lxcfs容器隔离技术实现原理分析之loadavg、cpuonline

lxcfs是什么

我们知道runc没有做到完全隔离/proc、/sys路径下的文件，所以容器内通过top、free等命令看到的数据都是物理机上的。对于习惯了虚机，物理机的同学来说不太友好，而且这些命令似乎也失去了本质意义。
lxcfs作用就是将容器内/proc、/sys文件与物理机隔离，让top等命令显示容器内真实数据。

说明

lxcfs是以用户空间文件系统（Filesystem in Userspace）为基础，以cgroup技术实现的用户空间的虚拟文件系统。先对fuse和cgroup有大致了解，看本文效果更好些。本文不介绍lxcfs的安装及使用，网上不乏这样的好文章。我们主要介绍下lxcfs对cpuonline、loadavg的现实，这两部分弄懂，其它也大体相同。

容器中读取lxcfs文件系统

lxcfs程序启动时会指定一个路径（如下图是/var/lib/lxcfs）作为挂载点，以后读取这个路径的下文件（cgroup、proc、sys）vfs都会调用内核fuse，fuse回调lxcfs实现的文件操作函数。容器内读取lxcfs文件系统中的数据时，通过gblic系统调用vfs接口然后转到fuse内核模块，内核模块fuse回调lxcfs程序中实现的回调函数，获取容器的cgroup，然后去宿主机对应cgroup下读取并计算后得到容器的实际mem、cpu等信息。lxcfs将物理机的cgroups挂载到运行时环境/run/lxcfs/controllers，但直接在物理机上看不见，因为程序中用unshare做了mounts namespace隔离。lxcfs程序中所有的cgroups信息都从/run/lxcfs/controllers下获得。

源码

因为工作中正好需要这两部分，所以主要介绍下cpuonline和loadavg的实现。nginx、java等程序根据cpu核心数启动相应个数的进程，cpuonline是相关系统调用的数据来源。没隔离导致的容器内获取到cpu核数是物理机的，本应该创建2个进程，实际却创建40个（容器2c，物理机40c），由于更多的上下文切换导致明显的性能下降。loadavg目前没看到有关的分析，这里也简单介绍下。

隔离效果

物理机40c128g

1、cpuonline
物理机
容器2c4g

2、loadavg
物理机

容器2c4g

可以看到cpuonline和load average都已经隔离

实现分析

注：cgropu的各个controller文件在main函数执行前打开，保存在fd_hierarchies中，后面使用直接掉openat，不是每次都要open、close文件。通过c语言的attribute((constructor)) 属性，声明collect_and_mount_subsystems这个函数
看下collect_and_mount_subsystems

static void __attribute__((constructor)) collect_and_mount_subsystems(void)
{
    FILE *f;
char *cret, *line = NULL;
char cwd[MAXPATHLEN];
    size_t len = 0;
int i, init_ns = -1;
bool found_unified = false;

if ((f = fopen("/proc/self/cgroup", "r")) == NULL) {
lxcfs_error("Error opening /proc/self/cgroup: %s\n", strerror(errno));
return;
}
// 读取宿主机上namespaces controller保存到hierarchies
while (getline(&line, &len, f) != -1) {
......
if (!store_hierarchy(line, p))
goto out;
}

/* Preserve initial namespace. */
    init_ns = preserve_mnt_ns(getpid());
if (init_ns < 0) {
lxcfs_error("%s\n", "Failed to preserve initial mount namespace.");
goto out;
}

    fd_hierarchies = malloc(sizeof(int) * num_hierarchies);
if (!fd_hierarchies) {
lxcfs_error("%s\n", strerror(errno));
goto out;
}

for (i = 0; i < num_hierarchies; i++)
        fd_hierarchies[i] = -1;

    cret = getcwd(cwd, MAXPATHLEN);
if (!cret)
lxcfs_debug("Could not retrieve current working directory: %s.\n", strerror(errno));

/* This function calls unshare(CLONE_NEWNS) our initial mount namespace
     * to privately mount lxcfs cgroups. */
// 关键是这里，将cgroup下各个控制模块，挂载到lxcfs进程的自由的mount ns下（/run/lxcfs/container）
if (!cgfs_setup_controllers()) {
lxcfs_error("%s\n", "Failed to setup private cgroup mounts for lxcfs.");
goto out;
}
......
}
static bool cgfs_setup_controllers(void)
{
// 主要调用unshare 创建私有的mount ns
if (!cgfs_prepare_mounts())
return false;

if (!cgfs_mount_hierarchies()) {
lxcfs_error("%s\n", "Failed to set up private lxcfs cgroup mounts.");
return false;
}

if (!permute_root())
return false;

return true;
}
static bool cgfs_mount_hierarchies(void)
{
char *target;
    size_t clen, len;
int i, ret;

for (i = 0; i < num_hierarchies; i++) {
char *controller = hierarchies[i];

        clen = strlen(controller);
        len = strlen(BASEDIR) + clen + 2;
        target = malloc(len);
if (!target)
return false;

        ret = snprintf(target, len, "%s/%s", BASEDIR, controller);
if (ret < 0 || ret >= len) {
free(target);
return false;
}
if (mkdir(target, 0755) < 0 && errno != EEXIST) {
free(target);
return false;
}
if (!strcmp(controller, "unified"))
            ret = mount("none", target, "cgroup2", 0, NULL);
else
            ret = mount(controller, target, "cgroup", 0, controller);
if (ret < 0) {
lxcfs_error("Failed mounting cgroup %s: %s\n", controller, strerror(errno));
free(target);
return false;
}
// 将所有cgroup controller 文件打开，保存文件描述符
        fd_hierarchies[i] = open(target, O_DIRECTORY);
if (fd_hierarchies[i] < 0) {
free(target);
return false;
}
free(target);
}
return true;
}

lxcfs.c main中主要是解析命令行参数，并调用fuse提供的fuse_main函数将lxcfs相关的文件操作注册，并传入挂载点。

......
if (!fuse_main(nargs, newargv, &lxcfs_ops, opts))
......
const struct fuse_operations lxcfs_ops = {
.getattr = lxcfs_getattr,
.readlink = NULL,
.getdir = NULL,
.mknod = NULL,
.mkdir = lxcfs_mkdir,
.unlink = NULL,
.rmdir = lxcfs_rmdir,
.symlink = NULL,
.rename = NULL,
.link = NULL,
.chmod = lxcfs_chmod,
.chown = lxcfs_chown,
.truncate = lxcfs_truncate,
.utime = NULL,

.open = lxcfs_open,
.read = lxcfs_read,
.release = lxcfs_release,
.write = lxcfs_write,

.statfs = NULL,
.flush = lxcfs_flush,
.fsync = lxcfs_fsync,

.setxattr = NULL,
.getxattr = NULL,
.listxattr = NULL,
.removexattr = NULL,

.opendir = lxcfs_opendir,
.readdir = lxcfs_readdir,
.releasedir = lxcfs_releasedir,

.fsyncdir = NULL,
.init = NULL,
.destroy = NULL,
.access = lxcfs_access,
.create = NULL,
.ftruncate = NULL,
.fgetattr = NULL,
};

cpuonline
1、cpuonline信息在/sys/devices/system/cpu/路径下，lxcfs将对/sys（当然这里使用任何路径都可以）的操作注册到fuse
lxcfs.c：

const struct fuse_operations lxcfs_ops = {
......
.open = lxcfs_open,
.read = lxcfs_read,
.release = lxcfs_release,
.write = lxcfs_write,
......
}

static int lxcfs_read(const char *path, char *buf, size_t size, off_t offset,
struct fuse_file_info *fi)
{
int ret;
if (strncmp(path, "/cgroup", 7) == 0) {
up_users();
        ret = do_cg_read(path, buf, size, offset, fi);
down_users();
return ret;
}
if (strncmp(path, "/proc", 5) == 0) {
up_users();
        ret = do_proc_read(path, buf, size, offset, fi);
down_users();
return ret;
}
if (strncmp(path, "/sys", 4) == 0) {
up_users();
        ret = do_sys_read(path, buf, size, offset, fi);
down_users();
return ret;
}

return -EINVAL;
}

static int do_sys_read(const char *path, char *buf, size_t size, off_t offset,
struct fuse_file_info *fi)
{
int (*sys_read)(const char *path, char *buf, size_t size, off_t offset,
struct fuse_file_info *fi);
char *error;

dlerror();    /* Clear any existing error */
    sys_read = (int (*)(const char *, char *, size_t, off_t, struct fuse_file_info *)) dlsym(dlopen_handle, "sys_read");
    error = dlerror();
if (error != NULL) {
lxcfs_error("%s\n", error);
return -1;
}

return sys_read(path, buf, size, offset, fi);
}

文件操作相关代码（bindings.c、sysfs_fuse.c，cpuset.c）被封装成liblxcfs.so动态库，供lxcfs.c调用。上面do_sys_read通过dlsym获取liblxcfs.so动态库中的sys_read函数。

2、接着看下读cpuonline的过程
sysfs_fuse.c：

int sys_read(const char *path, char *buf, size_t size, off_t offset,
struct fuse_file_info *fi)
{
struct file_info *f = (struct file_info *)fi->fh;

switch (f->type) {
//cpuonline 模块，type在open时设置，这里不做过多介绍，主要看下
//sys_devices_system_cpu_online_read函数的实现
case LXC_TYPE_SYS_DEVICES_SYSTEM_CPU_ONLINE:  
return sys_devices_system_cpu_online_read(buf, size, offset, fi);
case LXC_TYPE_SYS_DEVICES:
case LXC_TYPE_SYS_DEVICES_SYSTEM:
case LXC_TYPE_SYS_DEVICES_SYSTEM_CPU:
default:
return -EINVAL;
}
}

static int sys_devices_system_cpu_online_read(char *buf, size_t size,
                          off_t offset,
struct fuse_file_info *fi)
{
//获取上下文信息，主要是读取cuponline进程（例如cat /sys/devices/system/cpu/online的cat进程，以下简称“调用进程”）的进程id
struct fuse_context *fc = fuse_get_context();
struct file_info *d = (struct file_info *)fi->fh;
char *cache = d->buf;
char *cg;
char *cpuset = NULL;
bool use_view;

int max_cpus = 0;
    pid_t initpid;
    ssize_t total_len = 0;

if (offset) {
if (!d->cached)
return 0;
if (offset > d->size)
return -EINVAL;
int left = d->size - offset;
        total_len = left > size ? size : left;
memcpy(buf, cache + offset, total_len);
return total_len;
}

//获取容器中1号进程在物理机上的进程id；initpid返回为0时，说明调用进程是物理上的进程
    initpid = lookup_initpid_in_store(fc->pid);
if (initpid <= 0)
        initpid = fc->pid;
//获取容器1号进程的cgroup
//例如：docker/368adedeb87172d68388cee9818e873d73503a5b1d1d2a6b47fbd053f6d68601
    cg = get_pid_cgroup(initpid, "cpuset");
if (!cg)
return read_file("/sys/devices/system/cpu/online", buf, size, d);
prune_init_slice(cg);

    cpuset = get_cpuset(cg);
if (!cpuset)
goto err;
// 检查cpu、 cpuacct 控制器是否存在，不存在直接返回物理机cpuonine信息
    use_view = use_cpuview(cg);

if (use_view)
// 获取容器真正可使用的cpu个数，如果容器没配置cpu quota（默认-1），则直接返回物理信息
        max_cpus = max_cpu_count(cg);

if (max_cpus == 0)
return read_file("/sys/devices/system/cpu/online", buf, size, d);
if (max_cpus > 1)
        total_len = snprintf(d->buf, d->buflen, "0-%d\n", max_cpus - 1);
else
        total_len = snprintf(d->buf, d->buflen, "0\n");
if (total_len < 0 || total_len >= d->buflen) {
lxcfs_error("%s\n", "failed to write to cache");
return 0;
}

    d->size = (int)total_len;
    d->cached = 1;

if (total_len > size)
        total_len = size;

memcpy(buf, d->buf, total_len);
err:
free(cpuset);
free(cg);
return total_len;
}
/*
 * Return the maximum number of visible CPUs based on CPU quotas.
 * If there is no quota set, zero is returned.
 */
int max_cpu_count(const char *cg)
{
int rv, nprocs;
int64_t cfs_quota, cfs_period;
int nr_cpus_in_cpuset = 0;
char *cpuset = NULL;
// 读取物理机上容器cpu的quota值
if (!read_cpu_cfs_param(cg, "quota", &cfs_quota))
return 0;
// 读取物理机上容器cpu的period值
if (!read_cpu_cfs_param(cg, "period", &cfs_period))
return 0;

    cpuset = get_cpuset(cg);
if (cpuset)
        nr_cpus_in_cpuset = cpu_number_in_cpuset(cpuset);

if (cfs_quota <= 0 || cfs_period <= 0){
if (nr_cpus_in_cpuset > 0)
return nr_cpus_in_cpuset;

return 0;
}

// 容器何用的cpu计算
    rv = cfs_quota / cfs_period;

/* In case quota/period does not yield a whole number, add one CPU for
     * the remainder.这里的意思是限制cpu为0.5和，视图效果为1核。1.5 即 2
     */
if ((cfs_quota % cfs_period) > 0)
        rv += 1;

/*获取可用的cpu核数sysconf(_SC_NPROCESSORS_ONLN)*/
    nprocs = get_nprocs();

if (rv > nprocs)
        rv = nprocs;

/* use min value in cpu quota and cpuset */
if (nr_cpus_in_cpuset > 0 && nr_cpus_in_cpuset < rv)
        rv = nr_cpus_in_cpuset;

return rv;
}
// 看下quota是怎么获取的
/*
 * Read cgroup CPU quota parameters from `cpu.cfs_quota_us` or `cpu.cfs_period_us`,
 * depending on `param`. Parameter value is returned throuh `value`.
 */
static bool read_cpu_cfs_param(const char *cg, const char *param, int64_t *value)
{
bool rv = false;
char file[11 + 6 + 1]; // cpu.cfs__us + quota/period + \0
char *str = NULL;

sprintf(file, "cpu.cfs_%s_us", param);

// 重点是这里
if (!cgfs_get_value("cpu", cg, file, &str))
goto err;
......
}
bool cgfs_get_value(const char *controller, const char *cgroup, const char *file, char **value)
{
int ret, fd, cfd;
    size_t len;
char *fnam, *tmpc;
// 获取cpu controller文件描述符，到之前说过fd_hierarchies中查
    tmpc = find_mounted_controller(controller, &cfd);
if (!tmpc)
return false;

/* Make sure we pass a relative path to *at() family of functions.
     * . + /cgroup + / + file + \0
     */
    len = strlen(cgroup) + strlen(file) + 3;
    fnam = alloca(len);
    ret = snprintf(fnam, len, "%s%s/%s", *cgroup == '/' ? "." : "", cgroup, file);
if (ret < 0 || (size_t)ret >= len)
return false;
// fd也就是 /run/lxcfs/controllers/cpu/docker/dockerid/cpu.cfs_quota_us
    fd = openat(cfd, fnam, O_RDONLY);
if (fd < 0)
return false;
// 读值cfs_quota_us
*value = slurp_file(fnam, fd);
return *value != NULL;
}

/*
 * Traverse the hash table and update it.
 */
void *load_begin(void *arg)
{

......
while (1) {
if (loadavg_stop == 1)
return NULL;

        time1 = clock();
for (i = 0; i < LOAD_SIZE; i++) {
pthread_mutex_lock(&load_hash[i].lock);
if (load_hash[i].next == NULL) {
pthread_mutex_unlock(&load_hash[i].lock);
continue;
}
            f = load_hash[i].next;
            first_node = 1;
while (f) {
......
// 更新负载
                sum = refresh_load(f, path);
if (sum == 0) {
                    f = del_node(f, i);
} else {
out:                    f = f->next;
}
free(path);
......
}

if (loadavg_stop == 1)
return NULL;

        time2 = clock();
usleep(FLUSH_TIME * 1000000 - (int)((time2 - time1) * 1000000 / CLOCKS_PER_SEC));
}
}

主要分析下refresh_load

/*
 * Return 0 means that container p->cg is closed.
 * Return -1 means that error occurred in refresh.
 * Positive num equals the total number of pid.
 */
static int refresh_load(struct load_node *p, char *path)
{
    FILE *f = NULL;
char **idbuf;
char proc_path[256];
int i, ret, run_pid = 0, total_pid = 0, last_pid = 0;
char *line = NULL;
    size_t linelen = 0;
int sum, length;
    DIR *dp;
struct dirent *file;

do {
        idbuf = malloc(sizeof(char *));
} while (!idbuf);
// 这里从/sys/fs/cgroup/cpu/docker/containerid/cgroup.procs to find the process pid.那容器内进程的pid
    sum = calc_pid(&idbuf, path, DEPTH_DIR, 0, p->cfd);
/*  normal exit  */
if (sum == 0)
goto out;

for (i = 0; i < sum; i++) {
/*clean up '\n' */
        length = strlen(idbuf[i])-1;
        idbuf[i][length] = '\0';
        ret = snprintf(proc_path, 256, "/proc/%s/task", idbuf[i]);
if (ret < 0 || ret > 255) {
lxcfs_error("%s\n", "snprintf() failed in refresh_load.");
            i = sum;
            sum = -1;
goto err_out;
}

        dp = opendir(proc_path);
if (!dp) {
lxcfs_error("%s\n", "Open proc_path failed in refresh_load.");
continue;
}
// 遍历/proc//task 目录（一个进程中创建的每个线程，/proc//task 中会创建一个相应的目录），查找状态为R或者D的task
while ((file = readdir(dp)) != NULL) {
if (strncmp(file->d_name, ".", 1) == 0)
continue;
if (strncmp(file->d_name, "..", 1) == 0)
continue;
            total_pid++;
/* We make the biggest pid become last_pid.*/
            ret = atof(file->d_name);
            last_pid = (ret > last_pid) ? ret : last_pid;

            ret = snprintf(proc_path, 256, "/proc/%s/task/%s/status", idbuf[i], file->d_name);
if (ret < 0 || ret > 255) {
lxcfs_error("%s\n", "snprintf() failed in refresh_load.");
                i = sum;
                sum = -1;
closedir(dp);
goto err_out;
}
            f = fopen(proc_path, "r");
if (f != NULL) {
while (getline(&line, &linelen, f) != -1) {
/* Find State */
if ((line[0] == 'S') && (line[1] == 't'))
break;
}
if ((line[7] == 'R') || (line[7] == 'D'))
                run_pid++;
fclose(f);
}
}
closedir(dp);
}
/*Calculate the loadavg.*/
// 获取到活跃的task数量后，是时候表演真正的技术了（计算平均负载）。计算公式与内核一致：load(t) = load(t-1) e-5/60 + n (1 - e-5/60)
// 具体含义可以参考：[https://www.helpsystems.com/resources/guideshow-it-works](https://w/unix-load-average-part-1-ww.helpsystems.com/resources/guides/unix-load-average-part-1-how-it-works)

    p->avenrun[0] = calc_load(p->avenrun[0], EXP_1, run_pid);
    p->avenrun[1] = calc_load(p->avenrun[1], EXP_5, run_pid);
    p->avenrun[2] = calc_load(p->avenrun[2], EXP_15, run_pid);
    p->run_pid = run_pid;
    p->total_pid = total_pid;
    p->last_pid = last_pid;

free(line);
err_out:
for (; i > 0; i--)
free(idbuf[i-1]);
out:
free(idbuf);
return sum;
}

2、读取loadavg
负载计算明白了，读就简单了。这里要注意下的是，load_hash，哈希表中的数据是容器第一次读/proc/loadavg时插入的（毕竟没办法事先知道容器的进程cgroup）。

static int proc_loadavg_read(char *buf, size_t size, off_t offset,
struct fuse_file_info *fi)
{
struct fuse_context *fc = fuse_get_context();
struct file_info *d = (struct file_info *)fi->fh;
    pid_t initpid;
    char *cg;
    size_t total_len = 0;
    char *cache = d->buf;
struct load_node *n;
    int hash;
    int cfd, rv = 0;
    unsigned long a, b, c;

if (offset) {
if (offset > d->size)
return -EINVAL;
if (!d->cached)
return 0;
        int left = d->size - offset;
        total_len = left > size ? size : left;
memcpy(buf, cache + offset, total_len);
return total_len;
}
if (!loadavg)
return read_file("/proc/loadavg", buf, size, d);

    initpid = lookup_initpid_in_store(fc->pid);
if (initpid <= 0)
        initpid = fc->pid;
    cg = get_pid_cgroup(initpid, "cpu");
if (!cg)
return read_file("/proc/loadavg", buf, size, d);

prune_init_slice(cg);
    hash = calc_hash(cg) % LOAD_SIZE;
// 根据cgroup在hash表查找node
    n = locate_node(cg, hash);

/* First time */
// 第一读时，先把节点信息插到hash边
if (n == NULL) {
if (!find_mounted_controller("cpu", &cfd)) {
/*
             * In locate_node() above, pthread_rwlock_unlock() isn't used
             * because delete is not allowed before read has ended.
             */
pthread_rwlock_unlock(&load_hash[hash].rdlock);
            rv = 0;
            goto err;
}
do {
            n = malloc(sizeof(struct load_node));
} while (!n);

do {
            n->cg = malloc(strlen(cg)+1);
} while (!n->cg);
strcpy(n->cg, cg);
        n->avenrun[0] = 0;
        n->avenrun[1] = 0;
        n->avenrun[2] = 0;
        n->run_pid = 0;
        n->total_pid = 1;
        n->last_pid = initpid;
        n->cfd = cfd;
insert_node(&n, hash);
}
// 第二次以后开始从daemon的计算结果中读取
    a = n->avenrun[0] + (FIXED_1/200);
    b = n->avenrun[1] + (FIXED_1/200);
    c = n->avenrun[2] + (FIXED_1/200);
    total_len = snprintf(d->buf, d->buflen, "%lu.%02lu %lu.%02lu %lu.%02lu %d/%d %d\n",
LOAD_INT(a), LOAD_FRAC(a),
LOAD_INT(b), LOAD_FRAC(b),
LOAD_INT(c), LOAD_FRAC(c),
        n->run_pid, n->total_pid, n->last_pid);
pthread_rwlock_unlock(&load_hash[hash].rdlock);
if (total_len < 0 || total_len >=  d->buflen) {
lxcfs_error("%s\n", "Failed to write to cache");
        rv = 0;
        goto err;
}
    d->size = (int)total_len;
    d->cached = 1;

if (total_len > size)
        total_len = size;
memcpy(buf, d->buf, total_len);
    rv = total_len;

err:
free(cg);
return rv;
}

参考文章

• https://www.helpsystems.com/resources/guides/unix-load-average-part-1-how-it-works
• https://github.com/libfuse/libfuse
• https://github.com/lxc/lxcfs
• https://www.helpsystems.com/resources/guides/unix-load-average-part-1-how-it-works

关注我们