {
449			task_lock(q);	/* Protect ->comm from prctl() */
450			pr_err("Kill process %d (%s) sharing same memory\n",
451				task_pid_nr(q), q->comm);
452			task_unlock(q);
453			force_sig(SIGKILL, q);
454		}
455
456	set_tsk_thread_flag(p, TIF_MEMDIE);
457	force_sig(SIGKILL, p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
458
459	return 0;
460}
461#undef K
462
463static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
464			    unsigned int points, unsigned long totalpages,
465			    struct mem_cgroup *mem, nodemask_t *nodemask,
466			    const char *message)
467{
468	struct task_struct *victim = p;
469	struct task_struct *child;
470	struct task_struct *t = p;
471	unsigned int victim_points = 0;
472
473	if (printk_ratelimit())
474		dump_header(p, gfp_mask, order, mem, nodemask);
 
 
475
476	/*
477	 * If the task is already exiting, don't alarm the sysadmin or kill
478	 * its children or threads, just set TIF_MEMDIE so it can die quickly
 
 
479	 */
480	if (p->flags & PF_EXITING) {
481		set_tsk_thread_flag(p, TIF_MEMDIE);
482		return 0;
483	}
484
485	task_lock(p);
486	pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
487		message, task_pid_nr(p), p->comm, points);
488	task_unlock(p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
489
490	/*
491	 * If any of p's children has a different mm and is eligible for kill,
492	 * the one with the highest oom_badness() score is sacrificed for its
493	 * parent.  This attempts to lose the minimal amount of work done while
494	 * still freeing memory.
495	 */
496	do {
497		list_for_each_entry(child, &t->children, sibling) {
498			unsigned int child_points;
499
500			if (child->mm == p->mm)
501				continue;
502			/*
503			 * oom_badness() returns 0 if the thread is unkillable
504			 */
505			child_points = oom_badness(child, mem, nodemask,
506								totalpages);
507			if (child_points > victim_points) {
508				victim = child;
509				victim_points = child_points;
510			}
 
 
 
511		}
512	} while_each_thread(p, t);
513
514	return oom_kill_task(victim, mem);
 
 
 
 
515}
516
517/*
518 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
519 */
520static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
521				int order, const nodemask_t *nodemask)
522{
523	if (likely(!sysctl_panic_on_oom))
 
524		return;
525	if (sysctl_panic_on_oom != 2) {
526		/*
527		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
528		 * does not panic for cpuset, mempolicy, or memcg allocation
529		 * failures.
530		 */
531		if (constraint != CONSTRAINT_NONE)
532			return;
533	}
534	read_lock(&tasklist_lock);
535	dump_header(NULL, gfp_mask, order, NULL, nodemask);
536	read_unlock(&tasklist_lock);
537	panic("Out of memory: %s panic_on_oom is enabled\n",
538		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
539}
540
541#ifdef CONFIG_CGROUP_MEM_RES_CTLR
542void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
543{
544	unsigned long limit;
545	unsigned int points = 0;
546	struct task_struct *p;
 
 
 
 
 
 
547
548	/*
549	 * If current has a pending SIGKILL, then automatically select it.  The
550	 * goal is to allow it to allocate so that it may quickly exit and free
551	 * its memory.
552	 */
553	if (fatal_signal_pending(current)) {
554		set_thread_flag(TIF_MEMDIE);
 
 
 
 
 
 
 
 
 
 
555		return;
 
 
 
 
 
556	}
557
558	check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
559	limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
560	read_lock(&tasklist_lock);
561retry:
562	p = select_bad_process(&points, limit, mem, NULL);
563	if (!p || PTR_ERR(p) == -1UL)
564		goto out;
565
566	if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
567				"Memory cgroup out of memory"))
568		goto retry;
569out:
570	read_unlock(&tasklist_lock);
571}
572#endif
573
574static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 
 
 
 
 
575
576int register_oom_notifier(struct notifier_block *nb)
 
 
 
 
 
 
 
577{
578	return blocking_notifier_chain_register(&oom_notify_list, nb);
 
579}
580EXPORT_SYMBOL_GPL(register_oom_notifier);
581
582int unregister_oom_notifier(struct notifier_block *nb)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
583{
584	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
585}
586EXPORT_SYMBOL_GPL(unregister_oom_notifier);
587
588/*
589 * Try to acquire the OOM killer lock for the zones in zonelist.  Returns zero
590 * if a parallel OOM killing is already taking place that includes a zone in
591 * the zonelist.  Otherwise, locks all zones in the zonelist and returns 1.
 
 
592 */
593int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
594{
595	struct zoneref *z;
596	struct zone *zone;
597	int ret = 1;
598
599	spin_lock(&zone_scan_lock);
600	for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
601		if (zone_is_oom_locked(zone)) {
602			ret = 0;
603			goto out;
604		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
605	}
606
607	for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
608		/*
609		 * Lock each zone in the zonelist under zone_scan_lock so a
610		 * parallel invocation of try_set_zonelist_oom() doesn't succeed
611		 * when it shouldn't.
612		 */
613		zone_set_flag(zone, ZONE_OOM_LOCKED);
 
 
614	}
 
615
616out:
617	spin_unlock(&zone_scan_lock);
618	return ret;
 
 
619}
 
620
621/*
622 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
623 * allocation attempts with zonelists containing them may now recall the OOM
624 * killer, if necessary.
625 */
626void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
627{
628	struct zoneref *z;
629	struct zone *zone;
630
631	spin_lock(&zone_scan_lock);
632	for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
633		zone_clear_flag(zone, ZONE_OOM_LOCKED);
634	}
635	spin_unlock(&zone_scan_lock);
636}
637
638/*
639 * Try to acquire the oom killer lock for all system zones.  Returns zero if a
640 * parallel oom killing is taking place, otherwise locks all zones and returns
641 * non-zero.
642 */
643static int try_set_system_oom(void)
644{
645	struct zone *zone;
646	int ret = 1;
 
 
647
648	spin_lock(&zone_scan_lock);
649	for_each_populated_zone(zone)
650		if (zone_is_oom_locked(zone)) {
651			ret = 0;
652			goto out;
653		}
654	for_each_populated_zone(zone)
655		zone_set_flag(zone, ZONE_OOM_LOCKED);
656out:
657	spin_unlock(&zone_scan_lock);
658	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
659}
660
661/*
662 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
663 * attempts or page faults may now recall the oom killer, if necessary.
664 */
665static void clear_system_oom(void)
666{
667	struct zone *zone;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
668
669	spin_lock(&zone_scan_lock);
670	for_each_populated_zone(zone)
671		zone_clear_flag(zone, ZONE_OOM_LOCKED);
672	spin_unlock(&zone_scan_lock);
673}
 
674
675/**
676 * out_of_memory - kill the "best" process when we run out of memory
677 * @zonelist: zonelist pointer
678 * @gfp_mask: memory allocation flags
679 * @order: amount of memory being requested as a power of 2
680 * @nodemask: nodemask passed to page allocator
681 *
682 * If we run out of memory, we have the choice between either
683 * killing a random task (bad), letting the system crash (worse)
684 * OR try to be smart about which process to kill. Note that we
685 * don't have to be perfect here, we just have to be good.
686 */
687void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
688		int order, nodemask_t *nodemask)
689{
690	const nodemask_t *mpol_mask;
691	struct task_struct *p;
692	unsigned long totalpages;
693	unsigned long freed = 0;
694	unsigned int points;
695	enum oom_constraint constraint = CONSTRAINT_NONE;
696	int killed = 0;
697
698	blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
699	if (freed > 0)
700		/* Got some memory back in the last second. */
701		return;
 
 
702
703	/*
704	 * If current has a pending SIGKILL, then automatically select it.  The
705	 * goal is to allow it to allocate so that it may quickly exit and free
706	 * its memory.
707	 */
708	if (fatal_signal_pending(current)) {
709		set_thread_flag(TIF_MEMDIE);
710		return;
 
711	}
712
713	/*
 
 
 
 
 
 
 
 
 
 
714	 * Check if there were limitations on the allocation (only relevant for
715	 * NUMA) that may require different handling.
716	 */
717	constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
718						&totalpages);
719	mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
720	check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
721
722	read_lock(&tasklist_lock);
723	if (sysctl_oom_kill_allocating_task &&
724	    !oom_unkillable_task(current, NULL, nodemask) &&
725	    current->mm && !atomic_read(&current->mm->oom_disable_count)) {
726		/*
727		 * oom_kill_process() needs tasklist_lock held.  If it returns
728		 * non-zero, current could not be killed so we must fallback to
729		 * the tasklist scan.
730		 */
731		if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
732				NULL, nodemask,
733				"Out of memory (oom_kill_allocating_task)"))
734			goto out;
735	}
736
737retry:
738	p = select_bad_process(&points, totalpages, NULL, mpol_mask);
739	if (PTR_ERR(p) == -1UL)
740		goto out;
741
742	/* Found nothing?!?! Either we hang forever, or we panic. */
743	if (!p) {
744		dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
745		read_unlock(&tasklist_lock);
746		panic("Out of memory and no killable processes...\n");
 
 
747	}
748
749	if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
750				nodemask, "Out of memory"))
751		goto retry;
752	killed = 1;
753out:
754	read_unlock(&tasklist_lock);
755
756	/*
757	 * Give "p" a good chance of killing itself before we
758	 * retry to allocate memory unless "p" is current
759	 */
760	if (killed && !test_thread_flag(TIF_MEMDIE))
761		schedule_timeout_uninterruptible(1);
762}
763
764/*
765 * The pagefault handler calls here because it is out of memory, so kill a
766 * memory-hogging task.  If a populated zone has ZONE_OOM_LOCKED set, a parallel
767 * oom killing is already in progress so do nothing.  If a task is found with
768 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
769 */
770void pagefault_out_of_memory(void)
771{
772	if (try_set_system_oom()) {
773		out_of_memory(NULL, 0, 0, NULL);
774		clear_system_oom();
775	}
776	if (!test_thread_flag(TIF_MEMDIE))
777		schedule_timeout_uninterruptible(1);
 
 
 
 
 
 
 
 
 
778}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/oom_kill.c
   4 * 
   5 *  Copyright (C)  1998,2000  Rik van Riel
   6 *	Thanks go out to Claus Fischer for some serious inspiration and
   7 *	for goading me into coding this file...
   8 *  Copyright (C)  2010  Google, Inc.
   9 *	Rewritten by David Rientjes
  10 *
  11 *  The routines in this file are used to kill a process when
  12 *  we're seriously out of memory. This gets called from __alloc_pages()
  13 *  in mm/page_alloc.c when we really run out of memory.
  14 *
  15 *  Since we won't call these routines often (on a well-configured
  16 *  machine) this file will double as a 'coding guide' and a signpost
  17 *  for newbie kernel hackers. It features several pointers to major
  18 *  kernel subsystems and hints as to where to find out what things do.
  19 */
  20
  21#include <linux/oom.h>
  22#include <linux/mm.h>
  23#include <linux/err.h>
  24#include <linux/gfp.h>
  25#include <linux/sched.h>
  26#include <linux/sched/mm.h>
  27#include <linux/sched/coredump.h>
  28#include <linux/sched/task.h>
  29#include <linux/sched/debug.h>
  30#include <linux/swap.h>
  31#include <linux/timex.h>
  32#include <linux/jiffies.h>
  33#include <linux/cpuset.h>
  34#include <linux/export.h>
  35#include <linux/notifier.h>
  36#include <linux/memcontrol.h>
  37#include <linux/mempolicy.h>
  38#include <linux/security.h>
  39#include <linux/ptrace.h>
  40#include <linux/freezer.h>
  41#include <linux/ftrace.h>
  42#include <linux/ratelimit.h>
  43#include <linux/kthread.h>
  44#include <linux/init.h>
  45#include <linux/mmu_notifier.h>
  46
  47#include <asm/tlb.h>
  48#include "internal.h"
  49#include "slab.h"
  50
  51#define CREATE_TRACE_POINTS
  52#include <trace/events/oom.h>
  53
  54int sysctl_panic_on_oom;
  55int sysctl_oom_kill_allocating_task;
  56int sysctl_oom_dump_tasks = 1;
 
  57
  58/*
  59 * Serializes oom killer invocations (out_of_memory()) from all contexts to
  60 * prevent from over eager oom killing (e.g. when the oom killer is invoked
  61 * from different domains).
  62 *
  63 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
  64 * and mark_oom_victim
  65 */
  66DEFINE_MUTEX(oom_lock);
  67/* Serializes oom_score_adj and oom_score_adj_min updates */
  68DEFINE_MUTEX(oom_adj_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
  69
  70static inline bool is_memcg_oom(struct oom_control *oc)
  71{
  72	return oc->memcg != NULL;
  73}
  74
  75#ifdef CONFIG_NUMA
  76/**
  77 * oom_cpuset_eligible() - check task eligibility for kill
  78 * @start: task struct of which task to consider
  79 * @oc: pointer to struct oom_control
  80 *
  81 * Task eligibility is determined by whether or not a candidate task, @tsk,
  82 * shares the same mempolicy nodes as current if it is bound by such a policy
  83 * and whether or not it has the same set of allowed cpuset nodes.
  84 *
  85 * This function is assuming oom-killer context and 'current' has triggered
  86 * the oom-killer.
  87 */
  88static bool oom_cpuset_eligible(struct task_struct *start,
  89				struct oom_control *oc)
  90{
  91	struct task_struct *tsk;
  92	bool ret = false;
  93	const nodemask_t *mask = oc->nodemask;
  94
  95	if (is_memcg_oom(oc))
  96		return true;
  97
  98	rcu_read_lock();
  99	for_each_thread(start, tsk) {
 100		if (mask) {
 101			/*
 102			 * If this is a mempolicy constrained oom, tsk's
 103			 * cpuset is irrelevant.  Only return true if its
 104			 * mempolicy intersects current, otherwise it may be
 105			 * needlessly killed.
 106			 */
 107			ret = mempolicy_in_oom_domain(tsk, mask);
 
 108		} else {
 109			/*
 110			 * This is not a mempolicy constrained oom, so only
 111			 * check the mems of tsk's cpuset.
 112			 */
 113			ret = cpuset_mems_allowed_intersects(current, tsk);
 
 114		}
 115		if (ret)
 116			break;
 117	}
 118	rcu_read_unlock();
 119
 120	return ret;
 121}
 122#else
 123static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
 
 124{
 125	return true;
 126}
 127#endif /* CONFIG_NUMA */
 128
 129/*
 130 * The process p may have detached its own ->mm while exiting or through
 131 * kthread_use_mm(), but one or more of its subthreads may still have a valid
 132 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
 133 * task_lock() held.
 134 */
 135struct task_struct *find_lock_task_mm(struct task_struct *p)
 136{
 137	struct task_struct *t;
 138
 139	rcu_read_lock();
 140
 141	for_each_thread(p, t) {
 142		task_lock(t);
 143		if (likely(t->mm))
 144			goto found;
 145		task_unlock(t);
 146	}
 147	t = NULL;
 148found:
 149	rcu_read_unlock();
 150
 151	return t;
 152}
 153
 154/*
 155 * order == -1 means the oom kill is required by sysrq, otherwise only
 156 * for display purposes.
 157 */
 158static inline bool is_sysrq_oom(struct oom_control *oc)
 159{
 160	return oc->order == -1;
 161}
 162
 163/* return true if the task is not adequate as candidate victim task. */
 164static bool oom_unkillable_task(struct task_struct *p)
 
 165{
 166	if (is_global_init(p))
 167		return true;
 168	if (p->flags & PF_KTHREAD)
 169		return true;
 170	return false;
 171}
 172
 173/*
 174 * Check whether unreclaimable slab amount is greater than
 175 * all user memory(LRU pages).
 176 * dump_unreclaimable_slab() could help in the case that
 177 * oom due to too much unreclaimable slab used by kernel.
 178*/
 179static bool should_dump_unreclaim_slab(void)
 180{
 181	unsigned long nr_lru;
 182
 183	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
 184		 global_node_page_state(NR_INACTIVE_ANON) +
 185		 global_node_page_state(NR_ACTIVE_FILE) +
 186		 global_node_page_state(NR_INACTIVE_FILE) +
 187		 global_node_page_state(NR_ISOLATED_ANON) +
 188		 global_node_page_state(NR_ISOLATED_FILE) +
 189		 global_node_page_state(NR_UNEVICTABLE);
 190
 191	return (global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B) > nr_lru);
 192}
 193
 194/**
 195 * oom_badness - heuristic function to determine which candidate task to kill
 196 * @p: task struct of which task we should calculate
 197 * @totalpages: total present RAM allowed for page allocation
 198 *
 199 * The heuristic for determining which task to kill is made to be as simple and
 200 * predictable as possible.  The goal is to return the highest value for the
 201 * task consuming the most memory to avoid subsequent oom failures.
 202 */
 203long oom_badness(struct task_struct *p, unsigned long totalpages)
 
 204{
 205	long points;
 206	long adj;
 207
 208	if (oom_unkillable_task(p))
 209		return LONG_MIN;
 210
 211	p = find_lock_task_mm(p);
 212	if (!p)
 213		return LONG_MIN;
 214
 215	/*
 216	 * Do not even consider tasks which are explicitly marked oom
 217	 * unkillable or have been already oom reaped or the are in
 218	 * the middle of vfork
 219	 */
 220	adj = (long)p->signal->oom_score_adj;
 221	if (adj == OOM_SCORE_ADJ_MIN ||
 222			test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
 223			in_vfork(p)) {
 224		task_unlock(p);
 225		return LONG_MIN;
 226	}
 227
 228	/*
 
 
 
 
 
 
 
 229	 * The baseline for the badness score is the proportion of RAM that each
 230	 * task's rss, pagetable and swap space use.
 231	 */
 232	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
 233		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
 
 
 
 234	task_unlock(p);
 235
 236	/* Normalize to oom_score_adj units */
 237	adj *= totalpages / 1000;
 238	points += adj;
 239
 240	return points;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 241}
 242
 243static const char * const oom_constraint_text[] = {
 244	[CONSTRAINT_NONE] = "CONSTRAINT_NONE",
 245	[CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
 246	[CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
 247	[CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
 248};
 249
 250/*
 251 * Determine the type of allocation constraint.
 252 */
 253static enum oom_constraint constrained_alloc(struct oom_control *oc)
 
 
 
 254{
 255	struct zone *zone;
 256	struct zoneref *z;
 257	enum zone_type highest_zoneidx = gfp_zone(oc->gfp_mask);
 258	bool cpuset_limited = false;
 259	int nid;
 260
 261	if (is_memcg_oom(oc)) {
 262		oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
 263		return CONSTRAINT_MEMCG;
 264	}
 265
 266	/* Default to all available memory */
 267	oc->totalpages = totalram_pages() + total_swap_pages;
 268
 269	if (!IS_ENABLED(CONFIG_NUMA))
 270		return CONSTRAINT_NONE;
 271
 272	if (!oc->zonelist)
 273		return CONSTRAINT_NONE;
 274	/*
 275	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
 276	 * to kill current.We have to random task kill in this case.
 277	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
 278	 */
 279	if (oc->gfp_mask & __GFP_THISNODE)
 280		return CONSTRAINT_NONE;
 281
 282	/*
 283	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
 284	 * the page allocator means a mempolicy is in effect.  Cpuset policy
 285	 * is enforced in get_page_from_freelist().
 286	 */
 287	if (oc->nodemask &&
 288	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
 289		oc->totalpages = total_swap_pages;
 290		for_each_node_mask(nid, *oc->nodemask)
 291			oc->totalpages += node_present_pages(nid);
 292		return CONSTRAINT_MEMORY_POLICY;
 293	}
 294
 295	/* Check this allocation failure is caused by cpuset's wall function */
 296	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
 297			highest_zoneidx, oc->nodemask)
 298		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
 299			cpuset_limited = true;
 300
 301	if (cpuset_limited) {
 302		oc->totalpages = total_swap_pages;
 303		for_each_node_mask(nid, cpuset_current_mems_allowed)
 304			oc->totalpages += node_present_pages(nid);
 305		return CONSTRAINT_CPUSET;
 306	}
 307	return CONSTRAINT_NONE;
 308}
 309
 310static int oom_evaluate_task(struct task_struct *task, void *arg)
 
 
 311{
 312	struct oom_control *oc = arg;
 313	long points;
 314
 315	if (oom_unkillable_task(task))
 316		goto next;
 317
 318	/* p may not have freeable memory in nodemask */
 319	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
 320		goto next;
 321
 322	/*
 323	 * This task already has access to memory reserves and is being killed.
 324	 * Don't allow any other task to have access to the reserves unless
 325	 * the task has MMF_OOM_SKIP because chances that it would release
 326	 * any memory is quite low.
 327	 */
 328	if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
 329		if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
 330			goto next;
 331		goto abort;
 332	}
 333
 334	/*
 335	 * If task is allocating a lot of memory and has been marked to be
 336	 * killed first if it triggers an oom, then select it.
 337	 */
 338	if (oom_task_origin(task)) {
 339		points = LONG_MAX;
 340		goto select;
 341	}
 342
 343	points = oom_badness(task, oc->totalpages);
 344	if (points == LONG_MIN || points < oc->chosen_points)
 345		goto next;
 346
 347select:
 348	if (oc->chosen)
 349		put_task_struct(oc->chosen);
 350	get_task_struct(task);
 351	oc->chosen = task;
 352	oc->chosen_points = points;
 353next:
 354	return 0;
 355abort:
 356	if (oc->chosen)
 357		put_task_struct(oc->chosen);
 358	oc->chosen = (void *)-1UL;
 359	return 1;
 360}
 
 361
 362/*
 363 * Simple selection loop. We choose the process with the highest number of
 364 * 'points'. In case scan was aborted, oc->chosen is set to -1.
 
 
 365 */
 366static void select_bad_process(struct oom_control *oc)
 367{
 368	oc->chosen_points = LONG_MIN;
 369
 370	if (is_memcg_oom(oc))
 371		mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
 372	else {
 373		struct task_struct *p;
 374
 375		rcu_read_lock();
 376		for_each_process(p)
 377			if (oom_evaluate_task(p, oc))
 378				break;
 379		rcu_read_unlock();
 380	}
 381}
 382
 383static int dump_task(struct task_struct *p, void *arg)
 384{
 385	struct oom_control *oc = arg;
 386	struct task_struct *task;
 
 387
 388	if (oom_unkillable_task(p))
 389		return 0;
 390
 391	/* p may not have freeable memory in nodemask */
 392	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
 393		return 0;
 
 394
 395	task = find_lock_task_mm(p);
 396	if (!task) {
 397		/*
 398		 * All of p's threads have already detached their mm's. There's
 399		 * no need to report them; they can't be oom killed anyway.
 
 
 
 
 
 400		 */
 401		return 0;
 402	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 403
 404	pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu         %5hd %s\n",
 405		task->pid, from_kuid(&init_user_ns, task_uid(task)),
 406		task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
 407		mm_pgtables_bytes(task->mm),
 408		get_mm_counter(task->mm, MM_SWAPENTS),
 409		task->signal->oom_score_adj, task->comm);
 410	task_unlock(task);
 411
 412	return 0;
 413}
 414
 415/**
 416 * dump_tasks - dump current memory state of all system tasks
 417 * @oc: pointer to struct oom_control
 
 418 *
 419 * Dumps the current memory state of all eligible tasks.  Tasks not in the same
 420 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
 421 * are not shown.
 422 * State information includes task's pid, uid, tgid, vm size, rss,
 423 * pgtables_bytes, swapents, oom_score_adj value, and name.
 
 
 424 */
 425static void dump_tasks(struct oom_control *oc)
 426{
 427	pr_info("Tasks state (memory values in pages):\n");
 428	pr_info("[  pid  ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj name\n");
 
 
 
 
 
 429
 430	if (is_memcg_oom(oc))
 431		mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
 432	else {
 433		struct task_struct *p;
 434
 435		rcu_read_lock();
 436		for_each_process(p)
 437			dump_task(p, oc);
 438		rcu_read_unlock();
 
 
 
 
 
 
 
 439	}
 440}
 441
 442static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
 443{
 444	/* one line summary of the oom killer context. */
 445	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
 446			oom_constraint_text[oc->constraint],
 447			nodemask_pr_args(oc->nodemask));
 448	cpuset_print_current_mems_allowed();
 449	mem_cgroup_print_oom_context(oc->memcg, victim);
 450	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
 451		from_kuid(&init_user_ns, task_uid(victim)));
 452}
 453
 454static void dump_header(struct oom_control *oc, struct task_struct *p)
 455{
 456	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
 457		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
 458			current->signal->oom_score_adj);
 459	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
 460		pr_warn("COMPACTION is disabled!!!\n");
 461
 462	dump_stack();
 463	if (is_memcg_oom(oc))
 464		mem_cgroup_print_oom_meminfo(oc->memcg);
 465	else {
 466		show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
 467		if (should_dump_unreclaim_slab())
 468			dump_unreclaimable_slab();
 469	}
 470	if (sysctl_oom_dump_tasks)
 471		dump_tasks(oc);
 472	if (p)
 473		dump_oom_summary(oc, p);
 474}
 475
 476/*
 477 * Number of OOM victims in flight
 478 */
 479static atomic_t oom_victims = ATOMIC_INIT(0);
 480static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
 481
 482static bool oom_killer_disabled __read_mostly;
 483
 484#define K(x) ((x) << (PAGE_SHIFT-10))
 485
 486/*
 487 * task->mm can be NULL if the task is the exited group leader.  So to
 488 * determine whether the task is using a particular mm, we examine all the
 489 * task's threads: if one of those is using this mm then this task was also
 490 * using it.
 491 */
 492bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
 493{
 494	struct task_struct *t;
 
 495
 496	for_each_thread(p, t) {
 497		struct mm_struct *t_mm = READ_ONCE(t->mm);
 498		if (t_mm)
 499			return t_mm == mm;
 500	}
 501	return false;
 502}
 503
 504#ifdef CONFIG_MMU
 505/*
 506 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
 507 * victim (if that is possible) to help the OOM killer to move on.
 508 */
 509static struct task_struct *oom_reaper_th;
 510static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
 511static struct task_struct *oom_reaper_list;
 512static DEFINE_SPINLOCK(oom_reaper_lock);
 513
 514bool __oom_reap_task_mm(struct mm_struct *mm)
 515{
 516	struct vm_area_struct *vma;
 517	bool ret = true;
 
 518
 519	/*
 520	 * Tell all users of get_user/copy_from_user etc... that the content
 521	 * is no longer stable. No barriers really needed because unmapping
 522	 * should imply barriers already and the reader would hit a page fault
 523	 * if it stumbled over a reaped memory.
 524	 */
 525	set_bit(MMF_UNSTABLE, &mm->flags);
 526
 527	for (vma = mm->mmap ; vma; vma = vma->vm_next) {
 528		if (!can_madv_lru_vma(vma))
 529			continue;
 
 
 
 
 
 
 
 530
 531		/*
 532		 * Only anonymous pages have a good chance to be dropped
 533		 * without additional steps which we cannot afford as we
 534		 * are OOM already.
 535		 *
 536		 * We do not even care about fs backed pages because all
 537		 * which are reclaimable have already been reclaimed and
 538		 * we do not want to block exit_mmap by keeping mm ref
 539		 * count elevated without a good reason.
 540		 */
 541		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
 542			struct mmu_notifier_range range;
 543			struct mmu_gather tlb;
 544
 545			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
 546						vma, mm, vma->vm_start,
 547						vma->vm_end);
 548			tlb_gather_mmu(&tlb, mm);
 549			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
 550				tlb_finish_mmu(&tlb);
 551				ret = false;
 552				continue;
 553			}
 554			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
 555			mmu_notifier_invalidate_range_end(&range);
 556			tlb_finish_mmu(&tlb);
 557		}
 558	}
 559
 560	return ret;
 561}
 
 562
 563/*
 564 * Reaps the address space of the give task.
 565 *
 566 * Returns true on success and false if none or part of the address space
 567 * has been reclaimed and the caller should retry later.
 568 */
 569static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
 570{
 571	bool ret = true;
 572
 573	if (!mmap_read_trylock(mm)) {
 574		trace_skip_task_reaping(tsk->pid);
 575		return false;
 576	}
 577
 578	/*
 579	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
 580	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
 581	 * under mmap_lock for reading because it serializes against the
 582	 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
 583	 */
 584	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
 585		trace_skip_task_reaping(tsk->pid);
 586		goto out_unlock;
 587	}
 588
 589	trace_start_task_reaping(tsk->pid);
 590
 591	/* failed to reap part of the address space. Try again later */
 592	ret = __oom_reap_task_mm(mm);
 593	if (!ret)
 594		goto out_finish;
 595
 596	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
 597			task_pid_nr(tsk), tsk->comm,
 598			K(get_mm_counter(mm, MM_ANONPAGES)),
 599			K(get_mm_counter(mm, MM_FILEPAGES)),
 600			K(get_mm_counter(mm, MM_SHMEMPAGES)));
 601out_finish:
 602	trace_finish_task_reaping(tsk->pid);
 603out_unlock:
 604	mmap_read_unlock(mm);
 605
 606	return ret;
 607}
 608
 609#define MAX_OOM_REAP_RETRIES 10
 610static void oom_reap_task(struct task_struct *tsk)
 611{
 612	int attempts = 0;
 613	struct mm_struct *mm = tsk->signal->oom_mm;
 614
 615	/* Retry the mmap_read_trylock(mm) a few times */
 616	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
 617		schedule_timeout_idle(HZ/10);
 618
 619	if (attempts <= MAX_OOM_REAP_RETRIES ||
 620	    test_bit(MMF_OOM_SKIP, &mm->flags))
 621		goto done;
 622
 623	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
 624		task_pid_nr(tsk), tsk->comm);
 625	sched_show_task(tsk);
 626	debug_show_all_locks();
 627
 628done:
 629	tsk->oom_reaper_list = NULL;
 630
 631	/*
 632	 * Hide this mm from OOM killer because it has been either reaped or
 633	 * somebody can't call mmap_write_unlock(mm).
 634	 */
 635	set_bit(MMF_OOM_SKIP, &mm->flags);
 
 
 
 
 636
 637	/* Drop a reference taken by wake_oom_reaper */
 638	put_task_struct(tsk);
 639}
 640
 641static int oom_reaper(void *unused)
 642{
 643	while (true) {
 644		struct task_struct *tsk = NULL;
 645
 646		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
 647		spin_lock(&oom_reaper_lock);
 648		if (oom_reaper_list != NULL) {
 649			tsk = oom_reaper_list;
 650			oom_reaper_list = tsk->oom_reaper_list;
 651		}
 652		spin_unlock(&oom_reaper_lock);
 653
 654		if (tsk)
 655			oom_reap_task(tsk);
 656	}
 657
 658	return 0;
 659}
 660
 661static void wake_oom_reaper(struct task_struct *tsk)
 
 
 
 
 662{
 663	/* mm is already queued? */
 664	if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
 665		return;
 666
 667	get_task_struct(tsk);
 668
 669	spin_lock(&oom_reaper_lock);
 670	tsk->oom_reaper_list = oom_reaper_list;
 671	oom_reaper_list = tsk;
 672	spin_unlock(&oom_reaper_lock);
 673	trace_wake_reaper(tsk->pid);
 674	wake_up(&oom_reaper_wait);
 
 
 
 
 
 675}
 676
 677static int __init oom_init(void)
 
 678{
 679	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
 680	return 0;
 681}
 682subsys_initcall(oom_init)
 683#else
 684static inline void wake_oom_reaper(struct task_struct *tsk)
 685{
 686}
 687#endif /* CONFIG_MMU */
 688
 689/**
 690 * mark_oom_victim - mark the given task as OOM victim
 691 * @tsk: task to mark
 692 *
 693 * Has to be called with oom_lock held and never after
 694 * oom has been disabled already.
 695 *
 696 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
 697 * under task_lock or operate on the current).
 698 */
 699static void mark_oom_victim(struct task_struct *tsk)
 700{
 701	struct mm_struct *mm = tsk->mm;
 702
 703	WARN_ON(oom_killer_disabled);
 704	/* OOM killer might race with memcg OOM */
 705	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
 706		return;
 707
 708	/* oom_mm is bound to the signal struct life time. */
 709	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
 710		mmgrab(tsk->signal->oom_mm);
 711		set_bit(MMF_OOM_VICTIM, &mm->flags);
 712	}
 713
 714	/*
 715	 * Make sure that the task is woken up from uninterruptible sleep
 716	 * if it is frozen because OOM killer wouldn't be able to free
 717	 * any memory and livelock. freezing_slow_path will tell the freezer
 718	 * that TIF_MEMDIE tasks should be ignored.
 719	 */
 720	__thaw_task(tsk);
 721	atomic_inc(&oom_victims);
 722	trace_mark_victim(tsk->pid);
 
 
 
 
 723}
 
 724
 725/**
 726 * exit_oom_victim - note the exit of an OOM victim
 727 */
 728void exit_oom_victim(void)
 729{
 730	clear_thread_flag(TIF_MEMDIE);
 731
 732	if (!atomic_dec_return(&oom_victims))
 733		wake_up_all(&oom_victims_wait);
 734}
 735
 736/**
 737 * oom_killer_enable - enable OOM killer
 738 */
 739void oom_killer_enable(void)
 740{
 741	oom_killer_disabled = false;
 742	pr_info("OOM killer enabled.\n");
 743}
 
 744
 745/**
 746 * oom_killer_disable - disable OOM killer
 747 * @timeout: maximum timeout to wait for oom victims in jiffies
 748 *
 749 * Forces all page allocations to fail rather than trigger OOM killer.
 750 * Will block and wait until all OOM victims are killed or the given
 751 * timeout expires.
 752 *
 753 * The function cannot be called when there are runnable user tasks because
 754 * the userspace would see unexpected allocation failures as a result. Any
 755 * new usage of this function should be consulted with MM people.
 756 *
 757 * Returns true if successful and false if the OOM killer cannot be
 758 * disabled.
 759 */
 760bool oom_killer_disable(signed long timeout)
 761{
 762	signed long ret;
 763
 764	/*
 765	 * Make sure to not race with an ongoing OOM killer. Check that the
 766	 * current is not killed (possibly due to sharing the victim's memory).
 767	 */
 768	if (mutex_lock_killable(&oom_lock))
 769		return false;
 770	oom_killer_disabled = true;
 771	mutex_unlock(&oom_lock);
 772
 773	ret = wait_event_interruptible_timeout(oom_victims_wait,
 774			!atomic_read(&oom_victims), timeout);
 775	if (ret <= 0) {
 776		oom_killer_enable();
 777		return false;
 778	}
 779	pr_info("OOM killer disabled.\n");
 780
 781	return true;
 782}
 783
 784static inline bool __task_will_free_mem(struct task_struct *task)
 785{
 786	struct signal_struct *sig = task->signal;
 787
 788	/*
 789	 * A coredumping process may sleep for an extended period in exit_mm(),
 790	 * so the oom killer cannot assume that the process will promptly exit
 791	 * and release memory.
 792	 */
 793	if (sig->flags & SIGNAL_GROUP_COREDUMP)
 794		return false;
 795
 796	if (sig->flags & SIGNAL_GROUP_EXIT)
 797		return true;
 798
 799	if (thread_group_empty(task) && (task->flags & PF_EXITING))
 800		return true;
 801
 802	return false;
 803}
 
 804
 805/*
 806 * Checks whether the given task is dying or exiting and likely to
 807 * release its address space. This means that all threads and processes
 808 * sharing the same mm have to be killed or exiting.
 809 * Caller has to make sure that task->mm is stable (hold task_lock or
 810 * it operates on the current).
 811 */
 812static bool task_will_free_mem(struct task_struct *task)
 813{
 814	struct mm_struct *mm = task->mm;
 815	struct task_struct *p;
 816	bool ret = true;
 817
 818	/*
 819	 * Skip tasks without mm because it might have passed its exit_mm and
 820	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
 821	 * on that for now. We can consider find_lock_task_mm in future.
 822	 */
 823	if (!mm)
 824		return false;
 825
 826	if (!__task_will_free_mem(task))
 827		return false;
 828
 829	/*
 830	 * This task has already been drained by the oom reaper so there are
 831	 * only small chances it will free some more
 832	 */
 833	if (test_bit(MMF_OOM_SKIP, &mm->flags))
 834		return false;
 835
 836	if (atomic_read(&mm->mm_users) <= 1)
 837		return true;
 838
 839	/*
 840	 * Make sure that all tasks which share the mm with the given tasks
 841	 * are dying as well to make sure that a) nobody pins its mm and
 842	 * b) the task is also reapable by the oom reaper.
 843	 */
 844	rcu_read_lock();
 845	for_each_process(p) {
 846		if (!process_shares_mm(p, mm))
 847			continue;
 848		if (same_thread_group(task, p))
 849			continue;
 850		ret = __task_will_free_mem(p);
 851		if (!ret)
 852			break;
 853	}
 854	rcu_read_unlock();
 855
 856	return ret;
 857}
 858
 859static void __oom_kill_process(struct task_struct *victim, const char *message)
 860{
 861	struct task_struct *p;
 862	struct mm_struct *mm;
 863	bool can_oom_reap = true;
 864
 865	p = find_lock_task_mm(victim);
 866	if (!p) {
 867		pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
 868			message, task_pid_nr(victim), victim->comm);
 869		put_task_struct(victim);
 870		return;
 871	} else if (victim != p) {
 872		get_task_struct(p);
 873		put_task_struct(victim);
 874		victim = p;
 875	}
 876
 877	/* Get a reference to safely compare mm after task_unlock(victim) */
 878	mm = victim->mm;
 879	mmgrab(mm);
 880
 881	/* Raise event before sending signal: task reaper must see this */
 882	count_vm_event(OOM_KILL);
 883	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
 884
 885	/*
 886	 * We should send SIGKILL before granting access to memory reserves
 887	 * in order to prevent the OOM victim from depleting the memory
 888	 * reserves from the user space under its control.
 889	 */
 890	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
 891	mark_oom_victim(victim);
 892	pr_err("%s: Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB, UID:%u pgtables:%lukB oom_score_adj:%hd\n",
 893		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
 894		K(get_mm_counter(mm, MM_ANONPAGES)),
 895		K(get_mm_counter(mm, MM_FILEPAGES)),
 896		K(get_mm_counter(mm, MM_SHMEMPAGES)),
 897		from_kuid(&init_user_ns, task_uid(victim)),
 898		mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
 899	task_unlock(victim);
 900
 901	/*
 902	 * Kill all user processes sharing victim->mm in other thread groups, if
 903	 * any.  They don't get access to memory reserves, though, to avoid
 904	 * depletion of all memory.  This prevents mm->mmap_lock livelock when an
 905	 * oom killed thread cannot exit because it requires the semaphore and
 906	 * its contended by another thread trying to allocate memory itself.
 907	 * That thread will now get access to memory reserves since it has a
 908	 * pending fatal signal.
 909	 */
 910	rcu_read_lock();
 911	for_each_process(p) {
 912		if (!process_shares_mm(p, mm))
 913			continue;
 914		if (same_thread_group(p, victim))
 915			continue;
 916		if (is_global_init(p)) {
 917			can_oom_reap = false;
 918			set_bit(MMF_OOM_SKIP, &mm->flags);
 919			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
 920					task_pid_nr(victim), victim->comm,
 921					task_pid_nr(p), p->comm);
 922			continue;
 923		}
 924		/*
 925		 * No kthread_use_mm() user needs to read from the userspace so
 926		 * we are ok to reap it.
 
 927		 */
 928		if (unlikely(p->flags & PF_KTHREAD))
 929			continue;
 930		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
 931	}
 932	rcu_read_unlock();
 933
 934	if (can_oom_reap)
 935		wake_oom_reaper(victim);
 936
 937	mmdrop(mm);
 938	put_task_struct(victim);
 939}
 940#undef K
 941
 942/*
 943 * Kill provided task unless it's secured by setting
 944 * oom_score_adj to OOM_SCORE_ADJ_MIN.
 
 945 */
 946static int oom_kill_memcg_member(struct task_struct *task, void *message)
 947{
 948	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
 949	    !is_global_init(task)) {
 950		get_task_struct(task);
 951		__oom_kill_process(task, message);
 
 
 952	}
 953	return 0;
 954}
 955
 956static void oom_kill_process(struct oom_control *oc, const char *message)
 
 
 
 
 
 957{
 958	struct task_struct *victim = oc->chosen;
 959	struct mem_cgroup *oom_group;
 960	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
 961					      DEFAULT_RATELIMIT_BURST);
 962
 963	/*
 964	 * If the task is already exiting, don't alarm the sysadmin or kill
 965	 * its children or threads, just give it access to memory reserves
 966	 * so it can die quickly
 967	 */
 968	task_lock(victim);
 969	if (task_will_free_mem(victim)) {
 970		mark_oom_victim(victim);
 971		wake_oom_reaper(victim);
 972		task_unlock(victim);
 973		put_task_struct(victim);
 974		return;
 975	}
 976	task_unlock(victim);
 977
 978	if (__ratelimit(&oom_rs))
 979		dump_header(oc, victim);
 980
 981	/*
 982	 * Do we need to kill the entire memory cgroup?
 983	 * Or even one of the ancestor memory cgroups?
 984	 * Check this out before killing the victim task.
 985	 */
 986	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
 987
 988	__oom_kill_process(victim, message);
 989
 990	/*
 991	 * If necessary, kill all tasks in the selected memory cgroup.
 992	 */
 993	if (oom_group) {
 994		mem_cgroup_print_oom_group(oom_group);
 995		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
 996				      (void *)message);
 997		mem_cgroup_put(oom_group);
 998	}
 999}
1000
1001/*
1002 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
 
1003 */
1004static void check_panic_on_oom(struct oom_control *oc)
1005{
1006	if (likely(!sysctl_panic_on_oom))
1007		return;
1008	if (sysctl_panic_on_oom != 2) {
1009		/*
1010		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1011		 * does not panic for cpuset, mempolicy, or memcg allocation
1012		 * failures.
1013		 */
1014		if (oc->constraint != CONSTRAINT_NONE)
1015			return;
1016	}
1017	/* Do not panic for oom kills triggered by sysrq */
1018	if (is_sysrq_oom(oc))
1019		return;
1020	dump_header(oc, NULL);
1021	panic("Out of memory: %s panic_on_oom is enabled\n",
1022		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1023}
1024
1025static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1026
1027int register_oom_notifier(struct notifier_block *nb)
1028{
1029	return blocking_notifier_chain_register(&oom_notify_list, nb);
1030}
1031EXPORT_SYMBOL_GPL(register_oom_notifier);
1032
1033int unregister_oom_notifier(struct notifier_block *nb)
1034{
1035	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 
1036}
1037EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1038
1039/**
1040 * out_of_memory - kill the "best" process when we run out of memory
1041 * @oc: pointer to struct oom_control
 
 
 
1042 *
1043 * If we run out of memory, we have the choice between either
1044 * killing a random task (bad), letting the system crash (worse)
1045 * OR try to be smart about which process to kill. Note that we
1046 * don't have to be perfect here, we just have to be good.
1047 */
1048bool out_of_memory(struct oom_control *oc)
 
1049{
 
 
 
1050	unsigned long freed = 0;
1051
1052	if (oom_killer_disabled)
1053		return false;
1054
1055	if (!is_memcg_oom(oc)) {
1056		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1057		if (freed > 0)
1058			/* Got some memory back in the last second. */
1059			return true;
1060	}
1061
1062	/*
1063	 * If current has a pending SIGKILL or is exiting, then automatically
1064	 * select it.  The goal is to allow it to allocate so that it may
1065	 * quickly exit and free its memory.
1066	 */
1067	if (task_will_free_mem(current)) {
1068		mark_oom_victim(current);
1069		wake_oom_reaper(current);
1070		return true;
1071	}
1072
1073	/*
1074	 * The OOM killer does not compensate for IO-less reclaim.
1075	 * pagefault_out_of_memory lost its gfp context so we have to
1076	 * make sure exclude 0 mask - all other users should have at least
1077	 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1078	 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1079	 */
1080	if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1081		return true;
1082
1083	/*
1084	 * Check if there were limitations on the allocation (only relevant for
1085	 * NUMA and memcg) that may require different handling.
1086	 */
1087	oc->constraint = constrained_alloc(oc);
1088	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1089		oc->nodemask = NULL;
1090	check_panic_on_oom(oc);
1091
1092	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1093	    current->mm && !oom_unkillable_task(current) &&
1094	    oom_cpuset_eligible(current, oc) &&
1095	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1096		get_task_struct(current);
1097		oc->chosen = current;
1098		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1099		return true;
 
 
 
 
 
1100	}
1101
1102	select_bad_process(oc);
1103	/* Found nothing?!?! */
1104	if (!oc->chosen) {
1105		dump_header(oc, NULL);
1106		pr_warn("Out of memory and no killable processes...\n");
1107		/*
1108		 * If we got here due to an actual allocation at the
1109		 * system level, we cannot survive this and will enter
1110		 * an endless loop in the allocator. Bail out now.
1111		 */
1112		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1113			panic("System is deadlocked on memory\n");
1114	}
1115	if (oc->chosen && oc->chosen != (void *)-1UL)
1116		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1117				 "Memory cgroup out of memory");
1118	return !!oc->chosen;
 
 
 
 
 
 
 
 
 
 
1119}
1120
1121/*
1122 * The pagefault handler calls here because it is out of memory, so kill a
1123 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1124 * killing is already in progress so do nothing.
 
1125 */
1126void pagefault_out_of_memory(void)
1127{
1128	struct oom_control oc = {
1129		.zonelist = NULL,
1130		.nodemask = NULL,
1131		.memcg = NULL,
1132		.gfp_mask = 0,
1133		.order = 0,
1134	};
1135
1136	if (mem_cgroup_oom_synchronize(true))
1137		return;
1138
1139	if (!mutex_trylock(&oom_lock))
1140		return;
1141	out_of_memory(&oc);
1142	mutex_unlock(&oom_lock);
1143}