Android中基于CGroup的memory子系统HAL层分析-lmkd

Android在内存管理上于Linux有些小的区别，其中一个就是引入了lowmemorykiller。从lowmemorykiller.c位于drivers/staging/android也可知道，属于Android专有，没有进入Linux kernel的mainline。

lmkd，即Low Memory Killer Daemon，基于memory子系统和Kernel lowmemorykiller功能参数，选择一个合适的进程，然后kill进程，以达到释放内存的目的。所以也绕不开Kernel模块lowmemorykiller(drivers/staging/android/lowmemorykiller.c)。

在考虑一个系统服务的功能，不仅要分析其内部功能，还要对其输入(lmkd socket、memory子系统和lowmemory)和输出(kill)进行详细的分析，才能更好的理解整个lmkd建立的生态。

他们之间的关系可以简要概括如下：

lmkd相关模块关系

启动lmkd系统服务

在/etc/init/lmkd.rc中，启动lmkd系统服务，创建了lmkd socket，并且将lmkd设置为system-background类型的进程。

service lmkd /system/bin/lmkd
    class core
    group root readproc
    critical
    socket lmkd seqpacket 0660 system system
    writepid /dev/cpuset/system-background/tasks

lmkd框架分析

正如上图lmkd相关模块分析中所示，lmkd通过读取CGroup中memory子系统和lowmemory两个模块作为输入参数；输出是kill选定的进程。

正如所有的service一样，lmkd的起点也是main函数，lmkd的main函数很简单：

int main(int argc __unused, char **argv __unused) {
    struct sched_param param = {
            .sched_priority = 1,
    };

mlockall(MCL_FUTURE);  锁住该实时进程在物理内存上全部地址空间。这将阻止Linux将这个内存页调度到交换空间(swap space)，及时该进程已有一段时间没有访问这段空间。参见末尾参考资料。
    sched_setscheduler(0, SCHED_FIFO, &param);  设置lmkd调度类型为SCHED_FIFO的实时进程。
    if (!init())  初始化，主要是socket通信，epoll文件操作memory子系统sysfs
        mainloop();  epoll_wait处理epollfd

ALOGI("exiting");
    return 0;
}

下面来分析一下主要核心函数init：

static int init(void) {
    struct epoll_event epev;
    int i;
    int ret;

page_k = sysconf(_SC_PAGESIZE);
    if (page_k == -1)
        page_k = PAGE_SIZE;
    page_k /= 1024;

epollfd = epoll_create(MAX_EPOLL_EVENTS);  创建全局epoll文件句柄
    if (epollfd == -1) {
        ALOGE("epoll_create failed (errno=%d)", errno);
        return -1;
    }

ctrl_lfd = android_get_control_socket("lmkd");  打开lmkd socket文件句柄
    if (ctrl_lfd < 0) {
        ALOGE("get lmkd control socket failed");
        return -1;
    }

ret = listen(ctrl_lfd, 1);
    if (ret < 0) {
        ALOGE("lmkd control socket listen failed (errno=%d)", errno);
        return -1;
    }

epev.events = EPOLLIN;
    epev.data.ptr = (void *)ctrl_connect_handler;
    if (epoll_ctl(epollfd, EPOLL_CTL_ADD, ctrl_lfd, &epev) == -1) {  将lmkd socket加入epoll，处理函数问ctrl_connect_handler
        ALOGE("epoll_ctl for lmkd control socket failed (errno=%d)", errno);
        return -1;
    }
    maxevents++;

use_inkernel_interface = !access(INKERNEL_MINFREE_PATH, W_OK);

if (use_inkernel_interface) {
        ALOGI("Using in-kernel low memory killer interface");
    } else {
        ret = init_mp(MEMPRESSURE_WATCH_LEVEL, (void *)&mp_event); 处理memory pressure相关
        if (ret)
            ALOGE("Kernel does not support memory pressure events or in-kernel low memory killer");
    }

for (i = 0; i <= ADJTOSLOT(OOM_SCORE_ADJ_MAX); i++) {
        procadjslot_list[i].next = &procadjslot_list[i];
        procadjslot_list[i].prev = &procadjslot_list[i];
    }

return 0;
}

1.创建epollfd文件，MAX_EPOLL_EVENTS为3,；

2.连接到lmkd socket，并将文件句柄加到epollfd，EPOLLIN的句柄函数问ctrl_connect_handler。

3.init_mp初始化memory pressure相关参数，创建一个用于事件通知的文件句柄，加入到epollfd，EPOLLIN的处理函数为mp_event。

init_mp将memory.presure_level的句柄，和创建用于本进程事件通知的evfd，然后和levelstr一起写入cgroup.event_control。

static int init_mp(char *levelstr, void *event_handler)
{
    int mpfd;
    int evfd;
    int evctlfd;
    char buf[256];
    struct epoll_event epev;
    int ret;

mpfd = open(MEMCG_SYSFS_PATH "memory.pressure_level", O_RDONLY | O_CLOEXEC);
    if (mpfd < 0) {
        ALOGI("No kernel memory.pressure_level support (errno=%d)", errno);
        goto err_open_mpfd;
    }

evctlfd = open(MEMCG_SYSFS_PATH "cgroup.event_control", O_WRONLY | O_CLOEXEC);
    if (evctlfd < 0) {
        ALOGI("No kernel memory cgroup event control (errno=%d)", errno);
        goto err_open_evctlfd;
    }

evfd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC); 参见末尾参考资料，eventfd用于创建本进程事件通知的文件句柄。
    if (evfd < 0) {
        ALOGE("eventfd failed for level %s; errno=%d", levelstr, errno);
        goto err_eventfd;
    }

ret = snprintf(buf, sizeof(buf), "%d %d %s", evfd, mpfd, levelstr); ？？？
    if (ret >= (ssize_t)sizeof(buf)) {
        ALOGE("cgroup.event_control line overflow for level %s", levelstr);
        goto err;
    }

ret = write(evctlfd, buf, strlen(buf) + 1);
    if (ret == -1) {
        ALOGE("cgroup.event_control write failed for level %s; errno=%d",
              levelstr, errno);
        goto err;
    }

epev.events = EPOLLIN;
    epev.data.ptr = event_handler;
    ret = epoll_ctl(epollfd, EPOLL_CTL_ADD, evfd, &epev);
    if (ret == -1) {
        ALOGE("epoll_ctl for level %s failed; errno=%d", levelstr, errno);
        goto err;
    }
    maxevents++;
    mpevfd = evfd;
    return 0;

err:
    close(evfd);
err_eventfd:
    close(evctlfd);
err_open_evctlfd:
    close(mpfd);
err_open_mpfd:
    return -1;
}

ctrl_connect_handler是lmkd socket相关句柄函数，accept之后又会创建ctrl_dfd句柄。如果是EPOLLHUP，则关闭ctrl_dfd；如果是EPOLLIN，则会根据cmd类型进行不同处理。

static void ctrl_data_handler(uint32_t events) {
    if (events & EPOLLHUP) {
        ALOGI("ActivityManager disconnected");
        if (!ctrl_dfd_reopened)
            ctrl_data_close();
    } else if (events & EPOLLIN) {
        ctrl_command_handler();
    }
}

LMK_TARGET类型对应cmd_targt，用于设置"/sys/module/lowmemorykiller/parameters/minfree"和"/sys/module/lowmemorykiller/parameters/adj"。

LMK_PROCPRIO类型对应cmd_procprio，用于写入/proc/xxx/oom_score_adj，并将pid加入pidhash表中。

LMK_PROCREMOVE类型对应cmd_procremove，用于将pid从pidhash中移除。

在vmpressure上报low事件后，lmkd就会触发mp_event处理memory pressure相关事件。mp_event就是low的处理函数，通过kill进程来释放内存空间。

static void mp_event(uint32_t events __unused) {
    int ret;
    unsigned long long evcount;
    struct sysmeminfo mi;
    int other_free;
    int other_file;
    int killed_size;
    bool first = true;

ret = read(mpevfd, &evcount, sizeof(evcount));
    if (ret < 0)
        ALOGE("Error reading memory pressure event fd; errno=%d",
              errno);

if (time(NULL) - kill_lasttime < KILL_TIMEOUT)
        return;

while (zoneinfo_parse(&mi) < 0) {
        // Failed to read /proc/zoneinfo, assume ENOMEM and kill something
        find_and_kill_process(0, 0, true);
    }  解析/proc/zoneinfo，主要解析nr_free_pages、nr_file_pages、nr_shmem、high、protection:。

other_free = mi.nr_free_pages - mi.totalreserve_pages;
    other_file = mi.nr_file_pages - mi.nr_shmem;

基于zoneinfo解析，计算出other_free和other_file两个参数，用于选取待kill的进程。

do {
        killed_size = find_and_kill_process(other_free, other_file, first);这是最核心的地方。
        if (killed_size > 0) {
            first = false;
            other_free += killed_size;
            other_file += killed_size;
        }
    } while (killed_size > 0);循环释放，直到killed_size<=0，也即满足了最低内存需求。
}

find_and_kill_process根据other_free和other_file两个参数，确定在哪个adj组中寻找进程。然后寻找最近使用进程kill。

static int find_and_kill_process(int other_free, int other_file, bool first)
{
    int i;
    int min_score_adj = OOM_SCORE_ADJ_MAX + 1;
    int minfree = 0;
    int killed_size = 0;

for (i = 0; i < lowmem_targets_size; i++) {
        minfree = lowmem_minfree[i];
        if (other_free < minfree && other_file < minfree) {
            min_score_adj = lowmem_adj[i];
            break;
        }
    }

lowmem_minfree和lowmem_adj是从/sys/module/lowmemorykiller/parameters/minfree和/sys/module/lowmemorykiller/parameters/adj中解析出来的。释放内存以达到最低使用内存，adj从0到906，每一个adj都有对应的最低内存，逐级释放。

0,100,200,300,900,906
18432,23040,27648,32256,55296,80640

if (min_score_adj == OOM_SCORE_ADJ_MAX + 1)
        return 0;

for (i = OOM_SCORE_ADJ_MAX; i >= min_score_adj; i--) {
        struct proc *procp;

retry:
        procp = proc_adj_lru(i);  在procadjslot_list寻找最近使用的proc

if (procp) {
            killed_size = kill_one_process(procp, other_free, other_file, minfree, min_score_adj, first);  杀死procp指定的进程，返回释放的内存大小。
            if (killed_size < 0) {
                goto retry;
            } else {
                return killed_size;
            }
        }
    }

return 0;
}

lowmemorykiller分析

lowmemorykiller作为内核一个module，输入参数有如下：

/sys/module/lowmemorykiller/parameters/adj  0,100,200,300,900,906
/sys/module/lowmemorykiller/parameters/cost  32
/sys/module/lowmemorykiller/parameters/debug_level  1
/sys/module/lowmemorykiller/parameters/minfree  18432,23040,27648,32256,55296,80640

adj文件包含oom_adj的阈值，minfree存放着对应的阈值，以page为单位。当对应的minfree值达到，则进程的oom_adj如果大于这个值将被杀掉。

ProcessList.java中定义的mOomAdj的值通过writeLmkd写入sysfs节点，和上面对应：

private final int[] mOomAdj = new int[] {
        FOREGROUND_APP_ADJ, VISIBLE_APP_ADJ, PERCEPTIBLE_APP_ADJ,
        BACKUP_APP_ADJ, CACHED_APP_MIN_ADJ, CACHED_APP_MAX_ADJ
};

// These are the low-end OOM level limits.  This is appropriate for an
// HVGA or smaller phone with less than 512MB.  Values are in KB.
private final int[] mOomMinFreeLow = new int[] {
        12288, 18432, 24576,
        36864, 43008, 49152
};
// These are the high-end OOM level limits.  This is appropriate for a
// 1280x800 or larger screen with around 1GB RAM.  Values are in KB.
private final int[] mOomMinFreeHigh = new int[] {
        73728, 92160, 110592,
        129024, 147456, 184320
};

在frameworks/base/services/core/java/com/android/server/am/ProcessList.java中定义了，不同类型进程对应的adj值：

static final int CACHED_APP_MAX_ADJ = 906;
static final int CACHED_APP_MIN_ADJ = 900;

static final int SERVICE_B_ADJ = 800;

static final int PREVIOUS_APP_ADJ = 700;

static final int HOME_APP_ADJ = 600;

static final int SERVICE_ADJ = 500;

static final int HEAVY_WEIGHT_APP_ADJ = 400;

static final int BACKUP_APP_ADJ = 300;

static final int PERCEPTIBLE_APP_ADJ = 200;

static final int VISIBLE_APP_ADJ = 100;
static final int VISIBLE_APP_LAYER_MAX = PERCEPTIBLE_APP_ADJ - VISIBLE_APP_ADJ - 1;

static final int FOREGROUND_APP_ADJ = 0;

static final int PERSISTENT_SERVICE_ADJ = -700;

static final int PERSISTENT_PROC_ADJ = -800;

static final int SYSTEM_ADJ = -900;

static final int NATIVE_ADJ = -1000;

lowmem_init是整个模块的入口，主要注册一个shrinker，lowmem_shrinker。shrinker是内核内存回收机制。

static struct shrinker lowmem_shrinker = {
    .scan_objects = lowmem_scan,  如果count_objects返回值不为0，则被调用。
    .count_objects = lowmem_count,  返回缓存中可被释放的内存大小。
    .seeks = DEFAULT_SEEKS * 16
};

lowmem_scan是shrinker的核心：

static unsigned long lowmem_scan(struct shrinker *s, struct shrink_control *sc)
{
    struct task_struct *tsk;
    struct task_struct *selected = NULL;
    unsigned long rem = 0;
    int tasksize;
    int i;
    short min_score_adj = OOM_SCORE_ADJ_MAX + 1;
    int minfree = 0;
    int selected_tasksize = 0;
    short selected_oom_score_adj;
    int array_size = ARRAY_SIZE(lowmem_adj);
    int other_free = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
    int other_file = global_page_state(NR_FILE_PAGES) -
                        global_page_state(NR_SHMEM) -
                        total_swapcache_pages();

if (lowmem_adj_size < array_size)
        array_size = lowmem_adj_size;
    if (lowmem_minfree_size < array_size)
        array_size = lowmem_minfree_size;
    for (i = 0; i < array_size; i++) {
        minfree = lowmem_minfree[i];
        if (other_free < minfree && other_file < minfree) {
            min_score_adj = lowmem_adj[i];  确定min_score_adj，从adj小的开始，也即内存最紧张的adj开始。直到找到other_free/other_file都小于minfree的adj。比这个adj大的进程都可以释放。
            break;
        }
    }

lowmem_print(3, "lowmem_scan %lu, %x, ofree %d %d, ma %hd\n",
            sc->nr_to_scan, sc->gfp_mask, other_free,
            other_file, min_score_adj);

if (min_score_adj == OOM_SCORE_ADJ_MAX + 1) {
        lowmem_print(5, "lowmem_scan %lu, %x, return 0\n",
                 sc->nr_to_scan, sc->gfp_mask);
        return 0;
    }

selected_oom_score_adj = min_score_adj;

rcu_read_lock();
    for_each_process(tsk) {  遍历所有进程
        struct task_struct *p;
        short oom_score_adj;

if (tsk->flags & PF_KTHREAD)
            continue;

p = find_lock_task_mm(tsk);
        if (!p)
            continue;

if (test_tsk_thread_flag(p, TIF_MEMDIE) &&
            time_before_eq(jiffies, lowmem_deathpending_timeout)) {
            task_unlock(p);
            rcu_read_unlock();
            return 0;
        }
        oom_score_adj = p->signal->oom_score_adj;
        if (oom_score_adj < min_score_adj) {  跳过高优先级的adj，adj小的优先级高。
            task_unlock(p);
            continue;
        }
        tasksize = get_mm_rss(p->mm);
        task_unlock(p);
        if (tasksize <= 0)
            continue;
        if (selected) {
            if (oom_score_adj < selected_oom_score_adj)  跳过高优先级的adj，adj小的优先级高。
                continue;
            if (oom_score_adj == selected_oom_score_adj &&
                tasksize <= selected_tasksize)  如果adj和选中优先级相同，则选用tasksize大的进程，能释放更多空间。
                continue;
        }
        selected = p;
        selected_tasksize = tasksize;
        selected_oom_score_adj = oom_score_adj;
        lowmem_print(2, "select '%s' (%d), adj %hd, size %d, to kill\n",
                 p->comm, p->pid, oom_score_adj, tasksize);
    }  所以总的原则是对所有oom_score_adj大于等于min_score_adj的进程，选取tasksize最大的进进程。也即根据进程的重要性(oom_adj)和释放量(tasksize)进行选取。
    if (selected) {
        long cache_size = other_file * (long)(PAGE_SIZE / 1024);
        long cache_limit = minfree * (long)(PAGE_SIZE / 1024);
        long free = other_free * (long)(PAGE_SIZE / 1024);

task_lock(selected);
        send_sig(SIGKILL, selected, 0);  发送SIGKILL信号到选定的进程
        /*
         * FIXME: lowmemorykiller shouldn't abuse global OOM killer
         * infrastructure. There is no real reason why the selected
         * task should have access to the memory reserves.
         */
        if (selected->mm)
            mark_oom_victim(selected);
        task_unlock(selected);
        trace_lowmemory_kill(selected, cache_size, cache_limit, free);
        lowmem_print(1, "Killing '%s' (%d), adj %hd,\n" \
                    "   to free %ldkB on behalf of '%s' (%d) because\n" \
                    "   cache %ldkB is below limit %ldkB for oom_score_adj %hd\n" \
                    "   Free memory is %ldkB above reserved\n",
                 selected->comm, selected->pid,
                 selected_oom_score_adj,
                 selected_tasksize * (long)(PAGE_SIZE / 1024),
                 current->comm, current->pid,
                 cache_size, cache_limit,
                 min_score_adj,
                 free);
        lowmem_deathpending_timeout = jiffies + HZ;
        rem += selected_tasksize;
    }

lowmem_print(4, "lowmem_scan %lu, %x, return %lu\n",
             sc->nr_to_scan, sc->gfp_mask, rem);
    rcu_read_unlock();
    return rem;
}

每一个进程都有oom_adj/oom_score/oom_score_adj节点，

oom_adj  -13
oom_score  0
oom_score_adj –800

oom_adj=oom_score_adj*17/1000=800*17/1000=13.6

CGroup memory子系统参数详解

要理解memory.pressure_level，就要从何为Memory Pressure开始。

pressure_level通知可以被用来监控内存分配代价；基于不同的pressure_level，采取不同的策略管理内存资源。有以下三种pressure_level：

low：系统会采取回收内存给新的内存分配。

medium：系统会使用swap、换出活动文件缓存等方式来腾空内存

critical：表示系统此时已经OOM或者内核OOM即将触发，应用应该尽可能采取措施腾出内存空间。

pressure_level出发后产生的events会向上传播，直到被处理。比如三个cgroup：A->B->C。A、B、C都有事件监听器，此时C触发了memory pressure。这种情况下，C会受到通知，而A和B则不会。这是为了避免此类消息广播，进而打断系统。

memory.pressure_level只是被用来设置eventfd，节点的读写操作都没有实现，所以在sysfs中无从获得信息。下面是一个使用示例：

• 使用eventfd创建一个evfd句柄
• 打开memory.pressure_level节点mpfd
• 将“<evfd> <mpfd> <level>”组成的字符串写入cgroup.event_control

那么如果memory pressure达到一定level(low/medium/critical)，相关应用就会通过eventfd被通知到。下面是lmkd中的一个实现：

static int init_mp(char *levelstr, void *event_handler)
{
…

mpfd = open(MEMCG_SYSFS_PATH "memory.pressure_level", O_RDONLY | O_CLOEXEC);
    evctlfd = open(MEMCG_SYSFS_PATH "cgroup.event_control", O_WRONLY | O_CLOEXEC);
    evfd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
    ret = snprintf(buf, sizeof(buf), "%d %d %s", evfd, mpfd, levelstr);
    ret = write(evctlfd, buf, strlen(buf) + 1);

epev.events = EPOLLIN;
    epev.data.ptr = event_handler;
    ret = epoll_ctl(epollfd, EPOLL_CTL_ADD, evfd, &epev);
}

所以重点就转到分析cgroup.event_control

static struct cftype mem_cgroup_legacy_files[] = {
    {
        .name = "cgroup.event_control",        /* XXX: for compat */
        .write = memcg_write_event_control,
        .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
    },

}

memcg_write_event_control解析lmkd写入的字符串，然后注册cgroup的事件处理函数。

static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
                     char *buf, size_t nbytes, loff_t off)
{
    struct cgroup_subsys_state *css = of_css(of);
    struct mem_cgroup *memcg = mem_cgroup_from_css(css);
    struct mem_cgroup_event *event;
    struct cgroup_subsys_state *cfile_css;
    unsigned int efd, cfd;
    struct fd efile;
    struct fd cfile;
    const char *name;
    char *endp;
    int ret;

buf = strstrip(buf);

efd = simple_strtoul(buf, &endp, 10);  解析出eventfd文件句柄
    if (*endp != ' ')
        return -EINVAL;
    buf = endp + 1;

cfd = simple_strtoul(buf, &endp, 10);  解析出字符串的第二个参数句柄
    if ((*endp != ' ') && (*endp != '\0'))
        return -EINVAL;
    buf = endp + 1;  解析出第三个参数

event = kzalloc(sizeof(*event), GFP_KERNEL);
    if (!event)
        return -ENOMEM;

event->memcg = memcg;
    INIT_LIST_HEAD(&event->list);
    init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
    init_waitqueue_func_entry(&event->wait, memcg_event_wake);
    INIT_WORK(&event->remove, memcg_event_remove);

efile = fdget(efd);
    if (!efile.file) {
        ret = -EBADF;
        goto out_kfree;
    }

event->eventfd = eventfd_ctx_fileget(efile.file);
    if (IS_ERR(event->eventfd)) {
        ret = PTR_ERR(event->eventfd);
        goto out_put_efile;
    }

cfile = fdget(cfd);
    if (!cfile.file) {
        ret = -EBADF;
        goto out_put_eventfd;
    }

/* the process need read permission on control file */
    /* AV: shouldn't we check that it's been opened for read instead? */
    ret = inode_permission(file_inode(cfile.file), MAY_READ);
    if (ret < 0)
        goto out_put_cfile;

/*
     * Determine the event callbacks and set them in @event.  This used
     * to be done via struct cftype but cgroup core no longer knows
     * about these events.  The following is crude but the whole thing
     * is for compatibility anyway.
     *
     * DO NOT ADD NEW FILES.
     */
    name = cfile.file->f_path.dentry->d_name.name;

if (!strcmp(name, "memory.usage_in_bytes")) {  根据第二个参数文件名，选择不同注册/去注册函数。
        event->register_event = mem_cgroup_usage_register_event;
        event->unregister_event = mem_cgroup_usage_unregister_event;
    } else if (!strcmp(name, "memory.oom_control")) {
        event->register_event = mem_cgroup_oom_register_event;
        event->unregister_event = mem_cgroup_oom_unregister_event;
    } else if (!strcmp(name, "memory.pressure_level")) {
        event->register_event = vmpressure_register_event;
        event->unregister_event = vmpressure_unregister_event;
    } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
        event->register_event = memsw_cgroup_usage_register_event;
        event->unregister_event = memsw_cgroup_usage_unregister_event;
    } else {
        ret = -EINVAL;
        goto out_put_cfile;
    }

/*
     * Verify @cfile should belong to @css.  Also, remaining events are
     * automatically removed on cgroup destruction but the removal is
     * asynchronous, so take an extra ref on @css.
     */
    cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
                           &memory_cgrp_subsys);
    ret = -EINVAL;
    if (IS_ERR(cfile_css))
        goto out_put_cfile;
    if (cfile_css != css) {
        css_put(cfile_css);
        goto out_put_cfile;
    }

ret = event->register_event(memcg, event->eventfd, buf);  执行注册
    if (ret)
        goto out_put_css;

efile.file->f_op->poll(efile.file, &event->pt);

spin_lock(&memcg->event_list_lock);
    list_add(&event->list, &memcg->event_list);
    spin_unlock(&memcg->event_list_lock);

fdput(cfile);
    fdput(efile);

return nbytes;

out_put_css:
    css_put(css);
out_put_cfile:
    fdput(cfile);
out_put_eventfd:
    eventfd_ctx_put(event->eventfd);
out_put_efile:
    fdput(efile);
out_kfree:
    kfree(event);

return ret;
}

vmpressure_register_event会将vmpressure通知和eventfs绑定，这样lmkd就会收到vmpressure的通知了。

memcg：需要关注vmpressure通知的CGroup子系统memory

eventfd：接收vmpressure通知的eventfd句柄

args：设置pressure_level参数

int vmpressure_register_event(struct mem_cgroup *memcg,
                  struct eventfd_ctx *eventfd, const char *args)
{
    struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
    struct vmpressure_event *ev;
    int level;

for (level = 0; level < VMPRESSURE_NUM_LEVELS; level++) {
        if (!strcmp(vmpressure_str_levels[level], args))  检查pressure_level有效性：low/medium/critical。
            break;
    }

if (level >= VMPRESSURE_NUM_LEVELS)
        return -EINVAL;

ev = kzalloc(sizeof(*ev), GFP_KERNEL);
    if (!ev)
        return -ENOMEM;

ev->efd = eventfd;
    ev->level = level;

mutex_lock(&vmpr->events_lock);
    list_add(&ev->node, &vmpr->events);
    mutex_unlock(&vmpr->events_lock);

return 0;
}

关于Memory Pressure深度阅读参考：Documents/cgroups/memory.txt 第11小节 Memory Pressure

这里有涉及到一个概念vmpressure。应用不会去关注系统有多少可用空间，但是作为一个整体的系统如果能对内存紧缺进行通知，并让应用采取相关措施以减少内存分配。vmpressure就是这样一种机制，通过vmpressure内核能够通知用户空间，系统当前处于何种memory pressure等级。

应用？

整个框架提供的配置参数就是应用的切入点：

1. 根据内存大小？屏幕分辨率？…情况配置不同的minfree值。

2. 增加adj个数，增加lowmemorykiller的控制粒度；或者修改adj大小，改变不同类型进程的优先级。

3. memory pressure的levelstr，low？medium？critical？进行不同的处理？

4. 修改vmpressure触发不同level的条件？

参考资料

mlockall/munlockall：http://pubs.opengroup.org/onlinepubs/007908799/xsh/mlockall.html

mlockall函数：http://blog.csdn.net/zhjutao/article/details/8652252

event：http://www.man7.org/linux/man-pages/man2/eventfd.2.html

Memory Pressure：https://linux-mm.org/Memory_pressure

vmpressure_fd：https://lwn.net/Articles/524742/