oom_kill.c

   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/task.h>
  28#include <linux/sched/debug.h>
  29#include <linux/swap.h>
  30#include <linux/syscalls.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#include <linux/cred.h>
  47#include <linux/nmi.h>
  48
  49#include <asm/tlb.h>
  50#include "internal.h"
  51#include "slab.h"
  52
  53#define CREATE_TRACE_POINTS
  54#include <trace/events/oom.h>
  55
  56static int sysctl_panic_on_oom;
  57static int sysctl_oom_kill_allocating_task;
  58static int sysctl_oom_dump_tasks = 1;
  59
  60/*
  61 * Serializes oom killer invocations (out_of_memory()) from all contexts to
  62 * prevent from over eager oom killing (e.g. when the oom killer is invoked
  63 * from different domains).
  64 *
  65 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
  66 * and mark_oom_victim
  67 */
  68DEFINE_MUTEX(oom_lock);
  69/* Serializes oom_score_adj and oom_score_adj_min updates */
  70DEFINE_MUTEX(oom_adj_mutex);
  71
  72static inline bool is_memcg_oom(struct oom_control *oc)
  73{
  74	return oc->memcg != NULL;
  75}
  76
  77#ifdef CONFIG_NUMA
  78/**
  79 * oom_cpuset_eligible() - check task eligibility for kill
  80 * @start: task struct of which task to consider
  81 * @oc: pointer to struct oom_control
  82 *
  83 * Task eligibility is determined by whether or not a candidate task, @tsk,
  84 * shares the same mempolicy nodes as current if it is bound by such a policy
  85 * and whether or not it has the same set of allowed cpuset nodes.
  86 *
  87 * This function is assuming oom-killer context and 'current' has triggered
  88 * the oom-killer.
  89 */
  90static bool oom_cpuset_eligible(struct task_struct *start,
  91				struct oom_control *oc)
  92{
  93	struct task_struct *tsk;
  94	bool ret = false;
  95	const nodemask_t *mask = oc->nodemask;
  96
  97	rcu_read_lock();
  98	for_each_thread(start, tsk) {
  99		if (mask) {
 100			/*
 101			 * If this is a mempolicy constrained oom, tsk's
 102			 * cpuset is irrelevant.  Only return true if its
 103			 * mempolicy intersects current, otherwise it may be
 104			 * needlessly killed.
 105			 */
 106			ret = mempolicy_in_oom_domain(tsk, mask);
 107		} else {
 108			/*
 109			 * This is not a mempolicy constrained oom, so only
 110			 * check the mems of tsk's cpuset.
 111			 */
 112			ret = cpuset_mems_allowed_intersects(current, tsk);
 113		}
 114		if (ret)
 115			break;
 116	}
 117	rcu_read_unlock();
 118
 119	return ret;
 120}
 121#else
 122static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
 123{
 124	return true;
 125}
 126#endif /* CONFIG_NUMA */
 127
 128/*
 129 * The process p may have detached its own ->mm while exiting or through
 130 * kthread_use_mm(), but one or more of its subthreads may still have a valid
 131 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
 132 * task_lock() held.
 133 */
 134struct task_struct *find_lock_task_mm(struct task_struct *p)
 135{
 136	struct task_struct *t;
 137
 138	rcu_read_lock();
 139
 140	for_each_thread(p, t) {
 141		task_lock(t);
 142		if (likely(t->mm))
 143			goto found;
 144		task_unlock(t);
 145	}
 146	t = NULL;
 147found:
 148	rcu_read_unlock();
 149
 150	return t;
 151}
 152
 153/*
 154 * order == -1 means the oom kill is required by sysrq, otherwise only
 155 * for display purposes.
 156 */
 157static inline bool is_sysrq_oom(struct oom_control *oc)
 158{
 159	return oc->order == -1;
 160}
 161
 162/* return true if the task is not adequate as candidate victim task. */
 163static bool oom_unkillable_task(struct task_struct *p)
 164{
 165	if (is_global_init(p))
 166		return true;
 167	if (p->flags & PF_KTHREAD)
 168		return true;
 169	return false;
 170}
 171
 172/*
 173 * Check whether unreclaimable slab amount is greater than
 174 * all user memory(LRU pages).
 175 * dump_unreclaimable_slab() could help in the case that
 176 * oom due to too much unreclaimable slab used by kernel.
 177*/
 178static bool should_dump_unreclaim_slab(void)
 179{
 180	unsigned long nr_lru;
 181
 182	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
 183		 global_node_page_state(NR_INACTIVE_ANON) +
 184		 global_node_page_state(NR_ACTIVE_FILE) +
 185		 global_node_page_state(NR_INACTIVE_FILE) +
 186		 global_node_page_state(NR_ISOLATED_ANON) +
 187		 global_node_page_state(NR_ISOLATED_FILE) +
 188		 global_node_page_state(NR_UNEVICTABLE);
 189
 190	return (global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B) > nr_lru);
 191}
 192
 193/**
 194 * oom_badness - heuristic function to determine which candidate task to kill
 195 * @p: task struct of which task we should calculate
 196 * @totalpages: total present RAM allowed for page allocation
 197 *
 198 * The heuristic for determining which task to kill is made to be as simple and
 199 * predictable as possible.  The goal is to return the highest value for the
 200 * task consuming the most memory to avoid subsequent oom failures.
 201 */
 202long oom_badness(struct task_struct *p, unsigned long totalpages)
 203{
 204	long points;
 205	long adj;
 206
 207	if (oom_unkillable_task(p))
 208		return LONG_MIN;
 209
 210	p = find_lock_task_mm(p);
 211	if (!p)
 212		return LONG_MIN;
 213
 214	/*
 215	 * Do not even consider tasks which are explicitly marked oom
 216	 * unkillable or have been already oom reaped or the are in
 217	 * the middle of vfork
 218	 */
 219	adj = (long)p->signal->oom_score_adj;
 220	if (adj == OOM_SCORE_ADJ_MIN ||
 221			mm_flags_test(MMF_OOM_SKIP, p->mm) ||
 222			in_vfork(p)) {
 223		task_unlock(p);
 224		return LONG_MIN;
 225	}
 226
 227	/*
 228	 * The baseline for the badness score is the proportion of RAM that each
 229	 * task's rss, pagetable and swap space use.
 230	 */
 231	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
 232		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
 233	task_unlock(p);
 234
 235	/* Normalize to oom_score_adj units */
 236	adj *= totalpages / 1000;
 237	points += adj;
 238
 239	return points;
 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 highest_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			highest_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	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 (mm_flags_test(MMF_OOM_SKIP, task->signal->oom_mm))
 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 = LONG_MAX;
 339		goto select;
 340	}
 341
 342	points = oom_badness(task, oc->totalpages);
 343	if (points == LONG_MIN || 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	oc->chosen_points = LONG_MIN;
 368
 369	if (is_memcg_oom(oc))
 370		mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
 371	else {
 372		struct task_struct *p;
 373
 374		rcu_read_lock();
 375		for_each_process(p)
 376			if (oom_evaluate_task(p, oc))
 377				break;
 378		rcu_read_unlock();
 379	}
 380}
 381
 382static int dump_task(struct task_struct *p, void *arg)
 383{
 384	struct oom_control *oc = arg;
 385	struct task_struct *task;
 386
 387	if (oom_unkillable_task(p))
 388		return 0;
 389
 390	/* p may not have freeable memory in nodemask */
 391	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
 392		return 0;
 393
 394	task = find_lock_task_mm(p);
 395	if (!task) {
 396		/*
 397		 * All of p's threads have already detached their mm's. There's
 398		 * no need to report them; they can't be oom killed anyway.
 399		 */
 400		return 0;
 401	}
 402
 403	pr_info("[%7d] %5d %5d %8lu %8lu %8lu %8lu %9lu %8ld %8lu         %5hd %s\n",
 404		task->pid, from_kuid(&init_user_ns, task_uid(task)),
 405		task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
 406		get_mm_counter(task->mm, MM_ANONPAGES), get_mm_counter(task->mm, MM_FILEPAGES),
 407		get_mm_counter(task->mm, MM_SHMEMPAGES), 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 rss_anon rss_file rss_shmem 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		int i = 0;
 435
 436		rcu_read_lock();
 437		for_each_process(p) {
 438			/* Avoid potential softlockup warning */
 439			if ((++i & 1023) == 0)
 440				touch_softlockup_watchdog();
 441			dump_task(p, oc);
 442		}
 443		rcu_read_unlock();
 444	}
 445}
 446
 447static void dump_oom_victim(struct oom_control *oc, struct task_struct *victim)
 448{
 449	/* one line summary of the oom killer context. */
 450	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
 451			oom_constraint_text[oc->constraint],
 452			nodemask_pr_args(oc->nodemask));
 453	cpuset_print_current_mems_allowed();
 454	mem_cgroup_print_oom_context(oc->memcg, victim);
 455	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
 456		from_kuid(&init_user_ns, task_uid(victim)));
 457}
 458
 459static void dump_header(struct oom_control *oc)
 460{
 461	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
 462		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
 463			current->signal->oom_score_adj);
 464	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
 465		pr_warn("COMPACTION is disabled!!!\n");
 466
 467	dump_stack();
 468	if (is_memcg_oom(oc))
 469		mem_cgroup_print_oom_meminfo(oc->memcg);
 470	else {
 471		__show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask, gfp_zone(oc->gfp_mask));
 472		if (should_dump_unreclaim_slab())
 473			dump_unreclaimable_slab();
 474	}
 475	mem_cgroup_show_protected_memory(oc->memcg);
 476	if (sysctl_oom_dump_tasks)
 477		dump_tasks(oc);
 478}
 479
 480/*
 481 * Number of OOM victims in flight
 482 */
 483static atomic_t oom_victims = ATOMIC_INIT(0);
 484static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
 485
 486static bool oom_killer_disabled __read_mostly;
 487
 488/*
 489 * task->mm can be NULL if the task is the exited group leader.  So to
 490 * determine whether the task is using a particular mm, we examine all the
 491 * task's threads: if one of those is using this mm then this task was also
 492 * using it.
 493 */
 494bool process_shares_mm(const struct task_struct *p, const struct mm_struct *mm)
 495{
 496	const struct task_struct *t;
 497
 498	for_each_thread(p, t) {
 499		const struct mm_struct *t_mm = READ_ONCE(t->mm);
 500		if (t_mm)
 501			return t_mm == mm;
 502	}
 503	return false;
 504}
 505
 506#ifdef CONFIG_MMU
 507/*
 508 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
 509 * victim (if that is possible) to help the OOM killer to move on.
 510 */
 511static struct task_struct *oom_reaper_th;
 512static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
 513static struct task_struct *oom_reaper_list;
 514static DEFINE_SPINLOCK(oom_reaper_lock);
 515
 516static bool __oom_reap_task_mm(struct mm_struct *mm)
 517{
 518	struct vm_area_struct *vma;
 519	bool ret = true;
 520	MA_STATE(mas, &mm->mm_mt, ULONG_MAX, ULONG_MAX);
 521
 522	/*
 523	 * Tell all users of get_user/copy_from_user etc... that the content
 524	 * is no longer stable. No barriers really needed because unmapping
 525	 * should imply barriers already and the reader would hit a page fault
 526	 * if it stumbled over a reaped memory.
 527	 */
 528	mm_flags_set(MMF_UNSTABLE, mm);
 529
 530	/*
 531	 * It might start racing with the dying task and compete for shared
 532	 * resources - e.g. page table lock contention has been observed.
 533	 * Reduce those races by reaping the oom victim from the other end
 534	 * of the address space.
 535	 */
 536	mas_for_each_rev(&mas, vma, 0) {
 537		if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
 538			continue;
 539
 540		/*
 541		 * Only anonymous pages have a good chance to be dropped
 542		 * without additional steps which we cannot afford as we
 543		 * are OOM already.
 544		 *
 545		 * We do not even care about fs backed pages because all
 546		 * which are reclaimable have already been reclaimed and
 547		 * we do not want to block exit_mmap by keeping mm ref
 548		 * count elevated without a good reason.
 549		 */
 550		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
 551			struct mmu_notifier_range range;
 552			struct mmu_gather tlb;
 553
 554			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
 555						mm, vma->vm_start,
 556						vma->vm_end);
 557			tlb_gather_mmu(&tlb, mm);
 558			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
 559				tlb_finish_mmu(&tlb);
 560				ret = false;
 561				continue;
 562			}
 563			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
 564			mmu_notifier_invalidate_range_end(&range);
 565			tlb_finish_mmu(&tlb);
 566		}
 567	}
 568
 569	return ret;
 570}
 571
 572/*
 573 * Reaps the address space of the given task.
 574 *
 575 * Returns true on success and false if none or part of the address space
 576 * has been reclaimed and the caller should retry later.
 577 */
 578static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
 579{
 580	bool ret = true;
 581
 582	if (!mmap_read_trylock(mm)) {
 583		trace_skip_task_reaping(tsk->pid);
 584		return false;
 585	}
 586
 587	/*
 588	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
 589	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
 590	 * under mmap_lock for reading because it serializes against the
 591	 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
 592	 */
 593	if (mm_flags_test(MMF_OOM_SKIP, mm)) {
 594		trace_skip_task_reaping(tsk->pid);
 595		goto out_unlock;
 596	}
 597
 598	trace_start_task_reaping(tsk->pid);
 599
 600	/* failed to reap part of the address space. Try again later */
 601	ret = __oom_reap_task_mm(mm);
 602	if (!ret)
 603		goto out_finish;
 604
 605	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
 606			task_pid_nr(tsk), tsk->comm,
 607			K(get_mm_counter(mm, MM_ANONPAGES)),
 608			K(get_mm_counter(mm, MM_FILEPAGES)),
 609			K(get_mm_counter(mm, MM_SHMEMPAGES)));
 610out_finish:
 611	trace_finish_task_reaping(tsk->pid);
 612out_unlock:
 613	mmap_read_unlock(mm);
 614
 615	return ret;
 616}
 617
 618#define MAX_OOM_REAP_RETRIES 10
 619static void oom_reap_task(struct task_struct *tsk)
 620{
 621	int attempts = 0;
 622	struct mm_struct *mm = tsk->signal->oom_mm;
 623
 624	/* Retry the mmap_read_trylock(mm) a few times */
 625	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
 626		schedule_timeout_idle(HZ/10);
 627
 628	if (attempts <= MAX_OOM_REAP_RETRIES ||
 629	    mm_flags_test(MMF_OOM_SKIP, mm))
 630		goto done;
 631
 632	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
 633		task_pid_nr(tsk), tsk->comm);
 634	sched_show_task(tsk);
 635	debug_show_all_locks();
 636
 637done:
 638	tsk->oom_reaper_list = NULL;
 639
 640	/*
 641	 * Hide this mm from OOM killer because it has been either reaped or
 642	 * somebody can't call mmap_write_unlock(mm).
 643	 */
 644	mm_flags_set(MMF_OOM_SKIP, mm);
 645
 646	/* Drop a reference taken by queue_oom_reaper */
 647	put_task_struct(tsk);
 648}
 649
 650static int oom_reaper(void *unused)
 651{
 652	set_freezable();
 653
 654	while (true) {
 655		struct task_struct *tsk = NULL;
 656
 657		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
 658		spin_lock_irq(&oom_reaper_lock);
 659		if (oom_reaper_list != NULL) {
 660			tsk = oom_reaper_list;
 661			oom_reaper_list = tsk->oom_reaper_list;
 662		}
 663		spin_unlock_irq(&oom_reaper_lock);
 664
 665		if (tsk)
 666			oom_reap_task(tsk);
 667	}
 668
 669	return 0;
 670}
 671
 672static void wake_oom_reaper(struct timer_list *timer)
 673{
 674	struct task_struct *tsk = container_of(timer, struct task_struct,
 675			oom_reaper_timer);
 676	struct mm_struct *mm = tsk->signal->oom_mm;
 677	unsigned long flags;
 678
 679	/* The victim managed to terminate on its own - see exit_mmap */
 680	if (mm_flags_test(MMF_OOM_SKIP, mm)) {
 681		put_task_struct(tsk);
 682		return;
 683	}
 684
 685	spin_lock_irqsave(&oom_reaper_lock, flags);
 686	tsk->oom_reaper_list = oom_reaper_list;
 687	oom_reaper_list = tsk;
 688	spin_unlock_irqrestore(&oom_reaper_lock, flags);
 689	trace_wake_reaper(tsk->pid);
 690	wake_up(&oom_reaper_wait);
 691}
 692
 693/*
 694 * Give the OOM victim time to exit naturally before invoking the oom_reaping.
 695 * The timers timeout is arbitrary... the longer it is, the longer the worst
 696 * case scenario for the OOM can take. If it is too small, the oom_reaper can
 697 * get in the way and release resources needed by the process exit path.
 698 * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
 699 * before the exit path is able to wake the futex waiters.
 700 */
 701#define OOM_REAPER_DELAY (2*HZ)
 702static void queue_oom_reaper(struct task_struct *tsk)
 703{
 704	/* mm is already queued? */
 705	if (mm_flags_test_and_set(MMF_OOM_REAP_QUEUED, tsk->signal->oom_mm))
 706		return;
 707
 708	get_task_struct(tsk);
 709	timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
 710	tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
 711	add_timer(&tsk->oom_reaper_timer);
 712}
 713
 714#ifdef CONFIG_SYSCTL
 715static const struct ctl_table vm_oom_kill_table[] = {
 716	{
 717		.procname	= "panic_on_oom",
 718		.data		= &sysctl_panic_on_oom,
 719		.maxlen		= sizeof(sysctl_panic_on_oom),
 720		.mode		= 0644,
 721		.proc_handler	= proc_dointvec_minmax,
 722		.extra1		= SYSCTL_ZERO,
 723		.extra2		= SYSCTL_TWO,
 724	},
 725	{
 726		.procname	= "oom_kill_allocating_task",
 727		.data		= &sysctl_oom_kill_allocating_task,
 728		.maxlen		= sizeof(sysctl_oom_kill_allocating_task),
 729		.mode		= 0644,
 730		.proc_handler	= proc_dointvec,
 731	},
 732	{
 733		.procname	= "oom_dump_tasks",
 734		.data		= &sysctl_oom_dump_tasks,
 735		.maxlen		= sizeof(sysctl_oom_dump_tasks),
 736		.mode		= 0644,
 737		.proc_handler	= proc_dointvec,
 738	},
 739};
 740#endif
 741
 742static int __init oom_init(void)
 743{
 744	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
 745#ifdef CONFIG_SYSCTL
 746	register_sysctl_init("vm", vm_oom_kill_table);
 747#endif
 748	return 0;
 749}
 750subsys_initcall(oom_init)
 751#else
 752static inline void queue_oom_reaper(struct task_struct *tsk)
 753{
 754}
 755#endif /* CONFIG_MMU */
 756
 757/**
 758 * mark_oom_victim - mark the given task as OOM victim
 759 * @tsk: task to mark
 760 *
 761 * Has to be called with oom_lock held and never after
 762 * oom has been disabled already.
 763 *
 764 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
 765 * under task_lock or operate on the current).
 766 */
 767static void mark_oom_victim(struct task_struct *tsk)
 768{
 769	const struct cred *cred;
 770	struct mm_struct *mm = tsk->mm;
 771
 772	WARN_ON(oom_killer_disabled);
 773	/* OOM killer might race with memcg OOM */
 774	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
 775		return;
 776
 777	/* oom_mm is bound to the signal struct life time. */
 778	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
 779		mmgrab(tsk->signal->oom_mm);
 780
 781	/*
 782	 * Make sure that the process is woken up from uninterruptible sleep
 783	 * if it is frozen because OOM killer wouldn't be able to free any
 784	 * memory and livelock. The freezer will thaw the tasks that are OOM
 785	 * victims regardless of the PM freezing and cgroup freezing states.
 786	 */
 787	thaw_process(tsk);
 788	atomic_inc(&oom_victims);
 789	cred = get_task_cred(tsk);
 790	trace_mark_victim(tsk, cred->uid.val);
 791	put_cred(cred);
 792}
 793
 794/**
 795 * exit_oom_victim - note the exit of an OOM victim
 796 */
 797void exit_oom_victim(void)
 798{
 799	clear_thread_flag(TIF_MEMDIE);
 800
 801	if (!atomic_dec_return(&oom_victims))
 802		wake_up_all(&oom_victims_wait);
 803}
 804
 805/**
 806 * oom_killer_enable - enable OOM killer
 807 */
 808void oom_killer_enable(void)
 809{
 810	oom_killer_disabled = false;
 811	pr_info("OOM killer enabled.\n");
 812}
 813
 814/**
 815 * oom_killer_disable - disable OOM killer
 816 * @timeout: maximum timeout to wait for oom victims in jiffies
 817 *
 818 * Forces all page allocations to fail rather than trigger OOM killer.
 819 * Will block and wait until all OOM victims are killed or the given
 820 * timeout expires.
 821 *
 822 * The function cannot be called when there are runnable user tasks because
 823 * the userspace would see unexpected allocation failures as a result. Any
 824 * new usage of this function should be consulted with MM people.
 825 *
 826 * Returns true if successful and false if the OOM killer cannot be
 827 * disabled.
 828 */
 829bool oom_killer_disable(signed long timeout)
 830{
 831	signed long ret;
 832
 833	/*
 834	 * Make sure to not race with an ongoing OOM killer. Check that the
 835	 * current is not killed (possibly due to sharing the victim's memory).
 836	 */
 837	if (mutex_lock_killable(&oom_lock))
 838		return false;
 839	oom_killer_disabled = true;
 840	mutex_unlock(&oom_lock);
 841
 842	ret = wait_event_interruptible_timeout(oom_victims_wait,
 843			!atomic_read(&oom_victims), timeout);
 844	if (ret <= 0) {
 845		oom_killer_enable();
 846		return false;
 847	}
 848	pr_info("OOM killer disabled.\n");
 849
 850	return true;
 851}
 852
 853static inline bool __task_will_free_mem(struct task_struct *task)
 854{
 855	struct signal_struct *sig = task->signal;
 856
 857	/*
 858	 * A coredumping process may sleep for an extended period in
 859	 * coredump_task_exit(), so the oom killer cannot assume that
 860	 * the process will promptly exit and release memory.
 861	 */
 862	if (sig->core_state)
 863		return false;
 864
 865	if (sig->flags & SIGNAL_GROUP_EXIT)
 866		return true;
 867
 868	if (thread_group_empty(task) && (task->flags & PF_EXITING))
 869		return true;
 870
 871	return false;
 872}
 873
 874/*
 875 * Checks whether the given task is dying or exiting and likely to
 876 * release its address space. This means that all threads and processes
 877 * sharing the same mm have to be killed or exiting.
 878 * Caller has to make sure that task->mm is stable (hold task_lock or
 879 * it operates on the current).
 880 */
 881static bool task_will_free_mem(struct task_struct *task)
 882{
 883	struct mm_struct *mm = task->mm;
 884	struct task_struct *p;
 885	bool ret = true;
 886
 887	/*
 888	 * Skip tasks without mm because it might have passed its exit_mm and
 889	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
 890	 * on that for now. We can consider find_lock_task_mm in future.
 891	 */
 892	if (!mm)
 893		return false;
 894
 895	if (!__task_will_free_mem(task))
 896		return false;
 897
 898	/*
 899	 * This task has already been drained by the oom reaper so there are
 900	 * only small chances it will free some more
 901	 */
 902	if (mm_flags_test(MMF_OOM_SKIP, mm))
 903		return false;
 904
 905	if (atomic_read(&mm->mm_users) <= 1)
 906		return true;
 907
 908	/*
 909	 * Make sure that all tasks which share the mm with the given tasks
 910	 * are dying as well to make sure that a) nobody pins its mm and
 911	 * b) the task is also reapable by the oom reaper.
 912	 */
 913	rcu_read_lock();
 914	for_each_process(p) {
 915		if (!process_shares_mm(p, mm))
 916			continue;
 917		if (same_thread_group(task, p))
 918			continue;
 919		ret = __task_will_free_mem(p);
 920		if (!ret)
 921			break;
 922	}
 923	rcu_read_unlock();
 924
 925	return ret;
 926}
 927
 928static void __oom_kill_process(struct task_struct *victim, const char *message)
 929{
 930	struct task_struct *p;
 931	struct mm_struct *mm;
 932	bool can_oom_reap = true;
 933
 934	p = find_lock_task_mm(victim);
 935	if (!p) {
 936		pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
 937			message, task_pid_nr(victim), victim->comm);
 938		put_task_struct(victim);
 939		return;
 940	} else if (victim != p) {
 941		get_task_struct(p);
 942		put_task_struct(victim);
 943		victim = p;
 944	}
 945
 946	/* Get a reference to safely compare mm after task_unlock(victim) */
 947	mm = victim->mm;
 948	mmgrab(mm);
 949
 950	/* Raise event before sending signal: task reaper must see this */
 951	count_vm_event(OOM_KILL);
 952	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
 953
 954	/*
 955	 * We should send SIGKILL before granting access to memory reserves
 956	 * in order to prevent the OOM victim from depleting the memory
 957	 * reserves from the user space under its control.
 958	 */
 959	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
 960	mark_oom_victim(victim);
 961	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",
 962		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
 963		K(get_mm_counter(mm, MM_ANONPAGES)),
 964		K(get_mm_counter(mm, MM_FILEPAGES)),
 965		K(get_mm_counter(mm, MM_SHMEMPAGES)),
 966		from_kuid(&init_user_ns, task_uid(victim)),
 967		mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
 968	task_unlock(victim);
 969
 970	/*
 971	 * Kill all user processes sharing victim->mm in other thread groups, if
 972	 * any.  They don't get access to memory reserves, though, to avoid
 973	 * depletion of all memory.  This prevents mm->mmap_lock livelock when an
 974	 * oom killed thread cannot exit because it requires the semaphore and
 975	 * its contended by another thread trying to allocate memory itself.
 976	 * That thread will now get access to memory reserves since it has a
 977	 * pending fatal signal.
 978	 */
 979	rcu_read_lock();
 980	for_each_process(p) {
 981		if (!process_shares_mm(p, mm))
 982			continue;
 983		if (same_thread_group(p, victim))
 984			continue;
 985		if (is_global_init(p)) {
 986			can_oom_reap = false;
 987			mm_flags_set(MMF_OOM_SKIP, mm);
 988			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
 989					task_pid_nr(victim), victim->comm,
 990					task_pid_nr(p), p->comm);
 991			continue;
 992		}
 993		/*
 994		 * No kthread_use_mm() user needs to read from the userspace so
 995		 * we are ok to reap it.
 996		 */
 997		if (unlikely(p->flags & PF_KTHREAD))
 998			continue;
 999		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
1000	}
1001	rcu_read_unlock();
1002
1003	if (can_oom_reap)
1004		queue_oom_reaper(victim);
1005
1006	mmdrop(mm);
1007	put_task_struct(victim);
1008}
1009
1010/*
1011 * Kill provided task unless it's secured by setting
1012 * oom_score_adj to OOM_SCORE_ADJ_MIN.
1013 */
1014static int oom_kill_memcg_member(struct task_struct *task, void *message)
1015{
1016	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1017	    !is_global_init(task)) {
1018		get_task_struct(task);
1019		__oom_kill_process(task, message);
1020	}
1021	return 0;
1022}
1023
1024static void oom_kill_process(struct oom_control *oc, const char *message)
1025{
1026	struct task_struct *victim = oc->chosen;
1027	struct mem_cgroup *oom_group;
1028	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1029					      DEFAULT_RATELIMIT_BURST);
1030
1031	/*
1032	 * If the task is already exiting, don't alarm the sysadmin or kill
1033	 * its children or threads, just give it access to memory reserves
1034	 * so it can die quickly
1035	 */
1036	task_lock(victim);
1037	if (task_will_free_mem(victim)) {
1038		mark_oom_victim(victim);
1039		queue_oom_reaper(victim);
1040		task_unlock(victim);
1041		put_task_struct(victim);
1042		return;
1043	}
1044	task_unlock(victim);
1045
1046	if (__ratelimit(&oom_rs)) {
1047		dump_header(oc);
1048		dump_oom_victim(oc, victim);
1049	}
1050
1051	/*
1052	 * Do we need to kill the entire memory cgroup?
1053	 * Or even one of the ancestor memory cgroups?
1054	 * Check this out before killing the victim task.
1055	 */
1056	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1057
1058	__oom_kill_process(victim, message);
1059
1060	/*
1061	 * If necessary, kill all tasks in the selected memory cgroup.
1062	 */
1063	if (oom_group) {
1064		memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1065		mem_cgroup_print_oom_group(oom_group);
1066		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1067				      (void *)message);
1068		mem_cgroup_put(oom_group);
1069	}
1070}
1071
1072/*
1073 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
1074 */
1075static void check_panic_on_oom(struct oom_control *oc)
1076{
1077	if (likely(!sysctl_panic_on_oom))
1078		return;
1079	if (sysctl_panic_on_oom != 2) {
1080		/*
1081		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1082		 * does not panic for cpuset, mempolicy, or memcg allocation
1083		 * failures.
1084		 */
1085		if (oc->constraint != CONSTRAINT_NONE)
1086			return;
1087	}
1088	/* Do not panic for oom kills triggered by sysrq */
1089	if (is_sysrq_oom(oc))
1090		return;
1091	dump_header(oc);
1092	panic("Out of memory: %s panic_on_oom is enabled\n",
1093		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1094}
1095
1096static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1097
1098int register_oom_notifier(struct notifier_block *nb)
1099{
1100	return blocking_notifier_chain_register(&oom_notify_list, nb);
1101}
1102EXPORT_SYMBOL_GPL(register_oom_notifier);
1103
1104int unregister_oom_notifier(struct notifier_block *nb)
1105{
1106	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
1107}
1108EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1109
1110/**
1111 * out_of_memory - kill the "best" process when we run out of memory
1112 * @oc: pointer to struct oom_control
1113 *
1114 * If we run out of memory, we have the choice between either
1115 * killing a random task (bad), letting the system crash (worse)
1116 * OR try to be smart about which process to kill. Note that we
1117 * don't have to be perfect here, we just have to be good.
1118 */
1119bool out_of_memory(struct oom_control *oc)
1120{
1121	unsigned long freed = 0;
1122
1123	if (oom_killer_disabled)
1124		return false;
1125
1126	if (!is_memcg_oom(oc)) {
1127		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1128		if (freed > 0 && !is_sysrq_oom(oc))
1129			/* Got some memory back in the last second. */
1130			return true;
1131	}
1132
1133	/*
1134	 * If current has a pending SIGKILL or is exiting, then automatically
1135	 * select it.  The goal is to allow it to allocate so that it may
1136	 * quickly exit and free its memory.
1137	 */
1138	if (task_will_free_mem(current)) {
1139		mark_oom_victim(current);
1140		queue_oom_reaper(current);
1141		return true;
1142	}
1143
1144	/*
1145	 * The OOM killer does not compensate for IO-less reclaim.
1146	 * But mem_cgroup_oom() has to invoke the OOM killer even
1147	 * if it is a GFP_NOFS allocation.
1148	 */
1149	if (!(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1150		return true;
1151
1152	/*
1153	 * Check if there were limitations on the allocation (only relevant for
1154	 * NUMA and memcg) that may require different handling.
1155	 */
1156	oc->constraint = constrained_alloc(oc);
1157	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1158		oc->nodemask = NULL;
1159	check_panic_on_oom(oc);
1160
1161	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1162	    current->mm && !oom_unkillable_task(current) &&
1163	    oom_cpuset_eligible(current, oc) &&
1164	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1165		get_task_struct(current);
1166		oc->chosen = current;
1167		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1168		return true;
1169	}
1170
1171	select_bad_process(oc);
1172	/* Found nothing?!?! */
1173	if (!oc->chosen) {
1174		dump_header(oc);
1175		pr_warn("Out of memory and no killable processes...\n");
1176		/*
1177		 * If we got here due to an actual allocation at the
1178		 * system level, we cannot survive this and will enter
1179		 * an endless loop in the allocator. Bail out now.
1180		 */
1181		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1182			panic("System is deadlocked on memory\n");
1183	}
1184	if (oc->chosen && oc->chosen != (void *)-1UL)
1185		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1186				 "Memory cgroup out of memory");
1187	return !!oc->chosen;
1188}
1189
1190/*
1191 * The pagefault handler calls here because some allocation has failed. We have
1192 * to take care of the memcg OOM here because this is the only safe context without
1193 * any locks held but let the oom killer triggered from the allocation context care
1194 * about the global OOM.
1195 */
1196void pagefault_out_of_memory(void)
1197{
1198	static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1199				      DEFAULT_RATELIMIT_BURST);
1200
1201	if (mem_cgroup_oom_synchronize(true))
1202		return;
1203
1204	if (fatal_signal_pending(current))
1205		return;
1206
1207	if (__ratelimit(&pfoom_rs))
1208		pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1209}
1210
1211SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1212{
1213#ifdef CONFIG_MMU
1214	struct mm_struct *mm = NULL;
1215	struct task_struct *task;
1216	struct task_struct *p;
1217	unsigned int f_flags;
1218	bool reap = false;
1219	long ret = 0;
1220
1221	if (flags)
1222		return -EINVAL;
1223
1224	task = pidfd_get_task(pidfd, &f_flags);
1225	if (IS_ERR(task))
1226		return PTR_ERR(task);
1227
1228	/*
1229	 * Make sure to choose a thread which still has a reference to mm
1230	 * during the group exit
1231	 */
1232	p = find_lock_task_mm(task);
1233	if (!p) {
1234		ret = -ESRCH;
1235		goto put_task;
1236	}
1237
1238	mm = p->mm;
1239	mmgrab(mm);
1240
1241	if (task_will_free_mem(p))
1242		reap = true;
1243	else {
1244		/* Error only if the work has not been done already */
1245		if (!mm_flags_test(MMF_OOM_SKIP, mm))
1246			ret = -EINVAL;
1247	}
1248	task_unlock(p);
1249
1250	if (!reap)
1251		goto drop_mm;
1252
1253	if (mmap_read_lock_killable(mm)) {
1254		ret = -EINTR;
1255		goto drop_mm;
1256	}
1257	/*
1258	 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1259	 * possible change in exit_mmap is seen
1260	 */
1261	if (!mm_flags_test(MMF_OOM_SKIP, mm) && !__oom_reap_task_mm(mm))
1262		ret = -EAGAIN;
1263	mmap_read_unlock(mm);
1264
1265drop_mm:
1266	mmdrop(mm);
1267put_task:
1268	put_task_struct(task);
1269	return ret;
1270#else
1271	return -ENOSYS;
1272#endif /* CONFIG_MMU */
1273}