init.c

#define _GNU_SOURCE
#include <errno.h>
#include <fcntl.h>
#include <ftw.h>
#include <getopt.h>
#include <libkmod.h>
#include <linux/random.h>  // RNDADDENTROPY
#include <net/if.h>
#include <netinet/ip.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mount.h>
#include <sys/resource.h>
#include <sys/socket.h>
#include <sys/stat.h>
#include <sys/sysmacros.h>
#include <sys/types.h>
#include <sys/utsname.h>
#include <sys/wait.h>
#include <unistd.h>

#include "../vsockexec/vsock.h"

#ifdef DEBUG
#ifdef USE_TCP
static const int tcpmode = 1;
#else
static const int tcpmode;
#endif
// vsockexec opens vsock connections for the specified stdio descriptors and
// then execs the specified process.


static int opentcp(unsigned short port)
{
    int s = socket(AF_INET, SOCK_STREAM, 0);
    if (s < 0)
    {
        return -1;
    }

    struct sockaddr_in addr = {0};
    addr.sin_family = AF_INET;
    addr.sin_port = htons(port);
    addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
    if (connect(s, (struct sockaddr *)&addr, sizeof(addr)) < 0)
    {
        return -1;
    }

    return s;
}
#endif

#define DEFAULT_PATH_ENV "PATH=/sbin:/usr/sbin:/bin:/usr/bin"
#define OPEN_FDS 15

const char *const default_envp[] = {
    DEFAULT_PATH_ENV,
    NULL,
};

// global kmod k_ctx so we can access it in the file tree traversal 
struct kmod_ctx *k_ctx;

// When nothing is passed, default to the LCOWv1 behavior.
const char *const default_argv[] = { "/bin/gcs", "-loglevel", "debug", "-logfile=/run/gcs/gcs.log" };
const char *const default_shell = "/bin/sh";
const char *const lib_modules = "/lib/modules";

struct Mount {
    const char *source, *target, *type;
    unsigned long flags;
    const void *data;
};

struct Mkdir {
    const char *path;
    mode_t mode;
};

struct Mknod {
    const char *path;
    mode_t mode;
    int major, minor;
};

struct Symlink {
    const char *linkpath, *target;
};

enum OpType {
    OpMount,
    OpMkdir,
    OpMknod,
    OpSymlink,
};

struct InitOp {
    enum OpType op;
    union {
        struct Mount mount;
        struct Mkdir mkdir;
        struct Mknod mknod;
        struct Symlink symlink;
    };
};

const struct InitOp ops[] = {
    // mount /proc (which should already exist)
    { OpMount, .mount = { "proc", "/proc", "proc", MS_NODEV | MS_NOSUID | MS_NOEXEC } },

    // add symlinks in /dev (which is already mounted)
    { OpSymlink, .symlink = { "/dev/fd", "/proc/self/fd" } },
    { OpSymlink, .symlink = { "/dev/stdin", "/proc/self/fd/0" } },
    { OpSymlink, .symlink = { "/dev/stdout", "/proc/self/fd/1" } },
    { OpSymlink, .symlink = { "/dev/stderr", "/proc/self/fd/2" } },

    // mount tmpfs on /run and /tmp (which should already exist)
    { OpMount, .mount = { "tmpfs", "/run", "tmpfs", MS_NODEV | MS_NOSUID | MS_NOEXEC, "mode=0755" } },
    { OpMount, .mount = { "tmpfs", "/tmp", "tmpfs", MS_NODEV | MS_NOSUID | MS_NOEXEC } },

    // mount shm and devpts
    { OpMkdir, .mkdir = { "/dev/shm", 0755 } },
    { OpMount, .mount = { "shm", "/dev/shm", "tmpfs", MS_NODEV | MS_NOSUID | MS_NOEXEC } },
    { OpMkdir, .mkdir = { "/dev/pts", 0755 } },
    { OpMount, .mount = { "devpts", "/dev/pts", "devpts", MS_NOSUID | MS_NOEXEC } },

    // mount /sys (which should already exist)
    { OpMount, .mount = { "sysfs", "/sys", "sysfs", MS_NODEV | MS_NOSUID | MS_NOEXEC } },
    { OpMount, .mount = { "cgroup_root", "/sys/fs/cgroup", "tmpfs", MS_NODEV | MS_NOSUID | MS_NOEXEC, "mode=0755" } },
};

void warn(const char *msg) {
    int error = errno;
    perror(msg);
    errno = error;
}

void warn2(const char *msg1, const char *msg2) {
    int error = errno;
    fputs(msg1, stderr);
    fputs(": ", stderr);
    errno = error;
    warn(msg2);
}

_Noreturn void dien() {
    #ifdef DEBUG
    printf("dien errno = %d", errno);
    #endif
    exit(errno);
}

_Noreturn void die(const char *msg) {
    warn(msg);
    dien();
}

_Noreturn void die2(const char *msg1, const char *msg2) {
    warn2(msg1, msg2);
    dien();
}

void init_rlimit() {
    // Set the hard limit for number of open fds much larger. The kernel sets
    // a limit of 4096 for historical reasons, and this limit is too low for
    // some software. According to the systemd developers, there is no downside
    // to a large hard limit in modern Linux kernels.
    //
    // Retain the small soft limit of 1024 for appcompat.
    struct rlimit rlim = {
        .rlim_cur = 1024,
        .rlim_max = 1024 * 1024,
    };
    if (setrlimit(RLIMIT_NOFILE, &rlim) < 0) {
        die("setrlimit(RLIMIT_NOFILE)");
    }
}

void init_dev() {
    if (mount("dev", "/dev", "devtmpfs", MS_NOSUID | MS_NOEXEC, NULL) < 0) {
        #ifdef DEBUG
        printf("mount - errno %d\n", errno);
        #endif
        warn2("mount", "/dev");
        // /dev will be already mounted if devtmpfs.mount = 1 on the kernel
        // command line or CONFIG_DEVTMPFS_MOUNT is set. Do not consider this
        // an error.
        if (errno != EBUSY) {
            dien();
        }
    }
}

void init_fs(const struct InitOp *ops, size_t count) {
    for (size_t i = 0; i < count; i++) {
        switch (ops[i].op) {
        case OpMount: {
            const struct Mount *m = &ops[i].mount;
            #ifdef DEBUG
            printf("OpMount src %s target %s type %s flags %lu data %p\n", m->source, m->target, m->type, m->flags, m->data);
            #endif
            if (mount(m->source, m->target, m->type, m->flags, m->data) < 0) {
                die2("mount", m->target);
            }
            break;
        }
        case OpMkdir: {
            const struct Mkdir *m = &ops[i].mkdir;
            #ifdef DEBUG
            printf("OpMkdir path %s mode %d\n", m->path, m->mode);
            #endif
            if (mkdir(m->path, m->mode) < 0) {
                warn2("mkdir", m->path);
                if (errno != EEXIST) {
                    dien();
                }
            }
            break;
        }
        case OpMknod: {
            const struct Mknod *n = &ops[i].mknod;
            #ifdef DEBUG
            printf("OpMknod path %s mode %d major %d minor %d\n", n->path, n->mode, n->major, n->minor);
            #endif
            if (mknod(n->path, n->mode, makedev(n->major, n->minor)) < 0) {
                warn2("mknod", n->path);
                if (errno != EEXIST) {
                    dien();
                }
            }
            break;
        }
        case OpSymlink: {
            const struct Symlink *sl = &ops[i].symlink;
            #ifdef DEBUG
            printf("OpSymlink targeg %s link %s\n", sl->target, sl->linkpath);
            #endif
            if (symlink(sl->target, sl->linkpath) < 0) {
                warn2("symlink", sl->linkpath);
                if (errno != EEXIST) {
                    dien();
                }
            }
            break;
        }
        }
    }
}

void init_cgroups() {
    const char *fpath = "/proc/cgroups";
    FILE *f = fopen(fpath, "r");
    if (f == NULL) {
        die2("fopen", fpath);
    }
    // Skip the first line.
    for (;;) {
        char c = fgetc(f);
        if (c == EOF || c == '\n') {
            break;
        }
    }
    for (;;) {
        static const char base_path[] = "/sys/fs/cgroup/";
        char path[sizeof(base_path) - 1 + 64];
        char* name = path + sizeof(base_path) - 1;
        int hier, groups, enabled;
        int r = fscanf(f, "%64s %d %d %d\n", name, &hier, &groups, &enabled);
        if (r == EOF) {
            break;
        }
        if (r != 4) {
            errno = errno ? : EINVAL;
            die2("fscanf", fpath);
        }
        if (enabled) {
            memcpy(path, base_path, sizeof(base_path) - 1);
            if (mkdir(path, 0755) < 0) {
                die2("mkdir", path);
            }
            if (mount(name, path, "cgroup", MS_NODEV | MS_NOSUID | MS_NOEXEC, name) < 0) {
                die2("mount", path);
            }
        }
    }
    fclose(f);
}

void init_network(const char *iface, int domain) {
    int s = socket(domain, SOCK_DGRAM, IPPROTO_IP);
    if (s < 0) {
        if (errno == EAFNOSUPPORT) {
            return;
        }
        die("socket");
    }

    struct ifreq request = {0};
    strncpy(request.ifr_name, iface, sizeof(request.ifr_name));
    if (ioctl(s, SIOCGIFFLAGS, &request) < 0) {
        die2("ioctl(SIOCGIFFLAGS)", iface);
    }

    request.ifr_flags |= IFF_UP | IFF_RUNNING;
    if (ioctl(s, SIOCSIFFLAGS, &request) < 0) {
        die2("ioctl(SIOCSIFFLAGS)", iface);
    }

    close(s);
}

// inject boot-time entropy after reading it from a vsock port
void init_entropy(int port) {
    int s = openvsock(VMADDR_CID_HOST, port);
    if (s < 0) {
        die("openvsock entropy");
    }

    int e = open("/dev/random", O_RDWR);
    if (e < 0) {
        die("open /dev/random");
    }

    struct {
        int entropy_count;
        int buf_size;
        char buf[4096];
    } buf;

    for (;;) {
        ssize_t n = read(s, buf.buf, sizeof(buf.buf));
        if (n < 0) {
            die("read entropy");
        }

        if (n == 0) {
            break;
        }

        buf.entropy_count = n * 8; // in bits
        buf.buf_size = n; // in bytes
        if (ioctl(e, RNDADDENTROPY, &buf) < 0) {
            die("ioctl(RNDADDENTROPY)");
        }
    }

    close(s);
    close(e);
}

pid_t launch(int argc, char **argv) {
    int pid = fork();
    if (pid != 0) {
        if (pid < 0) {
            die("fork");
        }

        return pid;
    }

    // Unblock signals before execing.
    sigset_t set;
    sigfillset(&set);
    sigprocmask(SIG_UNBLOCK, &set, 0);

    // Create a session and process group.
    setsid();
    setpgid(0, 0);

    // Terminate the arguments and exec.
    char **argvn = alloca(sizeof(argv[0]) * (argc + 1));
    memcpy(argvn, argv, sizeof(argv[0]) * argc);
    argvn[argc] = NULL;
    if (putenv(DEFAULT_PATH_ENV)) { // Specify the PATH used for execvpe
        die("putenv");
    }
    execvpe(argvn[0], argvn, (char**)default_envp);
    die2("execvpe", argvn[0]);
}

int reap_until(pid_t until_pid) {
    for (;;) {
        int status;
        pid_t pid = wait(&status);
        if (pid < 0) {
            die("wait");
        }

        if (pid == until_pid) {
            // The initial child process died. Pass through the exit status.
            if (WIFEXITED(status)) {
                if (WEXITSTATUS(status) != 0) {
                    fputs("child exited with error\n", stderr);
                }
                return WEXITSTATUS(status);
            }
            fputs("child exited by signal: ", stderr);
            fputs(strsignal(WTERMSIG(status)), stderr);
            fputs("\n", stderr);
            return 128 + WTERMSIG(status);
        }
    }
}

// load_module gets the module from the absolute path to the module and then 
// inserts into the kernel. 
int load_module(struct kmod_ctx *ctx, const char *module_path) {
    struct kmod_module *mod = NULL; 
    int err;

    #ifdef DEBUG
    printf("loading module: %s\n", module_path);
    #endif

    err = kmod_module_new_from_path(ctx, module_path, &mod);
    if (err < 0) {
        return err; 
    }

    err = kmod_module_probe_insert_module(mod, 0, NULL, NULL, NULL, NULL); 
    if (err < 0) {
        kmod_module_unref(mod); 
        return err;
    }

    kmod_module_unref(mod); 
    return 0;
}

// parse_tree_entry is called by ftw for each directory and file in the file tree.
// If this entry is a file and has a .ko file extension, attempt to load into kernel.
int parse_tree_entry(const char *fpath, const struct stat *sb, int typeflag) {
    int result;
    const char *ext; 

    if (typeflag != FTW_F) {
        // do nothing if this isn't a file 
        return 0; 
    }
        
    ext = strrchr(fpath, '.');
    if (!ext || ext == fpath) {
        // no file extension found in the filepath
        return 0; 
    }

    if ((result = strcmp(ext, ".ko")) != 0) {
        // file does not have .ko extension so it is not a kernel module
        return 0; 
    }

    // print warning if we fail to load the module, but don't fail fn so 
    // we keep trying to load the rest of the modules. 
    result = load_module(k_ctx, fpath);
    if (result != 0) {
        warn2("failed to load module", fpath);
    }
    return 0;
}

// load_all_modules finds the modules in the image and loads them using kmod, 
// which accounts for ordering requirements. 
void load_all_modules() {
    int max_path = 256;
    char modules_dir[max_path];
    struct utsname uname_data;
    int ret; 

    // get information on the running kernel 
    ret = uname(&uname_data);
    if (ret != 0) {
        die("failed to get kernel information");
    }
    
    // create the absolute path of the modules directory this looks 
    // like /lib/modules/<uname.release>
    ret = snprintf(modules_dir, max_path, "%s/%s", lib_modules, uname_data.release); 
    if (ret < 0) {
        die("failed to create the modules directory path");
    } else if (ret > max_path) {
        die("modules directory buffer larger than expected"); 
    }

    if (k_ctx == NULL) {
        k_ctx = kmod_new(NULL, NULL);
        if (k_ctx == NULL) {
            die("failed to create kmod context");
        }
    }

    kmod_load_resources(k_ctx);
    ret = ftw(modules_dir, parse_tree_entry, OPEN_FDS); 
    if (ret < 0) {
        kmod_unref(k_ctx);
        die("failed to load kmod resources");
    } else if (ret != 0) {
        // Don't fail on error from walking the file tree and loading modules right now. 
        // ftw may return an error if the modules directory doesn't exist, which
        // may be the case for some images. Additionally, we don't currently support
        // using a denylist when loading modules, so we may try to load modules 
        // that cannot be loaded until later, such as nvidia modules which fail to 
        // load if no device is present. 
        warn("error adding modules");
    }

    kmod_unref(k_ctx);
}

#ifdef DEBUG
int debug_main(int argc, char **argv) {
    unsigned int ports[3] = {2056, 2056, 2056};
    int sockets[3] = {-1, -1, -1};

    for (int i = 0; i < 3; i++)
    {
        if (ports[i] != 0)
        {
            int j;
            for (j = 0; j < i; j++)
            {
                if (ports[i] == ports[j])
                {
                    int s = dup(sockets[j]);
                    if (s < 0)
                    {
                        perror("dup");
                        return 1;
                    }
                    sockets[i] = s;
                    break;
                }
            }

            if (j == i)
            {
                int s = tcpmode ? opentcp(ports[i]) : openvsock(VMADDR_CID_HOST, ports[i]);
                if (s < 0)
                {
                    fprintf(stderr, "connect: port %u: %s", ports[i], strerror(errno));
                    return 1;
                }
                sockets[i] = s;
            }
        }
    }

    for (int i = 0; i < 3; i++)
    {
        if (sockets[i] >= 0)
        {
            dup2(sockets[i], i);
            close(sockets[i]);
        }
    }
}
#endif

int main(int argc, char **argv) {
    #ifdef DEBUG
    debug_main(argc, argv);
    printf("Running init\n");
    #endif
    char *debug_shell = NULL;
    int entropy_port = 0;
    if (argc <= 1) {
        argv = (char **)default_argv;
        argc = sizeof(default_argv) / sizeof(default_argv[0]);
        optind = 0;
        debug_shell = (char*)default_shell;
    } else {
        for (int opt; (opt = getopt(argc, argv, "+d:e:")) >= 0; ) {
            switch (opt) {
            case 'd':
                debug_shell = optarg;
                break;

            case 'e':
                entropy_port = atoi(optarg);
                #ifdef DEBUG
                printf("entropy port %d\n", entropy_port);
                #endif
                if (entropy_port == 0) {
                    fputs("invalid entropy port\n", stderr);
                    exit(1);
                }

                break;

            default:
                exit(1);
            }
        }
    }

    char **child_argv = argv + optind;
    int child_argc = argc - optind;

    // Block all signals in init. SIGCHLD will still cause wait() to return.
    sigset_t set;
    #ifdef DEBUG
    printf("sigfillset(&set)\n");
    #endif
    sigfillset(&set);

    #ifdef DEBUG
    printf("sigfillset\n");
    #endif
    sigprocmask(SIG_BLOCK, &set, 0);

    #ifdef DEBUG
    printf("init_rlimit\n");
    #endif
    init_rlimit();

    #ifdef DEBUG
    printf("init_dev\n");
    #endif
    init_dev();

    #ifdef DEBUG
    printf("init_fs\n");
    #endif
    init_fs(ops, sizeof(ops) / sizeof(ops[0]));

    #ifdef DEBUG
    printf("init_cgroups\n");
    #endif
    init_cgroups();

    #ifdef DEBUG
    printf("init_network\n");
    #endif
    init_network("lo", AF_INET);
    init_network("lo", AF_INET6);
    if (entropy_port != 0) {
        init_entropy(entropy_port);
    }

    #ifdef DEBUG
    printf("loading modules\n");
    #endif
    load_all_modules();

    pid_t pid = launch(child_argc, child_argv);
    if (debug_shell != NULL) {
        // The debug shell takes over as the primary child.
        pid = launch(1, &debug_shell);
    }

    // Reap until the initial child process dies.
    return reap_until(pid);
}