OOM Killer 持续更新中

虚拟地址空间的概念都门儿清，Linux 认为用户进程调用malloc申请了动态内存后不见得马上就会使用这段内存进行写读操作。
Linux使用了拖到最后的分配机制——用时分配机制。
但是，即使是用时分配，也不见得总有内存可分配。当进程太多，显得内存（加上swap）不足时，就有问题了。所有进程使用的内存量PrM是一定的，而系统物理内存量PyM本身又是有限的，也是一定的。

当 用时分配机制导致  PrM > PyM  时，Linux的机制是无奈但又必须Kill一些重要性低的进程来回收内存，即OOM Killer（Out of memory killer）。
 

 

以arm64 arch 为例：

从文件entry.S开始，此文件的作用有必要搞懂：异常向量表。
arch/arm64/kernel/entry.S

el0_ia:
...
bl do_el0_ia_bp_hardening  //bl指令是转移到子程序执行，
                            //并且事先保存当前位置的下一条指令地址到R14寄存器（lr），
                            //于是可以通过将lr的值mov到pc寄存器，来达到函数返回的效果。
arch/arm6/mm/fault.c :: asmlinkage void __exception do_el0_ia_bp_hardening(){do_mem_abort()}
same file fault.c:: void  do_mem_abort(){   struct fault_info *inf=esr_to_fault_info(esr)  inf->fun()；}
same file fault.c:: static const struct fault_info fault_info[]={
{},{},...,{do_translation_fault,SIGSEGV,SEGV_MAPERR,"level 1 translation fault"},...{};

}

kernel/arch/arm64/mm/fault.c中 进行线性地址到物理地址的转换时，第一级页表包含的入口地址有误，
此时do_translation_fault(){do_page_fault(addr,esr,regs);}
 do_page_fault(){当内存不足时会调用：pagefault_out_of_memory();}
void  pagefault_out_of_memory(void){out_of_memory(&oc)}
bool  out_of_memory(struct oom_control *oc){select_bad_process(oc)}
static int oom_evaluate_task(struct task_struct *task, void *arg){oom_badness(task,NULL,oc-->nodemask, oc->totalpages);}
 

 

函数oom_badness()的注释说的比较清楚，如下链接并复制中。

 

 

 

 

 

 

 

 

 

 

 

 

 

 

https://github.com/torvalds/linux/blob/master/mm/oom_kill.c

 

// SPDX-License-Identifier: GPL-2.0-only
/*
 *  linux/mm/oom_kill.c
 * 
 *  Copyright (C)  1998,2000  Rik van Riel
 *    Thanks go out to Claus Fischer for some serious inspiration and
 *    for goading me into coding this file...
 *  Copyright (C)  2010  Google, Inc.
 *    Rewritten by David Rientjes
 *
 *  The routines in this file are used to kill a process when
 *  we're seriously out of memory. This gets called from __alloc_pages()
 *  in mm/page_alloc.c when we really run out of memory.
 *
 *  Since we won't call these routines often (on a well-configured
 *  machine) this file will double as a 'coding guide' and a signpost
 *  for newbie kernel hackers. It features several pointers to major
 *  kernel subsystems and hints as to where to find out what things do.
 */

#include <linux/oom.h>
#include <linux/mm.h>
#include <linux/err.h>
#include <linux/gfp.h>
#include <linux/sched.h>
#include <linux/sched/mm.h>
#include <linux/sched/coredump.h>
#include <linux/sched/task.h>
#include <linux/swap.h>
#include <linux/timex.h>
#include <linux/jiffies.h>
#include <linux/cpuset.h>
#include <linux/export.h>
#include <linux/notifier.h>
#include <linux/memcontrol.h>
#include <linux/mempolicy.h>
#include <linux/security.h>
#include <linux/ptrace.h>
#include <linux/freezer.h>
#include <linux/ftrace.h>
#include <linux/ratelimit.h>
#include <linux/kthread.h>
#include <linux/init.h>
#include <linux/mmu_notifier.h>

#include <asm/tlb.h>
#include "internal.h"
#include "slab.h"

#define CREATE_TRACE_POINTS
#include <trace/events/oom.h>

int sysctl_panic_on_oom;
int sysctl_oom_kill_allocating_task;
int sysctl_oom_dump_tasks = 1;

/*
 * Serializes oom killer invocations (out_of_memory()) from all contexts to
 * prevent from over eager oom killing (e.g. when the oom killer is invoked
 * from different domains).
 *
 * oom_killer_disable() relies on this lock to stabilize oom_killer_disabled
 * and mark_oom_victim
 */
DEFINE_MUTEX(oom_lock);

#ifdef CONFIG_NUMA
/**
 * has_intersects_mems_allowed() - check task eligiblity for kill
 * @start: task struct of which task to consider
 * @mask: nodemask passed to page allocator for mempolicy ooms
 *
 * Task eligibility is determined by whether or not a candidate task, @tsk,
 * shares the same mempolicy nodes as current if it is bound by such a policy
 * and whether or not it has the same set of allowed cpuset nodes.
 */
static bool has_intersects_mems_allowed(struct task_struct *start,
                    const nodemask_t *mask)
{
    struct task_struct *tsk;
    bool ret = false;

    rcu_read_lock();
    for_each_thread(start, tsk) {
        if (mask) {
            /*
             * If this is a mempolicy constrained oom, tsk's
             * cpuset is irrelevant.  Only return true if its
             * mempolicy intersects current, otherwise it may be
             * needlessly killed.
             */
            ret = mempolicy_nodemask_intersects(tsk, mask);
        } else {
            /*
             * This is not a mempolicy constrained oom, so only
             * check the mems of tsk's cpuset.
             */
            ret = cpuset_mems_allowed_intersects(current, tsk);
        }
        if (ret)
            break;
    }
    rcu_read_unlock();

    return ret;
}
#else
static bool has_intersects_mems_allowed(struct task_struct *tsk,
                    const nodemask_t *mask)
{
    return true;
}
#endif /* CONFIG_NUMA */

/*
 * The process p may have detached its own ->mm while exiting or through
 * use_mm(), but one or more of its subthreads may still have a valid
 * pointer.  Return p, or any of its subthreads with a valid ->mm, with
 * task_lock() held.
 */
struct task_struct *find_lock_task_mm(struct task_struct *p)
{
    struct task_struct *t;

    rcu_read_lock();

    for_each_thread(p, t) {
        task_lock(t);
        if (likely(t->mm))
            goto found;
        task_unlock(t);
    }
    t = NULL;
found:
    rcu_read_unlock();

    return t;
}

/*
 * order == -1 means the oom kill is required by sysrq, otherwise only
 * for display purposes.
 */
static inline bool is_sysrq_oom(struct oom_control *oc)
{
    return oc->order == -1;
}

static inline bool is_memcg_oom(struct oom_control *oc)
{
    return oc->memcg != NULL;
}

/* return true if the task is not adequate as candidate victim task. */
static bool oom_unkillable_task(struct task_struct *p,
        struct mem_cgroup *memcg, const nodemask_t *nodemask)
{
    if (is_global_init(p))
        return true;
    if (p->flags & PF_KTHREAD)
        return true;

    /* When mem_cgroup_out_of_memory() and p is not member of the group */
    if (memcg && !task_in_mem_cgroup(p, memcg))
        return true;

    /* p may not have freeable memory in nodemask */
    if (!has_intersects_mems_allowed(p, nodemask))
        return true;

    return false;
}

/*
 * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
 * than all user memory (LRU pages)
 */
static bool is_dump_unreclaim_slabs(void)
{
    unsigned long nr_lru;

    nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
         global_node_page_state(NR_INACTIVE_ANON) +
         global_node_page_state(NR_ACTIVE_FILE) +
         global_node_page_state(NR_INACTIVE_FILE) +
         global_node_page_state(NR_ISOLATED_ANON) +
         global_node_page_state(NR_ISOLATED_FILE) +
         global_node_page_state(NR_UNEVICTABLE);

    return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
}

/**
 * oom_badness - heuristic function to determine which candidate task to kill
 * @p: task struct of which task we should calculate
 * @totalpages: total present RAM allowed for page allocation
 * @memcg: task's memory controller, if constrained
 * @nodemask: nodemask passed to page allocator for mempolicy ooms
 *
 * The heuristic for determining which task to kill is made to be as simple and
 * predictable as possible.  The goal is to return the highest value for the
 * task consuming the most memory to avoid subsequent oom failures.
 */
unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
              const nodemask_t *nodemask, unsigned long totalpages)
{
    long points;
    long adj;

    if (oom_unkillable_task(p, memcg, nodemask))
        return 0;

    p = find_lock_task_mm(p);
    if (!p)
        return 0;

    /*
     * Do not even consider tasks which are explicitly marked oom
     * unkillable or have been already oom reaped or the are in
     * the middle of vfork
     */
    adj = (long)p->signal->oom_score_adj;
    if (adj == OOM_SCORE_ADJ_MIN ||
            test_bit(MMF_OOM_SKIP, &p->mm->flags) ||
            in_vfork(p)) {
        task_unlock(p);
        return 0;
    }

    /*
     * The baseline for the badness score is the proportion of RAM that each
     * task's rss, pagetable and swap space use.
     */
    points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) +
        mm_pgtables_bytes(p->mm) / PAGE_SIZE;
    task_unlock(p);

    /* Normalize to oom_score_adj units */
    adj *= totalpages / 1000;
    points += adj;

    /*
     * Never return 0 for an eligible task regardless of the root bonus and
     * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here).
     */
    return points > 0 ? points : 1;
}

static const char * const oom_constraint_text[] = {
    [CONSTRAINT_NONE] = "CONSTRAINT_NONE",
    [CONSTRAINT_CPUSET] = "CONSTRAINT_CPUSET",
    [CONSTRAINT_MEMORY_POLICY] = "CONSTRAINT_MEMORY_POLICY",
    [CONSTRAINT_MEMCG] = "CONSTRAINT_MEMCG",
};

/*
 * Determine the type of allocation constraint.
 */
static enum oom_constraint constrained_alloc(struct oom_control *oc)
{
    struct zone *zone;
    struct zoneref *z;
    enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
    bool cpuset_limited = false;
    int nid;

    if (is_memcg_oom(oc)) {
        oc->totalpages = mem_cgroup_get_max(oc->memcg) ?: 1;
        return CONSTRAINT_MEMCG;
    }

    /* Default to all available memory */
    oc->totalpages = totalram_pages() + total_swap_pages;

    if (!IS_ENABLED(CONFIG_NUMA))
        return CONSTRAINT_NONE;

    if (!oc->zonelist)
        return CONSTRAINT_NONE;
    /*
     * Reach here only when __GFP_NOFAIL is used. So, we should avoid
     * to kill current.We have to random task kill in this case.
     * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
     */
    if (oc->gfp_mask & __GFP_THISNODE)
        return CONSTRAINT_NONE;

    /*
     * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
     * the page allocator means a mempolicy is in effect.  Cpuset policy
     * is enforced in get_page_from_freelist().
     */
    if (oc->nodemask &&
        !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
        oc->totalpages = total_swap_pages;
        for_each_node_mask(nid, *oc->nodemask)
            oc->totalpages += node_spanned_pages(nid);
        return CONSTRAINT_MEMORY_POLICY;
    }

    /* Check this allocation failure is caused by cpuset's wall function */
    for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
            high_zoneidx, oc->nodemask)
        if (!cpuset_zone_allowed(zone, oc->gfp_mask))
            cpuset_limited = true;

    if (cpuset_limited) {
        oc->totalpages = total_swap_pages;
        for_each_node_mask(nid, cpuset_current_mems_allowed)
            oc->totalpages += node_spanned_pages(nid);
        return CONSTRAINT_CPUSET;
    }
    return CONSTRAINT_NONE;
}

static int oom_evaluate_task(struct task_struct *task, void *arg)
{
    struct oom_control *oc = arg;
    unsigned long points;

    if (oom_unkillable_task(task, NULL, oc->nodemask))
        goto next;

    /*
     * This task already has access to memory reserves and is being killed.
     * Don't allow any other task to have access to the reserves unless
     * the task has MMF_OOM_SKIP because chances that it would release
     * any memory is quite low.
     */
    if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) {
        if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags))
            goto next;
        goto abort;
    }

    /*
     * If task is allocating a lot of memory and has been marked to be
     * killed first if it triggers an oom, then select it.
     */
    if (oom_task_origin(task)) {
        points = ULONG_MAX;
        goto select;
    }

    points = oom_badness(task, NULL, oc->nodemask, oc->totalpages);
    if (!points || points < oc->chosen_points)
        goto next;

    /* Prefer thread group leaders for display purposes */
    if (points == oc->chosen_points && thread_group_leader(oc->chosen))
        goto next;
select:
    if (oc->chosen)
        put_task_struct(oc->chosen);
    get_task_struct(task);
    oc->chosen = task;
    oc->chosen_points = points;
next:
    return 0;
abort:
    if (oc->chosen)
        put_task_struct(oc->chosen);
    oc->chosen = (void *)-1UL;
    return 1;
}

/*
 * Simple selection loop. We choose the process with the highest number of
 * 'points'. In case scan was aborted, oc->chosen is set to -1.
 */
static void select_bad_process(struct oom_control *oc)
{
    if (is_memcg_oom(oc))
        mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc);
    else {
        struct task_struct *p;

        rcu_read_lock();
        for_each_process(p)
            if (oom_evaluate_task(p, oc))
                break;
        rcu_read_unlock();
    }

    oc->chosen_points = oc->chosen_points * 1000 / oc->totalpages;
}

/**
 * dump_tasks - dump current memory state of all system tasks
 * @memcg: current's memory controller, if constrained
 * @nodemask: nodemask passed to page allocator for mempolicy ooms
 *
 * Dumps the current memory state of all eligible tasks.  Tasks not in the same
 * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
 * are not shown.
 * State information includes task's pid, uid, tgid, vm size, rss,
 * pgtables_bytes, swapents, oom_score_adj value, and name.
 */
static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
{
    struct task_struct *p;
    struct task_struct *task;

    pr_info("Tasks state (memory values in pages):\n");
    pr_info("[  pid  ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj name\n");
    rcu_read_lock();
    for_each_process(p) {
        if (oom_unkillable_task(p, memcg, nodemask))
            continue;

        task = find_lock_task_mm(p);
        if (!task) {
            /*
             * This is a kthread or all of p's threads have already
             * detached their mm's.  There's no need to report
             * them; they can't be oom killed anyway.
             */
            continue;
        }

        pr_info("[%7d] %5d %5d %8lu %8lu %8ld %8lu         %5hd %s\n",
            task->pid, from_kuid(&init_user_ns, task_uid(task)),
            task->tgid, task->mm->total_vm, get_mm_rss(task->mm),
            mm_pgtables_bytes(task->mm),
            get_mm_counter(task->mm, MM_SWAPENTS),
            task->signal->oom_score_adj, task->comm);
        task_unlock(task);
    }
    rcu_read_unlock();
}

static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
{
    /* one line summary of the oom killer context. */
    pr_info("oom-kill:constraint=%s,nodemask=%*pbl",
            oom_constraint_text[oc->constraint],
            nodemask_pr_args(oc->nodemask));
    cpuset_print_current_mems_allowed();
    mem_cgroup_print_oom_context(oc->memcg, victim);
    pr_cont(",task=%s,pid=%d,uid=%d\n", victim->comm, victim->pid,
        from_kuid(&init_user_ns, task_uid(victim)));
}

static void dump_header(struct oom_control *oc, struct task_struct *p)
{
    pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
        current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
            current->signal->oom_score_adj);
    if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order)
        pr_warn("COMPACTION is disabled!!!\n");

    dump_stack();
    if (is_memcg_oom(oc))
        mem_cgroup_print_oom_meminfo(oc->memcg);
    else {
        show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
        if (is_dump_unreclaim_slabs())
            dump_unreclaimable_slab();
    }
    if (sysctl_oom_dump_tasks)
        dump_tasks(oc->memcg, oc->nodemask);
    if (p)
        dump_oom_summary(oc, p);
}

/*
 * Number of OOM victims in flight
 */
static atomic_t oom_victims = ATOMIC_INIT(0);
static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait);

static bool oom_killer_disabled __read_mostly;

#define K(x) ((x) << (PAGE_SHIFT-10))

/*
 * task->mm can be NULL if the task is the exited group leader.  So to
 * determine whether the task is using a particular mm, we examine all the
 * task's threads: if one of those is using this mm then this task was also
 * using it.
 */
bool process_shares_mm(struct task_struct *p, struct mm_struct *mm)
{
    struct task_struct *t;

    for_each_thread(p, t) {
        struct mm_struct *t_mm = READ_ONCE(t->mm);
        if (t_mm)
            return t_mm == mm;
    }
    return false;
}

#ifdef CONFIG_MMU
/*
 * OOM Reaper kernel thread which tries to reap the memory used by the OOM
 * victim (if that is possible) to help the OOM killer to move on.
 */
static struct task_struct *oom_reaper_th;
static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait);
static struct task_struct *oom_reaper_list;
static DEFINE_SPINLOCK(oom_reaper_lock);

bool __oom_reap_task_mm(struct mm_struct *mm)
{
    struct vm_area_struct *vma;
    bool ret = true;

    /*
     * Tell all users of get_user/copy_from_user etc... that the content
     * is no longer stable. No barriers really needed because unmapping
     * should imply barriers already and the reader would hit a page fault
     * if it stumbled over a reaped memory.
     */
    set_bit(MMF_UNSTABLE, &mm->flags);

    for (vma = mm->mmap ; vma; vma = vma->vm_next) {
        if (!can_madv_dontneed_vma(vma))
            continue;

        /*
         * Only anonymous pages have a good chance to be dropped
         * without additional steps which we cannot afford as we
         * are OOM already.
         *
         * We do not even care about fs backed pages because all
         * which are reclaimable have already been reclaimed and
         * we do not want to block exit_mmap by keeping mm ref
         * count elevated without a good reason.
         */
        if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) {
            struct mmu_notifier_range range;
            struct mmu_gather tlb;

            mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0,
                        vma, mm, vma->vm_start,
                        vma->vm_end);
            tlb_gather_mmu(&tlb, mm, range.start, range.end);
            if (mmu_notifier_invalidate_range_start_nonblock(&range)) {
                tlb_finish_mmu(&tlb, range.start, range.end);
                ret = false;
                continue;
            }
            unmap_page_range(&tlb, vma, range.start, range.end, NULL);
            mmu_notifier_invalidate_range_end(&range);
            tlb_finish_mmu(&tlb, range.start, range.end);
        }
    }

    return ret;
}

/*
 * Reaps the address space of the give task.
 *
 * Returns true on success and false if none or part of the address space
 * has been reclaimed and the caller should retry later.
 */
static bool oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm)
{
    bool ret = true;

    if (!down_read_trylock(&mm->mmap_sem)) {
        trace_skip_task_reaping(tsk->pid);
        return false;
    }

    /*
     * MMF_OOM_SKIP is set by exit_mmap when the OOM reaper can't
     * work on the mm anymore. The check for MMF_OOM_SKIP must run
     * under mmap_sem for reading because it serializes against the
     * down_write();up_write() cycle in exit_mmap().
     */
    if (test_bit(MMF_OOM_SKIP, &mm->flags)) {
        trace_skip_task_reaping(tsk->pid);
        goto out_unlock;
    }

    trace_start_task_reaping(tsk->pid);

    /* failed to reap part of the address space. Try again later */
    ret = __oom_reap_task_mm(mm);
    if (!ret)
        goto out_finish;

    pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
            task_pid_nr(tsk), tsk->comm,
            K(get_mm_counter(mm, MM_ANONPAGES)),
            K(get_mm_counter(mm, MM_FILEPAGES)),
            K(get_mm_counter(mm, MM_SHMEMPAGES)));
out_finish:
    trace_finish_task_reaping(tsk->pid);
out_unlock:
    up_read(&mm->mmap_sem);

    return ret;
}

#define MAX_OOM_REAP_RETRIES 10
static void oom_reap_task(struct task_struct *tsk)
{
    int attempts = 0;
    struct mm_struct *mm = tsk->signal->oom_mm;

    /* Retry the down_read_trylock(mmap_sem) a few times */
    while (attempts++ < MAX_OOM_REAP_RETRIES && !oom_reap_task_mm(tsk, mm))
        schedule_timeout_idle(HZ/10);

    if (attempts <= MAX_OOM_REAP_RETRIES ||
        test_bit(MMF_OOM_SKIP, &mm->flags))
        goto done;

    pr_info("oom_reaper: unable to reap pid:%d (%s)\n",
        task_pid_nr(tsk), tsk->comm);
    debug_show_all_locks();

done:
    tsk->oom_reaper_list = NULL;

    /*
     * Hide this mm from OOM killer because it has been either reaped or
     * somebody can't call up_write(mmap_sem).
     */
    set_bit(MMF_OOM_SKIP, &mm->flags);

    /* Drop a reference taken by wake_oom_reaper */
    put_task_struct(tsk);
}

static int oom_reaper(void *unused)
{
    while (true) {
        struct task_struct *tsk = NULL;

        wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL);
        spin_lock(&oom_reaper_lock);
        if (oom_reaper_list != NULL) {
            tsk = oom_reaper_list;
            oom_reaper_list = tsk->oom_reaper_list;
        }
        spin_unlock(&oom_reaper_lock);

        if (tsk)
            oom_reap_task(tsk);
    }

    return 0;
}

static void wake_oom_reaper(struct task_struct *tsk)
{
    /* mm is already queued? */
    if (test_and_set_bit(MMF_OOM_REAP_QUEUED, &tsk->signal->oom_mm->flags))
        return;

    get_task_struct(tsk);

    spin_lock(&oom_reaper_lock);
    tsk->oom_reaper_list = oom_reaper_list;
    oom_reaper_list = tsk;
    spin_unlock(&oom_reaper_lock);
    trace_wake_reaper(tsk->pid);
    wake_up(&oom_reaper_wait);
}

static int __init oom_init(void)
{
    oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper");
    return 0;
}
subsys_initcall(oom_init)
#else
static inline void wake_oom_reaper(struct task_struct *tsk)
{
}
#endif /* CONFIG_MMU */

/**
 * mark_oom_victim - mark the given task as OOM victim
 * @tsk: task to mark
 *
 * Has to be called with oom_lock held and never after
 * oom has been disabled already.
 *
 * tsk->mm has to be non NULL and caller has to guarantee it is stable (either
 * under task_lock or operate on the current).
 */
static void mark_oom_victim(struct task_struct *tsk)
{
    struct mm_struct *mm = tsk->mm;

    WARN_ON(oom_killer_disabled);
    /* OOM killer might race with memcg OOM */
    if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
        return;

    /* oom_mm is bound to the signal struct life time. */
    if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) {
        mmgrab(tsk->signal->oom_mm);
        set_bit(MMF_OOM_VICTIM, &mm->flags);
    }

    /*
     * Make sure that the task is woken up from uninterruptible sleep
     * if it is frozen because OOM killer wouldn't be able to free
     * any memory and livelock. freezing_slow_path will tell the freezer
     * that TIF_MEMDIE tasks should be ignored.
     */
    __thaw_task(tsk);
    atomic_inc(&oom_victims);
    trace_mark_victim(tsk->pid);
}

/**
 * exit_oom_victim - note the exit of an OOM victim
 */
void exit_oom_victim(void)
{
    clear_thread_flag(TIF_MEMDIE);

    if (!atomic_dec_return(&oom_victims))
        wake_up_all(&oom_victims_wait);
}

/**
 * oom_killer_enable - enable OOM killer
 */
void oom_killer_enable(void)
{
    oom_killer_disabled = false;
    pr_info("OOM killer enabled.\n");
}

/**
 * oom_killer_disable - disable OOM killer
 * @timeout: maximum timeout to wait for oom victims in jiffies
 *
 * Forces all page allocations to fail rather than trigger OOM killer.
 * Will block and wait until all OOM victims are killed or the given
 * timeout expires.
 *
 * The function cannot be called when there are runnable user tasks because
 * the userspace would see unexpected allocation failures as a result. Any
 * new usage of this function should be consulted with MM people.
 *
 * Returns true if successful and false if the OOM killer cannot be
 * disabled.
 */
bool oom_killer_disable(signed long timeout)
{
    signed long ret;

    /*
     * Make sure to not race with an ongoing OOM killer. Check that the
     * current is not killed (possibly due to sharing the victim's memory).
     */
    if (mutex_lock_killable(&oom_lock))
        return false;
    oom_killer_disabled = true;
    mutex_unlock(&oom_lock);

    ret = wait_event_interruptible_timeout(oom_victims_wait,
            !atomic_read(&oom_victims), timeout);
    if (ret <= 0) {
        oom_killer_enable();
        return false;
    }
    pr_info("OOM killer disabled.\n");

    return true;
}

static inline bool __task_will_free_mem(struct task_struct *task)
{
    struct signal_struct *sig = task->signal;

    /*
     * A coredumping process may sleep for an extended period in exit_mm(),
     * so the oom killer cannot assume that the process will promptly exit
     * and release memory.
     */
    if (sig->flags & SIGNAL_GROUP_COREDUMP)
        return false;

    if (sig->flags & SIGNAL_GROUP_EXIT)
        return true;

    if (thread_group_empty(task) && (task->flags & PF_EXITING))
        return true;

    return false;
}

/*
 * Checks whether the given task is dying or exiting and likely to
 * release its address space. This means that all threads and processes
 * sharing the same mm have to be killed or exiting.
 * Caller has to make sure that task->mm is stable (hold task_lock or
 * it operates on the current).
 */
static bool task_will_free_mem(struct task_struct *task)
{
    struct mm_struct *mm = task->mm;
    struct task_struct *p;
    bool ret = true;

    /*
     * Skip tasks without mm because it might have passed its exit_mm and
     * exit_oom_victim. oom_reaper could have rescued that but do not rely
     * on that for now. We can consider find_lock_task_mm in future.
     */
    if (!mm)
        return false;

    if (!__task_will_free_mem(task))
        return false;

    /*
     * This task has already been drained by the oom reaper so there are
     * only small chances it will free some more
     */
    if (test_bit(MMF_OOM_SKIP, &mm->flags))
        return false;

    if (atomic_read(&mm->mm_users) <= 1)
        return true;

    /*
     * Make sure that all tasks which share the mm with the given tasks
     * are dying as well to make sure that a) nobody pins its mm and
     * b) the task is also reapable by the oom reaper.
     */
    rcu_read_lock();
    for_each_process(p) {
        if (!process_shares_mm(p, mm))
            continue;
        if (same_thread_group(task, p))
            continue;
        ret = __task_will_free_mem(p);
        if (!ret)
            break;
    }
    rcu_read_unlock();

    return ret;
}

static void __oom_kill_process(struct task_struct *victim, const char *message)
{
    struct task_struct *p;
    struct mm_struct *mm;
    bool can_oom_reap = true;

    p = find_lock_task_mm(victim);
    if (!p) {
        put_task_struct(victim);
        return;
    } else if (victim != p) {
        get_task_struct(p);
        put_task_struct(victim);
        victim = p;
    }

    /* Get a reference to safely compare mm after task_unlock(victim) */
    mm = victim->mm;
    mmgrab(mm);

    /* Raise event before sending signal: task reaper must see this */
    count_vm_event(OOM_KILL);
    memcg_memory_event_mm(mm, MEMCG_OOM_KILL);

    /*
     * We should send SIGKILL before granting access to memory reserves
     * in order to prevent the OOM victim from depleting the memory
     * reserves from the user space under its control.
     */
    do_send_sig_info(SIGKILL, SEND_SIG_PRIV, victim, PIDTYPE_TGID);
    mark_oom_victim(victim);
    pr_err("%s: Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n",
        message, task_pid_nr(victim), victim->comm,
        K(victim->mm->total_vm),
        K(get_mm_counter(victim->mm, MM_ANONPAGES)),
        K(get_mm_counter(victim->mm, MM_FILEPAGES)),
        K(get_mm_counter(victim->mm, MM_SHMEMPAGES)));
    task_unlock(victim);

    /*
     * Kill all user processes sharing victim->mm in other thread groups, if
     * any.  They don't get access to memory reserves, though, to avoid
     * depletion of all memory.  This prevents mm->mmap_sem livelock when an
     * oom killed thread cannot exit because it requires the semaphore and
     * its contended by another thread trying to allocate memory itself.
     * That thread will now get access to memory reserves since it has a
     * pending fatal signal.
     */
    rcu_read_lock();
    for_each_process(p) {
        if (!process_shares_mm(p, mm))
            continue;
        if (same_thread_group(p, victim))
            continue;
        if (is_global_init(p)) {
            can_oom_reap = false;
            set_bit(MMF_OOM_SKIP, &mm->flags);
            pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n",
                    task_pid_nr(victim), victim->comm,
                    task_pid_nr(p), p->comm);
            continue;
        }
        /*
         * No use_mm() user needs to read from the userspace so we are
         * ok to reap it.
         */
        if (unlikely(p->flags & PF_KTHREAD))
            continue;
        do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_TGID);
    }
    rcu_read_unlock();

    if (can_oom_reap)
        wake_oom_reaper(victim);

    mmdrop(mm);
    put_task_struct(victim);
}
#undef K

/*
 * Kill provided task unless it's secured by setting
 * oom_score_adj to OOM_SCORE_ADJ_MIN.
 */
static int oom_kill_memcg_member(struct task_struct *task, void *message)
{
    if (task->signal->oom_score_adj != OOM_SCORE_ADJ_MIN &&
        !is_global_init(task)) {
        get_task_struct(task);
        __oom_kill_process(task, message);
    }
    return 0;
}

static void oom_kill_process(struct oom_control *oc, const char *message)
{
    struct task_struct *victim = oc->chosen;
    struct mem_cgroup *oom_group;
    static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL,
                          DEFAULT_RATELIMIT_BURST);

    /*
     * If the task is already exiting, don't alarm the sysadmin or kill
     * its children or threads, just give it access to memory reserves
     * so it can die quickly
     */
    task_lock(victim);
    if (task_will_free_mem(victim)) {
        mark_oom_victim(victim);
        wake_oom_reaper(victim);
        task_unlock(victim);
        put_task_struct(victim);
        return;
    }
    task_unlock(victim);

    if (__ratelimit(&oom_rs))
        dump_header(oc, victim);

    /*
     * Do we need to kill the entire memory cgroup?
     * Or even one of the ancestor memory cgroups?
     * Check this out before killing the victim task.
     */
    oom_group = mem_cgroup_get_oom_group(victim, oc->memcg);

    __oom_kill_process(victim, message);

    /*
     * If necessary, kill all tasks in the selected memory cgroup.
     */
    if (oom_group) {
        mem_cgroup_print_oom_group(oom_group);
        mem_cgroup_scan_tasks(oom_group, oom_kill_memcg_member,
                      (void*)message);
        mem_cgroup_put(oom_group);
    }
}

/*
 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
 */
static void check_panic_on_oom(struct oom_control *oc)
{
    if (likely(!sysctl_panic_on_oom))
        return;
    if (sysctl_panic_on_oom != 2) {
        /*
         * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
         * does not panic for cpuset, mempolicy, or memcg allocation
         * failures.
         */
        if (oc->constraint != CONSTRAINT_NONE)
            return;
    }
    /* Do not panic for oom kills triggered by sysrq */
    if (is_sysrq_oom(oc))
        return;
    dump_header(oc, NULL);
    panic("Out of memory: %s panic_on_oom is enabled\n",
        sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
}

static BLOCKING_NOTIFIER_HEAD(oom_notify_list);

int register_oom_notifier(struct notifier_block *nb)
{
    return blocking_notifier_chain_register(&oom_notify_list, nb);
}
EXPORT_SYMBOL_GPL(register_oom_notifier);

int unregister_oom_notifier(struct notifier_block *nb)
{
    return blocking_notifier_chain_unregister(&oom_notify_list, nb);
}
EXPORT_SYMBOL_GPL(unregister_oom_notifier);

/**
 * out_of_memory - kill the "best" process when we run out of memory
 * @oc: pointer to struct oom_control
 *
 * If we run out of memory, we have the choice between either
 * killing a random task (bad), letting the system crash (worse)
 * OR try to be smart about which process to kill. Note that we
 * don't have to be perfect here, we just have to be good.
 */
bool out_of_memory(struct oom_control *oc)
{
    unsigned long freed = 0;

    if (oom_killer_disabled)
        return false;

    if (!is_memcg_oom(oc)) {
        blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
        if (freed > 0)
            /* Got some memory back in the last second. */
            return true;
    }

    /*
     * If current has a pending SIGKILL or is exiting, then automatically
     * select it.  The goal is to allow it to allocate so that it may
     * quickly exit and free its memory.
     */
    if (task_will_free_mem(current)) {
        mark_oom_victim(current);
        wake_oom_reaper(current);
        return true;
    }

    /*
     * The OOM killer does not compensate for IO-less reclaim.
     * pagefault_out_of_memory lost its gfp context so we have to
     * make sure exclude 0 mask - all other users should have at least
     * ___GFP_DIRECT_RECLAIM to get here.
     */
    if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS))
        return true;

    /*
     * Check if there were limitations on the allocation (only relevant for
     * NUMA and memcg) that may require different handling.
     */
    oc->constraint = constrained_alloc(oc);
    if (oc->constraint != CONSTRAINT_MEMORY_POLICY)
        oc->nodemask = NULL;
    check_panic_on_oom(oc);

    if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task &&
        current->mm && !oom_unkillable_task(current, NULL, oc->nodemask) &&
        current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
        get_task_struct(current);
        oc->chosen = current;
        oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)");
        return true;
    }

    select_bad_process(oc);
    /* Found nothing?!?! */
    if (!oc->chosen) {
        dump_header(oc, NULL);
        pr_warn("Out of memory and no killable processes...\n");
        /*
         * If we got here due to an actual allocation at the
         * system level, we cannot survive this and will enter
         * an endless loop in the allocator. Bail out now.
         */
        if (!is_sysrq_oom(oc) && !is_memcg_oom(oc))
            panic("System is deadlocked on memory\n");
    }
    if (oc->chosen && oc->chosen != (void *)-1UL)
        oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" :
                 "Memory cgroup out of memory");
    return !!oc->chosen;
}

/*
 * The pagefault handler calls here because it is out of memory, so kill a
 * memory-hogging task. If oom_lock is held by somebody else, a parallel oom
 * killing is already in progress so do nothing.
 */
void pagefault_out_of_memory(void)
{
    struct oom_control oc = {
        .zonelist = NULL,
        .nodemask = NULL,
        .memcg = NULL,
        .gfp_mask = 0,
        .order = 0,
    };

    if (mem_cgroup_oom_synchronize(true))
        return;

    if (!mutex_trylock(&oom_lock))
        return;
    out_of_memory(&oc);
    mutex_unlock(&oom_lock);
}