Linux Audio

Check our new training course

Loading...
v5.4
   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/swap.h>
  30#include <linux/timex.h>
  31#include <linux/jiffies.h>
  32#include <linux/cpuset.h>
  33#include <linux/export.h>
  34#include <linux/notifier.h>
  35#include <linux/memcontrol.h>
  36#include <linux/mempolicy.h>
  37#include <linux/security.h>
  38#include <linux/ptrace.h>
  39#include <linux/freezer.h>
  40#include <linux/ftrace.h>
  41#include <linux/ratelimit.h>
  42#include <linux/kthread.h>
  43#include <linux/init.h>
  44#include <linux/mmu_notifier.h>
  45
  46#include <asm/tlb.h>
  47#include "internal.h"
  48#include "slab.h"
  49
  50#define CREATE_TRACE_POINTS
  51#include <trace/events/oom.h>
  52
  53int sysctl_panic_on_oom;
  54int sysctl_oom_kill_allocating_task;
  55int sysctl_oom_dump_tasks = 1;
 
  56
  57/*
  58 * Serializes oom killer invocations (out_of_memory()) from all contexts to
  59 * prevent from over eager oom killing (e.g. when the oom killer is invoked
  60 * from different domains).
  61 *
  62 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
  63 * and mark_oom_victim
 
  64 */
  65DEFINE_MUTEX(oom_lock);
 
 
 
 
 
 
 
 
 
  66
  67static inline bool is_memcg_oom(struct oom_control *oc)
 
 
 
 
 
 
 
 
  68{
  69	return oc->memcg != NULL;
 
 
 
 
 
 
 
 
 
  70}
  71
  72#ifdef CONFIG_NUMA
  73/**
  74 * oom_cpuset_eligible() - check task eligiblity for kill
  75 * @start: task struct of which task to consider
  76 * @oc: pointer to struct oom_control
  77 *
  78 * Task eligibility is determined by whether or not a candidate task, @tsk,
  79 * shares the same mempolicy nodes as current if it is bound by such a policy
  80 * and whether or not it has the same set of allowed cpuset nodes.
  81 *
  82 * This function is assuming oom-killer context and 'current' has triggered
  83 * the oom-killer.
  84 */
  85static bool oom_cpuset_eligible(struct task_struct *start,
  86				struct oom_control *oc)
  87{
  88	struct task_struct *tsk;
  89	bool ret = false;
  90	const nodemask_t *mask = oc->nodemask;
  91
  92	if (is_memcg_oom(oc))
  93		return true;
  94
  95	rcu_read_lock();
  96	for_each_thread(start, tsk) {
  97		if (mask) {
  98			/*
  99			 * If this is a mempolicy constrained oom, tsk's
 100			 * cpuset is irrelevant.  Only return true if its
 101			 * mempolicy intersects current, otherwise it may be
 102			 * needlessly killed.
 103			 */
 104			ret = mempolicy_nodemask_intersects(tsk, mask);
 
 105		} else {
 106			/*
 107			 * This is not a mempolicy constrained oom, so only
 108			 * check the mems of tsk's cpuset.
 109			 */
 110			ret = cpuset_mems_allowed_intersects(current, tsk);
 
 111		}
 112		if (ret)
 113			break;
 114	}
 115	rcu_read_unlock();
 116
 117	return ret;
 118}
 119#else
 120static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
 
 121{
 122	return true;
 123}
 124#endif /* CONFIG_NUMA */
 125
 126/*
 127 * The process p may have detached its own ->mm while exiting or through
 128 * use_mm(), but one or more of its subthreads may still have a valid
 129 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
 130 * task_lock() held.
 131 */
 132struct task_struct *find_lock_task_mm(struct task_struct *p)
 133{
 134	struct task_struct *t;
 135
 136	rcu_read_lock();
 137
 138	for_each_thread(p, t) {
 139		task_lock(t);
 140		if (likely(t->mm))
 141			goto found;
 142		task_unlock(t);
 143	}
 144	t = NULL;
 145found:
 146	rcu_read_unlock();
 147
 148	return t;
 149}
 150
 151/*
 152 * order == -1 means the oom kill is required by sysrq, otherwise only
 153 * for display purposes.
 154 */
 155static inline bool is_sysrq_oom(struct oom_control *oc)
 156{
 157	return oc->order == -1;
 158}
 159
 160/* return true if the task is not adequate as candidate victim task. */
 161static bool oom_unkillable_task(struct task_struct *p)
 
 162{
 163	if (is_global_init(p))
 164		return true;
 165	if (p->flags & PF_KTHREAD)
 166		return true;
 167	return false;
 168}
 169
 170/*
 171 * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
 172 * than all user memory (LRU pages)
 173 */
 174static bool is_dump_unreclaim_slabs(void)
 175{
 176	unsigned long nr_lru;
 177
 178	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
 179		 global_node_page_state(NR_INACTIVE_ANON) +
 180		 global_node_page_state(NR_ACTIVE_FILE) +
 181		 global_node_page_state(NR_INACTIVE_FILE) +
 182		 global_node_page_state(NR_ISOLATED_ANON) +
 183		 global_node_page_state(NR_ISOLATED_FILE) +
 184		 global_node_page_state(NR_UNEVICTABLE);
 185
 186	return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
 187}
 188
 189/**
 190 * oom_badness - heuristic function to determine which candidate task to kill
 191 * @p: task struct of which task we should calculate
 192 * @totalpages: total present RAM allowed for page allocation
 193 *
 194 * The heuristic for determining which task to kill is made to be as simple and
 195 * predictable as possible.  The goal is to return the highest value for the
 196 * task consuming the most memory to avoid subsequent oom failures.
 197 */
 198unsigned long oom_badness(struct task_struct *p, unsigned long totalpages)
 
 199{
 200	long points;
 201	long adj;
 202
 203	if (oom_unkillable_task(p))
 204		return 0;
 205
 206	p = find_lock_task_mm(p);
 207	if (!p)
 208		return 0;
 209
 210	/*
 211	 * Do not even consider tasks which are explicitly marked oom
 212	 * unkillable or have been already oom reaped or the are in
 213	 * the middle of vfork
 214	 */
 215	adj = (long)p->signal->oom_score_adj;
 216	if (adj == OOM_SCORE_ADJ_MIN ||
 217			test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
 218			in_vfork(p)) {
 219		task_unlock(p);
 220		return 0;
 221	}
 222
 223	/*
 224	 * The baseline for the badness score is the proportion of RAM that each
 225	 * task's rss, pagetable and swap space use.
 226	 */
 227	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
 228		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
 229	task_unlock(p);
 230
 
 
 
 
 
 
 
 231	/* Normalize to oom_score_adj units */
 232	adj *= totalpages / 1000;
 233	points += adj;
 234
 235	/*
 236	 * Never return 0 for an eligible task regardless of the root bonus and
 237	 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
 238	 */
 239	return points > 0 ? points : 1;
 240}
 241
 242static const char * const oom_constraint_text[] = {
 243	[CONSTRAINT_NONE] = "CONSTRAINT_NONE",
 244	[CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
 245	[CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
 246	[CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
 247};
 248
 249/*
 250 * Determine the type of allocation constraint.
 251 */
 252static enum oom_constraint constrained_alloc(struct oom_control *oc)
 
 
 
 253{
 254	struct zone *zone;
 255	struct zoneref *z;
 256	enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
 257	bool cpuset_limited = false;
 258	int nid;
 259
 260	if (is_memcg_oom(oc)) {
 261		oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
 262		return CONSTRAINT_MEMCG;
 263	}
 264
 265	/* Default to all available memory */
 266	oc->totalpages = totalram_pages() + total_swap_pages;
 267
 268	if (!IS_ENABLED(CONFIG_NUMA))
 269		return CONSTRAINT_NONE;
 270
 271	if (!oc->zonelist)
 272		return CONSTRAINT_NONE;
 273	/*
 274	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
 275	 * to kill current.We have to random task kill in this case.
 276	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
 277	 */
 278	if (oc->gfp_mask & __GFP_THISNODE)
 279		return CONSTRAINT_NONE;
 280
 281	/*
 282	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
 283	 * the page allocator means a mempolicy is in effect.  Cpuset policy
 284	 * is enforced in get_page_from_freelist().
 285	 */
 286	if (oc->nodemask &&
 287	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
 288		oc->totalpages = total_swap_pages;
 289		for_each_node_mask(nid, *oc->nodemask)
 290			oc->totalpages += node_present_pages(nid);
 291		return CONSTRAINT_MEMORY_POLICY;
 292	}
 293
 294	/* Check this allocation failure is caused by cpuset's wall function */
 295	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
 296			high_zoneidx, oc->nodemask)
 297		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
 298			cpuset_limited = true;
 299
 300	if (cpuset_limited) {
 301		oc->totalpages = total_swap_pages;
 302		for_each_node_mask(nid, cpuset_current_mems_allowed)
 303			oc->totalpages += node_present_pages(nid);
 304		return CONSTRAINT_CPUSET;
 305	}
 306	return CONSTRAINT_NONE;
 307}
 308
 309static int oom_evaluate_task(struct task_struct *task, void *arg)
 
 
 310{
 311	struct oom_control *oc = arg;
 312	unsigned long points;
 313
 314	if (oom_unkillable_task(task))
 315		goto next;
 316
 317	/* p may not have freeable memory in nodemask */
 318	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
 319		goto next;
 320
 321	/*
 322	 * This task already has access to memory reserves and is being killed.
 323	 * Don't allow any other task to have access to the reserves unless
 324	 * the task has MMF_OOM_SKIP because chances that it would release
 325	 * any memory is quite low.
 326	 */
 327	if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
 328		if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
 329			goto next;
 330		goto abort;
 331	}
 332
 333	/*
 334	 * If task is allocating a lot of memory and has been marked to be
 335	 * killed first if it triggers an oom, then select it.
 336	 */
 337	if (oom_task_origin(task)) {
 338		points = ULONG_MAX;
 339		goto select;
 340	}
 341
 342	points = oom_badness(task, oc->totalpages);
 343	if (!points || points < oc->chosen_points)
 344		goto next;
 345
 346select:
 347	if (oc->chosen)
 348		put_task_struct(oc->chosen);
 349	get_task_struct(task);
 350	oc->chosen = task;
 351	oc->chosen_points = points;
 352next:
 353	return 0;
 354abort:
 355	if (oc->chosen)
 356		put_task_struct(oc->chosen);
 357	oc->chosen = (void *)-1UL;
 358	return 1;
 359}
 
 360
 361/*
 362 * Simple selection loop. We choose the process with the highest number of
 363 * 'points'. In case scan was aborted, oc->chosen is set to -1.
 
 
 364 */
 365static void select_bad_process(struct oom_control *oc)
 
 
 366{
 367	if (is_memcg_oom(oc))
 368		mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
 369	else {
 370		struct task_struct *p;
 371
 372		rcu_read_lock();
 373		for_each_process(p)
 374			if (oom_evaluate_task(p, oc))
 375				break;
 376		rcu_read_unlock();
 377	}
 378}
 379
 380static int dump_task(struct task_struct *p, void *arg)
 381{
 382	struct oom_control *oc = arg;
 383	struct task_struct *task;
 384
 385	if (oom_unkillable_task(p))
 386		return 0;
 387
 388	/* p may not have freeable memory in nodemask */
 389	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
 390		return 0;
 
 391
 392	task = find_lock_task_mm(p);
 393	if (!task) {
 394		/*
 395		 * This is a kthread or all of p's threads have already
 396		 * detached their mm's.  There's no need to report
 397		 * them; they can't be oom killed anyway.
 
 
 
 
 398		 */
 399		return 0;
 400	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 401
 402	pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu         %5hd %s\n",
 403		task->pid, from_kuid(&init_user_ns, task_uid(task)),
 404		task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
 405		mm_pgtables_bytes(task->mm),
 406		get_mm_counter(task->mm, MM_SWAPENTS),
 407		task->signal->oom_score_adj, task->comm);
 408	task_unlock(task);
 409
 410	return 0;
 
 411}
 412
 413/**
 414 * dump_tasks - dump current memory state of all system tasks
 415 * @oc: pointer to struct oom_control
 
 416 *
 417 * Dumps the current memory state of all eligible tasks.  Tasks not in the same
 418 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
 419 * are not shown.
 420 * State information includes task's pid, uid, tgid, vm size, rss,
 421 * pgtables_bytes, swapents, oom_score_adj value, and name.
 
 
 422 */
 423static void dump_tasks(struct oom_control *oc)
 424{
 425	pr_info("Tasks state (memory values in pages):\n");
 426	pr_info("[  pid  ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj name\n");
 427
 428	if (is_memcg_oom(oc))
 429		mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
 430	else {
 431		struct task_struct *p;
 432
 433		rcu_read_lock();
 434		for_each_process(p)
 435			dump_task(p, oc);
 436		rcu_read_unlock();
 
 
 
 
 
 
 
 
 
 
 
 
 437	}
 438}
 439
 440static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
 441{
 442	/* one line summary of the oom killer context. */
 443	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
 444			oom_constraint_text[oc->constraint],
 445			nodemask_pr_args(oc->nodemask));
 446	cpuset_print_current_mems_allowed();
 447	mem_cgroup_print_oom_context(oc->memcg, victim);
 448	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
 449		from_kuid(&init_user_ns, task_uid(victim)));
 450}
 451
 452static void dump_header(struct oom_control *oc, struct task_struct *p)
 453{
 454	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
 455		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
 456			current->signal->oom_score_adj);
 457	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
 458		pr_warn("COMPACTION is disabled!!!\n");
 459
 460	dump_stack();
 461	if (is_memcg_oom(oc))
 462		mem_cgroup_print_oom_meminfo(oc->memcg);
 463	else {
 464		show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
 465		if (is_dump_unreclaim_slabs())
 466			dump_unreclaimable_slab();
 467	}
 468	if (sysctl_oom_dump_tasks)
 469		dump_tasks(oc);
 470	if (p)
 471		dump_oom_summary(oc, p);
 472}
 473
 474/*
 475 * Number of OOM victims in flight
 476 */
 477static atomic_t oom_victims = ATOMIC_INIT(0);
 478static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
 479
 480static bool oom_killer_disabled __read_mostly;
 481
 482#define K(x) ((x) << (PAGE_SHIFT-10))
 
 
 
 
 
 
 
 
 
 
 
 
 483
 484/*
 485 * task->mm can be NULL if the task is the exited group leader.  So to
 486 * determine whether the task is using a particular mm, we examine all the
 487 * task's threads: if one of those is using this mm then this task was also
 488 * using it.
 489 */
 490bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
 491{
 492	struct task_struct *t;
 493
 494	for_each_thread(p, t) {
 495		struct mm_struct *t_mm = READ_ONCE(t->mm);
 496		if (t_mm)
 497			return t_mm == mm;
 498	}
 499	return false;
 500}
 501
 502#ifdef CONFIG_MMU
 503/*
 504 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
 505 * victim (if that is possible) to help the OOM killer to move on.
 506 */
 507static struct task_struct *oom_reaper_th;
 508static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
 509static struct task_struct *oom_reaper_list;
 510static DEFINE_SPINLOCK(oom_reaper_lock);
 511
 512bool __oom_reap_task_mm(struct mm_struct *mm)
 513{
 514	struct vm_area_struct *vma;
 515	bool ret = true;
 516
 517	/*
 518	 * Tell all users of get_user/copy_from_user etc... that the content
 519	 * is no longer stable. No barriers really needed because unmapping
 520	 * should imply barriers already and the reader would hit a page fault
 521	 * if it stumbled over a reaped memory.
 522	 */
 523	set_bit(MMF_UNSTABLE, &mm->flags);
 
 
 524
 525	for (vma = mm->mmap ; vma; vma = vma->vm_next) {
 526		if (!can_madv_lru_vma(vma))
 527			continue;
 528
 529		/*
 530		 * Only anonymous pages have a good chance to be dropped
 531		 * without additional steps which we cannot afford as we
 532		 * are OOM already.
 533		 *
 534		 * We do not even care about fs backed pages because all
 535		 * which are reclaimable have already been reclaimed and
 536		 * we do not want to block exit_mmap by keeping mm ref
 537		 * count elevated without a good reason.
 538		 */
 539		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
 540			struct mmu_notifier_range range;
 541			struct mmu_gather tlb;
 542
 543			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
 544						vma, mm, vma->vm_start,
 545						vma->vm_end);
 546			tlb_gather_mmu(&tlb, mm, range.start, range.end);
 547			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
 548				tlb_finish_mmu(&tlb, range.start, range.end);
 549				ret = false;
 550				continue;
 
 
 
 
 
 
 
 
 551			}
 552			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
 553			mmu_notifier_invalidate_range_end(&range);
 554			tlb_finish_mmu(&tlb, range.start, range.end);
 555		}
 556	}
 557
 558	return ret;
 559}
 560
 561/*
 562 * Reaps the address space of the give task.
 563 *
 564 * Returns true on success and false if none or part of the address space
 565 * has been reclaimed and the caller should retry later.
 566 */
 567static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
 568{
 569	bool ret = true;
 570
 571	if (!down_read_trylock(&mm->mmap_sem)) {
 572		trace_skip_task_reaping(tsk->pid);
 573		return false;
 574	}
 575
 576	/*
 577	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
 578	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
 579	 * under mmap_sem for reading because it serializes against the
 580	 * down_write();up_write() cycle in exit_mmap().
 581	 */
 582	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
 583		trace_skip_task_reaping(tsk->pid);
 584		goto out_unlock;
 585	}
 586
 587	trace_start_task_reaping(tsk->pid);
 588
 589	/* failed to reap part of the address space. Try again later */
 590	ret = __oom_reap_task_mm(mm);
 591	if (!ret)
 592		goto out_finish;
 593
 594	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
 595			task_pid_nr(tsk), tsk->comm,
 596			K(get_mm_counter(mm, MM_ANONPAGES)),
 597			K(get_mm_counter(mm, MM_FILEPAGES)),
 598			K(get_mm_counter(mm, MM_SHMEMPAGES)));
 599out_finish:
 600	trace_finish_task_reaping(tsk->pid);
 601out_unlock:
 602	up_read(&mm->mmap_sem);
 603
 604	return ret;
 605}
 606
 607#define MAX_OOM_REAP_RETRIES 10
 608static void oom_reap_task(struct task_struct *tsk)
 609{
 610	int attempts = 0;
 611	struct mm_struct *mm = tsk->signal->oom_mm;
 612
 613	/* Retry the down_read_trylock(mmap_sem) a few times */
 614	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
 615		schedule_timeout_idle(HZ/10);
 616
 617	if (attempts <= MAX_OOM_REAP_RETRIES ||
 618	    test_bit(MMF_OOM_SKIP, &mm->flags))
 619		goto done;
 620
 621	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
 622		task_pid_nr(tsk), tsk->comm);
 623	debug_show_all_locks();
 624
 625done:
 626	tsk->oom_reaper_list = NULL;
 
 
 
 
 
 627
 628	/*
 629	 * Hide this mm from OOM killer because it has been either reaped or
 630	 * somebody can't call up_write(mmap_sem).
 
 
 
 
 
 631	 */
 632	set_bit(MMF_OOM_SKIP, &mm->flags);
 633
 634	/* Drop a reference taken by wake_oom_reaper */
 635	put_task_struct(tsk);
 636}
 637
 638static int oom_reaper(void *unused)
 639{
 640	while (true) {
 641		struct task_struct *tsk = NULL;
 642
 643		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
 644		spin_lock(&oom_reaper_lock);
 645		if (oom_reaper_list != NULL) {
 646			tsk = oom_reaper_list;
 647			oom_reaper_list = tsk->oom_reaper_list;
 648		}
 649		spin_unlock(&oom_reaper_lock);
 650
 651		if (tsk)
 652			oom_reap_task(tsk);
 653	}
 654
 655	return 0;
 
 656}
 
 657
 658static void wake_oom_reaper(struct task_struct *tsk)
 659{
 660	/* mm is already queued? */
 661	if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
 662		return;
 663
 664	get_task_struct(tsk);
 665
 666	spin_lock(&oom_reaper_lock);
 667	tsk->oom_reaper_list = oom_reaper_list;
 668	oom_reaper_list = tsk;
 669	spin_unlock(&oom_reaper_lock);
 670	trace_wake_reaper(tsk->pid);
 671	wake_up(&oom_reaper_wait);
 672}
 673
 674static int __init oom_init(void)
 675{
 676	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
 677	return 0;
 678}
 679subsys_initcall(oom_init)
 680#else
 681static inline void wake_oom_reaper(struct task_struct *tsk)
 682{
 683}
 684#endif /* CONFIG_MMU */
 685
 686/**
 687 * mark_oom_victim - mark the given task as OOM victim
 688 * @tsk: task to mark
 689 *
 690 * Has to be called with oom_lock held and never after
 691 * oom has been disabled already.
 692 *
 693 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
 694 * under task_lock or operate on the current).
 695 */
 696static void mark_oom_victim(struct task_struct *tsk)
 
 697{
 698	struct mm_struct *mm = tsk->mm;
 699
 700	WARN_ON(oom_killer_disabled);
 701	/* OOM killer might race with memcg OOM */
 702	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
 703		return;
 704
 705	/* oom_mm is bound to the signal struct life time. */
 706	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
 707		mmgrab(tsk->signal->oom_mm);
 708		set_bit(MMF_OOM_VICTIM, &mm->flags);
 
 
 
 709	}
 710
 711	/*
 712	 * Make sure that the task is woken up from uninterruptible sleep
 713	 * if it is frozen because OOM killer wouldn't be able to free
 714	 * any memory and livelock. freezing_slow_path will tell the freezer
 715	 * that TIF_MEMDIE tasks should be ignored.
 716	 */
 717	__thaw_task(tsk);
 718	atomic_inc(&oom_victims);
 719	trace_mark_victim(tsk->pid);
 720}
 721
 722/**
 723 * exit_oom_victim - note the exit of an OOM victim
 724 */
 725void exit_oom_victim(void)
 726{
 727	clear_thread_flag(TIF_MEMDIE);
 728
 729	if (!atomic_dec_return(&oom_victims))
 730		wake_up_all(&oom_victims_wait);
 731}
 732
 733/**
 734 * oom_killer_enable - enable OOM killer
 735 */
 736void oom_killer_enable(void)
 737{
 738	oom_killer_disabled = false;
 739	pr_info("OOM killer enabled.\n");
 740}
 741
 742/**
 743 * oom_killer_disable - disable OOM killer
 744 * @timeout: maximum timeout to wait for oom victims in jiffies
 745 *
 746 * Forces all page allocations to fail rather than trigger OOM killer.
 747 * Will block and wait until all OOM victims are killed or the given
 748 * timeout expires.
 749 *
 750 * The function cannot be called when there are runnable user tasks because
 751 * the userspace would see unexpected allocation failures as a result. Any
 752 * new usage of this function should be consulted with MM people.
 753 *
 754 * Returns true if successful and false if the OOM killer cannot be
 755 * disabled.
 756 */
 757bool oom_killer_disable(signed long timeout)
 758{
 759	signed long ret;
 
 
 760
 761	/*
 762	 * Make sure to not race with an ongoing OOM killer. Check that the
 763	 * current is not killed (possibly due to sharing the victim's memory).
 
 764	 */
 765	if (mutex_lock_killable(&oom_lock))
 766		return false;
 767	oom_killer_disabled = true;
 768	mutex_unlock(&oom_lock);
 769
 770	ret = wait_event_interruptible_timeout(oom_victims_wait,
 771			!atomic_read(&oom_victims), timeout);
 772	if (ret <= 0) {
 773		oom_killer_enable();
 774		return false;
 775	}
 776	pr_info("OOM killer disabled.\n");
 777
 778	return true;
 
 
 
 
 
 
 
 779}
 
 780
 781static inline bool __task_will_free_mem(struct task_struct *task)
 782{
 783	struct signal_struct *sig = task->signal;
 784
 785	/*
 786	 * A coredumping process may sleep for an extended period in exit_mm(),
 787	 * so the oom killer cannot assume that the process will promptly exit
 788	 * and release memory.
 789	 */
 790	if (sig->flags & SIGNAL_GROUP_COREDUMP)
 791		return false;
 792
 793	if (sig->flags & SIGNAL_GROUP_EXIT)
 794		return true;
 795
 796	if (thread_group_empty(task) && (task->flags & PF_EXITING))
 797		return true;
 798
 799	return false;
 
 
 800}
 
 801
 802/*
 803 * Checks whether the given task is dying or exiting and likely to
 804 * release its address space. This means that all threads and processes
 805 * sharing the same mm have to be killed or exiting.
 806 * Caller has to make sure that task->mm is stable (hold task_lock or
 807 * it operates on the current).
 808 */
 809static bool task_will_free_mem(struct task_struct *task)
 810{
 811	struct mm_struct *mm = task->mm;
 812	struct task_struct *p;
 813	bool ret = true;
 814
 815	/*
 816	 * Skip tasks without mm because it might have passed its exit_mm and
 817	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
 818	 * on that for now. We can consider find_lock_task_mm in future.
 819	 */
 820	if (!mm)
 821		return false;
 822
 823	if (!__task_will_free_mem(task))
 824		return false;
 825
 826	/*
 827	 * This task has already been drained by the oom reaper so there are
 828	 * only small chances it will free some more
 829	 */
 830	if (test_bit(MMF_OOM_SKIP, &mm->flags))
 831		return false;
 832
 833	if (atomic_read(&mm->mm_users) <= 1)
 834		return true;
 835
 836	/*
 837	 * Make sure that all tasks which share the mm with the given tasks
 838	 * are dying as well to make sure that a) nobody pins its mm and
 839	 * b) the task is also reapable by the oom reaper.
 840	 */
 841	rcu_read_lock();
 842	for_each_process(p) {
 843		if (!process_shares_mm(p, mm))
 844			continue;
 845		if (same_thread_group(task, p))
 846			continue;
 847		ret = __task_will_free_mem(p);
 848		if (!ret)
 849			break;
 850	}
 851	rcu_read_unlock();
 852
 853	return ret;
 854}
 
 855
 856static void __oom_kill_process(struct task_struct *victim, const char *message)
 
 
 
 
 
 857{
 858	struct task_struct *p;
 859	struct mm_struct *mm;
 860	bool can_oom_reap = true;
 861
 862	p = find_lock_task_mm(victim);
 863	if (!p) {
 864		put_task_struct(victim);
 865		return;
 866	} else if (victim != p) {
 867		get_task_struct(p);
 868		put_task_struct(victim);
 869		victim = p;
 870	}
 871
 872	/* Get a reference to safely compare mm after task_unlock(victim) */
 873	mm = victim->mm;
 874	mmgrab(mm);
 875
 876	/* Raise event before sending signal: task reaper must see this */
 877	count_vm_event(OOM_KILL);
 878	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
 879
 880	/*
 881	 * We should send SIGKILL before granting access to memory reserves
 882	 * in order to prevent the OOM victim from depleting the memory
 883	 * reserves from the user space under its control.
 884	 */
 885	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
 886	mark_oom_victim(victim);
 887	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",
 888		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
 889		K(get_mm_counter(mm, MM_ANONPAGES)),
 890		K(get_mm_counter(mm, MM_FILEPAGES)),
 891		K(get_mm_counter(mm, MM_SHMEMPAGES)),
 892		from_kuid(&init_user_ns, task_uid(victim)),
 893		mm_pgtables_bytes(mm), victim->signal->oom_score_adj);
 894	task_unlock(victim);
 895
 896	/*
 897	 * Kill all user processes sharing victim->mm in other thread groups, if
 898	 * any.  They don't get access to memory reserves, though, to avoid
 899	 * depletion of all memory.  This prevents mm->mmap_sem livelock when an
 900	 * oom killed thread cannot exit because it requires the semaphore and
 901	 * its contended by another thread trying to allocate memory itself.
 902	 * That thread will now get access to memory reserves since it has a
 903	 * pending fatal signal.
 904	 */
 905	rcu_read_lock();
 906	for_each_process(p) {
 907		if (!process_shares_mm(p, mm))
 908			continue;
 909		if (same_thread_group(p, victim))
 910			continue;
 911		if (is_global_init(p)) {
 912			can_oom_reap = false;
 913			set_bit(MMF_OOM_SKIP, &mm->flags);
 914			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
 915					task_pid_nr(victim), victim->comm,
 916					task_pid_nr(p), p->comm);
 917			continue;
 918		}
 919		/*
 920		 * No use_mm() user needs to read from the userspace so we are
 921		 * ok to reap it.
 
 922		 */
 923		if (unlikely(p->flags & PF_KTHREAD))
 924			continue;
 925		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
 926	}
 927	rcu_read_unlock();
 928
 929	if (can_oom_reap)
 930		wake_oom_reaper(victim);
 931
 932	mmdrop(mm);
 933	put_task_struct(victim);
 934}
 935#undef K
 936
 937/*
 938 * Kill provided task unless it's secured by setting
 939 * oom_score_adj to OOM_SCORE_ADJ_MIN.
 
 940 */
 941static int oom_kill_memcg_member(struct task_struct *task, void *message)
 942{
 943	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
 944	    !is_global_init(task)) {
 945		get_task_struct(task);
 946		__oom_kill_process(task, message);
 947	}
 948	return 0;
 949}
 950
 951static void oom_kill_process(struct oom_control *oc, const char *message)
 952{
 953	struct task_struct *victim = oc->chosen;
 954	struct mem_cgroup *oom_group;
 955	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
 956					      DEFAULT_RATELIMIT_BURST);
 957
 958	/*
 959	 * If the task is already exiting, don't alarm the sysadmin or kill
 960	 * its children or threads, just give it access to memory reserves
 961	 * so it can die quickly
 962	 */
 963	task_lock(victim);
 964	if (task_will_free_mem(victim)) {
 965		mark_oom_victim(victim);
 966		wake_oom_reaper(victim);
 967		task_unlock(victim);
 968		put_task_struct(victim);
 969		return;
 970	}
 971	task_unlock(victim);
 972
 973	if (__ratelimit(&oom_rs))
 974		dump_header(oc, victim);
 975
 976	/*
 977	 * Do we need to kill the entire memory cgroup?
 978	 * Or even one of the ancestor memory cgroups?
 979	 * Check this out before killing the victim task.
 980	 */
 981	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
 982
 983	__oom_kill_process(victim, message);
 984
 985	/*
 986	 * If necessary, kill all tasks in the selected memory cgroup.
 987	 */
 988	if (oom_group) {
 989		mem_cgroup_print_oom_group(oom_group);
 990		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
 991				      (void*)message);
 992		mem_cgroup_put(oom_group);
 993	}
 
 994}
 995
 996/*
 997 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
 
 
 998 */
 999static void check_panic_on_oom(struct oom_control *oc)
1000{
1001	if (likely(!sysctl_panic_on_oom))
1002		return;
1003	if (sysctl_panic_on_oom != 2) {
1004		/*
1005		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1006		 * does not panic for cpuset, mempolicy, or memcg allocation
1007		 * failures.
1008		 */
1009		if (oc->constraint != CONSTRAINT_NONE)
1010			return;
1011	}
1012	/* Do not panic for oom kills triggered by sysrq */
1013	if (is_sysrq_oom(oc))
1014		return;
1015	dump_header(oc, NULL);
1016	panic("Out of memory: %s panic_on_oom is enabled\n",
1017		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1018}
1019
1020static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1021
1022int register_oom_notifier(struct notifier_block *nb)
1023{
1024	return blocking_notifier_chain_register(&oom_notify_list, nb);
 
 
 
 
 
 
 
 
1025}
1026EXPORT_SYMBOL_GPL(register_oom_notifier);
1027
1028int unregister_oom_notifier(struct notifier_block *nb)
 
 
 
 
1029{
1030	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 
 
 
 
 
1031}
1032EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1033
1034/**
1035 * out_of_memory - kill the "best" process when we run out of memory
1036 * @oc: pointer to struct oom_control
 
 
 
 
1037 *
1038 * If we run out of memory, we have the choice between either
1039 * killing a random task (bad), letting the system crash (worse)
1040 * OR try to be smart about which process to kill. Note that we
1041 * don't have to be perfect here, we just have to be good.
1042 */
1043bool out_of_memory(struct oom_control *oc)
 
1044{
 
 
 
1045	unsigned long freed = 0;
1046
1047	if (oom_killer_disabled)
1048		return false;
1049
1050	if (!is_memcg_oom(oc)) {
1051		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1052		if (freed > 0)
1053			/* Got some memory back in the last second. */
1054			return true;
1055	}
1056
1057	/*
1058	 * If current has a pending SIGKILL or is exiting, then automatically
1059	 * select it.  The goal is to allow it to allocate so that it may
1060	 * quickly exit and free its memory.
1061	 */
1062	if (task_will_free_mem(current)) {
1063		mark_oom_victim(current);
1064		wake_oom_reaper(current);
1065		return true;
1066	}
1067
1068	/*
1069	 * The OOM killer does not compensate for IO-less reclaim.
1070	 * pagefault_out_of_memory lost its gfp context so we have to
1071	 * make sure exclude 0 mask - all other users should have at least
1072	 * ___GFP_DIRECT_RECLAIM to get here. But mem_cgroup_oom() has to
1073	 * invoke the OOM killer even if it is a GFP_NOFS allocation.
1074	 */
1075	if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1076		return true;
1077
1078	/*
1079	 * Check if there were limitations on the allocation (only relevant for
1080	 * NUMA and memcg) that may require different handling.
1081	 */
1082	oc->constraint = constrained_alloc(oc);
1083	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1084		oc->nodemask = NULL;
1085	check_panic_on_oom(oc);
1086
1087	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1088	    current->mm && !oom_unkillable_task(current) &&
1089	    oom_cpuset_eligible(current, oc) &&
1090	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1091		get_task_struct(current);
1092		oc->chosen = current;
1093		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1094		return true;
 
 
 
 
 
 
 
 
 
1095	}
1096
1097	select_bad_process(oc);
1098	/* Found nothing?!?! */
1099	if (!oc->chosen) {
1100		dump_header(oc, NULL);
1101		pr_warn("Out of memory and no killable processes...\n");
1102		/*
1103		 * If we got here due to an actual allocation at the
1104		 * system level, we cannot survive this and will enter
1105		 * an endless loop in the allocator. Bail out now.
1106		 */
1107		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1108			panic("System is deadlocked on memory\n");
1109	}
1110	if (oc->chosen && oc->chosen != (void *)-1UL)
1111		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1112				 "Memory cgroup out of memory");
1113	return !!oc->chosen;
 
 
 
 
 
1114}
1115
1116/*
1117 * The pagefault handler calls here because it is out of memory, so kill a
1118 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
1119 * killing is already in progress so do nothing.
 
1120 */
1121void pagefault_out_of_memory(void)
1122{
1123	struct oom_control oc = {
1124		.zonelist = NULL,
1125		.nodemask = NULL,
1126		.memcg = NULL,
1127		.gfp_mask = 0,
1128		.order = 0,
1129	};
1130
1131	if (mem_cgroup_oom_synchronize(true))
1132		return;
1133
1134	if (!mutex_trylock(&oom_lock))
1135		return;
1136	out_of_memory(&oc);
1137	mutex_unlock(&oom_lock);
1138}
v3.5.6
 
  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/export.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#include <linux/freezer.h>
 36#include <linux/ftrace.h>
 37#include <linux/ratelimit.h>
 
 
 
 
 
 
 
 38
 39#define CREATE_TRACE_POINTS
 40#include <trace/events/oom.h>
 41
 42int sysctl_panic_on_oom;
 43int sysctl_oom_kill_allocating_task;
 44int sysctl_oom_dump_tasks = 1;
 45static DEFINE_SPINLOCK(zone_scan_lock);
 46
 47/*
 48 * compare_swap_oom_score_adj() - compare and swap current's oom_score_adj
 49 * @old_val: old oom_score_adj for compare
 50 * @new_val: new oom_score_adj for swap
 51 *
 52 * Sets the oom_score_adj value for current to @new_val iff its present value is
 53 * @old_val.  Usually used to reinstate a previous value to prevent racing with
 54 * userspacing tuning the value in the interim.
 55 */
 56void compare_swap_oom_score_adj(int old_val, int new_val)
 57{
 58	struct sighand_struct *sighand = current->sighand;
 59
 60	spin_lock_irq(&sighand->siglock);
 61	if (current->signal->oom_score_adj == old_val)
 62		current->signal->oom_score_adj = new_val;
 63	trace_oom_score_adj_update(current);
 64	spin_unlock_irq(&sighand->siglock);
 65}
 66
 67/**
 68 * test_set_oom_score_adj() - set current's oom_score_adj and return old value
 69 * @new_val: new oom_score_adj value
 70 *
 71 * Sets the oom_score_adj value for current to @new_val with proper
 72 * synchronization and returns the old value.  Usually used to temporarily
 73 * set a value, save the old value in the caller, and then reinstate it later.
 74 */
 75int test_set_oom_score_adj(int new_val)
 76{
 77	struct sighand_struct *sighand = current->sighand;
 78	int old_val;
 79
 80	spin_lock_irq(&sighand->siglock);
 81	old_val = current->signal->oom_score_adj;
 82	current->signal->oom_score_adj = new_val;
 83	trace_oom_score_adj_update(current);
 84	spin_unlock_irq(&sighand->siglock);
 85
 86	return old_val;
 87}
 88
 89#ifdef CONFIG_NUMA
 90/**
 91 * has_intersects_mems_allowed() - check task eligiblity for kill
 92 * @tsk: task struct of which task to consider
 93 * @mask: nodemask passed to page allocator for mempolicy ooms
 94 *
 95 * Task eligibility is determined by whether or not a candidate task, @tsk,
 96 * shares the same mempolicy nodes as current if it is bound by such a policy
 97 * and whether or not it has the same set of allowed cpuset nodes.
 
 
 
 98 */
 99static bool has_intersects_mems_allowed(struct task_struct *tsk,
100					const nodemask_t *mask)
101{
102	struct task_struct *start = tsk;
 
 
103
104	do {
 
 
 
 
105		if (mask) {
106			/*
107			 * If this is a mempolicy constrained oom, tsk's
108			 * cpuset is irrelevant.  Only return true if its
109			 * mempolicy intersects current, otherwise it may be
110			 * needlessly killed.
111			 */
112			if (mempolicy_nodemask_intersects(tsk, mask))
113				return true;
114		} else {
115			/*
116			 * This is not a mempolicy constrained oom, so only
117			 * check the mems of tsk's cpuset.
118			 */
119			if (cpuset_mems_allowed_intersects(current, tsk))
120				return true;
121		}
122	} while_each_thread(start, tsk);
 
 
 
123
124	return false;
125}
126#else
127static bool has_intersects_mems_allowed(struct task_struct *tsk,
128					const nodemask_t *mask)
129{
130	return true;
131}
132#endif /* CONFIG_NUMA */
133
134/*
135 * The process p may have detached its own ->mm while exiting or through
136 * use_mm(), but one or more of its subthreads may still have a valid
137 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
138 * task_lock() held.
139 */
140struct task_struct *find_lock_task_mm(struct task_struct *p)
141{
142	struct task_struct *t = p;
 
 
143
144	do {
145		task_lock(t);
146		if (likely(t->mm))
147			return t;
148		task_unlock(t);
149	} while_each_thread(p, t);
 
 
 
150
151	return NULL;
 
 
 
 
 
 
 
 
 
152}
153
154/* return true if the task is not adequate as candidate victim task. */
155static bool oom_unkillable_task(struct task_struct *p,
156		const struct mem_cgroup *memcg, const nodemask_t *nodemask)
157{
158	if (is_global_init(p))
159		return true;
160	if (p->flags & PF_KTHREAD)
161		return true;
 
 
162
163	/* When mem_cgroup_out_of_memory() and p is not member of the group */
164	if (memcg && !task_in_mem_cgroup(p, memcg))
165		return true;
 
 
 
 
166
167	/* p may not have freeable memory in nodemask */
168	if (!has_intersects_mems_allowed(p, nodemask))
169		return true;
 
 
 
 
170
171	return false;
172}
173
174/**
175 * oom_badness - heuristic function to determine which candidate task to kill
176 * @p: task struct of which task we should calculate
177 * @totalpages: total present RAM allowed for page allocation
178 *
179 * The heuristic for determining which task to kill is made to be as simple and
180 * predictable as possible.  The goal is to return the highest value for the
181 * task consuming the most memory to avoid subsequent oom failures.
182 */
183unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
184			  const nodemask_t *nodemask, unsigned long totalpages)
185{
186	long points;
187	long adj;
188
189	if (oom_unkillable_task(p, memcg, nodemask))
190		return 0;
191
192	p = find_lock_task_mm(p);
193	if (!p)
194		return 0;
195
196	adj = p->signal->oom_score_adj;
197	if (adj == OOM_SCORE_ADJ_MIN) {
 
 
 
 
 
 
 
198		task_unlock(p);
199		return 0;
200	}
201
202	/*
203	 * The baseline for the badness score is the proportion of RAM that each
204	 * task's rss, pagetable and swap space use.
205	 */
206	points = get_mm_rss(p->mm) + p->mm->nr_ptes +
207		 get_mm_counter(p->mm, MM_SWAPENTS);
208	task_unlock(p);
209
210	/*
211	 * Root processes get 3% bonus, just like the __vm_enough_memory()
212	 * implementation used by LSMs.
213	 */
214	if (has_capability_noaudit(p, CAP_SYS_ADMIN))
215		adj -= 30;
216
217	/* Normalize to oom_score_adj units */
218	adj *= totalpages / 1000;
219	points += adj;
220
221	/*
222	 * Never return 0 for an eligible task regardless of the root bonus and
223	 * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
224	 */
225	return points > 0 ? points : 1;
226}
227
 
 
 
 
 
 
 
228/*
229 * Determine the type of allocation constraint.
230 */
231#ifdef CONFIG_NUMA
232static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
233				gfp_t gfp_mask, nodemask_t *nodemask,
234				unsigned long *totalpages)
235{
236	struct zone *zone;
237	struct zoneref *z;
238	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
239	bool cpuset_limited = false;
240	int nid;
241
 
 
 
 
 
242	/* Default to all available memory */
243	*totalpages = totalram_pages + total_swap_pages;
 
 
 
244
245	if (!zonelist)
246		return CONSTRAINT_NONE;
247	/*
248	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
249	 * to kill current.We have to random task kill in this case.
250	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
251	 */
252	if (gfp_mask & __GFP_THISNODE)
253		return CONSTRAINT_NONE;
254
255	/*
256	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
257	 * the page allocator means a mempolicy is in effect.  Cpuset policy
258	 * is enforced in get_page_from_freelist().
259	 */
260	if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
261		*totalpages = total_swap_pages;
262		for_each_node_mask(nid, *nodemask)
263			*totalpages += node_spanned_pages(nid);
 
264		return CONSTRAINT_MEMORY_POLICY;
265	}
266
267	/* Check this allocation failure is caused by cpuset's wall function */
268	for_each_zone_zonelist_nodemask(zone, z, zonelist,
269			high_zoneidx, nodemask)
270		if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
271			cpuset_limited = true;
272
273	if (cpuset_limited) {
274		*totalpages = total_swap_pages;
275		for_each_node_mask(nid, cpuset_current_mems_allowed)
276			*totalpages += node_spanned_pages(nid);
277		return CONSTRAINT_CPUSET;
278	}
279	return CONSTRAINT_NONE;
280}
281#else
282static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
283				gfp_t gfp_mask, nodemask_t *nodemask,
284				unsigned long *totalpages)
285{
286	*totalpages = totalram_pages + total_swap_pages;
287	return CONSTRAINT_NONE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
288}
289#endif
290
291/*
292 * Simple selection loop. We chose the process with the highest
293 * number of 'points'. We expect the caller will lock the tasklist.
294 *
295 * (not docbooked, we don't want this one cluttering up the manual)
296 */
297static struct task_struct *select_bad_process(unsigned int *ppoints,
298		unsigned long totalpages, struct mem_cgroup *memcg,
299		const nodemask_t *nodemask, bool force_kill)
300{
301	struct task_struct *g, *p;
302	struct task_struct *chosen = NULL;
303	unsigned long chosen_points = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
304
305	do_each_thread(g, p) {
306		unsigned int points;
307
308		if (p->exit_state)
309			continue;
310		if (oom_unkillable_task(p, memcg, nodemask))
311			continue;
312
 
 
313		/*
314		 * This task already has access to memory reserves and is
315		 * being killed. Don't allow any other task access to the
316		 * memory reserve.
317		 *
318		 * Note: this may have a chance of deadlock if it gets
319		 * blocked waiting for another task which itself is waiting
320		 * for memory. Is there a better alternative?
321		 */
322		if (test_tsk_thread_flag(p, TIF_MEMDIE)) {
323			if (unlikely(frozen(p)))
324				__thaw_task(p);
325			if (!force_kill)
326				return ERR_PTR(-1UL);
327		}
328		if (!p->mm)
329			continue;
330
331		if (p->flags & PF_EXITING) {
332			/*
333			 * If p is the current task and is in the process of
334			 * releasing memory, we allow the "kill" to set
335			 * TIF_MEMDIE, which will allow it to gain access to
336			 * memory reserves.  Otherwise, it may stall forever.
337			 *
338			 * The loop isn't broken here, however, in case other
339			 * threads are found to have already been oom killed.
340			 */
341			if (p == current) {
342				chosen = p;
343				chosen_points = ULONG_MAX;
344			} else if (!force_kill) {
345				/*
346				 * If this task is not being ptraced on exit,
347				 * then wait for it to finish before killing
348				 * some other task unnecessarily.
349				 */
350				if (!(p->group_leader->ptrace & PT_TRACE_EXIT))
351					return ERR_PTR(-1UL);
352			}
353		}
354
355		points = oom_badness(p, memcg, nodemask, totalpages);
356		if (points > chosen_points) {
357			chosen = p;
358			chosen_points = points;
359		}
360	} while_each_thread(g, p);
 
361
362	*ppoints = chosen_points * 1000 / totalpages;
363	return chosen;
364}
365
366/**
367 * dump_tasks - dump current memory state of all system tasks
368 * @memcg: current's memory controller, if constrained
369 * @nodemask: nodemask passed to page allocator for mempolicy ooms
370 *
371 * Dumps the current memory state of all eligible tasks.  Tasks not in the same
372 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
373 * are not shown.
374 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
375 * value, oom_score_adj value, and name.
376 *
377 * Call with tasklist_lock read-locked.
378 */
379static void dump_tasks(const struct mem_cgroup *memcg, const nodemask_t *nodemask)
380{
381	struct task_struct *p;
382	struct task_struct *task;
383
384	pr_info("[ pid ]   uid  tgid total_vm      rss cpu oom_adj oom_score_adj name\n");
385	for_each_process(p) {
386		if (oom_unkillable_task(p, memcg, nodemask))
387			continue;
388
389		task = find_lock_task_mm(p);
390		if (!task) {
391			/*
392			 * This is a kthread or all of p's threads have already
393			 * detached their mm's.  There's no need to report
394			 * them; they can't be oom killed anyway.
395			 */
396			continue;
397		}
398
399		pr_info("[%5d] %5d %5d %8lu %8lu %3u     %3d         %5d %s\n",
400			task->pid, from_kuid(&init_user_ns, task_uid(task)),
401			task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
402			task_cpu(task), task->signal->oom_adj,
403			task->signal->oom_score_adj, task->comm);
404		task_unlock(task);
405	}
406}
407
408static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
409			struct mem_cgroup *memcg, const nodemask_t *nodemask)
410{
411	task_lock(current);
412	pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
413		"oom_adj=%d, oom_score_adj=%d\n",
414		current->comm, gfp_mask, order, current->signal->oom_adj,
415		current->signal->oom_score_adj);
416	cpuset_print_task_mems_allowed(current);
417	task_unlock(current);
 
 
 
 
 
 
 
 
 
 
418	dump_stack();
419	mem_cgroup_print_oom_info(memcg, p);
420	show_mem(SHOW_MEM_FILTER_NODES);
 
 
 
 
 
421	if (sysctl_oom_dump_tasks)
422		dump_tasks(memcg, nodemask);
 
 
423}
424
 
 
 
 
 
 
 
 
425#define K(x) ((x) << (PAGE_SHIFT-10))
426static void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
427			     unsigned int points, unsigned long totalpages,
428			     struct mem_cgroup *memcg, nodemask_t *nodemask,
429			     const char *message)
430{
431	struct task_struct *victim = p;
432	struct task_struct *child;
433	struct task_struct *t = p;
434	struct mm_struct *mm;
435	unsigned int victim_points = 0;
436	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
437					      DEFAULT_RATELIMIT_BURST);
438
439	/*
440	 * If the task is already exiting, don't alarm the sysadmin or kill
441	 * its children or threads, just set TIF_MEMDIE so it can die quickly
442	 */
443	if (p->flags & PF_EXITING) {
444		set_tsk_thread_flag(p, TIF_MEMDIE);
445		return;
 
 
 
 
 
 
 
446	}
 
 
447
448	if (__ratelimit(&oom_rs))
449		dump_header(p, gfp_mask, order, memcg, nodemask);
 
 
 
 
 
 
 
450
451	task_lock(p);
452	pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
453		message, task_pid_nr(p), p->comm, points);
454	task_unlock(p);
455
456	/*
457	 * If any of p's children has a different mm and is eligible for kill,
458	 * the one with the highest oom_badness() score is sacrificed for its
459	 * parent.  This attempts to lose the minimal amount of work done while
460	 * still freeing memory.
461	 */
462	do {
463		list_for_each_entry(child, &t->children, sibling) {
464			unsigned int child_points;
465
466			if (child->mm == p->mm)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
467				continue;
468			/*
469			 * oom_badness() returns 0 if the thread is unkillable
470			 */
471			child_points = oom_badness(child, memcg, nodemask,
472								totalpages);
473			if (child_points > victim_points) {
474				victim = child;
475				victim_points = child_points;
476			}
 
 
 
477		}
478	} while_each_thread(p, t);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
479
480	victim = find_lock_task_mm(victim);
481	if (!victim)
482		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
483
484	/* mm cannot safely be dereferenced after task_unlock(victim) */
485	mm = victim->mm;
486	pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
487		task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
488		K(get_mm_counter(victim->mm, MM_ANONPAGES)),
489		K(get_mm_counter(victim->mm, MM_FILEPAGES)));
490	task_unlock(victim);
491
492	/*
493	 * Kill all user processes sharing victim->mm in other thread groups, if
494	 * any.  They don't get access to memory reserves, though, to avoid
495	 * depletion of all memory.  This prevents mm->mmap_sem livelock when an
496	 * oom killed thread cannot exit because it requires the semaphore and
497	 * its contended by another thread trying to allocate memory itself.
498	 * That thread will now get access to memory reserves since it has a
499	 * pending fatal signal.
500	 */
501	for_each_process(p)
502		if (p->mm == mm && !same_thread_group(p, victim) &&
503		    !(p->flags & PF_KTHREAD)) {
504			if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
505				continue;
 
 
 
 
 
506
507			task_lock(p);	/* Protect ->comm from prctl() */
508			pr_err("Kill process %d (%s) sharing same memory\n",
509				task_pid_nr(p), p->comm);
510			task_unlock(p);
511			do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
512		}
 
 
 
 
 
513
514	set_tsk_thread_flag(victim, TIF_MEMDIE);
515	do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
516}
517#undef K
518
519/*
520 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
521 */
522static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
523				int order, const nodemask_t *nodemask)
524{
525	if (likely(!sysctl_panic_on_oom))
 
 
 
 
526		return;
527	if (sysctl_panic_on_oom != 2) {
528		/*
529		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
530		 * does not panic for cpuset, mempolicy, or memcg allocation
531		 * failures.
532		 */
533		if (constraint != CONSTRAINT_NONE)
534			return;
535	}
536	read_lock(&tasklist_lock);
537	dump_header(NULL, gfp_mask, order, NULL, nodemask);
538	read_unlock(&tasklist_lock);
539	panic("Out of memory: %s panic_on_oom is enabled\n",
540		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
541}
542
543#ifdef CONFIG_CGROUP_MEM_RES_CTLR
544void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
545			      int order)
 
 
 
 
 
 
 
 
 
 
 
 
 
546{
547	unsigned long limit;
548	unsigned int points = 0;
549	struct task_struct *p;
550
551	/*
552	 * If current has a pending SIGKILL, then automatically select it.  The
553	 * goal is to allow it to allocate so that it may quickly exit and free
554	 * its memory.
555	 */
556	if (fatal_signal_pending(current)) {
557		set_thread_flag(TIF_MEMDIE);
558		return;
 
 
 
 
 
 
 
559	}
 
560
561	check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
562	limit = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
563	read_lock(&tasklist_lock);
564	p = select_bad_process(&points, limit, memcg, NULL, false);
565	if (p && PTR_ERR(p) != -1UL)
566		oom_kill_process(p, gfp_mask, order, points, limit, memcg, NULL,
567				 "Memory cgroup out of memory");
568	read_unlock(&tasklist_lock);
569}
570#endif
571
572static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
573
574int register_oom_notifier(struct notifier_block *nb)
575{
576	return blocking_notifier_chain_register(&oom_notify_list, nb);
577}
578EXPORT_SYMBOL_GPL(register_oom_notifier);
579
580int unregister_oom_notifier(struct notifier_block *nb)
 
 
 
 
 
 
 
581{
582	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
583}
584EXPORT_SYMBOL_GPL(unregister_oom_notifier);
585
586/*
587 * Try to acquire the OOM killer lock for the zones in zonelist.  Returns zero
588 * if a parallel OOM killing is already taking place that includes a zone in
589 * the zonelist.  Otherwise, locks all zones in the zonelist and returns 1.
590 */
591int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
592{
593	struct zoneref *z;
594	struct zone *zone;
595	int ret = 1;
596
597	spin_lock(&zone_scan_lock);
598	for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
599		if (zone_is_oom_locked(zone)) {
600			ret = 0;
601			goto out;
602		}
 
 
603	}
604
605	for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
606		/*
607		 * Lock each zone in the zonelist under zone_scan_lock so a
608		 * parallel invocation of try_set_zonelist_oom() doesn't succeed
609		 * when it shouldn't.
610		 */
611		zone_set_flag(zone, ZONE_OOM_LOCKED);
 
 
612	}
 
613
614out:
615	spin_unlock(&zone_scan_lock);
616	return ret;
 
 
617}
 
618
619/*
620 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
621 * allocation attempts with zonelists containing them may now recall the OOM
622 * killer, if necessary.
623 */
624void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
625{
626	struct zoneref *z;
627	struct zone *zone;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
628
629	spin_lock(&zone_scan_lock);
630	for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
631		zone_clear_flag(zone, ZONE_OOM_LOCKED);
 
 
 
 
 
632	}
633	spin_unlock(&zone_scan_lock);
634}
635
636/*
637 * Try to acquire the oom killer lock for all system zones.  Returns zero if a
638 * parallel oom killing is taking place, otherwise locks all zones and returns
639 * non-zero.
640 */
641static int try_set_system_oom(void)
642{
643	struct zone *zone;
644	int ret = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
645
646	spin_lock(&zone_scan_lock);
647	for_each_populated_zone(zone)
648		if (zone_is_oom_locked(zone)) {
649			ret = 0;
650			goto out;
651		}
652	for_each_populated_zone(zone)
653		zone_set_flag(zone, ZONE_OOM_LOCKED);
654out:
655	spin_unlock(&zone_scan_lock);
656	return ret;
657}
 
658
659/*
660 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
661 * attempts or page faults may now recall the oom killer, if necessary.
662 */
663static void clear_system_oom(void)
664{
665	struct zone *zone;
666
667	spin_lock(&zone_scan_lock);
668	for_each_populated_zone(zone)
669		zone_clear_flag(zone, ZONE_OOM_LOCKED);
670	spin_unlock(&zone_scan_lock);
671}
 
672
673/**
674 * out_of_memory - kill the "best" process when we run out of memory
675 * @zonelist: zonelist pointer
676 * @gfp_mask: memory allocation flags
677 * @order: amount of memory being requested as a power of 2
678 * @nodemask: nodemask passed to page allocator
679 * @force_kill: true if a task must be killed, even if others are exiting
680 *
681 * If we run out of memory, we have the choice between either
682 * killing a random task (bad), letting the system crash (worse)
683 * OR try to be smart about which process to kill. Note that we
684 * don't have to be perfect here, we just have to be good.
685 */
686void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
687		int order, nodemask_t *nodemask, bool force_kill)
688{
689	const nodemask_t *mpol_mask;
690	struct task_struct *p;
691	unsigned long totalpages;
692	unsigned long freed = 0;
693	unsigned int points;
694	enum oom_constraint constraint = CONSTRAINT_NONE;
695	int killed = 0;
696
697	blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
698	if (freed > 0)
699		/* Got some memory back in the last second. */
700		return;
 
 
701
702	/*
703	 * If current has a pending SIGKILL, then automatically select it.  The
704	 * goal is to allow it to allocate so that it may quickly exit and free
705	 * its memory.
706	 */
707	if (fatal_signal_pending(current)) {
708		set_thread_flag(TIF_MEMDIE);
709		return;
 
710	}
711
712	/*
 
 
 
 
 
 
 
 
 
 
713	 * Check if there were limitations on the allocation (only relevant for
714	 * NUMA) that may require different handling.
715	 */
716	constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
717						&totalpages);
718	mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
719	check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
720
721	read_lock(&tasklist_lock);
722	if (sysctl_oom_kill_allocating_task &&
723	    !oom_unkillable_task(current, NULL, nodemask) &&
724	    current->mm) {
725		oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
726				 nodemask,
727				 "Out of memory (oom_kill_allocating_task)");
728		goto out;
729	}
730
731	p = select_bad_process(&points, totalpages, NULL, mpol_mask,
732			       force_kill);
733	/* Found nothing?!?! Either we hang forever, or we panic. */
734	if (!p) {
735		dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
736		read_unlock(&tasklist_lock);
737		panic("Out of memory and no killable processes...\n");
738	}
739	if (PTR_ERR(p) != -1UL) {
740		oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
741				 nodemask, "Out of memory");
742		killed = 1;
 
 
 
 
 
 
 
 
 
743	}
744out:
745	read_unlock(&tasklist_lock);
746
747	/*
748	 * Give "p" a good chance of killing itself before we
749	 * retry to allocate memory unless "p" is current
750	 */
751	if (killed && !test_thread_flag(TIF_MEMDIE))
752		schedule_timeout_uninterruptible(1);
753}
754
755/*
756 * The pagefault handler calls here because it is out of memory, so kill a
757 * memory-hogging task.  If a populated zone has ZONE_OOM_LOCKED set, a parallel
758 * oom killing is already in progress so do nothing.  If a task is found with
759 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
760 */
761void pagefault_out_of_memory(void)
762{
763	if (try_set_system_oom()) {
764		out_of_memory(NULL, 0, 0, NULL, false);
765		clear_system_oom();
766	}
767	if (!test_thread_flag(TIF_MEMDIE))
768		schedule_timeout_uninterruptible(1);
 
 
 
 
 
 
 
 
 
769}