1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/oom_kill.c
   4 * 
   5 *  Copyright (C)  1998,2000  Rik van Riel
   6 *	Thanks go out to Claus Fischer for some serious inspiration and
   7 *	for goading me into coding this file...
   8 *  Copyright (C)  2010  Google, Inc.
   9 *	Rewritten by David Rientjes
  10 *
  11 *  The routines in this file are used to kill a process when
  12 *  we're seriously out of memory. This gets called from __alloc_pages()
  13 *  in mm/page_alloc.c when we really run out of memory.
  14 *
  15 *  Since we won't call these routines often (on a well-configured
  16 *  machine) this file will double as a 'coding guide' and a signpost
  17 *  for newbie kernel hackers. It features several pointers to major
  18 *  kernel subsystems and hints as to where to find out what things do.
  19 */
  20
  21#include <linux/oom.h>
  22#include <linux/mm.h>
  23#include <linux/err.h>
  24#include <linux/gfp.h>
  25#include <linux/sched.h>
  26#include <linux/sched/mm.h>
  27#include <linux/sched/coredump.h>
  28#include <linux/sched/task.h>
  29#include <linux/sched/debug.h>
  30#include <linux/swap.h>
  31#include <linux/syscalls.h>
  32#include <linux/timex.h>
  33#include <linux/jiffies.h>
  34#include <linux/cpuset.h>
  35#include <linux/export.h>
  36#include <linux/notifier.h>
  37#include <linux/memcontrol.h>
  38#include <linux/mempolicy.h>
  39#include <linux/security.h>
  40#include <linux/ptrace.h>
  41#include <linux/freezer.h>
  42#include <linux/ftrace.h>
  43#include <linux/ratelimit.h>
  44#include <linux/kthread.h>
  45#include <linux/init.h>
  46#include <linux/mmu_notifier.h>
  47
  48#include <asm/tlb.h>
  49#include "internal.h"
  50#include "slab.h"
  51
  52#define CREATE_TRACE_POINTS
  53#include <trace/events/oom.h>
  54
  55static int sysctl_panic_on_oom;
  56static int sysctl_oom_kill_allocating_task;
  57static int sysctl_oom_dump_tasks = 1;
 
  58
  59/*
  60 * Serializes oom killer invocations (out_of_memory()) from all contexts to
  61 * prevent from over eager oom killing (e.g. when the oom killer is invoked
  62 * from different domains).
  63 *
  64 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
  65 * and mark_oom_victim
 
  66 */
  67DEFINE_MUTEX(oom_lock);
  68/* Serializes oom_score_adj and oom_score_adj_min updates */
  69DEFINE_MUTEX(oom_adj_mutex);
 
 
 
 
 
 
 
  70
  71static inline bool is_memcg_oom(struct oom_control *oc)
 
 
 
 
 
 
 
 
  72{
  73	return oc->memcg != NULL;
 
 
 
 
 
 
 
 
 
  74}
  75
  76#ifdef CONFIG_NUMA
  77/**
  78 * oom_cpuset_eligible() - check task eligibility for kill
  79 * @start: task struct of which task to consider
  80 * @oc: pointer to struct oom_control
  81 *
  82 * Task eligibility is determined by whether or not a candidate task, @tsk,
  83 * shares the same mempolicy nodes as current if it is bound by such a policy
  84 * and whether or not it has the same set of allowed cpuset nodes.
  85 *
  86 * This function is assuming oom-killer context and 'current' has triggered
  87 * the oom-killer.
  88 */
  89static bool oom_cpuset_eligible(struct task_struct *start,
  90				struct oom_control *oc)
  91{
  92	struct task_struct *tsk;
  93	bool ret = false;
  94	const nodemask_t *mask = oc->nodemask;
  95
  96	rcu_read_lock();
  97	for_each_thread(start, tsk) {
  98		if (mask) {
  99			/*
 100			 * If this is a mempolicy constrained oom, tsk's
 101			 * cpuset is irrelevant.  Only return true if its
 102			 * mempolicy intersects current, otherwise it may be
 103			 * needlessly killed.
 104			 */
 105			ret = mempolicy_in_oom_domain(tsk, mask);
 
 106		} else {
 107			/*
 108			 * This is not a mempolicy constrained oom, so only
 109			 * check the mems of tsk's cpuset.
 110			 */
 111			ret = cpuset_mems_allowed_intersects(current, tsk);
 
 112		}
 113		if (ret)
 114			break;
 115	}
 116	rcu_read_unlock();
 117
 118	return ret;
 119}
 120#else
 121static bool oom_cpuset_eligible(struct task_struct *tsk, struct oom_control *oc)
 
 122{
 123	return true;
 124}
 125#endif /* CONFIG_NUMA */
 126
 127/*
 128 * The process p may have detached its own ->mm while exiting or through
 129 * kthread_use_mm(), but one or more of its subthreads may still have a valid
 130 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
 131 * task_lock() held.
 132 */
 133struct task_struct *find_lock_task_mm(struct task_struct *p)
 134{
 135	struct task_struct *t;
 136
 137	rcu_read_lock();
 138
 139	for_each_thread(p, t) {
 140		task_lock(t);
 141		if (likely(t->mm))
 142			goto found;
 143		task_unlock(t);
 144	}
 145	t = NULL;
 146found:
 147	rcu_read_unlock();
 148
 149	return t;
 150}
 151
 152/*
 153 * order == -1 means the oom kill is required by sysrq, otherwise only
 154 * for display purposes.
 155 */
 156static inline bool is_sysrq_oom(struct oom_control *oc)
 157{
 158	return oc->order == -1;
 159}
 160
 161/* return true if the task is not adequate as candidate victim task. */
 162static bool oom_unkillable_task(struct task_struct *p)
 
 163{
 164	if (is_global_init(p))
 165		return true;
 166	if (p->flags & PF_KTHREAD)
 167		return true;
 168	return false;
 169}
 170
 171/*
 172 * Check whether unreclaimable slab amount is greater than
 173 * all user memory(LRU pages).
 174 * dump_unreclaimable_slab() could help in the case that
 175 * oom due to too much unreclaimable slab used by kernel.
 176*/
 177static bool should_dump_unreclaim_slab(void)
 178{
 179	unsigned long nr_lru;
 180
 181	nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
 182		 global_node_page_state(NR_INACTIVE_ANON) +
 183		 global_node_page_state(NR_ACTIVE_FILE) +
 184		 global_node_page_state(NR_INACTIVE_FILE) +
 185		 global_node_page_state(NR_ISOLATED_ANON) +
 186		 global_node_page_state(NR_ISOLATED_FILE) +
 187		 global_node_page_state(NR_UNEVICTABLE);
 188
 189	return (global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B) > nr_lru);
 
 
 
 
 190}
 191
 192/**
 193 * oom_badness - heuristic function to determine which candidate task to kill
 194 * @p: task struct of which task we should calculate
 195 * @totalpages: total present RAM allowed for page allocation
 196 *
 197 * The heuristic for determining which task to kill is made to be as simple and
 198 * predictable as possible.  The goal is to return the highest value for the
 199 * task consuming the most memory to avoid subsequent oom failures.
 200 */
 201long oom_badness(struct task_struct *p, unsigned long totalpages)
 
 202{
 203	long points;
 204	long adj;
 205
 206	if (oom_unkillable_task(p))
 207		return LONG_MIN;
 208
 209	p = find_lock_task_mm(p);
 210	if (!p)
 211		return LONG_MIN;
 212
 213	/*
 214	 * Do not even consider tasks which are explicitly marked oom
 215	 * unkillable or have been already oom reaped or the are in
 216	 * the middle of vfork
 217	 */
 218	adj = (long)p->signal->oom_score_adj;
 219	if (adj == OOM_SCORE_ADJ_MIN ||
 220			test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
 221			in_vfork(p)) {
 222		task_unlock(p);
 223		return LONG_MIN;
 224	}
 225
 226	/*
 227	 * The baseline for the badness score is the proportion of RAM that each
 228	 * task's rss, pagetable and swap space use.
 229	 */
 230	points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
 231		mm_pgtables_bytes(p->mm) / PAGE_SIZE;
 232	task_unlock(p);
 233
 
 
 
 
 
 
 
 234	/* Normalize to oom_score_adj units */
 235	adj *= totalpages / 1000;
 236	points += adj;
 237
 238	return points;
 
 
 
 
 239}
 240
 241static const char * const oom_constraint_text[] = {
 242	[CONSTRAINT_NONE] = "CONSTRAINT_NONE",
 243	[CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
 244	[CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
 245	[CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
 246};
 247
 248/*
 249 * Determine the type of allocation constraint.
 250 */
 251static enum oom_constraint constrained_alloc(struct oom_control *oc)
 
 
 
 252{
 253	struct zone *zone;
 254	struct zoneref *z;
 255	enum zone_type highest_zoneidx = gfp_zone(oc->gfp_mask);
 256	bool cpuset_limited = false;
 257	int nid;
 258
 259	if (is_memcg_oom(oc)) {
 260		oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
 261		return CONSTRAINT_MEMCG;
 262	}
 263
 264	/* Default to all available memory */
 265	oc->totalpages = totalram_pages() + total_swap_pages;
 266
 267	if (!IS_ENABLED(CONFIG_NUMA))
 268		return CONSTRAINT_NONE;
 269
 270	if (!oc->zonelist)
 271		return CONSTRAINT_NONE;
 272	/*
 273	 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
 274	 * to kill current.We have to random task kill in this case.
 275	 * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
 276	 */
 277	if (oc->gfp_mask & __GFP_THISNODE)
 278		return CONSTRAINT_NONE;
 279
 280	/*
 281	 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
 282	 * the page allocator means a mempolicy is in effect.  Cpuset policy
 283	 * is enforced in get_page_from_freelist().
 284	 */
 285	if (oc->nodemask &&
 286	    !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
 287		oc->totalpages = total_swap_pages;
 288		for_each_node_mask(nid, *oc->nodemask)
 289			oc->totalpages += node_present_pages(nid);
 290		return CONSTRAINT_MEMORY_POLICY;
 291	}
 292
 293	/* Check this allocation failure is caused by cpuset's wall function */
 294	for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
 295			highest_zoneidx, oc->nodemask)
 296		if (!cpuset_zone_allowed(zone, oc->gfp_mask))
 297			cpuset_limited = true;
 298
 299	if (cpuset_limited) {
 300		oc->totalpages = total_swap_pages;
 301		for_each_node_mask(nid, cpuset_current_mems_allowed)
 302			oc->totalpages += node_present_pages(nid);
 303		return CONSTRAINT_CPUSET;
 304	}
 305	return CONSTRAINT_NONE;
 306}
 307
 308static int oom_evaluate_task(struct task_struct *task, void *arg)
 
 
 309{
 310	struct oom_control *oc = arg;
 311	long points;
 312
 313	if (oom_unkillable_task(task))
 314		goto next;
 315
 316	/* p may not have freeable memory in nodemask */
 317	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
 318		goto next;
 319
 320	/*
 321	 * This task already has access to memory reserves and is being killed.
 322	 * Don't allow any other task to have access to the reserves unless
 323	 * the task has MMF_OOM_SKIP because chances that it would release
 324	 * any memory is quite low.
 325	 */
 326	if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
 327		if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
 328			goto next;
 329		goto abort;
 330	}
 331
 332	/*
 333	 * If task is allocating a lot of memory and has been marked to be
 334	 * killed first if it triggers an oom, then select it.
 335	 */
 336	if (oom_task_origin(task)) {
 337		points = LONG_MAX;
 338		goto select;
 339	}
 340
 341	points = oom_badness(task, oc->totalpages);
 342	if (points == LONG_MIN || points < oc->chosen_points)
 343		goto next;
 344
 345select:
 346	if (oc->chosen)
 347		put_task_struct(oc->chosen);
 348	get_task_struct(task);
 349	oc->chosen = task;
 350	oc->chosen_points = points;
 351next:
 352	return 0;
 353abort:
 354	if (oc->chosen)
 355		put_task_struct(oc->chosen);
 356	oc->chosen = (void *)-1UL;
 357	return 1;
 358}
 
 359
 360/*
 361 * Simple selection loop. We choose the process with the highest number of
 362 * 'points'. In case scan was aborted, oc->chosen is set to -1.
 
 
 363 */
 364static void select_bad_process(struct oom_control *oc)
 
 
 365{
 366	oc->chosen_points = LONG_MIN;
 
 
 367
 368	if (is_memcg_oom(oc))
 369		mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
 370	else {
 371		struct task_struct *p;
 372
 373		rcu_read_lock();
 374		for_each_process(p)
 375			if (oom_evaluate_task(p, oc))
 376				break;
 377		rcu_read_unlock();
 378	}
 379}
 380
 381static int dump_task(struct task_struct *p, void *arg)
 382{
 383	struct oom_control *oc = arg;
 384	struct task_struct *task;
 385
 386	if (oom_unkillable_task(p))
 387		return 0;
 388
 389	/* p may not have freeable memory in nodemask */
 390	if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
 391		return 0;
 
 392
 393	task = find_lock_task_mm(p);
 394	if (!task) {
 395		/*
 396		 * All of p's threads have already detached their mm's. There's
 397		 * no need to report them; they can't be oom killed anyway.
 
 
 
 
 
 398		 */
 399		return 0;
 400	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 401
 402	pr_info("[%7d] %5d %5d %8lu %8lu %8lu %8lu %9lu %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		get_mm_counter(task->mm, MM_ANONPAGES), get_mm_counter(task->mm, MM_FILEPAGES),
 406		get_mm_counter(task->mm, MM_SHMEMPAGES), mm_pgtables_bytes(task->mm),
 407		get_mm_counter(task->mm, MM_SWAPENTS),
 408		task->signal->oom_score_adj, task->comm);
 409	task_unlock(task);
 410
 411	return 0;
 
 412}
 413
 414/**
 415 * dump_tasks - dump current memory state of all system tasks
 416 * @oc: pointer to struct oom_control
 
 417 *
 418 * Dumps the current memory state of all eligible tasks.  Tasks not in the same
 419 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
 420 * are not shown.
 421 * State information includes task's pid, uid, tgid, vm size, rss,
 422 * pgtables_bytes, swapents, oom_score_adj value, and name.
 
 
 423 */
 424static void dump_tasks(struct oom_control *oc)
 425{
 426	pr_info("Tasks state (memory values in pages):\n");
 427	pr_info("[  pid  ]   uid  tgid total_vm      rss rss_anon rss_file rss_shmem pgtables_bytes swapents oom_score_adj name\n");
 428
 429	if (is_memcg_oom(oc))
 430		mem_cgroup_scan_tasks(oc->memcg, dump_task, oc);
 431	else {
 432		struct task_struct *p;
 433
 434		rcu_read_lock();
 435		for_each_process(p)
 436			dump_task(p, oc);
 437		rcu_read_unlock();
 438	}
 439}
 440
 441static void dump_oom_victim(struct oom_control *oc, struct task_struct *victim)
 442{
 443	/* one line summary of the oom killer context. */
 444	pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
 445			oom_constraint_text[oc->constraint],
 446			nodemask_pr_args(oc->nodemask));
 447	cpuset_print_current_mems_allowed();
 448	mem_cgroup_print_oom_context(oc->memcg, victim);
 449	pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
 450		from_kuid(&init_user_ns, task_uid(victim)));
 451}
 452
 453static void dump_header(struct oom_control *oc)
 454{
 455	pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
 456		current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
 457			current->signal->oom_score_adj);
 458	if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
 459		pr_warn("COMPACTION is disabled!!!\n");
 460
 461	dump_stack();
 462	if (is_memcg_oom(oc))
 463		mem_cgroup_print_oom_meminfo(oc->memcg);
 464	else {
 465		__show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask, gfp_zone(oc->gfp_mask));
 466		if (should_dump_unreclaim_slab())
 467			dump_unreclaimable_slab();
 468	}
 469	if (sysctl_oom_dump_tasks)
 470		dump_tasks(oc);
 471}
 472
 473/*
 474 * Number of OOM victims in flight
 475 */
 476static atomic_t oom_victims = ATOMIC_INIT(0);
 477static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);
 478
 479static bool oom_killer_disabled __read_mostly;
 480
 481/*
 482 * task->mm can be NULL if the task is the exited group leader.  So to
 483 * determine whether the task is using a particular mm, we examine all the
 484 * task's threads: if one of those is using this mm then this task was also
 485 * using it.
 486 */
 487bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
 488{
 489	struct task_struct *t;
 490
 491	for_each_thread(p, t) {
 492		struct mm_struct *t_mm = READ_ONCE(t->mm);
 493		if (t_mm)
 494			return t_mm == mm;
 495	}
 496	return false;
 497}
 498
 499#ifdef CONFIG_MMU
 500/*
 501 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
 502 * victim (if that is possible) to help the OOM killer to move on.
 503 */
 504static struct task_struct *oom_reaper_th;
 505static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
 506static struct task_struct *oom_reaper_list;
 507static DEFINE_SPINLOCK(oom_reaper_lock);
 508
 509static bool __oom_reap_task_mm(struct mm_struct *mm)
 510{
 511	struct vm_area_struct *vma;
 512	bool ret = true;
 513	VMA_ITERATOR(vmi, mm, 0);
 514
 515	/*
 516	 * Tell all users of get_user/copy_from_user etc... that the content
 517	 * is no longer stable. No barriers really needed because unmapping
 518	 * should imply barriers already and the reader would hit a page fault
 519	 * if it stumbled over a reaped memory.
 520	 */
 521	set_bit(MMF_UNSTABLE, &mm->flags);
 522
 523	for_each_vma(vmi, vma) {
 524		if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
 525			continue;
 526
 527		/*
 528		 * Only anonymous pages have a good chance to be dropped
 529		 * without additional steps which we cannot afford as we
 530		 * are OOM already.
 531		 *
 532		 * We do not even care about fs backed pages because all
 533		 * which are reclaimable have already been reclaimed and
 534		 * we do not want to block exit_mmap by keeping mm ref
 535		 * count elevated without a good reason.
 536		 */
 537		if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
 538			struct mmu_notifier_range range;
 539			struct mmu_gather tlb;
 540
 541			mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
 542						mm, vma->vm_start,
 543						vma->vm_end);
 544			tlb_gather_mmu(&tlb, mm);
 545			if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
 546				tlb_finish_mmu(&tlb);
 547				ret = false;
 548				continue;
 549			}
 550			unmap_page_range(&tlb, vma, range.start, range.end, NULL);
 551			mmu_notifier_invalidate_range_end(&range);
 552			tlb_finish_mmu(&tlb);
 553		}
 554	}
 555
 556	return ret;
 557}
 558
 559/*
 560 * Reaps the address space of the give task.
 561 *
 562 * Returns true on success and false if none or part of the address space
 563 * has been reclaimed and the caller should retry later.
 564 */
 565static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
 566{
 567	bool ret = true;
 568
 569	if (!mmap_read_trylock(mm)) {
 570		trace_skip_task_reaping(tsk->pid);
 571		return false;
 572	}
 573
 574	/*
 575	 * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
 576	 * work on the mm anymore. The check for MMF_OOM_SKIP must run
 577	 * under mmap_lock for reading because it serializes against the
 578	 * mmap_write_lock();mmap_write_unlock() cycle in exit_mmap().
 579	 */
 580	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
 581		trace_skip_task_reaping(tsk->pid);
 582		goto out_unlock;
 583	}
 584
 585	trace_start_task_reaping(tsk->pid);
 586
 587	/* failed to reap part of the address space. Try again later */
 588	ret = __oom_reap_task_mm(mm);
 589	if (!ret)
 590		goto out_finish;
 591
 592	pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
 593			task_pid_nr(tsk), tsk->comm,
 594			K(get_mm_counter(mm, MM_ANONPAGES)),
 595			K(get_mm_counter(mm, MM_FILEPAGES)),
 596			K(get_mm_counter(mm, MM_SHMEMPAGES)));
 597out_finish:
 598	trace_finish_task_reaping(tsk->pid);
 599out_unlock:
 600	mmap_read_unlock(mm);
 601
 602	return ret;
 603}
 604
 605#define MAX_OOM_REAP_RETRIES 10
 606static void oom_reap_task(struct task_struct *tsk)
 607{
 608	int attempts = 0;
 609	struct mm_struct *mm = tsk->signal->oom_mm;
 
 
 
 
 
 
 
 
 610
 611	/* Retry the mmap_read_trylock(mm) a few times */
 612	while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
 613		schedule_timeout_idle(HZ/10);
 614
 615	if (attempts <= MAX_OOM_REAP_RETRIES ||
 616	    test_bit(MMF_OOM_SKIP, &mm->flags))
 617		goto done;
 618
 619	pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
 620		task_pid_nr(tsk), tsk->comm);
 621	sched_show_task(tsk);
 622	debug_show_all_locks();
 623
 624done:
 625	tsk->oom_reaper_list = NULL;
 
 
 
 
 
 626
 627	/*
 628	 * Hide this mm from OOM killer because it has been either reaped or
 629	 * somebody can't call mmap_write_unlock(mm).
 
 
 
 
 
 630	 */
 631	set_bit(MMF_OOM_SKIP, &mm->flags);
 632
 633	/* Drop a reference taken by queue_oom_reaper */
 634	put_task_struct(tsk);
 635}
 636
 637static int oom_reaper(void *unused)
 638{
 639	set_freezable();
 640
 641	while (true) {
 642		struct task_struct *tsk = NULL;
 643
 644		wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
 645		spin_lock_irq(&oom_reaper_lock);
 646		if (oom_reaper_list != NULL) {
 647			tsk = oom_reaper_list;
 648			oom_reaper_list = tsk->oom_reaper_list;
 649		}
 650		spin_unlock_irq(&oom_reaper_lock);
 651
 652		if (tsk)
 653			oom_reap_task(tsk);
 654	}
 655
 656	return 0;
 657}
 658
 659static void wake_oom_reaper(struct timer_list *timer)
 660{
 661	struct task_struct *tsk = container_of(timer, struct task_struct,
 662			oom_reaper_timer);
 663	struct mm_struct *mm = tsk->signal->oom_mm;
 664	unsigned long flags;
 665
 666	/* The victim managed to terminate on its own - see exit_mmap */
 667	if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
 668		put_task_struct(tsk);
 669		return;
 670	}
 671
 672	spin_lock_irqsave(&oom_reaper_lock, flags);
 673	tsk->oom_reaper_list = oom_reaper_list;
 674	oom_reaper_list = tsk;
 675	spin_unlock_irqrestore(&oom_reaper_lock, flags);
 676	trace_wake_reaper(tsk->pid);
 677	wake_up(&oom_reaper_wait);
 678}
 
 679
 680/*
 681 * Give the OOM victim time to exit naturally before invoking the oom_reaping.
 682 * The timers timeout is arbitrary... the longer it is, the longer the worst
 683 * case scenario for the OOM can take. If it is too small, the oom_reaper can
 684 * get in the way and release resources needed by the process exit path.
 685 * e.g. The futex robust list can sit in Anon|Private memory that gets reaped
 686 * before the exit path is able to wake the futex waiters.
 687 */
 688#define OOM_REAPER_DELAY (2*HZ)
 689static void queue_oom_reaper(struct task_struct *tsk)
 690{
 691	/* mm is already queued? */
 692	if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
 693		return;
 694
 695	get_task_struct(tsk);
 696	timer_setup(&tsk->oom_reaper_timer, wake_oom_reaper, 0);
 697	tsk->oom_reaper_timer.expires = jiffies + OOM_REAPER_DELAY;
 698	add_timer(&tsk->oom_reaper_timer);
 699}
 700
 701#ifdef CONFIG_SYSCTL
 702static struct ctl_table vm_oom_kill_table[] = {
 703	{
 704		.procname	= "panic_on_oom",
 705		.data		= &sysctl_panic_on_oom,
 706		.maxlen		= sizeof(sysctl_panic_on_oom),
 707		.mode		= 0644,
 708		.proc_handler	= proc_dointvec_minmax,
 709		.extra1		= SYSCTL_ZERO,
 710		.extra2		= SYSCTL_TWO,
 711	},
 712	{
 713		.procname	= "oom_kill_allocating_task",
 714		.data		= &sysctl_oom_kill_allocating_task,
 715		.maxlen		= sizeof(sysctl_oom_kill_allocating_task),
 716		.mode		= 0644,
 717		.proc_handler	= proc_dointvec,
 718	},
 719	{
 720		.procname	= "oom_dump_tasks",
 721		.data		= &sysctl_oom_dump_tasks,
 722		.maxlen		= sizeof(sysctl_oom_dump_tasks),
 723		.mode		= 0644,
 724		.proc_handler	= proc_dointvec,
 725	},
 726	{}
 727};
 728#endif
 729
 730static int __init oom_init(void)
 731{
 732	oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
 733#ifdef CONFIG_SYSCTL
 734	register_sysctl_init("vm", vm_oom_kill_table);
 735#endif
 736	return 0;
 737}
 738subsys_initcall(oom_init)
 739#else
 740static inline void queue_oom_reaper(struct task_struct *tsk)
 741{
 742}
 743#endif /* CONFIG_MMU */
 744
 745/**
 746 * mark_oom_victim - mark the given task as OOM victim
 747 * @tsk: task to mark
 748 *
 749 * Has to be called with oom_lock held and never after
 750 * oom has been disabled already.
 751 *
 752 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
 753 * under task_lock or operate on the current).
 754 */
 755static void mark_oom_victim(struct task_struct *tsk)
 756{
 757	struct mm_struct *mm = tsk->mm;
 758
 759	WARN_ON(oom_killer_disabled);
 760	/* OOM killer might race with memcg OOM */
 761	if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
 762		return;
 763
 764	/* oom_mm is bound to the signal struct life time. */
 765	if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm))
 766		mmgrab(tsk->signal->oom_mm);
 767
 768	/*
 769	 * Make sure that the task is woken up from uninterruptible sleep
 770	 * if it is frozen because OOM killer wouldn't be able to free
 771	 * any memory and livelock. freezing_slow_path will tell the freezer
 772	 * that TIF_MEMDIE tasks should be ignored.
 773	 */
 774	__thaw_task(tsk);
 775	atomic_inc(&oom_victims);
 776	trace_mark_victim(tsk->pid);
 777}
 778
 779/**
 780 * exit_oom_victim - note the exit of an OOM victim
 781 */
 782void exit_oom_victim(void)
 783{
 784	clear_thread_flag(TIF_MEMDIE);
 785
 786	if (!atomic_dec_return(&oom_victims))
 787		wake_up_all(&oom_victims_wait);
 
 
 
 
 
 
 788}
 
 789
 790/**
 791 * oom_killer_enable - enable OOM killer
 792 */
 793void oom_killer_enable(void)
 794{
 795	oom_killer_disabled = false;
 796	pr_info("OOM killer enabled.\n");
 797}
 798
 799/**
 800 * oom_killer_disable - disable OOM killer
 801 * @timeout: maximum timeout to wait for oom victims in jiffies
 802 *
 803 * Forces all page allocations to fail rather than trigger OOM killer.
 804 * Will block and wait until all OOM victims are killed or the given
 805 * timeout expires.
 806 *
 807 * The function cannot be called when there are runnable user tasks because
 808 * the userspace would see unexpected allocation failures as a result. Any
 809 * new usage of this function should be consulted with MM people.
 810 *
 811 * Returns true if successful and false if the OOM killer cannot be
 812 * disabled.
 813 */
 814bool oom_killer_disable(signed long timeout)
 815{
 816	signed long ret;
 817
 818	/*
 819	 * Make sure to not race with an ongoing OOM killer. Check that the
 820	 * current is not killed (possibly due to sharing the victim's memory).
 821	 */
 822	if (mutex_lock_killable(&oom_lock))
 823		return false;
 824	oom_killer_disabled = true;
 825	mutex_unlock(&oom_lock);
 826
 827	ret = wait_event_interruptible_timeout(oom_victims_wait,
 828			!atomic_read(&oom_victims), timeout);
 829	if (ret <= 0) {
 830		oom_killer_enable();
 831		return false;
 832	}
 833	pr_info("OOM killer disabled.\n");
 834
 835	return true;
 836}
 
 837
 838static inline bool __task_will_free_mem(struct task_struct *task)
 839{
 840	struct signal_struct *sig = task->signal;
 841
 842	/*
 843	 * A coredumping process may sleep for an extended period in
 844	 * coredump_task_exit(), so the oom killer cannot assume that
 845	 * the process will promptly exit and release memory.
 846	 */
 847	if (sig->core_state)
 848		return false;
 849
 850	if (sig->flags & SIGNAL_GROUP_EXIT)
 851		return true;
 852
 853	if (thread_group_empty(task) && (task->flags & PF_EXITING))
 854		return true;
 855
 856	return false;
 857}
 
 858
 859/*
 860 * Checks whether the given task is dying or exiting and likely to
 861 * release its address space. This means that all threads and processes
 862 * sharing the same mm have to be killed or exiting.
 863 * Caller has to make sure that task->mm is stable (hold task_lock or
 864 * it operates on the current).
 865 */
 866static bool task_will_free_mem(struct task_struct *task)
 867{
 868	struct mm_struct *mm = task->mm;
 869	struct task_struct *p;
 870	bool ret = true;
 871
 872	/*
 873	 * Skip tasks without mm because it might have passed its exit_mm and
 874	 * exit_oom_victim. oom_reaper could have rescued that but do not rely
 875	 * on that for now. We can consider find_lock_task_mm in future.
 876	 */
 877	if (!mm)
 878		return false;
 879
 880	if (!__task_will_free_mem(task))
 881		return false;
 882
 883	/*
 884	 * This task has already been drained by the oom reaper so there are
 885	 * only small chances it will free some more
 886	 */
 887	if (test_bit(MMF_OOM_SKIP, &mm->flags))
 888		return false;
 889
 890	if (atomic_read(&mm->mm_users) <= 1)
 891		return true;
 892
 893	/*
 894	 * Make sure that all tasks which share the mm with the given tasks
 895	 * are dying as well to make sure that a) nobody pins its mm and
 896	 * b) the task is also reapable by the oom reaper.
 897	 */
 898	rcu_read_lock();
 899	for_each_process(p) {
 900		if (!process_shares_mm(p, mm))
 901			continue;
 902		if (same_thread_group(task, p))
 903			continue;
 904		ret = __task_will_free_mem(p);
 905		if (!ret)
 906			break;
 907	}
 908	rcu_read_unlock();
 909
 910	return ret;
 911}
 912
 913static void __oom_kill_process(struct task_struct *victim, const char *message)
 914{
 915	struct task_struct *p;
 916	struct mm_struct *mm;
 917	bool can_oom_reap = true;
 918
 919	p = find_lock_task_mm(victim);
 920	if (!p) {
 921		pr_info("%s: OOM victim %d (%s) is already exiting. Skip killing the task\n",
 922			message, task_pid_nr(victim), victim->comm);
 923		put_task_struct(victim);
 924		return;
 925	} else if (victim != p) {
 926		get_task_struct(p);
 927		put_task_struct(victim);
 928		victim = p;
 929	}
 930
 931	/* Get a reference to safely compare mm after task_unlock(victim) */
 932	mm = victim->mm;
 933	mmgrab(mm);
 934
 935	/* Raise event before sending signal: task reaper must see this */
 936	count_vm_event(OOM_KILL);
 937	memcg_memory_event_mm(mm, MEMCG_OOM_KILL);
 938
 939	/*
 940	 * We should send SIGKILL before granting access to memory reserves
 941	 * in order to prevent the OOM victim from depleting the memory
 942	 * reserves from the user space under its control.
 943	 */
 944	do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
 945	mark_oom_victim(victim);
 946	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",
 947		message, task_pid_nr(victim), victim->comm, K(mm->total_vm),
 948		K(get_mm_counter(mm, MM_ANONPAGES)),
 949		K(get_mm_counter(mm, MM_FILEPAGES)),
 950		K(get_mm_counter(mm, MM_SHMEMPAGES)),
 951		from_kuid(&init_user_ns, task_uid(victim)),
 952		mm_pgtables_bytes(mm) >> 10, victim->signal->oom_score_adj);
 953	task_unlock(victim);
 954
 955	/*
 956	 * Kill all user processes sharing victim->mm in other thread groups, if
 957	 * any.  They don't get access to memory reserves, though, to avoid
 958	 * depletion of all memory.  This prevents mm->mmap_lock livelock when an
 959	 * oom killed thread cannot exit because it requires the semaphore and
 960	 * its contended by another thread trying to allocate memory itself.
 961	 * That thread will now get access to memory reserves since it has a
 962	 * pending fatal signal.
 963	 */
 964	rcu_read_lock();
 965	for_each_process(p) {
 966		if (!process_shares_mm(p, mm))
 967			continue;
 968		if (same_thread_group(p, victim))
 969			continue;
 970		if (is_global_init(p)) {
 971			can_oom_reap = false;
 972			set_bit(MMF_OOM_SKIP, &mm->flags);
 973			pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
 974					task_pid_nr(victim), victim->comm,
 975					task_pid_nr(p), p->comm);
 976			continue;
 977		}
 978		/*
 979		 * No kthread_use_mm() user needs to read from the userspace so
 980		 * we are ok to reap it.
 
 981		 */
 982		if (unlikely(p->flags & PF_KTHREAD))
 983			continue;
 984		do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
 985	}
 986	rcu_read_unlock();
 987
 988	if (can_oom_reap)
 989		queue_oom_reaper(victim);
 990
 991	mmdrop(mm);
 992	put_task_struct(victim);
 
 993}
 994
 995/*
 996 * Kill provided task unless it's secured by setting
 997 * oom_score_adj to OOM_SCORE_ADJ_MIN.
 
 998 */
 999static int oom_kill_memcg_member(struct task_struct *task, void *message)
1000{
1001	if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
1002	    !is_global_init(task)) {
1003		get_task_struct(task);
1004		__oom_kill_process(task, message);
1005	}
1006	return 0;
1007}
1008
1009static void oom_kill_process(struct oom_control *oc, const char *message)
1010{
1011	struct task_struct *victim = oc->chosen;
1012	struct mem_cgroup *oom_group;
1013	static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1014					      DEFAULT_RATELIMIT_BURST);
1015
1016	/*
1017	 * If the task is already exiting, don't alarm the sysadmin or kill
1018	 * its children or threads, just give it access to memory reserves
1019	 * so it can die quickly
1020	 */
1021	task_lock(victim);
1022	if (task_will_free_mem(victim)) {
1023		mark_oom_victim(victim);
1024		queue_oom_reaper(victim);
1025		task_unlock(victim);
1026		put_task_struct(victim);
1027		return;
1028	}
1029	task_unlock(victim);
1030
1031	if (__ratelimit(&oom_rs)) {
1032		dump_header(oc);
1033		dump_oom_victim(oc, victim);
1034	}
1035
1036	/*
1037	 * Do we need to kill the entire memory cgroup?
1038	 * Or even one of the ancestor memory cgroups?
1039	 * Check this out before killing the victim task.
1040	 */
1041	oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);
1042
1043	__oom_kill_process(victim, message);
1044
1045	/*
1046	 * If necessary, kill all tasks in the selected memory cgroup.
1047	 */
1048	if (oom_group) {
1049		memcg_memory_event(oom_group, MEMCG_OOM_GROUP_KILL);
1050		mem_cgroup_print_oom_group(oom_group);
1051		mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
1052				      (void *)message);
1053		mem_cgroup_put(oom_group);
1054	}
 
1055}
1056
1057/*
1058 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
 
 
1059 */
1060static void check_panic_on_oom(struct oom_control *oc)
1061{
1062	if (likely(!sysctl_panic_on_oom))
1063		return;
1064	if (sysctl_panic_on_oom != 2) {
1065		/*
1066		 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
1067		 * does not panic for cpuset, mempolicy, or memcg allocation
1068		 * failures.
1069		 */
1070		if (oc->constraint != CONSTRAINT_NONE)
1071			return;
1072	}
1073	/* Do not panic for oom kills triggered by sysrq */
1074	if (is_sysrq_oom(oc))
1075		return;
1076	dump_header(oc);
1077	panic("Out of memory: %s panic_on_oom is enabled\n",
1078		sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
1079}
1080
1081static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
1082
1083int register_oom_notifier(struct notifier_block *nb)
1084{
1085	return blocking_notifier_chain_register(&oom_notify_list, nb);
 
 
 
 
 
 
 
 
1086}
1087EXPORT_SYMBOL_GPL(register_oom_notifier);
1088
1089int unregister_oom_notifier(struct notifier_block *nb)
 
 
 
 
1090{
1091	return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 
 
 
 
 
1092}
1093EXPORT_SYMBOL_GPL(unregister_oom_notifier);
1094
1095/**
1096 * out_of_memory - kill the "best" process when we run out of memory
1097 * @oc: pointer to struct oom_control
 
 
 
 
1098 *
1099 * If we run out of memory, we have the choice between either
1100 * killing a random task (bad), letting the system crash (worse)
1101 * OR try to be smart about which process to kill. Note that we
1102 * don't have to be perfect here, we just have to be good.
1103 */
1104bool out_of_memory(struct oom_control *oc)
 
1105{
 
 
 
1106	unsigned long freed = 0;
1107
1108	if (oom_killer_disabled)
1109		return false;
1110
1111	if (!is_memcg_oom(oc)) {
1112		blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
1113		if (freed > 0 && !is_sysrq_oom(oc))
1114			/* Got some memory back in the last second. */
1115			return true;
1116	}
1117
1118	/*
1119	 * If current has a pending SIGKILL or is exiting, then automatically
1120	 * select it.  The goal is to allow it to allocate so that it may
1121	 * quickly exit and free its memory.
1122	 */
1123	if (task_will_free_mem(current)) {
1124		mark_oom_victim(current);
1125		queue_oom_reaper(current);
1126		return true;
1127	}
1128
1129	/*
1130	 * The OOM killer does not compensate for IO-less reclaim.
1131	 * But mem_cgroup_oom() has to invoke the OOM killer even
1132	 * if it is a GFP_NOFS allocation.
1133	 */
1134	if (!(oc->gfp_mask & __GFP_FS) && !is_memcg_oom(oc))
1135		return true;
1136
1137	/*
1138	 * Check if there were limitations on the allocation (only relevant for
1139	 * NUMA and memcg) that may require different handling.
1140	 */
1141	oc->constraint = constrained_alloc(oc);
1142	if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
1143		oc->nodemask = NULL;
1144	check_panic_on_oom(oc);
1145
1146	if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
1147	    current->mm && !oom_unkillable_task(current) &&
1148	    oom_cpuset_eligible(current, oc) &&
1149	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
1150		get_task_struct(current);
1151		oc->chosen = current;
1152		oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
1153		return true;
 
 
 
 
 
 
 
 
 
1154	}
1155
1156	select_bad_process(oc);
1157	/* Found nothing?!?! */
1158	if (!oc->chosen) {
1159		dump_header(oc);
1160		pr_warn("Out of memory and no killable processes...\n");
1161		/*
1162		 * If we got here due to an actual allocation at the
1163		 * system level, we cannot survive this and will enter
1164		 * an endless loop in the allocator. Bail out now.
1165		 */
1166		if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
1167			panic("System is deadlocked on memory\n");
1168	}
1169	if (oc->chosen && oc->chosen != (void *)-1UL)
1170		oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
1171				 "Memory cgroup out of memory");
1172	return !!oc->chosen;
 
 
 
 
 
1173}
1174
1175/*
1176 * The pagefault handler calls here because some allocation has failed. We have
1177 * to take care of the memcg OOM here because this is the only safe context without
1178 * any locks held but let the oom killer triggered from the allocation context care
1179 * about the global OOM.
1180 */
1181void pagefault_out_of_memory(void)
1182{
1183	static DEFINE_RATELIMIT_STATE(pfoom_rs, DEFAULT_RATELIMIT_INTERVAL,
1184				      DEFAULT_RATELIMIT_BURST);
1185
1186	if (mem_cgroup_oom_synchronize(true))
1187		return;
1188
1189	if (fatal_signal_pending(current))
1190		return;
1191
1192	if (__ratelimit(&pfoom_rs))
1193		pr_warn("Huh VM_FAULT_OOM leaked out to the #PF handler. Retrying PF\n");
1194}
1195
1196SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
1197{
1198#ifdef CONFIG_MMU
1199	struct mm_struct *mm = NULL;
1200	struct task_struct *task;
1201	struct task_struct *p;
1202	unsigned int f_flags;
1203	bool reap = false;
1204	long ret = 0;
1205
1206	if (flags)
1207		return -EINVAL;
1208
1209	task = pidfd_get_task(pidfd, &f_flags);
1210	if (IS_ERR(task))
1211		return PTR_ERR(task);
1212
1213	/*
1214	 * Make sure to choose a thread which still has a reference to mm
1215	 * during the group exit
1216	 */
1217	p = find_lock_task_mm(task);
1218	if (!p) {
1219		ret = -ESRCH;
1220		goto put_task;
1221	}
1222
1223	mm = p->mm;
1224	mmgrab(mm);
1225
1226	if (task_will_free_mem(p))
1227		reap = true;
1228	else {
1229		/* Error only if the work has not been done already */
1230		if (!test_bit(MMF_OOM_SKIP, &mm->flags))
1231			ret = -EINVAL;
1232	}
1233	task_unlock(p);
1234
1235	if (!reap)
1236		goto drop_mm;
1237
1238	if (mmap_read_lock_killable(mm)) {
1239		ret = -EINTR;
1240		goto drop_mm;
1241	}
1242	/*
1243	 * Check MMF_OOM_SKIP again under mmap_read_lock protection to ensure
1244	 * possible change in exit_mmap is seen
1245	 */
1246	if (!test_bit(MMF_OOM_SKIP, &mm->flags) && !__oom_reap_task_mm(mm))
1247		ret = -EAGAIN;
1248	mmap_read_unlock(mm);
1249
1250drop_mm:
1251	mmdrop(mm);
1252put_task:
1253	put_task_struct(task);
1254	return ret;
1255#else
1256	return -ENOSYS;
1257#endif /* CONFIG_MMU */
1258}